allow for setting beta schedules of unets differently in the decoder, as what was used in the paper was cosine, cosine, linear

* Converted decoder trainer to use accelerate

* Fixed issue where metric evaluation would hang on distributed mode

* Implemented functional saving
Loading still fails due to some issue with the optimizer

* Fixed issue with loading decoders

* Fixed issue with tracker config

* Fixed issue with amp
Updated logging to be more logical

* Saving checkpoint now saves position in training as well
Fixed an issue with running out of gpu space due to loading weights into the gpu twice

* Fixed ema for distributed training

* Fixed isue where get_pkg_version was reintroduced

* Changed decoder trainer to upload config as a file

Fixed issue where loading best would error

README.md

@@ -27,10 +27,27 @@ As of 5/23/22, it is no longer SOTA. SOTA will be <a href="https://github.com/lu
 - <a href="https://twitter.com/Buntworthy/status/1529475416775434240?t=0GEge3Kr9I36cjcUVCQUTg">Justin Pinkney</a> successfully trained the diffusion prior in the repository for his CLIP to Stylegan2 text-to-image application
 
 ## Pre-Trained Models
+
 - LAION is training prior models. Checkpoints are available on <a href="https://huggingface.co/zenglishuci/conditioned-prior">🤗huggingface</a> and the training statistics are available on <a href="https://wandb.ai/nousr_laion/conditioned-prior/reports/LAION-DALLE2-PyTorch-Prior--VmlldzoyMDI2OTIx">🐝WANDB</a>.
 - Decoder - <a href="https://wandb.ai/veldrovive/dalle2_train_decoder/runs/jkrtg0so?workspace=user-veldrovive">In-progress test run</a> 🚧
+- Decoder - <a href="https://wandb.ai/veldrovive/dalle2_train_decoder/runs/3d5rytsa?workspace=">Another test run with sparse attention</a>
 - DALL-E 2 🚧
 
+## Appreciation
+
+This library would not have gotten to this working state without the help of
+
+- <a href="https://github.com/nousr">Zion</a> for the distributed training code for the diffusion prior
+- <a href="https://github.com/Veldrovive">Aidan</a> for the distributed training code for the decoder as well as the dataloaders
+- <a href="https://github.com/krish240574">Kumar</a> for working on the initial diffusion training script
+- <a href="https://github.com/rom1504">Romain</a> for the pull request reviews and project management
+- <a href="https://github.com/Ciaohe">He Cao</a> and <a href="https://github.com/xiankgx">xiankgx</a> for the Q&A and for identifying of critical bugs
+- <a href="https://github.com/crowsonkb">Katherine</a> for her advice
+- <a href="https://stability.ai/">Stability AI</a> for the generous sponsorship
+- <a href="https://huggingface.co">🤗 Huggingface</a> and in particular <a href="https://github.com/sgugger">Sylvain</a> for the <a href="https://github.com/huggingface/accelerate">Accelerate</a> library
+
+... and many others. Thank you! 🙏
+
 ## Install
 
 ```bash
@@ -1041,19 +1058,6 @@ Once built, images will be saved to the same directory the command is invoked
 
 <a href="https://github.com/lucidrains/stylegan2-pytorch">template</a>
 
-## Appreciation
-
-This library would not have gotten to this working state without the help of
-
-- <a href="https://github.com/nousr">Zion</a> and <a href="https://github.com/krish240574">Kumar</a> for the diffusion training script
-- <a href="https://github.com/Veldrovive">Aidan</a> for the decoder training script and dataloaders
-- <a href="https://github.com/rom1504">Romain</a> for the pull request reviews and project management
-- <a href="https://github.com/Ciaohe">He Cao</a> and <a href="https://github.com/xiankgx">xiankgx</a> for the Q&A and for identifying of critical bugs
-- <a href="https://github.com/crowsonkb">Katherine</a> for her advice
-- <a href="https://stability.ai/">Stability AI</a> for the generous sponsorship
-
-... and many others. Thank you! 🙏
-
 ## Todo
 
 - [x] finish off gaussian diffusion class for latent embedding - allow for prediction of epsilon
@@ -1207,4 +1211,14 @@ This library would not have gotten to this working state without the help of
 }
 ```
 
+```bibtex
+@article{Choi2022PerceptionPT,
+    title   = {Perception Prioritized Training of Diffusion Models},
+    author  = {Jooyoung Choi and Jungbeom Lee and Chaehun Shin and Sungwon Kim and Hyunwoo J. Kim and Sung-Hoon Yoon},
+    journal = {ArXiv},
+    year    = {2022},
+    volume  = {abs/2204.00227}
+}
+```
+
 *Creating noise from data is easy; creating data from noise is generative modeling.* - <a href="https://arxiv.org/abs/2011.13456">Yang Song's paper</a>

dalle2_pytorch/dalle2_pytorch.py

@@ -1,7 +1,6 @@
 import math
 import random
 from tqdm import tqdm
-from inspect import isfunction
 from functools import partial, wraps
 from contextlib import contextmanager
 from collections import namedtuple
@@ -12,7 +11,7 @@ import torch.nn.functional as F
 from torch import nn, einsum
 import torchvision.transforms as T
 
-from einops import rearrange, repeat
+from einops import rearrange, repeat, reduce
 from einops.layers.torch import Rearrange
 from einops_exts import rearrange_many, repeat_many, check_shape
 from einops_exts.torch import EinopsToAndFrom
@@ -57,7 +56,7 @@ def maybe(fn):
 def default(val, d):
     if exists(val):
         return val
-    return d() if isfunction(d) else d
+    return d() if callable(d) else d
 
 def cast_tuple(val, length = 1):
     if isinstance(val, list):
@@ -314,11 +313,6 @@ def extract(a, t, x_shape):
     out = a.gather(-1, t)
     return out.reshape(b, *((1,) * (len(x_shape) - 1)))
 
-def noise_like(shape, device, repeat=False):
-    repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
-    noise = lambda: torch.randn(shape, device=device)
-    return repeat_noise() if repeat else noise()
-
 def meanflat(x):
     return x.mean(dim = tuple(range(1, len(x.shape))))
 
@@ -373,7 +367,7 @@ def quadratic_beta_schedule(timesteps):
     scale = 1000 / timesteps
     beta_start = scale * 0.0001
     beta_end = scale * 0.02
-    return torch.linspace(beta_start**2, beta_end**2, timesteps, dtype = torch.float64) ** 2
+    return torch.linspace(beta_start**0.5, beta_end**0.5, timesteps, dtype = torch.float64) ** 2
 
 
 def sigmoid_beta_schedule(timesteps):
@@ -384,8 +378,8 @@ def sigmoid_beta_schedule(timesteps):
     return torch.sigmoid(betas) * (beta_end - beta_start) + beta_start
 
 
-class BaseGaussianDiffusion(nn.Module):
-    def __init__(self, *, beta_schedule, timesteps, loss_type):
+class NoiseScheduler(nn.Module):
+    def __init__(self, *, beta_schedule, timesteps, loss_type, p2_loss_weight_gamma = 0., p2_loss_weight_k = 1):
         super().__init__()
 
         if beta_schedule == "cosine":
@@ -450,6 +444,11 @@ class BaseGaussianDiffusion(nn.Module):
         register_buffer('posterior_mean_coef1', betas * torch.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod))
         register_buffer('posterior_mean_coef2', (1. - alphas_cumprod_prev) * torch.sqrt(alphas) / (1. - alphas_cumprod))
 
+        # p2 loss reweighting
+
+        self.has_p2_loss_reweighting = p2_loss_weight_gamma > 0.
+        register_buffer('p2_loss_weight', (p2_loss_weight_k + alphas_cumprod / (1 - alphas_cumprod)) ** -p2_loss_weight_gamma)
+
     def q_posterior(self, x_start, x_t, t):
         posterior_mean = (
             extract(self.posterior_mean_coef1, t, x_t.shape) * x_start +
@@ -473,11 +472,10 @@ class BaseGaussianDiffusion(nn.Module):
             extract(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
         )
 
-    def sample(self, *args, **kwargs):
-        raise NotImplementedError
-
-    def forward(self, *args, **kwargs):
-        raise NotImplementedError
+    def p2_reweigh_loss(self, loss, times):
+        if not self.has_p2_loss_reweighting:
+            return loss
+        return loss * extract(self.p2_loss_weight, times, loss.shape)
 
 # diffusion prior
 
@@ -688,8 +686,7 @@ class Attention(nn.Module):
 
         # attention
 
-        sim = sim - sim.amax(dim = -1, keepdim = True).detach()
-        attn = sim.softmax(dim = -1)
+        attn = sim.softmax(dim = -1, dtype = torch.float32)
         attn = self.dropout(attn)
 
         # aggregate values
@@ -863,7 +860,7 @@ class DiffusionPriorNetwork(nn.Module):
 
         return pred_image_embed
 
-class DiffusionPrior(BaseGaussianDiffusion):
+class DiffusionPrior(nn.Module):
     def __init__(
         self,
         net,
@@ -884,7 +881,9 @@ class DiffusionPrior(BaseGaussianDiffusion):
         image_embed_scale = None,           # this is for scaling the l2-normed image embedding, so it is more suitable for gaussian diffusion, as outlined by Katherine (@crowsonkb) https://github.com/lucidrains/DALLE2-pytorch/issues/60#issue-1226116132
         clip_adapter_overrides = dict()
     ):
-        super().__init__(
+        super().__init__()
+
+        self.noise_scheduler = NoiseScheduler(
             beta_schedule = beta_schedule,
             timesteps = timesteps,
             loss_type = loss_type
@@ -924,6 +923,13 @@ class DiffusionPrior(BaseGaussianDiffusion):
         self.training_clamp_l2norm = training_clamp_l2norm
         self.init_image_embed_l2norm = init_image_embed_l2norm
 
+        # device tracker
+        self.register_buffer('_dummy', torch.tensor([True]), persistent = False)
+
+    @property
+    def device(self):
+        return self._dummy.device
+
     def p_mean_variance(self, x, t, text_cond, clip_denoised = False, cond_scale = 1.):
         assert not (cond_scale != 1. and not self.can_classifier_guidance), 'the model was not trained with conditional dropout, and thus one cannot use classifier free guidance (cond_scale anything other than 1)'
 
@@ -934,7 +940,7 @@ class DiffusionPrior(BaseGaussianDiffusion):
             # not 100% sure of this above line - for any spectators, let me know in the github issues (or through a pull request) if you know how to correctly do this
             # i'll be rereading https://arxiv.org/abs/2111.14822, where i think a similar approach is taken
         else:
-            x_recon = self.predict_start_from_noise(x, t = t, noise = pred)
+            x_recon = self.noise_scheduler.predict_start_from_noise(x, t = t, noise = pred)
 
         if clip_denoised and not self.predict_x_start:
             x_recon.clamp_(-1., 1.)
@@ -942,21 +948,21 @@ class DiffusionPrior(BaseGaussianDiffusion):
         if self.predict_x_start and self.sampling_clamp_l2norm:
             x_recon = l2norm(x_recon) * self.image_embed_scale
 
-        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
+        model_mean, posterior_variance, posterior_log_variance = self.noise_scheduler.q_posterior(x_start=x_recon, x_t=x, t=t)
         return model_mean, posterior_variance, posterior_log_variance
 
     @torch.no_grad()
-    def p_sample(self, x, t, text_cond = None, clip_denoised = True, repeat_noise = False, cond_scale = 1.):
+    def p_sample(self, x, t, text_cond = None, clip_denoised = True, cond_scale = 1.):
         b, *_, device = *x.shape, x.device
         model_mean, _, model_log_variance = self.p_mean_variance(x = x, t = t, text_cond = text_cond, clip_denoised = clip_denoised, cond_scale = cond_scale)
-        noise = noise_like(x.shape, device, repeat_noise)
+        noise = torch.randn_like(x)
         # no noise when t == 0
         nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
         return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
 
     @torch.no_grad()
     def p_sample_loop(self, shape, text_cond, cond_scale = 1.):
-        device = self.betas.device
+        device = self.device
 
         b = shape[0]
         image_embed = torch.randn(shape, device=device)
@@ -964,7 +970,7 @@ class DiffusionPrior(BaseGaussianDiffusion):
         if self.init_image_embed_l2norm:
             image_embed = l2norm(image_embed) * self.image_embed_scale
 
-        for i in tqdm(reversed(range(0, self.num_timesteps)), desc='sampling loop time step', total=self.num_timesteps):
+        for i in tqdm(reversed(range(0, self.noise_scheduler.num_timesteps)), desc='sampling loop time step', total=self.noise_scheduler.num_timesteps):
             times = torch.full((b,), i, device = device, dtype = torch.long)
             image_embed = self.p_sample(image_embed, times, text_cond = text_cond, cond_scale = cond_scale)
 
@@ -973,7 +979,7 @@ class DiffusionPrior(BaseGaussianDiffusion):
     def p_losses(self, image_embed, times, text_cond, noise = None):
         noise = default(noise, lambda: torch.randn_like(image_embed))
 
-        image_embed_noisy = self.q_sample(x_start = image_embed, t = times, noise = noise)
+        image_embed_noisy = self.noise_scheduler.q_sample(x_start = image_embed, t = times, noise = noise)
 
         pred = self.net(
             image_embed_noisy,
@@ -987,7 +993,7 @@ class DiffusionPrior(BaseGaussianDiffusion):
 
         target = noise if not self.predict_x_start else image_embed
 
-        loss = self.loss_fn(pred, target)
+        loss = self.noise_scheduler.loss_fn(pred, target)
         return loss
 
     @torch.no_grad()
@@ -998,7 +1004,7 @@ class DiffusionPrior(BaseGaussianDiffusion):
 
         img = torch.randn(shape, device = device)
 
-        for i in tqdm(reversed(range(0, self.num_timesteps)), desc = 'sampling loop time step', total = self.num_timesteps):
+        for i in tqdm(reversed(range(0, self.noise_scheduler.num_timesteps)), desc = 'sampling loop time step', total = self.noise_scheduler.num_timesteps):
             img = self.p_sample(img, torch.full((batch_size,), i, device = device, dtype = torch.long), text_cond = text_cond, cond_scale = cond_scale)
         return img
 
@@ -1070,7 +1076,7 @@ class DiffusionPrior(BaseGaussianDiffusion):
         # timestep conditioning from ddpm
 
         batch, device = image_embed.shape[0], image_embed.device
-        times = torch.randint(0, self.num_timesteps, (batch,), device = device, dtype = torch.long)
+        times = torch.randint(0, self.noise_scheduler.num_timesteps, (batch,), device = device, dtype = torch.long)
 
