simplify more

take care of gradient accumulation automatically for researchers, by passing in a max_batch_size on the decoder or diffusion prior trainer forward
take care of backwards within trainer classes for diffusion prior and decoder, readying to take care of gradient accumulation as well (plus, unsure if loss should be backwards within autocast block)
2026-02-12 19:44:26 +01:00 · 2022-05-14 17:13:13 -07:00 · 2022-05-14 17:04:09 -07:00 · 2022-05-14 15:49:24 -07:00 · 2022-05-14 14:23:52 -07:00 · 2022-05-14 13:57:00 -07:00
8 changed files with 897 additions and 194 deletions
--- a/README.md
+++ b/README.md
@@ -508,7 +508,7 @@ To use a pretrained OpenAI CLIP, simply import `OpenAIClipAdapter` and pass it i
 import torch
 from dalle2_pytorch import DALLE2, DiffusionPriorNetwork, DiffusionPrior, Unet, Decoder, OpenAIClipAdapter

-# openai pretrained clip - defaults to ViT/B-32
+# openai pretrained clip - defaults to ViT-B/32

 clip = OpenAIClipAdapter()

@@ -732,8 +732,8 @@ clip = CLIP(

 # mock data

-text = torch.randint(0, 49408, (4, 256)).cuda()
-images = torch.randn(4, 3, 256, 256).cuda()
+text = torch.randint(0, 49408, (32, 256)).cuda()
+images = torch.randn(32, 3, 256, 256).cuda()

 # decoder (with unet)

@@ -774,8 +774,12 @@ decoder_trainer = DecoderTrainer(
 )

 for unet_number in (1, 2):
-    loss = decoder_trainer(images, text = text, unet_number = unet_number)  # use the decoder_trainer forward
-    loss.backward()
+    loss = decoder_trainer(
+        images,
+        text = text,
+        unet_number = unet_number, # which unet to train on
+        max_batch_size = 4         # gradient accumulation - this sets the maximum batch size in which to do forward and backwards pass - for this example 32 / 4 == 8 times
+    )

    decoder_trainer.update(unet_number) # update the specific unet as well as its exponential moving average

@@ -786,6 +790,67 @@ mock_image_embed = torch.randn(4, 512).cuda()
 images = decoder_trainer.sample(mock_image_embed, text = text) # (4, 3, 256, 256)
 ```

+### Diffusion Prior Training
+
+Similarly, one can use the `DiffusionPriorTrainer` to automatically instantiate and keep track of an exponential moving averaged prior.
+
+```python
+import torch
+from dalle2_pytorch import DALLE2, DiffusionPriorNetwork, DiffusionPrior, DiffusionPriorTrainer, Unet, Decoder, CLIP
+
+clip = CLIP(
+    dim_text = 512,
+    dim_image = 512,
+    dim_latent = 512,
+    num_text_tokens = 49408,
+    text_enc_depth = 6,
+    text_seq_len = 256,
+    text_heads = 8,
+    visual_enc_depth = 6,
+    visual_image_size = 256,
+    visual_patch_size = 32,
+    visual_heads = 8
+).cuda()
+
+# mock data
+
+text = torch.randint(0, 49408, (4, 256)).cuda()
+images = torch.randn(4, 3, 256, 256).cuda()
+
+# prior networks (with transformer)
+
+prior_network = DiffusionPriorNetwork(
+    dim = 512,
+    depth = 6,
+    dim_head = 64,
+    heads = 8
+).cuda()
+
+diffusion_prior = DiffusionPrior(
+    net = prior_network,
+    clip = clip,
+    timesteps = 100,
+    cond_drop_prob = 0.2
+).cuda()
+
+diffusion_prior_trainer = DiffusionPriorTrainer(
+    diffusion_prior,
+    lr = 3e-4,
+    wd = 1e-2,
+    ema_beta = 0.99,
+    ema_update_after_step = 1000,
+    ema_update_every = 10,
+)
+
+loss = diffusion_prior_trainer(text, images)
+diffusion_prior_trainer.update()  # this will update the optimizer as well as the exponential moving averaged diffusion prior
+
+# after much of the above three lines in a loop
+# you can sample from the exponential moving average of the diffusion prior identically to how you do so for DiffusionPrior
+
+image_embeds = diffusion_prior_trainer.sample(text) # (4, 512) - exponential moving averaged image embeddings
+```
+
 ### Decoder Dataloaders

 In order to make loading data simple and efficient, we include some general dataloaders that can be used to train portions of the network.
@@ -840,7 +905,7 @@ Please note that the script internally passes text_embed and image_embed to the
 ### Usage 