         # scale image embed (Katherine)
 
@@ -1085,8 +1091,9 @@ class DiffusionPrior(BaseGaussianDiffusion):
 def Upsample(dim):
     return nn.ConvTranspose2d(dim, dim, 4, 2, 1)
 
-def Downsample(dim):
-    return nn.Conv2d(dim, dim, 4, 2, 1)
+def Downsample(dim, *, dim_out = None):
+    dim_out = default(dim_out, dim)
+    return nn.Conv2d(dim, dim_out, 4, 2, 1)
 
 class SinusoidalPosEmb(nn.Module):
     def __init__(self, dim):
@@ -1234,8 +1241,7 @@ class CrossAttention(nn.Module):
             mask = rearrange(mask, 'b j -> b 1 1 j')
             sim = sim.masked_fill(~mask, max_neg_value)
 
-        sim = sim - sim.amax(dim = -1, keepdim = True).detach()
-        attn = sim.softmax(dim = -1)
+        attn = sim.softmax(dim = -1, dtype = torch.float32)
 
         out = einsum('b h i j, b h j d -> b h i d', attn, v)
         out = rearrange(out, 'b h n d -> b n (h d)')
@@ -1352,6 +1358,7 @@ class Unet(nn.Module):
         init_cross_embed_kernel_sizes = (3, 7, 15),
         cross_embed_downsample = False,
         cross_embed_downsample_kernel_sizes = (2, 4),
+        memory_efficient = False,
         **kwargs
     ):
         super().__init__()
@@ -1371,7 +1378,7 @@ class Unet(nn.Module):
         self.channels_out = default(channels_out, channels)
 
         init_channels = channels if not lowres_cond else channels * 2 # in cascading diffusion, one concats the low resolution image, blurred, for conditioning the higher resolution synthesis
-        init_dim = default(init_dim, dim // 3 * 2)
+        init_dim = default(init_dim, dim)
 
         self.init_conv = CrossEmbedLayer(init_channels, dim_out = init_dim, kernel_sizes = init_cross_embed_kernel_sizes, stride = 1)
 
@@ -1428,6 +1435,7 @@ class Unet(nn.Module):
         # for classifier free guidance
 
         self.null_image_embed = nn.Parameter(torch.randn(1, num_image_tokens, cond_dim))
+        self.null_image_hiddens = nn.Parameter(torch.randn(1, time_cond_dim))
 
         self.max_text_len = max_text_len
         self.null_text_embed = nn.Parameter(torch.randn(1, max_text_len, cond_dim))
@@ -1461,10 +1469,11 @@ class Unet(nn.Module):
             layer_cond_dim = cond_dim if not is_first else None
 
             self.downs.append(nn.ModuleList([
-                ResnetBlock(dim_in, dim_out, time_cond_dim = time_cond_dim, groups = groups),
+                downsample_klass(dim_in, dim_out = dim_out) if memory_efficient else None,
+                ResnetBlock(dim_out if memory_efficient else dim_in, dim_out, time_cond_dim = time_cond_dim, groups = groups),
                 Residual(LinearAttention(dim_out, **attn_kwargs)) if sparse_attn else nn.Identity(),
                 nn.ModuleList([ResnetBlock(dim_out, dim_out, cond_dim = layer_cond_dim, time_cond_dim = time_cond_dim, groups = groups) for _ in range(layer_num_resnet_blocks)]),
-                downsample_klass(dim_out) if not is_last else nn.Identity()
+                downsample_klass(dim_out) if not is_last and not memory_efficient else None
             ]))
 
         mid_dim = dims[-1]
@@ -1473,19 +1482,19 @@ class Unet(nn.Module):
         self.mid_attn = EinopsToAndFrom('b c h w', 'b (h w) c', Residual(Attention(mid_dim, **attn_kwargs))) if attend_at_middle else None
         self.mid_block2 = ResnetBlock(mid_dim, mid_dim, cond_dim = cond_dim, time_cond_dim = time_cond_dim, groups = resnet_groups[-1])
 
-        for ind, ((dim_in, dim_out), groups, layer_num_resnet_blocks) in enumerate(zip(reversed(in_out[1:]), reversed(resnet_groups), reversed(num_resnet_blocks))):
-            is_last = ind >= (num_resolutions - 2)
+        for ind, ((dim_in, dim_out), groups, layer_num_resnet_blocks) in enumerate(zip(reversed(in_out), reversed(resnet_groups), reversed(num_resnet_blocks))):
+            is_last = ind >= (len(in_out) - 1)
             layer_cond_dim = cond_dim if not is_last else None
 
             self.ups.append(nn.ModuleList([
                 ResnetBlock(dim_out * 2, dim_in, cond_dim = layer_cond_dim, time_cond_dim = time_cond_dim, groups = groups),
                 Residual(LinearAttention(dim_in, **attn_kwargs)) if sparse_attn else nn.Identity(),
                 nn.ModuleList([ResnetBlock(dim_in, dim_in, cond_dim = layer_cond_dim, time_cond_dim = time_cond_dim, groups = groups)  for _ in range(layer_num_resnet_blocks)]),
-                Upsample(dim_in)
+                Upsample(dim_in) if not is_last or memory_efficient else nn.Identity()
             ]))
 
         self.final_conv = nn.Sequential(
-            ResnetBlock(dim, dim, groups = resnet_groups[0]),
+            ResnetBlock(dim * 2, dim, groups = resnet_groups[0]),
             nn.Conv2d(dim, self.channels_out, 1)
         )
 
@@ -1557,6 +1566,7 @@ class Unet(nn.Module):
         # initial convolution
 
         x = self.init_conv(x)
+        r = x.clone() # final residual
 
         # time conditioning
 
@@ -1565,19 +1575,28 @@ class Unet(nn.Module):
         time_tokens = self.to_time_tokens(time_hiddens)
         t = self.to_time_cond(time_hiddens)
 
-        # image embedding to be summed to time embedding
-        # discovered by @mhh0318 in the paper
-
-        if exists(image_embed) and exists(self.to_image_hiddens):
-            image_hiddens = self.to_image_hiddens(image_embed)
-            t = t + image_hiddens
-
         # conditional dropout
 
         image_keep_mask = prob_mask_like((batch_size,), 1 - image_cond_drop_prob, device = device)
         text_keep_mask = prob_mask_like((batch_size,), 1 - text_cond_drop_prob, device = device)
 
-        image_keep_mask, text_keep_mask = rearrange_many((image_keep_mask, text_keep_mask), 'b -> b 1 1')
+        text_keep_mask = rearrange(text_keep_mask, 'b -> b 1 1')
+
+        # image embedding to be summed to time embedding
+        # discovered by @mhh0318 in the paper
+
+        if exists(image_embed) and exists(self.to_image_hiddens):
+            image_hiddens = self.to_image_hiddens(image_embed)
+            image_keep_mask_hidden = rearrange(image_keep_mask, 'b -> b 1')
+            null_image_hiddens = self.null_image_hiddens.to(image_hiddens.dtype)
+
+            image_hiddens = torch.where(
+                image_keep_mask_hidden,
+                image_hiddens,
+                null_image_hiddens
+            )
+
+            t = t + image_hiddens
 
         # mask out image embedding depending on condition dropout
         # for classifier free guidance
@@ -1585,11 +1604,12 @@ class Unet(nn.Module):
         image_tokens = None
 
         if self.cond_on_image_embeds:
+            image_keep_mask_embed = rearrange(image_keep_mask, 'b -> b 1 1')
             image_tokens = self.image_to_tokens(image_embed)
             null_image_embed = self.null_image_embed.to(image_tokens.dtype) # for some reason pytorch AMP not working
 
             image_tokens = torch.where(
-                image_keep_mask,
+                image_keep_mask_embed,
                 image_tokens,
                 null_image_embed
             )
@@ -1644,7 +1664,10 @@ class Unet(nn.Module):
 
         hiddens = []
 
-        for init_block, sparse_attn, resnet_blocks, downsample in self.downs:
+        for pre_downsample, init_block, sparse_attn, resnet_blocks, post_downsample in self.downs:
+            if exists(pre_downsample):
+                x = pre_downsample(x)
+
             x = init_block(x, c, t)
             x = sparse_attn(x)
 
@@ -1652,7 +1675,9 @@ class Unet(nn.Module):
                 x = resnet_block(x, c, t)
 
             hiddens.append(x)
-            x = downsample(x)
+
+            if exists(post_downsample):
+                x = post_downsample(x)
 
         x = self.mid_block1(x, mid_c, t)
 
@@ -1662,7 +1687,7 @@ class Unet(nn.Module):
         x = self.mid_block2(x, mid_c, t)
 
         for init_block, sparse_attn, resnet_blocks, upsample in self.ups:
-            x = torch.cat((x, hiddens.pop()), dim=1)
+            x = torch.cat((x, hiddens.pop()), dim = 1)
             x = init_block(x, c, t)
             x = sparse_attn(x)
 
@@ -1671,6 +1696,7 @@ class Unet(nn.Module):
 
             x = upsample(x)
 
+        x = torch.cat((x, r), dim = 1)
         return self.final_conv(x)
 
 class LowresConditioner(nn.Module):
@@ -1719,7 +1745,7 @@ class LowresConditioner(nn.Module):
 
         return cond_fmap
 
-class Decoder(BaseGaussianDiffusion):
+class Decoder(nn.Module):
     def __init__(
         self,
         unet,
@@ -1732,7 +1758,7 @@ class Decoder(BaseGaussianDiffusion):
         image_cond_drop_prob = 0.1,
         text_cond_drop_prob = 0.5,
         loss_type = 'l2',
-        beta_schedule = 'cosine',
+        beta_schedule = None,
         predict_x_start = False,
         predict_x_start_for_latent_diffusion = False,
         image_sizes = None,                         # for cascading ddpm, image size at each stage
@@ -1750,13 +1776,11 @@ class Decoder(BaseGaussianDiffusion):
         unconditional = False,
         auto_normalize_img = True,                  # whether to take care of normalizing the image from [0, 1] to [-1, 1] and back automatically - you can turn this off if you want to pass in the [-1, 1] ranged image yourself from the dataloader
         use_dynamic_thres = False,                  # from the Imagen paper
-        dynamic_thres_percentile = 0.9
+        dynamic_thres_percentile = 0.9,
+        p2_loss_weight_gamma = 0.,                  # p2 loss weight, from https://arxiv.org/abs/2204.00227 - 0 is equivalent to weight of 1 across time - 1. is recommended
+        p2_loss_weight_k = 1
     ):
-        super().__init__(
-            beta_schedule = beta_schedule,
-            timesteps = timesteps,
-            loss_type = loss_type
-        )
+        super().__init__()
 
         self.unconditional = unconditional
 
@@ -1800,6 +1824,8 @@ class Decoder(BaseGaussianDiffusion):
         # while the rest of the unets are conditioned on the low resolution image produced by previous unet
 
         unets = cast_tuple(unet)
+        num_unets = len(unets)
+
         vaes = pad_tuple_to_length(cast_tuple(vae), len(unets), fillvalue = NullVQGanVAE(channels = self.channels))
 
         # whether to use learned variance, defaults to True for the first unet in the cascade, as in paper
@@ -1835,6 +1861,24 @@ class Decoder(BaseGaussianDiffusion):
             self.unets.append(one_unet)
             self.vaes.append(one_vae.copy_for_eval())
 
+        # create noise schedulers per unet
+
+        if not exists(beta_schedule):
+            beta_schedule = ('cosine', *(('cosine',) * max(num_unets - 2, 0)), *(('linear',) * int(num_unets > 1)))
+
+        self.noise_schedulers = nn.ModuleList([])
+
+        for unet_beta_schedule in beta_schedule:
+            noise_scheduler = NoiseScheduler(
+                beta_schedule = unet_beta_schedule,
+                timesteps = timesteps,
+                loss_type = loss_type,
+                p2_loss_weight_gamma = p2_loss_weight_gamma,
+                p2_loss_weight_k = p2_loss_weight_k
+            )
+
+            self.noise_schedulers.append(noise_scheduler)
+
         # unet image sizes
 
         image_sizes = default(image_sizes, (image_size,))
@@ -1884,6 +1928,14 @@ class Decoder(BaseGaussianDiffusion):
         self.normalize_img = normalize_neg_one_to_one if auto_normalize_img else identity
         self.unnormalize_img = unnormalize_zero_to_one if auto_normalize_img else identity
 
+        # device tracker
+
+        self.register_buffer('_dummy', torch.Tensor([True]), persistent = False)
+
+    @property
+    def device(self):
+        return self._dummy.device
+
     def get_unet(self, unet_number):
         assert 0 < unet_number <= len(self.unets)
         index = unet_number - 1
@@ -1907,7 +1959,7 @@ class Decoder(BaseGaussianDiffusion):
         for unet, device in zip(self.unets, devices):
             unet.to(device)
 
-    def p_mean_variance(self, unet, x, t, image_embed, text_encodings = None, text_mask = None, lowres_cond_img = None, clip_denoised = True, predict_x_start = False, learned_variance = False, cond_scale = 1., model_output = None):
+    def p_mean_variance(self, unet, x, t, image_embed, noise_scheduler, text_encodings = None, text_mask = None, lowres_cond_img = None, clip_denoised = True, predict_x_start = False, learned_variance = False, cond_scale = 1., model_output = None):
         assert not (cond_scale != 1. and not self.can_classifier_guidance), 'the decoder was not trained with conditional dropout, and thus one cannot use classifier free guidance (cond_scale anything other than 1)'
 
         pred = default(model_output, lambda: unet.forward_with_cond_scale(x, t, image_embed = image_embed, text_encodings = text_encodings, text_mask = text_mask, cond_scale = cond_scale, lowres_cond_img = lowres_cond_img))
@@ -1918,7 +1970,7 @@ class Decoder(BaseGaussianDiffusion):
         if predict_x_start:
             x_recon = pred
         else:
-            x_recon = self.predict_start_from_noise(x, t = t, noise = pred)
+            x_recon = noise_scheduler.predict_start_from_noise(x, t = t, noise = pred)
 
         if clip_denoised:
             # s is the threshold amount
@@ -1937,14 +1989,14 @@ class Decoder(BaseGaussianDiffusion):
             # clip by threshold, depending on whether static or dynamic
             x_recon = x_recon.clamp(-s, s) / s
 
-        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
+        model_mean, posterior_variance, posterior_log_variance = noise_scheduler.q_posterior(x_start=x_recon, x_t=x, t=t)
 
         if learned_variance:
             # if learned variance, posterio variance and posterior log variance are predicted by the network
             # by an interpolation of the max and min log beta values
             # eq 15 - https://arxiv.org/abs/2102.09672
-            min_log = extract(self.posterior_log_variance_clipped, t, x.shape)
-            max_log = extract(torch.log(self.betas), t, x.shape)
+            min_log = extract(noise_scheduler.posterior_log_variance_clipped, t, x.shape)
+            max_log = extract(torch.log(noise_scheduler.betas), t, x.shape)
             var_interp_frac = unnormalize_zero_to_one(var_interp_frac_unnormalized)
 
             if self.learned_variance_constrain_frac:
@@ -1956,17 +2008,17 @@ class Decoder(BaseGaussianDiffusion):
         return model_mean, posterior_variance, posterior_log_variance
 
     @torch.no_grad()
-    def p_sample(self, unet, x, t, image_embed, text_encodings = None, text_mask = None, cond_scale = 1., lowres_cond_img = None, predict_x_start = False, learned_variance = False, clip_denoised = True, repeat_noise = False):
+    def p_sample(self, unet, x, t, image_embed, noise_scheduler, text_encodings = None, text_mask = None, cond_scale = 1., lowres_cond_img = None, predict_x_start = False, learned_variance = False, clip_denoised = True):
         b, *_, device = *x.shape, x.device
-        model_mean, _, model_log_variance = self.p_mean_variance(unet, x = x, t = t, image_embed = image_embed, text_encodings = text_encodings, text_mask = text_mask, cond_scale = cond_scale, lowres_cond_img = lowres_cond_img, clip_denoised = clip_denoised, predict_x_start = predict_x_start, learned_variance = learned_variance)
-        noise = noise_like(x.shape, device, repeat_noise)
+        model_mean, _, model_log_variance = self.p_mean_variance(unet, x = x, t = t, image_embed = image_embed, text_encodings = text_encodings, text_mask = text_mask, cond_scale = cond_scale, lowres_cond_img = lowres_cond_img, clip_denoised = clip_denoised, predict_x_start = predict_x_start, noise_scheduler = noise_scheduler, learned_variance = learned_variance)
+        noise = torch.randn_like(x)
         # no noise when t == 0
         nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
         return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
 
     @torch.no_grad()
-    def p_sample_loop(self, unet, shape, image_embed, predict_x_start = False, learned_variance = False, clip_denoised = True, lowres_cond_img = None, text_encodings = None, text_mask = None, cond_scale = 1, is_latent_diffusion = False):
-        device = self.betas.device
+    def p_sample_loop(self, unet, shape, image_embed, noise_scheduler, predict_x_start = False, learned_variance = False, clip_denoised = True, lowres_cond_img = None, text_encodings = None, text_mask = None, cond_scale = 1, is_latent_diffusion = False):
+        device = self.device
 
         b = shape[0]
         img = torch.randn(shape, device = device)
@@ -1974,7 +2026,7 @@ class Decoder(BaseGaussianDiffusion):
         if not is_latent_diffusion:
             lowres_cond_img = maybe(self.normalize_img)(lowres_cond_img)
 
-        for i in tqdm(reversed(range(0, self.num_timesteps)), desc = 'sampling loop time step', total = self.num_timesteps):
+        for i in tqdm(reversed(range(0, noise_scheduler.num_timesteps)), desc = 'sampling loop time step', total = noise_scheduler.num_timesteps):
             img = self.p_sample(
                 unet,
                 img,
@@ -1985,6 +2037,7 @@ class Decoder(BaseGaussianDiffusion):
                 cond_scale = cond_scale,
                 lowres_cond_img = lowres_cond_img,
                 predict_x_start = predict_x_start,
+                noise_scheduler = noise_scheduler,
                 learned_variance = learned_variance,
                 clip_denoised = clip_denoised
             )
@@ -1992,7 +2045,7 @@ class Decoder(BaseGaussianDiffusion):
         unnormalize_img = self.unnormalize_img(img)
         return unnormalize_img
 
-    def p_losses(self, unet, x_start, times, *, image_embed, lowres_cond_img = None, text_encodings = None, text_mask = None, predict_x_start = False, noise = None, learned_variance = False, clip_denoised = False, is_latent_diffusion = False):
+    def p_losses(self, unet, x_start, times, *, image_embed, noise_scheduler, lowres_cond_img = None, text_encodings = None, text_mask = None, predict_x_start = False, noise = None, learned_variance = False, clip_denoised = False, is_latent_diffusion = False):
         noise = default(noise, lambda: torch.randn_like(x_start))
 
         # normalize to [-1, 1]
@@ -2003,7 +2056,7 @@ class Decoder(BaseGaussianDiffusion):
 
         # get x_t
 
-        x_noisy = self.q_sample(x_start = x_start, t = times, noise = noise)
+        x_noisy = noise_scheduler.q_sample(x_start = x_start, t = times, noise = noise)
 
         model_output = unet(
             x_noisy,
@@ -2023,7 +2076,12 @@ class Decoder(BaseGaussianDiffusion):
 
         target = noise if not predict_x_start else x_start
 
-        loss = self.loss_fn(pred, target)
+        loss = noise_scheduler.loss_fn(pred, target, reduction = 'none')
+        loss = reduce(loss, 'b ... -> b (...)', 'mean')
+
+        loss = noise_scheduler.p2_reweigh_loss(loss, times)
+
+        loss = loss.mean()
 
         if not learned_variance:
             # return simple loss if not using learned variance
@@ -2036,8 +2094,8 @@ class Decoder(BaseGaussianDiffusion):
 
         # if learning the variance, also include the extra weight kl loss
 
-        true_mean, _, true_log_variance_clipped = self.q_posterior(x_start = x_start, x_t = x_noisy, t = times)
-        model_mean, _, model_log_variance = self.p_mean_variance(unet, x = x_noisy, t = times, image_embed = image_embed, clip_denoised = clip_denoised, learned_variance = True, model_output = model_output)
+        true_mean, _, true_log_variance_clipped = noise_scheduler.q_posterior(x_start = x_start, x_t = x_noisy, t = times)
+        model_mean, _, model_log_variance = self.p_mean_variance(unet, x = x_noisy, t = times, image_embed = image_embed, noise_scheduler = noise_scheduler, clip_denoised = clip_denoised, learned_variance = True, model_output = model_output)
 
         # kl loss with detached model predicted mean, for stability reasons as in paper
 
@@ -2069,7 +2127,8 @@ class Decoder(BaseGaussianDiffusion):
         text_encodings = None,
         batch_size = 1,
         cond_scale = 1.,
-        stop_at_unet_number = None
+        stop_at_unet_number = None,
+        distributed = False,
     ):
         assert self.unconditional or exists(image_embed), 'image embed must be present on sampling from decoder unless if trained unconditionally'
 
@@ -2086,9 +2145,9 @@ class Decoder(BaseGaussianDiffusion):
         img = None
         is_cuda = next(self.parameters()).is_cuda
 
-        for unet_number, unet, vae, channel, image_size, predict_x_start, learned_variance in tqdm(zip(range(1, len(self.unets) + 1), self.unets, self.vaes, self.sample_channels, self.image_sizes, self.predict_x_start, self.learned_variance)):
+        for unet_number, unet, vae, channel, image_size, predict_x_start, learned_variance, noise_scheduler in tqdm(zip(range(1, len(self.unets) + 1), self.unets, self.vaes, self.sample_channels, self.image_sizes, self.predict_x_start, self.learned_variance, self.noise_schedulers)):
 
-            context = self.one_unet_in_gpu(unet = unet) if is_cuda else null_context()
+            context = self.one_unet_in_gpu(unet = unet) if is_cuda and not distributed else null_context()
 
             with context:
                 lowres_cond_img = None
@@ -2114,7 +2173,8 @@ class Decoder(BaseGaussianDiffusion):
                     learned_variance = learned_variance,
                     clip_denoised = not is_latent_diffusion,
                     lowres_cond_img = lowres_cond_img,
-                    is_latent_diffusion = is_latent_diffusion
+                    is_latent_diffusion = is_latent_diffusion,
+                    noise_scheduler = noise_scheduler
                 )
 
                 img = vae.decode(img)
@@ -2140,6 +2200,7 @@ class Decoder(BaseGaussianDiffusion):
         unet = self.get_unet(unet_number)
 
         vae                 = self.vaes[unet_index]
+        noise_scheduler     = self.noise_schedulers[unet_index]
         target_image_size   = self.image_sizes[unet_index]
         predict_x_start     = self.predict_x_start[unet_index]
         random_crop_size    = self.random_crop_sizes[unet_index]
@@ -2149,7 +2210,7 @@ class Decoder(BaseGaussianDiffusion):
         check_shape(image, 'b c h w', c = self.channels)
         assert h >= target_image_size and w >= target_image_size
 
-        times = torch.randint(0, self.num_timesteps, (b,), device = device, dtype = torch.long)
+        times = torch.randint(0, noise_scheduler.num_timesteps, (b,), device = device, dtype = torch.long)
 
         if not exists(image_embed) and not self.unconditional:
             assert exists(self.clip), 'if you want to derive CLIP image embeddings automatically, you must supply `clip` to the decoder on init'
@@ -2180,7 +2241,7 @@ class Decoder(BaseGaussianDiffusion):
             image = vae.encode(image)
             lowres_cond_img = maybe(vae.encode)(lowres_cond_img)
 
-        return self.p_losses(unet, image, times, image_embed = image_embed, text_encodings = text_encodings, text_mask = text_mask, lowres_cond_img = lowres_cond_img, predict_x_start = predict_x_start, learned_variance = learned_variance, is_latent_diffusion = is_latent_diffusion)
+        return self.p_losses(unet, image, times, image_embed = image_embed, text_encodings = text_encodings, text_mask = text_mask, lowres_cond_img = lowres_cond_img, predict_x_start = predict_x_start, learned_variance = learned_variance, is_latent_diffusion = is_latent_diffusion, noise_scheduler = noise_scheduler)
 
 # main class
 

dalle2_pytorch/dataloaders/decoder_loader.py

@@ -164,9 +164,6 @@ class ImageEmbeddingDataset(wds.DataPipeline, wds.compat.FluidInterface):
             # There may be webdataset shards that do not have a embedding shard associated with it. If we do not skip these, they would cause issues.
             self.append(skip_unassociated_shards(embeddings_url=embedding_folder_url, handler=handler))
 
-        self.append(wds.split_by_node)
-        self.append(wds.split_by_worker)
-
         self.append(wds.tarfile_to_samples(handler=handler))
         self.append(wds.decode("pilrgb", handler=handler))
         if embedding_folder_url is not None:

dalle2_pytorch/trackers.py

@@ -17,15 +17,15 @@ DEFAULT_DATA_PATH = './.tracker-data'
 def exists(val):
     return val is not None
 
-# load state dict functions
+# load file functions
 
-def load_wandb_state_dict(run_path, file_path, **kwargs):
+def load_wandb_file(run_path, file_path, **kwargs):
     wandb = import_or_print_error('wandb', '`pip install wandb` to use the wandb recall function')
     file_reference = wandb.restore(file_path, run_path=run_path)
-    return torch.load(file_reference.name)
+    return file_reference.name
 
-def load_local_state_dict(file_path, **kwargs):
-    return torch.load(file_path)
+def load_local_file(file_path, **kwargs):
+    return file_path
 
 # base class
 
@@ -55,12 +55,43 @@ class BaseTracker(nn.Module):
         """
         # TODO: Pull this into a dict or something similar so that we can add more sources without having a massive switch statement
         if recall_source == 'wandb':
-            return load_wandb_state_dict(*args, **kwargs)
+            return torch.load(load_wandb_file(*args, **kwargs))
         elif recall_source == 'local':
-            return load_local_state_dict(*args, **kwargs)
+            return torch.load(load_local_file(*args, **kwargs))
         else:
             raise ValueError('`recall_source` must be one of `wandb` or `local`')
 
+    def save_file(self, file_path, **kwargs):
+        raise NotImplementedError
+
+    def recall_file(self, recall_source, *args, **kwargs):
+        if recall_source == 'wandb':
+            return load_wandb_file(*args, **kwargs)
+        elif recall_source == 'local':
+            return load_local_file(*args, **kwargs)
+        else:
+            raise ValueError('`recall_source` must be one of `wandb` or `local`')
+
+# Tracker that no-ops all calls except for recall
+
+class DummyTracker(BaseTracker):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def init(self, config, **kwargs):
+        pass
+
+    def log(self, log, **kwargs):
+        pass
+
+    def log_images(self, images, **kwargs):
+        pass
+
+    def save_state_dict(self, state_dict, relative_path, **kwargs):
+        pass
+
+    def save_file(self, file_path, **kwargs):
+        pass
 
 # basic stdout class
 
@@ -76,6 +107,10 @@ class ConsoleTracker(BaseTracker):
     
     def save_state_dict(self, state_dict, relative_path, **kwargs):
         torch.save(state_dict, str(self.data_path / relative_path))
+    
+    def save_file(self, file_path, **kwargs):
+        # This is a no-op for local file systems since it is already saved locally
+        pass
 
 # basic wandb class
 
@@ -107,3 +142,11 @@ class WandbTracker(BaseTracker):
         full_path = str(self.data_path / relative_path)
         torch.save(state_dict, full_path)
         self.wandb.save(full_path, base_path = str(self.data_path))  # Upload and keep relative to data_path
+
+    def save_file(self, file_path, base_path=None, **kwargs):
+        """
+        Uploads a file from disk to wandb
+        """
+        if base_path is None:
+            base_path = self.data_path
+        self.wandb.save(str(file_path), base_path = str(base_path))

dalle2_pytorch/train_configs.py

@@ -173,7 +173,7 @@ class DecoderConfig(BaseModel):
     channels: int = 3
     timesteps: int = 1000
     loss_type: str = 'l2'
-    beta_schedule: str = 'cosine'
+    beta_schedule: ListOrTuple(str) = 'cosine'
     learned_variance: bool = True
     image_cond_drop_prob: float = 0.1
     text_cond_drop_prob: float = 0.5
@@ -261,6 +261,7 @@ class TrainDecoderConfig(BaseModel):
     evaluate: DecoderEvaluateConfig
     tracker: TrackerConfig
     load: DecoderLoadConfig
+    seed: int = 0
 
     @classmethod
     def from_json_path(cls, json_path):

dalle2_pytorch/trainer.py

@@ -14,6 +14,8 @@ from dalle2_pytorch.optimizer import get_optimizer
 from dalle2_pytorch.version import __version__
 from packaging import version
 
+from accelerate import Accelerator
+
 import numpy as np
 
 # helper functions
@@ -22,7 +24,9 @@ def exists(val):
     return val is not None
 
 def default(val, d):
-    return val if exists(val) else d
+    if exists(val):
+        return val
+    return d() if callable(d) else d
 
 def cast_tuple(val, length = 1):
     return val if isinstance(val, tuple) else ((val,) * length)
@@ -58,8 +62,15 @@ def num_to_groups(num, divisor):
         arr.append(remainder)
     return arr
 
-def get_pkg_version():
-    return __version__
+def clamp(value, min_value = None, max_value = None):
+    assert exists(min_value) or exists(max_value)
+    if exists(min_value):
+        value = max(value, min_value)
+
+    if exists(max_value):
+        value = min(value, max_value)
+
+    return value
 
 # decorators
 
@@ -182,13 +193,13 @@ class EMA(nn.Module):
     By adjusting the power, you can control how fast EMA will ramp up to your specified beta.
 
     @crowsonkb's notes on EMA Warmup:
-    
+
     If gamma=1 and power=1, implements a simple average. gamma=1, power=2/3 are
     good values for models you plan to train for a million or more steps (reaches decay
     factor 0.999 at 31.6K steps, 0.9999 at 1M steps), gamma=1, power=3/4 for models
     you plan to train for less (reaches decay factor 0.999 at 10K steps, 0.9999 at
     215.4k steps).
-    
+
     Args:
         inv_gamma (float): Inverse multiplicative factor of EMA warmup. Default: 1.
         power (float): Exponential factor of EMA warmup. Default: 1.
@@ -198,7 +209,7 @@ class EMA(nn.Module):
         self,
         model,
         beta = 0.9999,
-        update_after_step = 10000,
+        update_after_step = 100,
         update_every = 10,
         inv_gamma = 1.0,
         power = 2/3,
@@ -227,10 +238,17 @@ class EMA(nn.Module):
         for ma_param, current_param in zip(list(self.ema_model.parameters()), list(self.online_model.parameters())):
             ma_param.data.copy_(current_param.data)
 
+        for ma_buffer, current_buffer in zip(list(self.ema_model.buffers()), list(self.online_model.buffers())):
+            ma_buffer.data.copy_(current_buffer.data)
+
     def get_current_decay(self):
-        epoch = max(0, self.step.item() - self.update_after_step - 1)
+        epoch = clamp(self.step.item() - self.update_after_step - 1, min_value = 0)
         value = 1 - (1 + epoch / self.inv_gamma) ** - self.power
-        return 0. if epoch < 0 else min(self.beta, max(self.min_value, value))
+
+        if epoch <= 0:
+            return 0.
+
+        return clamp(value, min_value = self.min_value, max_value = self.beta)
 
     def update(self):
         step = self.step.item()
@@ -266,6 +284,7 @@ class EMA(nn.Module):
     def __call__(self, *args, **kwargs):
         return self.ema_model(*args, **kwargs)
 
+
 # diffusion prior trainer
 
 def prior_sample_in_chunks(fn):
@@ -289,88 +308,189 @@ class DiffusionPriorTrainer(nn.Module):
         max_grad_norm = None,
         amp = False,
         group_wd_params = True,
+        device = None,
+        accelerator = None,
         **kwargs
     ):
         super().__init__()
         assert isinstance(diffusion_prior, DiffusionPrior)
+        assert not exists(accelerator) or isinstance(accelerator, Accelerator)
+        assert exists(accelerator) or exists(device), "You must supply some method of obtaining a device."
         ema_kwargs, kwargs = groupby_prefix_and_trim('ema_', kwargs)
 
+        # assign some helpful member vars
+        self.accelerator = accelerator
+        self.device = accelerator.device if exists(accelerator) else device
+        self.text_conditioned = diffusion_prior.condition_on_text_encodings
+
+        # save model
+
         self.diffusion_prior = diffusion_prior
 
-        # exponential moving average
-
-        self.use_ema = use_ema
-        if self.use_ema:
-            self.ema_diffusion_prior = EMA(diffusion_prior, **ema_kwargs)
-
         # optimizer and mixed precision stuff
 
         self.amp = amp
 
         self.scaler = GradScaler(enabled = amp)
 
+        self.optim_kwargs = dict(lr=lr, wd=wd, eps=eps, group_wd_params=group_wd_params)
+
         self.optimizer = get_optimizer(
-            diffusion_prior.parameters(),
-            lr = lr,
-            wd = wd,
-            eps = eps,
-            group_wd_params = group_wd_params,
+            self.diffusion_prior.parameters(),
+            **self.optim_kwargs,
             **kwargs
         )
 
+        # distribute the model if using HFA
+        if exists(self.accelerator):
+            self.diffusion_prior, self.optimizer = self.accelerator.prepare(self.diffusion_prior, self.optimizer)
+
+        # exponential moving average stuff
+
+        self.use_ema = use_ema
+
+        if self.use_ema:
+            self.ema_diffusion_prior = EMA(self.unwrap_model(self.diffusion_prior), **ema_kwargs)
+
         # gradient clipping if needed
 
         self.max_grad_norm = max_grad_norm
 
+        # track steps internally
+
         self.register_buffer('step', torch.tensor([0]))
 
+    # accelerator wrappers
+
+    def print(self, msg):
+        if exists(self.accelerator):
+            self.accelerator.print(msg)
+        else:
+            print(msg)
+
+    def unwrap_model(self, model):
+        if exists(self.accelerator):
+            return self.accelerator.unwrap_model(model)
+        else:
+            return model
+
+    def wait_for_everyone(self):
+        if exists(self.accelerator):
+            self.accelerator.wait_for_everyone()
+
+    def is_main_process(self):
+        if exists(self.accelerator):
+            return self.accelerator.is_main_process
+        else:
+            return True
+
+    def clip_grad_norm_(self, *args):
+        if exists(self.accelerator):
+            return self.accelerator.clip_grad_norm_(*args)
+        else:
+            return torch.nn.utils.clip_grad_norm_(*args)
+
+    def backprop(self, x):
+        if exists(self.accelerator):
+            self.accelerator.backward(x)
+        else:
+            try:
+                x.backward()
+            except Exception as e:
+                self.print(f"Caught error in backprop call: {e}")
+
+    # utility
+
     def save(self, path, overwrite = True, **kwargs):
-        path = Path(path)
-        assert not (path.exists() and not overwrite)
-        path.parent.mkdir(parents = True, exist_ok = True)
+        # ensure we sync gradients before continuing
+        self.wait_for_everyone()
 
-        save_obj = dict(
-            scaler = self.scaler.state_dict(),
-            optimizer = self.optimizer.state_dict(),
-            model = self.diffusion_prior.state_dict(),
-            version = __version__,
-            step = self.step.item(),
-            **kwargs
-        )
+        # only save on the main process
+        if self.is_main_process():
+            self.print(f"Saving checkpoint at step: {self.step.item()}")
+            path = Path(path)
+            assert not (path.exists() and not overwrite)
+            path.parent.mkdir(parents = True, exist_ok = True)
 
-        if self.use_ema:
-            save_obj = {**save_obj, 'ema': self.ema_diffusion_prior.state_dict()}
+            save_obj = dict(
+                scaler = self.scaler.state_dict(),
+                optimizer = self.optimizer.state_dict(),
+                model = self.unwrap_model(self.diffusion_prior).state_dict(), # unwrap the model from distribution if applicable
+                version = version.parse(__version__),
+                step = self.step.item(),
+                **kwargs
+            )
 
-        torch.save(save_obj, str(path))
+            if self.use_ema:
+                save_obj = {
+                    **save_obj,
+                    'ema': self.ema_diffusion_prior.state_dict(),
+                    'ema_model': self.ema_diffusion_prior.ema_model.state_dict() # save the ema model specifically for easy ema-only reload
+                }
 
-    def load(self, path, only_model = False, strict = True):
+            torch.save(save_obj, str(path))
+
+    def load(self, path, overwrite_lr = True, strict = True):
+        """
+        Load a checkpoint of a diffusion prior trainer.
+
+        Will load the entire trainer, including the optimizer and EMA.
+
+        Params:
+            - path (str): a path to the DiffusionPriorTrainer checkpoint file
+            - overwrite_lr (bool): wether or not to overwrite the stored LR with the LR specified in the new trainer
+            - strict (bool): kwarg for `torch.nn.Module.load_state_dict`, will force an exact checkpoint match
+
+        Returns:
+            loaded_obj (dict): The loaded checkpoint dictionary
+        """
+
+        # all processes need to load checkpoint. no restriction here
         path = Path(path)
         assert path.exists()
 
-        loaded_obj = torch.load(str(path))
+        loaded_obj = torch.load(str(path), map_location=self.device)
 
         if version.parse(__version__) != loaded_obj['version']:
             print(f'loading saved diffusion prior at version {loaded_obj["version"]} but current package version is at {__version__}')
 
-        self.diffusion_prior.load_state_dict(loaded_obj['model'], strict = strict)
+        # unwrap the model when loading from checkpoint
+        self.unwrap_model(self.diffusion_prior).load_state_dict(loaded_obj['model'], strict = strict)
         self.step.copy_(torch.ones_like(self.step) * loaded_obj['step'])
 
-        if only_model:
-            return loaded_obj
-
         self.scaler.load_state_dict(loaded_obj['scaler'])
         self.optimizer.load_state_dict(loaded_obj['optimizer'])
 
+        if overwrite_lr:
+            new_lr = self.optim_kwargs["lr"]
+
+            self.print(f"Overriding LR to be {new_lr}")
+
+            for group in self.optimizer.param_groups:
+                group["lr"] = new_lr
+
         if self.use_ema:
             assert 'ema' in loaded_obj
             self.ema_diffusion_prior.load_state_dict(loaded_obj['ema'], strict = strict)
+            # below not be necessary, but I had a suspicion that this wasn't being loaded correctly
+            self.ema_diffusion_prior.ema_model.load_state_dict(loaded_obj["ema_model"])
+
+        # sync and inform
+        self.wait_for_everyone()
+        self.print(f"Loaded model")
 
         return loaded_obj
 
+    # model functionality
+
     def update(self):
+        # only continue with updates until all ranks finish
+        self.wait_for_everyone()
+
         if exists(self.max_grad_norm):
             self.scaler.unscale_(self.optimizer)
-            nn.utils.clip_grad_norm_(self.diffusion_prior.parameters(), self.max_grad_norm)
+            # utilize HFA clipping where applicable
+            self.clip_grad_norm_(self.diffusion_prior.parameters(), self.max_grad_norm)
 
         self.scaler.step(self.optimizer)
         self.scaler.update()
@@ -385,17 +505,32 @@ class DiffusionPriorTrainer(nn.Module):
     @cast_torch_tensor
     @prior_sample_in_chunks
     def p_sample_loop(self, *args, **kwargs):
-        return self.ema_diffusion_prior.ema_model.p_sample_loop(*args, **kwargs)
+        if self.use_ema:
+            return self.ema_diffusion_prior.ema_model.p_sample_loop(*args, **kwargs)
+        else:
+            return self.diffusion_prior.p_sample_loop(*args, **kwargs)
 
     @torch.no_grad()
     @cast_torch_tensor
     @prior_sample_in_chunks
     def sample(self, *args, **kwargs):
-        return self.ema_diffusion_prior.ema_model.sample(*args, **kwargs)
+        if self.use_ema:
+            return self.ema_diffusion_prior.ema_model.sample(*args, **kwargs)
+        else:
+            return self.diffusion_prior.sample(*args, **kwargs)
 
     @torch.no_grad()
     def sample_batch_size(self, *args, **kwargs):
-        return self.ema_diffusion_prior.ema_model.sample_batch_size(*args, **kwargs)
+        if self.use_ema:
+            return self.ema_diffusion_prior.ema_model.sample_batch_size(*args, **kwargs)
+        else:
+            return self.diffusion_prior.sample_batch_size(*args, **kwargs)
+
+    @torch.no_grad()
+    @cast_torch_tensor
+    @prior_sample_in_chunks
+    def embed_text(self, *args, **kwargs):
+        return self.unwrap_model(self.diffusion_prior).clip.embed_text(*args, **kwargs)
 
     @cast_torch_tensor
     def forward(
@@ -413,8 +548,10 @@ class DiffusionPriorTrainer(nn.Module):
 
             total_loss += loss.item()
 
+            # backprop with accelerate if applicable
+
             if self.training:
-                self.scaler.scale(loss).backward()
+                self.backprop(self.scaler.scale(loss))
 
         return total_loss
 
@@ -440,6 +577,7 @@ class DecoderTrainer(nn.Module):
     def __init__(
         self,
         decoder,
+        accelerator = None,
         use_ema = True,
         lr = 1e-4,
         wd = 1e-2,
@@ -453,8 +591,9 @@ class DecoderTrainer(nn.Module):
         assert isinstance(decoder, Decoder)
         ema_kwargs, kwargs = groupby_prefix_and_trim('ema_', kwargs)
 
-        self.decoder = decoder
-        self.num_unets = len(self.decoder.unets)
+        self.accelerator = default(accelerator, Accelerator)
+
+        self.num_unets = len(decoder.unets)
 
         self.use_ema = use_ema
         self.ema_unets = nn.ModuleList([])
@@ -466,7 +605,9 @@ class DecoderTrainer(nn.Module):
 
         lr, wd, eps = map(partial(cast_tuple, length = self.num_unets), (lr, wd, eps))
 
-        for ind, (unet, unet_lr, unet_wd, unet_eps) in enumerate(zip(self.decoder.unets, lr, wd, eps)):
+        optimizers = []
+
+        for unet, unet_lr, unet_wd, unet_eps in zip(decoder.unets, lr, wd, eps):
             optimizer = get_optimizer(
                 unet.parameters(),
                 lr = unet_lr,
@@ -476,19 +617,20 @@ class DecoderTrainer(nn.Module):
                 **kwargs
             )
 
-            setattr(self, f'optim{ind}', optimizer) # cannot use pytorch ModuleList for some reason with optimizers
+            optimizers.append(optimizer)
 
             if self.use_ema:
                 self.ema_unets.append(EMA(unet, **ema_kwargs))
 
-            scaler = GradScaler(enabled = amp)
-            setattr(self, f'scaler{ind}', scaler)
-
         # gradient clipping if needed
 
         self.max_grad_norm = max_grad_norm
 
         self.register_buffer('step', torch.tensor([0.]))
+        results = list(self.accelerator.prepare(decoder, *optimizers))
+        self.decoder = results.pop(0)
+        for opt_ind in range(len(optimizers)):
+            setattr(self, f'optim{opt_ind}', results.pop(0))
 
     def save(self, path, overwrite = True, **kwargs):
         path = Path(path)
@@ -496,47 +638,42 @@ class DecoderTrainer(nn.Module):
         path.parent.mkdir(parents = True, exist_ok = True)
 
         save_obj = dict(
-            model = self.decoder.state_dict(),
+            model = self.accelerator.unwrap_model(self.decoder).state_dict(),
             version = __version__,
             step = self.step.item(),
             **kwargs
         )
 
         for ind in range(0, self.num_unets):
-            scaler_key = f'scaler{ind}'
-            optimizer_key = f'scaler{ind}'
-            scaler = getattr(self, scaler_key)
+            optimizer_key = f'optim{ind}'
             optimizer = getattr(self, optimizer_key)
-            save_obj = {**save_obj, scaler_key: scaler.state_dict(), optimizer_key: optimizer.state_dict()}
+            save_obj = {**save_obj, optimizer_key: self.accelerator.unwrap_model(optimizer).state_dict()}
 
         if self.use_ema:
             save_obj = {**save_obj, 'ema': self.ema_unets.state_dict()}
 
-        torch.save(save_obj, str(path))
+        self.accelerator.save(save_obj, str(path))
 
     def load(self, path, only_model = False, strict = True):
         path = Path(path)
         assert path.exists()
 
-        loaded_obj = torch.load(str(path))
+        loaded_obj = torch.load(str(path), map_location = 'cpu')
 
-        if version.parse(__version__) != loaded_obj['version']:
-            print(f'loading saved decoder at version {loaded_obj["version"]}, but current package version is {get_pkg_version()}')
+        if version.parse(__version__) != version.parse(loaded_obj['version']):
+            self.accelerator.print(f'loading saved decoder at version {loaded_obj["version"]}, but current package version is {__version__}')
 
-        self.decoder.load_state_dict(loaded_obj['model'], strict = strict)
+        self.accelerator.unwrap_model(self.decoder).load_state_dict(loaded_obj['model'], strict = strict)
         self.step.copy_(torch.ones_like(self.step) * loaded_obj['step'])
 
         if only_model:
             return loaded_obj
 
         for ind in range(0, self.num_unets):
-            scaler_key = f'scaler{ind}'
-            optimizer_key = f'scaler{ind}'
-            scaler = getattr(self, scaler_key)
+            optimizer_key = f'optim{ind}'
             optimizer = getattr(self, optimizer_key)
 
-            scaler.load_state_dict(loaded_obj[scaler_key])
-            optimizer.load_state_dict(loaded_obj[optimizer_key])
+            self.accelerator.unwrap_model(optimizer).load_state_dict(loaded_obj[optimizer_key])
 
         if self.use_ema:
             assert 'ema' in loaded_obj
@@ -548,29 +685,18 @@ class DecoderTrainer(nn.Module):
     def unets(self):
         return nn.ModuleList([ema.ema_model for ema in self.ema_unets])
 
-    def scale(self, loss, *, unet_number):
-        assert 1 <= unet_number <= self.num_unets
-        index = unet_number - 1
-        scaler = getattr(self, f'scaler{index}')
-        return scaler.scale(loss)
-
     def update(self, unet_number = None):
         if self.num_unets == 1:
             unet_number = default(unet_number, 1)
 
         assert exists(unet_number) and 1 <= unet_number <= self.num_unets
         index = unet_number - 1
-        unet = self.decoder.unets[index]
 
         optimizer = getattr(self, f'optim{index}')
-        scaler = getattr(self, f'scaler{index}')
 
         if exists(self.max_grad_norm):
-            scaler.unscale_(optimizer)
-            nn.utils.clip_grad_norm_(unet.parameters(), self.max_grad_norm)
-
-        scaler.step(optimizer)
-        scaler.update()
+            self.accelerator.clip_grad_norm_(self.decoder.parameters(), self.max_grad_norm)  # Automatically unscales gradients
+        optimizer.step()
         optimizer.zero_grad()
 
         if self.use_ema:
@@ -583,15 +709,17 @@ class DecoderTrainer(nn.Module):
     @cast_torch_tensor
     @decoder_sample_in_chunks
     def sample(self, *args, **kwargs):
+        distributed = self.accelerator.num_processes > 1
+        base_decoder = self.accelerator.unwrap_model(self.decoder)
         if kwargs.pop('use_non_ema', False) or not self.use_ema:
-            return self.decoder.sample(*args, **kwargs)
+            return base_decoder.sample(*args, **kwargs, distributed = distributed)
 
-        trainable_unets = self.decoder.unets
-        self.decoder.unets = self.unets                  # swap in exponential moving averaged unets for sampling
+        trainable_unets = self.accelerator.unwrap_model(self.decoder).unets
+        base_decoder.unets = self.unets                  # swap in exponential moving averaged unets for sampling
 
-        output = self.decoder.sample(*args, **kwargs)
+        output = base_decoder.sample(*args, **kwargs, distributed = distributed)
 
-        self.decoder.unets = trainable_unets             # restore original training unets
+        base_decoder.unets = trainable_unets             # restore original training unets
 
         # cast the ema_model unets back to original device
         for ema in self.ema_unets:
@@ -613,13 +741,14 @@ class DecoderTrainer(nn.Module):
         total_loss = 0.
 
         for chunk_size_frac, (chunked_args, chunked_kwargs) in split_args_and_kwargs(*args, split_size = max_batch_size, **kwargs):
-            with autocast(enabled = self.amp):
+            # with autocast(enabled = self.amp):
+            with self.accelerator.autocast():
                 loss = self.decoder(*chunked_args, unet_number = unet_number, **chunked_kwargs)
                 loss = loss * chunk_size_frac
 
             total_loss += loss.item()
 
             if self.training:
-                self.scale(loss, unet_number = unet_number).backward()
+                self.accelerator.backward(loss)
 
         return total_loss

dalle2_pytorch/utils.py

@@ -1,4 +1,5 @@
 import time
+import importlib
 
 # time helpers
 

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '0.6.13'
+__version__ = '0.11.0'

dalle2_pytorch/vqgan_vae.py

@@ -68,8 +68,8 @@ def group_dict_by_key(cond, d):
         return_val[ind][key] = d[key]
     return (*return_val,)
 
-def string_begins_with(prefix, str):
-    return str.startswith(prefix)
+def string_begins_with(prefix, string_input):
+    return string_input.startswith(prefix)
 
 def group_by_key_prefix(prefix, d):
     return group_dict_by_key(partial(string_begins_with, prefix), d)

setup.py

@@ -24,6 +24,7 @@ setup(
     'text to image'
   ],
   install_requires=[
+    'accelerate',
     'click',
     'clip-anytorch',
     'coca-pytorch>=0.0.5',

train_decoder.py

@@ -1,7 +1,8 @@
-from dalle2_pytorch import Unet, Decoder
+from pathlib import Path
+
 from dalle2_pytorch.trainer import DecoderTrainer
 from dalle2_pytorch.dataloaders import create_image_embedding_dataloader
-from dalle2_pytorch.trackers import WandbTracker, ConsoleTracker
+from dalle2_pytorch.trackers import WandbTracker, ConsoleTracker, DummyTracker
 from dalle2_pytorch.train_configs import TrainDecoderConfig
 from dalle2_pytorch.utils import Timer, print_ribbon
 from dalle2_pytorch.dalle2_pytorch import resize_image_to
@@ -12,6 +13,8 @@ from torchmetrics.image.fid import FrechetInceptionDistance
 from torchmetrics.image.inception import InceptionScore
 from torchmetrics.image.kid import KernelInceptionDistance
 from torchmetrics.image.lpip import LearnedPerceptualImagePatchSimilarity
+from accelerate import Accelerator, DistributedDataParallelKwargs
+from accelerate.utils import dataclasses as accelerate_dataclasses
 import webdataset as wds
 import click
 
@@ -42,6 +45,7 @@ def create_dataloaders(
     train_prop = 0.75,
     val_prop = 0.15,
     test_prop = 0.10,
+    seed = 0,
     **kwargs
 ):
     """
@@ -52,7 +56,7 @@ def create_dataloaders(
     num_test = round(test_prop*len(available_shards))
     num_val = len(available_shards) - num_train - num_test
     assert num_train + num_test + num_val == len(available_shards), f"{num_train} + {num_test} + {num_val} = {num_train + num_test + num_val} != {len(available_shards)}"
-    train_split, test_split, val_split = torch.utils.data.random_split(available_shards, [num_train, num_test, num_val], generator=torch.Generator().manual_seed(0))
+    train_split, test_split, val_split = torch.utils.data.random_split(available_shards, [num_train, num_test, num_val], generator=torch.Generator().manual_seed(seed))
 
     # The shard number in the webdataset file names has a fixed width. We zero pad the shard numbers so they correspond to a filename.
     train_urls = [webdataset_base_url.format(str(shard).zfill(shard_width)) for shard in train_split]
@@ -117,7 +121,6 @@ def get_example_data(dataloader, device, n=5):
         captions.extend(list(txt))
         if len(images) >= n:
             break
-    print("Generated {} examples".format(len(images)))
     return list(zip(images[:n], embeddings[:n], captions[:n]))
 
 def generate_samples(trainer, example_data, text_prepend=""):
@@ -155,27 +158,34 @@ def evaluate_trainer(trainer, dataloader, device, n_evaluation_samples=1000, FID
     metrics = {}
     # Prepare the data
     examples = get_example_data(dataloader, device, n_evaluation_samples)
+    if len(examples) == 0:
+        print("No data to evaluate. Check that your dataloader has shards.")
+        return metrics
     real_images, generated_images, captions = generate_samples(trainer, examples)
     real_images = torch.stack(real_images).to(device=device, dtype=torch.float)
     generated_images = torch.stack(generated_images).to(device=device, dtype=torch.float)
     # Convert from [0, 1] to [0, 255] and from torch.float to torch.uint8
     int_real_images = real_images.mul(255).add(0.5).clamp(0, 255).type(torch.uint8)
     int_generated_images = generated_images.mul(255).add(0.5).clamp(0, 255).type(torch.uint8)
+
+    def null_sync(t, *args, **kwargs):
+        return [t]
+
     if exists(FID):
-        fid = FrechetInceptionDistance(**FID)
+        fid = FrechetInceptionDistance(**FID, dist_sync_fn=null_sync)
         fid.to(device=device)
         fid.update(int_real_images, real=True)
         fid.update(int_generated_images, real=False)
         metrics["FID"] = fid.compute().item()
     if exists(IS):
-        inception = InceptionScore(**IS)
+        inception = InceptionScore(**IS, dist_sync_fn=null_sync)
         inception.to(device=device)
         inception.update(int_real_images)
         is_mean, is_std = inception.compute()
         metrics["IS_mean"] = is_mean.item()
         metrics["IS_std"] = is_std.item()
     if exists(KID):
-        kernel_inception = KernelInceptionDistance(**KID)
+        kernel_inception = KernelInceptionDistance(**KID, dist_sync_fn=null_sync)
         kernel_inception.to(device=device)
         kernel_inception.update(int_real_images, real=True)
         kernel_inception.update(int_generated_images, real=False)
@@ -186,39 +196,47 @@ def evaluate_trainer(trainer, dataloader, device, n_evaluation_samples=1000, FID
         # Convert from [0, 1] to [-1, 1]
         renorm_real_images = real_images.mul(2).sub(1)
         renorm_generated_images = generated_images.mul(2).sub(1)
-        lpips = LearnedPerceptualImagePatchSimilarity(**LPIPS)
+        lpips = LearnedPerceptualImagePatchSimilarity(**LPIPS, dist_sync_fn=null_sync)
         lpips.to(device=device)
         lpips.update(renorm_real_images, renorm_generated_images)
         metrics["LPIPS"] = lpips.compute().item()
+
+    if trainer.accelerator.num_processes > 1:
+        # Then we should sync the metrics
+        metrics_order = sorted(metrics.keys())
+        metrics_tensor = torch.zeros(1, len(metrics), device=device, dtype=torch.float)
+        for i, metric_name in enumerate(metrics_order):
+            metrics_tensor[0, i] = metrics[metric_name]
+        metrics_tensor = trainer.accelerator.gather(metrics_tensor)
+        metrics_tensor = metrics_tensor.mean(dim=0)
+        for i, metric_name in enumerate(metrics_order):
+            metrics[metric_name] = metrics_tensor[i].item()
     return metrics
 
-def save_trainer(tracker, trainer, epoch, step, validation_losses, relative_paths):
+def save_trainer(tracker, trainer, epoch, sample, next_task, validation_losses, relative_paths):
     """
     Logs the model with an appropriate method depending on the tracker
     """
     if isinstance(relative_paths, str):
         relative_paths = [relative_paths]
-    trainer_state_dict = {}
-    trainer_state_dict["trainer"] = trainer.state_dict()
-    trainer_state_dict['epoch'] = epoch
-    trainer_state_dict['step'] = step
-    trainer_state_dict['validation_losses'] = validation_losses
     for relative_path in relative_paths:
-        tracker.save_state_dict(trainer_state_dict, relative_path)
+        local_path = str(tracker.data_path / relative_path)
+        trainer.save(local_path, epoch=epoch, sample=sample, next_task=next_task, validation_losses=validation_losses)
+        tracker.save_file(local_path)
     
 def recall_trainer(tracker, trainer, recall_source=None, **load_config):
     """
     Loads the model with an appropriate method depending on the tracker
     """
-    print(print_ribbon(f"Loading model from {recall_source}"))
-    state_dict = tracker.recall_state_dict(recall_source, **load_config)
-    trainer.load_state_dict(state_dict["trainer"])
-    print("Model loaded")
-    return state_dict["epoch"], state_dict["step"], state_dict["validation_losses"]
+    trainer.accelerator.print(print_ribbon(f"Loading model from {recall_source}"))
+    local_filepath = tracker.recall_file(recall_source, **load_config)
+    state_dict = trainer.load(local_filepath)
+    return state_dict.get("epoch", 0), state_dict.get("validation_losses", []), state_dict.get("next_task", "train"), state_dict.get("sample", 0)
 
 def train(
     dataloaders,
     decoder,
+    accelerator,
     tracker,
     inference_device,
     load_config=None,
@@ -237,17 +255,30 @@ def train(
     """
     Trains a decoder on a dataset.
     """
-    trainer = DecoderTrainer(  # TODO: Change the get_optimizer function so that it can take arbitrary named args so we can just put **kwargs as an argument here
+    is_master = accelerator.process_index == 0
+
+    trainer = DecoderTrainer(
+        accelerator,
         decoder,
         **kwargs
     )
+
     # Set up starting model and parameters based on a recalled state dict
-    start_step = 0
     start_epoch = 0
     validation_losses = []
+    next_task = 'train'
+    sample = 0
+    val_sample = 0
+    step = lambda: int(trainer.step.item())
 
     if exists(load_config) and exists(load_config.source):
-        start_epoch, start_step, validation_losses = recall_trainer(tracker, trainer, recall_source=load_config.source, **load_config)
+        start_epoch, validation_losses, next_task, recalled_sample = recall_trainer(tracker, trainer, recall_source=load_config.source, **load_config.dict())
+        if next_task == 'train':
+            sample = recalled_sample
+        if next_task == 'val':
+            val_sample = recalled_sample
+        accelerator.print(f"Loaded model from {load_config.source} on epoch {start_epoch} with minimum validation loss {min(validation_losses) if len(validation_losses) > 0 else 'N/A'}")
+        accelerator.print(f"Starting training from task {next_task} at sample {sample} and validation sample {val_sample}")
     trainer.to(device=inference_device)
 
     if not exists(unet_training_mask):
@@ -255,139 +286,185 @@ def train(
         unet_training_mask = [True] * trainer.num_unets
     assert len(unet_training_mask) == trainer.num_unets, f"The unet training mask should be the same length as the number of unets in the decoder. Got {len(unet_training_mask)} and {trainer.num_unets}"
 
-    print(print_ribbon("Generating Example Data", repeat=40))
-    print("This can take a while to load the shard lists...")
-    train_example_data = get_example_data(dataloaders["train_sampling"], inference_device, n_sample_images)
-    test_example_data = get_example_data(dataloaders["test_sampling"], inference_device, n_sample_images)
+    accelerator.print(print_ribbon("Generating Example Data", repeat=40))
+    accelerator.print("This can take a while to load the shard lists...")
+    if is_master:
+        train_example_data = get_example_data(dataloaders["train_sampling"], inference_device, n_sample_images)
+        accelerator.print("Generated training examples")
+        test_example_data = get_example_data(dataloaders["test_sampling"], inference_device, n_sample_images)
+        accelerator.print("Generated testing examples")
     
     send_to_device = lambda arr: [x.to(device=inference_device, dtype=torch.float) for x in arr]
-    step = start_step
 
+    sample_length_tensor = torch.zeros(1, dtype=torch.int, device=inference_device)
+    unet_losses_tensor = torch.zeros(TRAIN_CALC_LOSS_EVERY_ITERS, trainer.num_unets, dtype=torch.float, device=inference_device)
     for epoch in range(start_epoch, epochs):
-        print(print_ribbon(f"Starting epoch {epoch}", repeat=40))
+        accelerator.print(print_ribbon(f"Starting epoch {epoch}", repeat=40))
 
         timer = Timer()
+        last_sample = sample
+        last_snapshot = sample
 
-        sample = 0
-        last_sample = 0
-        last_snapshot = 0
+        if next_task == 'train':
+            for i, (img, emb) in enumerate(dataloaders["train"]):
+                # We want to count the total number of samples across all processes
+                sample_length_tensor[0] = len(img)
+                all_samples = accelerator.gather(sample_length_tensor)  # TODO: accelerator.reduce is broken when this was written. If it is fixed replace this.
+                total_samples = all_samples.sum().item()
+                sample += total_samples
+                img, emb = send_to_device((img, emb))
 
-        losses = []
+                trainer.train()
+                for unet in range(1, trainer.num_unets+1):
+                    # Check if this is a unet we are training
+                    if not unet_training_mask[unet-1]: # Unet index is the unet number - 1
+                        continue
 
-        for i, (img, emb) in enumerate(dataloaders["train"]):
-            step += 1
-            sample += img.shape[0]
-            img, emb = send_to_device((img, emb))
-            
-            trainer.train()
-            for unet in range(1, trainer.num_unets+1):
-                # Check if this is a unet we are training
-                if not unet_training_mask[unet-1]: # Unet index is the unet number - 1
-                    continue
+                    loss = trainer.forward(img, image_embed=emb, unet_number=unet)
+                    trainer.update(unet_number=unet)
+                    unet_losses_tensor[i % TRAIN_CALC_LOSS_EVERY_ITERS, unet-1] = loss
+                
+                samples_per_sec = (sample - last_sample) / timer.elapsed()
+                timer.reset()
+                last_sample = sample
 
-                loss = trainer.forward(img, image_embed=emb, unet_number=unet)
-                trainer.update(unet_number=unet)
-                losses.append(loss)
+                if i % TRAIN_CALC_LOSS_EVERY_ITERS == 0:
+                    # We want to average losses across all processes
+                    unet_all_losses = accelerator.gather(unet_losses_tensor)
+                    mask = unet_all_losses != 0
+                    unet_average_loss = (unet_all_losses * mask).sum(dim=0) / mask.sum(dim=0)
+                    loss_map = { f"Unet {index} Training Loss": loss.item() for index, loss in enumerate(unet_average_loss) if loss != 0 }
+                    log_data = {
+                        "Epoch": epoch,
+                        "Sample": sample,
+                        "Step": i,
+                        "Samples per second": samples_per_sec,
+                        **loss_map
+                    }
+                    # print(f"I am rank {accelerator.state.process_index}. Example weight: {trainer.decoder.state_dict()['module.unets.0.init_conv.convs.0.weight'][0,0,0,0]}")
+                    if is_master:
+                        tracker.log(log_data, step=step(), verbose=True)
 
-            samples_per_sec = (sample - last_sample) / timer.elapsed()
+                if is_master and last_snapshot + save_every_n_samples < sample:  # This will miss by some amount every time, but it's not a big deal... I hope
+                    # It is difficult to gather this kind of info on the accelerator, so we have to do it on the master
+                    print("Saving snapshot")
+                    last_snapshot = sample
+                    # We need to know where the model should be saved
+                    save_paths = []
+                    if save_latest:
+                        save_paths.append("latest.pth")
+                    if save_all:
+                        save_paths.append(f"checkpoints/epoch_{epoch}_step_{step()}.pth")
+                    save_trainer(tracker, trainer, epoch, sample, next_task, validation_losses, save_paths)
+                    if exists(n_sample_images) and n_sample_images > 0:
+                        trainer.eval()
+                        train_images, train_captions = generate_grid_samples(trainer, train_example_data, "Train: ")
+                        tracker.log_images(train_images, captions=train_captions, image_section="Train Samples", step=step())
+                
+                if epoch_samples is not None and sample >= epoch_samples:
+                    break
+            next_task = 'val'
+            sample = 0
 
-            timer.reset()
-            last_sample = sample
+        all_average_val_losses = None
+        if next_task == 'val':
+            trainer.eval()
+            accelerator.print(print_ribbon(f"Starting Validation {epoch}", repeat=40))
+            last_val_sample = val_sample
+            val_sample_length_tensor = torch.zeros(1, dtype=torch.int, device=inference_device)
+            average_val_loss_tensor = torch.zeros(1, trainer.num_unets, dtype=torch.float, device=inference_device)
+            timer = Timer()
+            accelerator.wait_for_everyone()
+            i = 0
+            for i, (img, emb, txt) in enumerate(dataloaders["val"]):
+                val_sample_length_tensor[0] = len(img)
+                all_samples = accelerator.gather(val_sample_length_tensor)
+                total_samples = all_samples.sum().item()
+                val_sample += total_samples
+                img, emb = send_to_device((img, emb))
 
-            if i % TRAIN_CALC_LOSS_EVERY_ITERS == 0:
-                average_loss = sum(losses) / len(losses)
-                log_data = {
-                    "Training loss": average_loss,
-                    "Epoch": epoch,
-                    "Sample": sample,
-                    "Step": i,
-                    "Samples per second": samples_per_sec
-                }
-                tracker.log(log_data, step=step, verbose=True)
-                losses = []
+                for unet in range(1, len(decoder.unets)+1):
+                    if not unet_training_mask[unet-1]: # Unet index is the unet number - 1
+                        # No need to evaluate an unchanging unet
+                        continue
+                    
+                    loss = trainer.forward(img.float(), image_embed=emb.float(), unet_number=unet)
+                    average_val_loss_tensor[0, unet-1] += loss
 
-            if last_snapshot + save_every_n_samples < sample:  # This will miss by some amount every time, but it's not a big deal... I hope
-                last_snapshot = sample
-                # We need to know where the model should be saved
+                if i % VALID_CALC_LOSS_EVERY_ITERS == 0:
+                    samples_per_sec = (val_sample - last_val_sample) / timer.elapsed()
+                    timer.reset()
+                    last_val_sample = val_sample
+                    accelerator.print(f"Epoch {epoch}/{epochs} Val Step {i} -  Sample {val_sample} - {samples_per_sec:.2f} samples/sec")
+                    accelerator.print(f"Loss: {(average_val_loss_tensor / (i+1))}")
+                    accelerator.print("")
+                
+                if validation_samples is not None and val_sample >= validation_samples:
+                    break
+            print(f"Rank {accelerator.state.process_index} finished validation after {i} steps")
+            accelerator.wait_for_everyone()
+            average_val_loss_tensor /= i+1
+            # Gather all the average loss tensors
+            all_average_val_losses = accelerator.gather(average_val_loss_tensor)
+            if is_master:
+                unet_average_val_loss = all_average_val_losses.mean(dim=0)
+                val_loss_map = { f"Unet {index} Validation Loss": loss.item() for index, loss in enumerate(unet_average_val_loss) if loss != 0 }
+                tracker.log(val_loss_map, step=step(), verbose=True)
+            next_task = 'eval'
+
+        if next_task == 'eval':
+            if exists(evaluate_config):
+                accelerator.print(print_ribbon(f"Starting Evaluation {epoch}", repeat=40))
+                evaluation = evaluate_trainer(trainer, dataloaders["val"], inference_device, **evaluate_config.dict())
+                if is_master:
+                    tracker.log(evaluation, step=step(), verbose=True)
+            next_task = 'sample'
+            val_sample = 0
+
+        if next_task == 'sample':
+            if is_master:
+                # Generate examples and save the model if we are the master
+                # Generate sample images
+                print(print_ribbon(f"Sampling Set {epoch}", repeat=40))
+                test_images, test_captions = generate_grid_samples(trainer, test_example_data, "Test: ")
+                train_images, train_captions = generate_grid_samples(trainer, train_example_data, "Train: ")
+                tracker.log_images(test_images, captions=test_captions, image_section="Test Samples", step=step())
+                tracker.log_images(train_images, captions=train_captions, image_section="Train Samples", step=step())
+
+                print(print_ribbon(f"Starting Saving {epoch}", repeat=40))
+                # Get the same paths
                 save_paths = []
                 if save_latest:
                     save_paths.append("latest.pth")
-                if save_all:
-                    save_paths.append(f"checkpoints/epoch_{epoch}_step_{step}.pth")
+                if all_average_val_losses is not None:
+                    average_loss = all_average_val_losses.mean(dim=0).item()
+                    if save_best and (len(validation_losses) == 0 or average_loss < min(validation_losses)):
+                        save_paths.append("best.pth")
+                    validation_losses.append(average_loss)
+                save_trainer(tracker, trainer, epoch, sample, next_task, validation_losses, save_paths)
+            next_task = 'train'
 
-                save_trainer(tracker, trainer, epoch, step, validation_losses, save_paths)
-
-                if exists(n_sample_images) and n_sample_images > 0:
-                    trainer.eval()
-                    train_images, train_captions = generate_grid_samples(trainer, train_example_data, "Train: ")
-                    tracker.log_images(train_images, captions=train_captions, image_section="Train Samples", step=step)
-
-            if exists(epoch_samples) and sample >= epoch_samples:
-                break
-
-        trainer.eval()
-        print(print_ribbon(f"Starting Validation {epoch}", repeat=40))
-        with torch.no_grad():
-            sample = 0
-            average_loss = 0
-            timer = Timer()
-            for i, (img, emb, *_) in enumerate(dataloaders["val"]):
-                sample += img.shape[0]
-                img, emb = send_to_device((img, emb))
-                
-                for unet in range(1, len(decoder.unets)+1):
-                    loss = trainer.forward(img.float(), image_embed=emb.float(), unet_number=unet)
-                    average_loss += loss
-
-                if i % VALID_CALC_LOSS_EVERY_ITERS == 0:
-                    print(f"Epoch {epoch}/{epochs} - {sample / timer.elapsed():.2f} samples/sec")
-                    print(f"Loss: {average_loss / (i+1)}")
-                    print("")
-
-                if exists(validation_samples) and sample >= validation_samples:
-                    break
-
-            average_loss /= i+1
-            log_data = {
-                "Validation loss": average_loss
-            }
-            tracker.log(log_data, step=step, verbose=True)
-
-        # Compute evaluation metrics
-        if exists(evaluate_config):
-            print(print_ribbon(f"Starting Evaluation {epoch}", repeat=40))
-            evaluation = evaluate_trainer(trainer, dataloaders["val"], inference_device, **evaluate_config.dict())
-            tracker.log(evaluation, step=step, verbose=True)
-
-        # Generate sample images
-        print(print_ribbon(f"Sampling Set {epoch}", repeat=40))
-        test_images, test_captions = generate_grid_samples(trainer, test_example_data, "Test: ")
-        train_images, train_captions = generate_grid_samples(trainer, train_example_data, "Train: ")
-        tracker.log_images(test_images, captions=test_captions, image_section="Test Samples", step=step)
-        tracker.log_images(train_images, captions=train_captions, image_section="Train Samples", step=step)
-
-        print(print_ribbon(f"Starting Saving {epoch}", repeat=40))
-        # Get the same paths
-        save_paths = []
-        if save_latest:
-            save_paths.append("latest.pth")
-        if save_best and (len(validation_losses) == 0 or average_loss < min(validation_losses)):
-            save_paths.append("best.pth")
-        validation_losses.append(average_loss)
-        save_trainer(tracker, trainer, epoch, step, validation_losses, save_paths)
-
-def create_tracker(config, tracker_type=None, data_path=None, **kwargs):
+def create_tracker(accelerator, config, config_path, tracker_type=None, data_path=None):
     """
     Creates a tracker of the specified type and initializes special features based on the full config
     """
     tracker_config = config.tracker
-    init_config = {}
+    accelerator_config = {
+        "Distributed": accelerator.distributed_type != accelerate_dataclasses.DistributedType.NO,
+        "DistributedType": accelerator.distributed_type,
+        "NumProcesses": accelerator.num_processes,
+        "MixedPrecision": accelerator.mixed_precision
+    }
+    init_config = { "config": {**config.dict(), **accelerator_config} }
+    data_path = data_path or tracker_config.data_path
+    tracker_type = tracker_type or tracker_config.tracker_type
 
-    if exists(tracker_config.init_config):
-        init_config["config"] = tracker_config.init_config
-
-    if tracker_type == "console":
-        tracker = ConsoleTracker(**init_config)
+    if tracker_type == "dummy":
+        tracker = DummyTracker(data_path)
+        tracker.init(**init_config)
+    elif tracker_type == "console":
+        tracker = ConsoleTracker(data_path)
+        tracker.init(**init_config)
     elif tracker_type == "wandb":
         # We need to initialize the resume state here
         load_config = config.load
@@ -401,51 +478,63 @@ def create_tracker(config, tracker_type=None, data_path=None, **kwargs):
         init_config["project"] = tracker_config.wandb_project
         tracker = WandbTracker(data_path)
         tracker.init(**init_config)
+        tracker.save_file(str(config_path.absolute()), str(config_path.parent.absolute()))
     else:
         raise ValueError(f"Tracker type {tracker_type} not supported by decoder trainer")
     return tracker
     
-def initialize_training(config):
-    # Create the save path
-    if "cuda" in config.train.device:
-        assert torch.cuda.is_available(), "CUDA is not available"
-    device = torch.device(config.train.device)
-    torch.cuda.set_device(device)
-    all_shards = list(range(config.data.start_shard, config.data.end_shard + 1))
+def initialize_training(config, config_path):
+    # Make sure if we are not loading, distributed models are initialized to the same values
+    torch.manual_seed(config.seed)
 
+    # Set up accelerator for configurable distributed training
+    ddp_kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
+    accelerator = Accelerator(kwargs_handlers=[ddp_kwargs])
+    
+    # Set up data
+    all_shards = list(range(config.data.start_shard, config.data.end_shard + 1))
+    world_size = accelerator.num_processes
+    rank = accelerator.process_index
+    shards_per_process = len(all_shards) // world_size
+    assert shards_per_process > 0, "Not enough shards to split evenly"
+    my_shards = all_shards[rank * shards_per_process: (rank + 1) * shards_per_process]
     dataloaders = create_dataloaders (
-        available_shards=all_shards,
+        available_shards=my_shards,
         img_preproc = config.data.img_preproc,
         train_prop = config.data.splits.train,
         val_prop = config.data.splits.val,
         test_prop = config.data.splits.test,
         n_sample_images=config.train.n_sample_images,
-        **config.data.dict()
+        **config.data.dict(),
+        rank = rank,
+        seed = config.seed,
     )
 
-    decoder = config.decoder.create().to(device = device)
+    # Create the decoder model and print basic info
+    decoder = config.decoder.create()
     num_parameters = sum(p.numel() for p in decoder.parameters())
-    print(print_ribbon("Loaded Config", repeat=40))
-    print(f"Number of parameters: {num_parameters}")
 
-    tracker = create_tracker(config, **config.tracker.dict())
+    # Create and initialize the tracker if we are the master
+    tracker = create_tracker(accelerator, config, config_path) if rank == 0 else create_tracker(accelerator, config, config_path, tracker_type="dummy")
 
-    train(dataloaders, decoder, 
+    accelerator.print(print_ribbon("Loaded Config", repeat=40))
+    accelerator.print(f"Running training with {accelerator.num_processes} processes and {accelerator.distributed_type} distributed training")
+    accelerator.print(f"Number of parameters: {num_parameters}")
+    train(dataloaders, decoder, accelerator,
         tracker=tracker,
-        inference_device=device,
+        inference_device=accelerator.device,
         load_config=config.load,
         evaluate_config=config.evaluate,
         **config.train.dict(),
     )
-
+    
 # Create a simple click command line interface to load the config and start the training
 @click.command()
 @click.option("--config_file", default="./train_decoder_config.json", help="Path to config file")
 def main(config_file):
-    print("Recalling config from {}".format(config_file))
-    config = TrainDecoderConfig.from_json_path(config_file)
-    initialize_training(config)
-
+    config_file_path = Path(config_file)
+    config = TrainDecoderConfig.from_json_path(str(config_file_path))
+    initialize_training(config, config_path=config_file_path)
 
 if __name__ == "__main__":
     main()

train_diffusion_prior.py

@@ -1,77 +1,135 @@
-from pathlib import Path
+# TODO: add start, num_data_points, eval_every and group to config
+# TODO: switch back to repo's wandb
+
+START = 0
+NUM_DATA_POINTS = 250e6
+EVAL_EVERY = 1000
+GROUP = "distributed"
+
+import os
 import click
-import math
-import numpy as np
+import wandb
 
 import torch
-import clip
 from torch import nn
+from torch.utils.data import DataLoader
 
-from dalle2_pytorch.dataloaders import make_splits, get_reader
-from dalle2_pytorch import DiffusionPrior, DiffusionPriorNetwork, OpenAIClipAdapter
-from dalle2_pytorch.trainer import DiffusionPriorTrainer, load_diffusion_model, save_diffusion_model
+import numpy as np
 
-from dalle2_pytorch.trackers import ConsoleTracker, WandbTracker
-from dalle2_pytorch.utils import Timer, print_ribbon
+from accelerate import Accelerator
 
-from tqdm import tqdm
+from dalle2_pytorch.dataloaders import get_reader, make_splits
+from dalle2_pytorch.utils import Timer
+from dalle2_pytorch.train_configs import (
+    DiffusionPriorTrainConfig,
+    TrainDiffusionPriorConfig,
+)
+from dalle2_pytorch.trackers import BaseTracker, WandbTracker
+from dalle2_pytorch import DiffusionPriorTrainer
 
-# constants
 
-REPORT_METRICS_EVERY = 250 # for cosine similarity and other metric reporting during training
+# helpers
 
-tracker = WandbTracker()
 
-# helpers functions
+cos = nn.CosineSimilarity(dim=1, eps=1e-6)
+
 
 def exists(val):
-    val is not None
+    return val is not None
 
-# functions
 
-def eval_model(model, dataloader, text_conditioned, loss_type, device, phase="Validation",):
-    model.eval()
+def make_model(
+    prior_config, train_config, device: str = None, accelerator: Accelerator = None
+):
+    # create model from config
+    diffusion_prior = prior_config.create()
+
+    # instantiate the trainer
+    trainer = DiffusionPriorTrainer(
+        diffusion_prior=diffusion_prior,
+        lr=train_config.lr,
+        wd=train_config.wd,
+        max_grad_norm=train_config.max_grad_norm,
+        amp=train_config.amp,
+        use_ema=train_config.use_ema,
+        device=device,
+        accelerator=accelerator,
+    )
+
+    return trainer
+
+
+# eval functions
+
+
+def eval_model(
+    trainer: DiffusionPriorTrainer,
+    dataloader: DataLoader,
+    text_conditioned: bool,
+    loss_type: str,
+    tracker_context: str,
+    tracker: BaseTracker = None,
+    use_ema: bool = True,
+):
+    trainer.eval()
+    if trainer.is_main_process():
+        click.secho(f"Measuring performance on {tracker_context}", fg="green", blink=True)
 
     with torch.no_grad():
-        total_loss = 0.
-        total_samples = 0.
+        total_loss = 0.0
+        total_samples = 0.0
 
-        for image_embeddings, text_data in tqdm(dataloader):
-            image_embeddings = image_embeddings.to(device)
-            text_data = text_data.to(device)
+        for image_embeddings, text_data in dataloader:
+            image_embeddings = image_embeddings.to(trainer.device)
+            text_data = text_data.to(trainer.device)
 
             batches = image_embeddings.shape[0]
 
             input_args = dict(image_embed=image_embeddings)
+
             if text_conditioned:
-                input_args = dict(**input_args, text = text_data)
+                input_args = dict(**input_args, text=text_data)
             else:
                 input_args = dict(**input_args, text_embed=text_data)
 
-            loss = model(**input_args)
+            if use_ema:
+                loss = trainer.ema_diffusion_prior(**input_args)
+            else:
+                loss = trainer(**input_args)
 
             total_loss += loss * batches
             total_samples += batches
 
-        avg_loss = (total_loss / total_samples)
+        avg_loss = total_loss / total_samples
 
-        tracker.log({f'{phase} {loss_type}': avg_loss})
+        stats = {f"{tracker_context}-{loss_type}": avg_loss}
+        trainer.print(stats)
 
-def report_cosine_sims(diffusion_prior, dataloader, text_conditioned, device):
-    diffusion_prior.eval()
+        if exists(tracker):
+            tracker.log(stats, step=trainer.step.item() + 1)
 
-    cos = nn.CosineSimilarity(dim=1, eps=1e-6)
 
-    for test_image_embeddings, text_data in tqdm(dataloader):
-        test_image_embeddings = test_image_embeddings.to(device)
-        text_data = text_data.to(device)
+def report_cosine_sims(
+    trainer: DiffusionPriorTrainer,
+    dataloader: DataLoader,
+    text_conditioned: bool,
+    tracker: BaseTracker,
+    tracker_context: str = "validation",
+):
+    trainer.eval()
+    if trainer.is_main_process():
+        click.secho("Measuring Cosine-Similarity", fg="green", blink=True)
+
+    for test_image_embeddings, text_data in dataloader:
+        test_image_embeddings = test_image_embeddings.to(trainer.device)
+        text_data = text_data.to(trainer.device)
 
         # we are text conditioned, we produce an embedding from the tokenized text
         if text_conditioned:
-            text_embedding, text_encodings, text_mask = diffusion_prior.clip.embed_text(
-                text_data)
-            text_cond = dict(text_embed=text_embedding,
-                             text_encodings=text_encodings, mask=text_mask)
+            text_embedding, text_encodings, text_mask = trainer.embed_text(text_data)
+            text_cond = dict(
+                text_embed=text_embedding, text_encodings=text_encodings, mask=text_mask
+            )
         else:
             text_embedding = text_data
             text_cond = dict(text_embed=text_embedding)
@@ -82,8 +140,9 @@ def report_cosine_sims(diffusion_prior, dataloader, text_conditioned, device):
         # roll the text to simulate "unrelated" captions
         rolled_idx = torch.roll(torch.arange(text_embedding.shape[0]), 1)
         text_embed_shuffled = text_embed_shuffled[rolled_idx]
-        text_embed_shuffled = text_embed_shuffled / \
-            text_embed_shuffled.norm(dim=1, keepdim=True)
+        text_embed_shuffled = text_embed_shuffled / text_embed_shuffled.norm(
+            dim=1, keepdim=True
+        )
 
         if text_conditioned:
             text_encodings_shuffled = text_encodings[rolled_idx]
@@ -92,294 +151,276 @@ def report_cosine_sims(diffusion_prior, dataloader, text_conditioned, device):
             text_encodings_shuffled = None
             text_mask_shuffled = None
 
-        text_cond_shuffled = dict(text_embed=text_embed_shuffled,
-                                  text_encodings=text_encodings_shuffled, mask=text_mask_shuffled)
+        text_cond_shuffled = dict(
+            text_embed=text_embed_shuffled,
+            text_encodings=text_encodings_shuffled,
+            mask=text_mask_shuffled,
+        )
 
         # prepare the text embedding
         text_embed = text_embedding / text_embedding.norm(dim=1, keepdim=True)
 
         # prepare image embeddings
-        test_image_embeddings = test_image_embeddings / \
-            test_image_embeddings.norm(dim=1, keepdim=True)
+        test_image_embeddings = test_image_embeddings / test_image_embeddings.norm(
+            dim=1, keepdim=True
+        )
 
         # predict on the unshuffled text embeddings
-        predicted_image_embeddings = diffusion_prior.p_sample_loop(
-            test_image_embeddings.shape, text_cond)
-        predicted_image_embeddings = predicted_image_embeddings / \
-            predicted_image_embeddings.norm(dim=1, keepdim=True)
+        predicted_image_embeddings = trainer.p_sample_loop(
+            test_image_embeddings.shape, text_cond
+        )
+
+        predicted_image_embeddings = (
+            predicted_image_embeddings
+            / predicted_image_embeddings.norm(dim=1, keepdim=True)
+        )
 
         # predict on the shuffled embeddings
-        predicted_unrelated_embeddings = diffusion_prior.p_sample_loop(
-            test_image_embeddings.shape, text_cond_shuffled)
-        predicted_unrelated_embeddings = predicted_unrelated_embeddings / \
-            predicted_unrelated_embeddings.norm(dim=1, keepdim=True)
+        predicted_unrelated_embeddings = trainer.p_sample_loop(
+            test_image_embeddings.shape, text_cond_shuffled
+        )
+
+        predicted_unrelated_embeddings = (
+            predicted_unrelated_embeddings
+            / predicted_unrelated_embeddings.norm(dim=1, keepdim=True)
+        )
 
         # calculate similarities
-        original_similarity = cos(
-           text_embed, test_image_embeddings).cpu().numpy()
-        predicted_similarity = cos(
-           text_embed, predicted_image_embeddings).cpu().numpy()
-        unrelated_similarity = cos(
-           text_embed, predicted_unrelated_embeddings).cpu().numpy()
-        predicted_img_similarity = cos(
-           test_image_embeddings, predicted_image_embeddings).cpu().numpy()
-        tracker.log({"CosineSimilarity(text_embed,image_embed)": np.mean(original_similarity),
-            "CosineSimilarity(text_embed,predicted_image_embed)":np.mean(predicted_similarity),
-            "CosineSimilarity(orig_image_embed,predicted_image_embed)":np.mean(predicted_img_similarity),
-            "CosineSimilarity(text_embed,predicted_unrelated_embed)": np.mean(unrelated_similarity),
-            "Cosine similarity difference":np.mean(predicted_similarity - original_similarity)})
+        original_similarity = cos(text_embed, test_image_embeddings).cpu().numpy()
+        predicted_similarity = cos(text_embed, predicted_image_embeddings).cpu().numpy()
+        unrelated_similarity = (
+            cos(text_embed, predicted_unrelated_embeddings).cpu().numpy()
+        )
+        predicted_img_similarity = (
+            cos(test_image_embeddings, predicted_image_embeddings).cpu().numpy()
+        )
+
+        stats = {
+            f"{tracker_context}/baseline similarity": np.mean(original_similarity),
+            f"{tracker_context}/similarity with text": np.mean(predicted_similarity),
+            f"{tracker_context}/similarity with original image": np.mean(
+                predicted_img_similarity
+            ),
+            f"{tracker_context}/similarity with unrelated caption": np.mean(unrelated_similarity),
+            f"{tracker_context}/difference from baseline similarity": np.mean(
+                predicted_similarity - original_similarity
+            ),
+        }
+
+        for k, v in stats.items():
+            trainer.print(f"{tracker_context}/{k}: {v}")
+
+        if exists(tracker):
+            tracker.log(stats, step=trainer.step.item() + 1)
+
+
+# training script
+
+
+def train(
+    trainer: DiffusionPriorTrainer,
+    train_loader: DataLoader,
+    eval_loader: DataLoader,
+    test_loader: DataLoader,
+    config: DiffusionPriorTrainConfig,
+):
+    # distributed tracking with wandb
+    if trainer.accelerator.num_processes > 1:
+        os.environ["WANDB_START_METHOD"] = "thread"
+
+    tracker = wandb.init(
+        name=f"RANK:{trainer.device}",
+        entity=config.tracker.wandb_entity,
+        project=config.tracker.wandb_project,
+        config=config.dict(),
+        group=GROUP,
+    )
+
+    # sync after tracker init
+    trainer.wait_for_everyone()
+
+    # init a timer
+    timer = Timer()
+
+    # do training
+    for img, txt in train_loader:
+        trainer.train()
+        current_step = trainer.step.item() + 1
+
+        # place data on device
+        img = img.to(trainer.device)
+        txt = txt.to(trainer.device)
+
+        # pass to model
+        loss = trainer(text=txt, image_embed=img)
+
+        # display & log loss (will only print from main process)
+        trainer.print(f"Step {current_step}: Loss {loss}")
+
+        # perform backprop & apply EMA updates
+        trainer.update()
+
+        # track samples/sec/rank
+        samples_per_sec = img.shape[0] / timer.elapsed()
+
+        # samples seen
+        samples_seen = (
+            config.data.batch_size * trainer.accelerator.num_processes * current_step
+        )
+
+        # ema decay
+        ema_decay = trainer.ema_diffusion_prior.get_current_decay()
+
+        # Log on all processes for debugging
+        tracker.log(
+            {
+                "tracking/samples-sec": samples_per_sec,
+                "tracking/samples-seen": samples_seen,
+                "tracking/ema-decay": ema_decay,
+                "metrics/training-loss": loss,
+            },
+            step=current_step,
+        )
+
+        # Metric Tracking & Checkpointing (outside of timer's scope)
+        if current_step % EVAL_EVERY == 0:
+            eval_model(
+                trainer=trainer,
+                dataloader=eval_loader,
+                text_conditioned=config.prior.condition_on_text_encodings,
+                loss_type=config.prior.loss_type,
+                tracker_context="metrics/online-model-validation",
+                tracker=tracker,
+                use_ema=False,
+            )
+
+            eval_model(
+                trainer=trainer,
+                dataloader=eval_loader,
+                text_conditioned=config.prior.condition_on_text_encodings,
+                loss_type=config.prior.loss_type,
+                tracker_context="metrics/ema-model-validation",
+                tracker=tracker,
+                use_ema=True,
+            )
+
+            report_cosine_sims(
+                trainer=trainer,
+                dataloader=eval_loader,
+                text_conditioned=config.prior.condition_on_text_encodings,
+                tracker=tracker,
+                tracker_context="metrics",
+            )
+
+        if current_step % config.train.save_every == 0:
+            trainer.save(f"{config.tracker.data_path}/chkpt_step_{current_step}.pth")
+
+        # reset timer for next round
+        timer.reset()
+
+    # evaluate on test data
+
+    eval_model(
+        trainer=trainer,
+        dataloader=test_loader,
+        text_conditioned=config.prior.condition_on_text_encodings,
+        loss_type=config.prior.loss_type,
+        tracker_context="test",
+        tracker=tracker,
+    )
+
+    report_cosine_sims(
+        trainer,
+        test_loader,
+        config.prior.condition_on_text_encodings,
+        tracker,
+        tracker_context="test",
+    )
+
+
+def initialize_training(config, accelerator=None):
+    """
+    Parse the configuration file, and prepare everything necessary for training
+    """
+
+    # get a device
+
+    if accelerator:
+        device = accelerator.device
+        click.secho(f"Accelerating on: {device}", fg="yellow")
+    else:
+        if torch.cuda.is_available():
+            click.secho("GPU detected, defaulting to cuda:0", fg="yellow")
+            device = "cuda:0"
+        else:
+            click.secho("No GPU detected...using cpu", fg="yellow")
+            device = "cpu"
+
+    # make the trainer (will automatically distribute if possible & configured)
+
+    trainer = make_model(config.prior, config.train, device, accelerator).to(device)
+
+    # reload from chcekpoint
+
+    if config.load.resume == True:
+        click.secho(f"Loading checkpoint: {config.load.source}", fg="cyan")
+        trainer.load(config.load.source)
+
+    # fetch and prepare data
+
+    if trainer.is_main_process():
+        click.secho("Grabbing data from source", fg="blue", blink=True)
+
+    img_reader = get_reader(
+        text_conditioned=trainer.text_conditioned,
+        img_url=config.data.image_url,
+        meta_url=config.data.meta_url,
+    )
+
+    train_loader, eval_loader, test_loader = make_splits(
+        text_conditioned=trainer.text_conditioned,
+        batch_size=config.data.batch_size,
+        num_data_points=NUM_DATA_POINTS,
+        train_split=config.data.splits.train,
+        eval_split=config.data.splits.val,
+        image_reader=img_reader,
+        rank=accelerator.state.process_index if exists(accelerator) else 0,
+        world_size=accelerator.state.num_processes if exists(accelerator) else 1,
+        start=START,
+    )
+
+    # wait for everyone to load data before continuing
+    trainer.wait_for_everyone()
+
+    # start training
+    train(
+        trainer=trainer,
+        train_loader=train_loader,
+        eval_loader=eval_loader,
+        test_loader=test_loader,
+        config=config,
+    )
 
 
 @click.command()
-@click.option("--wandb-entity", default="laion")
-@click.option("--wandb-project", default="diffusion-prior")
-@click.option("--wandb-dataset", default="LAION-5B")
-@click.option("--wandb-arch", default="DiffusionPrior")
-@click.option("--image-embed-url", default="https://mystic.the-eye.eu/public/AI/cah/laion5b/embeddings/laion2B-en/img_emb/")
-@click.option("--text-embed-url", default="https://mystic.the-eye.eu/public/AI/cah/laion5b/embeddings/laion2B-en/text_emb/")
-@click.option("--meta-url", default="https://mystic.the-eye.eu/public/AI/cah/laion5b/embeddings/laion2B-en/laion2B-en-metadata/")
-@click.option("--learning-rate", default=1.1e-4)
-@click.option("--weight-decay", default=6.02e-2)
-@click.option("--dropout", default=5e-2)
-@click.option("--max-grad-norm", default=0.5)
-@click.option("--num-data-points", default=250e6)
-@click.option("--batch-size", default=320)
-@click.option("--num-epochs", default=5)
-@click.option("--image-embed-dim", default=768)
-@click.option("--train-percent", default=0.9)
-@click.option("--val-percent", default=1e-7)
-@click.option("--test-percent", default=0.0999999)
-@click.option("--dpn-depth", default=12)
-@click.option("--dpn-dim-head", default=64)
-@click.option("--dpn-heads", default=12)
-@click.option("--dp-condition-on-text-encodings", default=True)
-@click.option("--dp-timesteps", default=1000)
-@click.option("--dp-normformer", default=True)
-@click.option("--dp-cond-drop-prob", default=0.1)
-@click.option("--dp-loss-type", default="l2")
-@click.option("--clip", default="ViT-L/14")
-@click.option("--amp", default=False)
-@click.option("--save-interval", default=120)
-@click.option("--save-path", default="./diffusion_prior_checkpoints")
-@click.option("--pretrained-model-path", default=None)
-@click.option("--gpu-device", default=0)
-def train(
-    wandb_entity,
-    wandb_project,
-    wandb_dataset,
-    wandb_arch,
-    image_embed_url,
-    text_embed_url,
-    meta_url,
-    learning_rate,
-    weight_decay,
-    dropout,
-    max_grad_norm,
-    num_data_points,
-    batch_size,
-    num_epochs,
-    image_embed_dim,
-    train_percent,
-    val_percent,
-    test_percent,
-    dpn_depth,
-    dpn_dim_head,
-    dpn_heads,
-    dp_condition_on_text_encodings,
-    dp_timesteps,
-    dp_normformer,
-    dp_cond_drop_prob,
-    dp_loss_type,
-    clip,
-    amp,
-    save_interval,
-    save_path,
-    pretrained_model_path,
-    gpu_device
-):
-    config = {
-        "learning_rate": learning_rate,
-        "architecture": wandb_arch,
-        "dataset": wandb_dataset,
-        "weight_decay": weight_decay,
-        "max_gradient_clipping_norm": max_grad_norm,
-        "batch_size": batch_size,
-        "epochs": num_epochs,
-        "diffusion_prior_network": {
-            "depth": dpn_depth,
-            "dim_head": dpn_dim_head,
-            "heads": dpn_heads,
-            "normformer": dp_normformer
-        },
-        "diffusion_prior": {
-            "condition_on_text_encodings": dp_condition_on_text_encodings,
-            "timesteps": dp_timesteps,
-            "cond_drop_prob": dp_cond_drop_prob,
-            "loss_type": dp_loss_type,
-            "clip": clip
-        }
-    }
-
-    # Check if DPRIOR_PATH exists(saved model path)
-
-    DPRIOR_PATH = pretrained_model_path
-    RESUME = exists(DPRIOR_PATH)
-
-    if not RESUME:
-        tracker.init(
-            entity = wandb_entity,
-            project = wandb_project,
-            config = config
-        )
-
-    # Obtain the utilized device.
-
-    has_cuda = torch.cuda.is_available()
-    if has_cuda:
-        device = torch.device(f"cuda:{gpu_device}")
-        torch.cuda.set_device(device)
-
-    # Training loop
-    # diffusion prior network
-
-    prior_network = DiffusionPriorNetwork(
-        dim = image_embed_dim,
-        depth = dpn_depth,
-        dim_head = dpn_dim_head,
-        heads = dpn_heads,
-        attn_dropout = dropout,
-        ff_dropout = dropout,
-        normformer = dp_normformer
-    )
-
-    # Load clip model if text-conditioning
-    if dp_condition_on_text_encodings:
-        clip_adapter = OpenAIClipAdapter(clip)
+@click.option("--hfa", default=True)
+@click.option("--config_path", default="configs/prior.json")
+def main(hfa, config_path):
+    # start HFA if requested
+    if hfa:
+        accelerator = Accelerator()
     else:
-        clip_adapter = None
+        accelerator = None
 
-    # diffusion prior with text embeddings and image embeddings pre-computed
+    # load the configuration file on main process
+    if not exists(accelerator) or accelerator.is_main_process:
+        click.secho(f"Loading configuration from {config_path}", fg="green")
 
-    diffusion_prior = DiffusionPrior(
-        net = prior_network,
-        clip = clip_adapter,
-        image_embed_dim = image_embed_dim,
-        timesteps = dp_timesteps,
-        cond_drop_prob = dp_cond_drop_prob,
-        loss_type = dp_loss_type,
-        condition_on_text_encodings = dp_condition_on_text_encodings
-    )
+    config = TrainDiffusionPriorConfig.from_json_path(config_path)
 
-    # Load pre-trained model from DPRIOR_PATH
-
-    if RESUME:
-        diffusion_prior, loaded_obj = load_diffusion_model(DPRIOR_PATH, device)
-        tracker.init(entity = wandb_entity, project = wandb_project, config = config)
-
-    # diffusion prior trainer
-
-    trainer = DiffusionPriorTrainer(
-        diffusion_prior = diffusion_prior,
-        lr = learning_rate,
-        wd = weight_decay,
-        max_grad_norm = max_grad_norm,
-        amp = amp,
-    ).to(device)
-
-    # load optimizer and scaler
-
-    if RESUME:
-        trainer.optimizer.load_state_dict(loaded_obj['optimizer'])
-        trainer.scaler.load_state_dict(loaded_obj['scaler'])
-
-    # Create save_path if it doesn't exist
-
-    Path(save_path).mkdir(exist_ok = True, parents = True)
-
-    # Utilize wrapper to abstract away loader logic
-    print_ribbon("Downloading Embeddings")
-    reader_args = dict(text_conditioned=dp_condition_on_text_encodings, img_url=image_embed_url)
-
-    if dp_condition_on_text_encodings:
-        reader_args = dict(**reader_args, meta_url=meta_url)
-        img_reader = get_reader(**reader_args)
-        train_loader, eval_loader, test_loader = make_splits(
-            text_conditioned=dp_condition_on_text_encodings,
-            batch_size=batch_size,
-            num_data_points=num_data_points,
-            train_split=train_percent,
-            eval_split=val_percent,
-            image_reader=img_reader
-            )
-    else:
-        reader_args = dict(**reader_args, txt_url=text_embed_url)
-        img_reader, txt_reader = get_reader(**reader_args)
-        train_loader, eval_loader, test_loader = make_splits(
-            text_conditioned=dp_condition_on_text_encodings,
-            batch_size=batch_size,
-            num_data_points=num_data_points,
-            train_split=train_percent,
-            eval_split=val_percent,
-            image_reader=img_reader,
-            text_reader=txt_reader
-            )
-
-    ### Training code ###
-
-    step = 1
-    timer = Timer()
-    epochs = num_epochs
-
-    for _ in range(epochs):
-
-        for image, text in tqdm(train_loader):
-            diffusion_prior.train()
-
-            image = image.to(device)
-            text = text.to(device)
-
-            input_args = dict(image_embed=image)
-            if dp_condition_on_text_encodings:
-                input_args = dict(**input_args, text = text)
-            else:
-                input_args = dict(**input_args, text_embed=text)
-
-            loss = trainer(**input_args)
-
-            # Samples per second
-
-            samples_per_sec = batch_size * step / timer.elapsed()
-
-            # Save checkpoint every save_interval minutes
-            if(int(timer.elapsed()) >= 60 * save_interval):
-                timer.reset()
-
-                save_diffusion_model(
-                    save_path,
-                    diffusion_prior,
-                    trainer.optimizer,
-                    trainer.scaler,
-                    config,
-                    image_embed_dim)
-
-            # Log to wandb
-            tracker.log({"Training loss": loss,
-                        "Steps": step,
-                        "Samples per second": samples_per_sec})
-            # Log cosineSim(text_embed,predicted_image_embed) - cosineSim(text_embed,image_embed)
-            # Use NUM_TEST_EMBEDDINGS samples from the test set each time
-            # Get embeddings from the most recently saved model
-            if(step % REPORT_METRICS_EVERY) == 0:
-                report_cosine_sims(diffusion_prior, eval_loader, dp_condition_on_text_encodings, device=device)
-                ### Evaluate model(validation run) ###
-                eval_model(diffusion_prior, eval_loader, dp_condition_on_text_encodings, dp_loss_type, phase="Validation", device=device)
-
-            step += 1
-            trainer.update()
-
-    ### Test run ###
-    eval_model(diffusion_prior, test_loader, dp_condition_on_text_encodings, dp_loss_type, phase="Test")
+    # send config to get processed
+    initialize_training(config, accelerator)
 
 
 if __name__ == "__main__":
-    train()
+    main()