 ```bash
-$ pyhon train_diffusion_prior.py
+$ python train_diffusion_prior.py
 ```

 The most significant parameters for the script are as follows:
@@ -865,7 +930,39 @@ The most significant parameters for the script are as follows:

 ### Sample wandb run log

-Please find a sample wandb run log at : https://wandb.ai/laion/diffusion-prior/runs/aul0rhv5?workspace=
+Please find a sample wandb run log at : https://wandb.ai/laion/diffusion-prior/runs/1blxu24j
+
+### Loading and saving the Diffusion Prior model
+
+Two methods are provided, load_diffusion_model and save_diffusion_model, the names being self-explanatory. 
+
+## from dalle2_pytorch.train import load_diffusion_model, save_diffusion_model
+
+    load_diffusion_model(dprior_path, device) 
+
+        dprior_path : path to saved model(.pth)
+    
+        device      : the cuda device you're running on
+    
+    save_diffusion_model(save_path, model, optimizer, scaler, config, image_embed_dim)
+    
+        save_path : path to save at
+    
+        model     : object of Diffusion_Prior
+    
+        optimizer : optimizer object - see train_diffusion_prior.py for how to create one. 
+    
+            e.g: optimizer = get_optimizer(diffusion_prior.net.parameters(), wd=weight_decay, lr=learning_rate)
+    
+        scaler    : a GradScaler object.
+    
+            e.g: scaler = GradScaler(enabled=amp)
+    
+        config    : config object created in train_diffusion_prior.py - see file for example. 
+    
+        image_embed_dim - the dimension of the image_embedding
+    
+            e.g: 768

 ## CLI (wip)

@@ -904,20 +1001,26 @@ Once built, images will be saved to the same directory the command is invoked
 - [x] add convnext backbone for vqgan-vae (in addition to vit [vit-vqgan] + resnet)
 - [x] make sure DDPMs can be run with traditional resnet blocks (but leave convnext as an option for experimentation)
 - [x] make sure for the latter unets in the cascade, one can train on crops for learning super resolution (constrain the unet to be only convolutions in that case, or allow conv-like attention with rel pos bias)
+- [x] offer setting in diffusion prior to split time and image embeddings into multiple tokens, configurable, for more surface area during attention
+- [x] make sure resnet hyperparameters can be configurable across unet depth (groups and expansion factor)
+- [x] pull logic for training diffusion prior into a class DiffusionPriorTrainer, for eventual script based + CLI based training
+- [x] make sure the cascading ddpm in the repository can be trained unconditionally, offer a one-line CLI tool for training on a folder of images
+- [x] bring in cross-scale embedding from iclr paper https://github.com/lucidrains/vit-pytorch/blob/main/vit_pytorch/crossformer.py#L14
+- [x] cross embed layers for downsampling, as an option
 - [ ] become an expert with unets, cleanup unet code, make it fully configurable, port all learnings over to https://github.com/lucidrains/x-unet (test out unet² in ddpm repo) - consider https://github.com/lucidrains/uformer-pytorch attention-based unet
- [ ] copy the cascading ddpm code to a separate repo (perhaps https://github.com/lucidrains/denoising-diffusion-pytorch) as the main contribution of dalle2 really is just the prior network
 - [ ] transcribe code to Jax, which lowers the activation energy for distributed training, given access to TPUs
- [ ] pull logic for training diffusion prior into a class DiffusionPriorTrainer, for eventual script based + CLI based training
 - [ ] train on a toy task, offer in colab
 - [ ] think about how best to design a declarative training config that handles preencoding for prior and training of multiple networks in decoder
 - [ ] extend diffusion head to use diffusion-gan (potentially using lightweight-gan) to speed up inference
- [ ] bring in cross-scale embedding from iclr paper https://github.com/lucidrains/vit-pytorch/blob/main/vit_pytorch/crossformer.py#L14
 - [ ] figure out if possible to augment with external memory, as described in https://arxiv.org/abs/2204.11824
 - [ ] test out grid attention in cascading ddpm locally, decide whether to keep or remove
 - [ ] use an experimental tracker agnostic setup, as done <a href="https://github.com/lucidrains/tf-bind-transformer#simple-trainer-class-for-fine-tuning">here</a>
 - [ ] interface out the vqgan-vae so a pretrained one can be pulled off the shelf to validate latent diffusion + DALL-E2
 - [ ] make sure FILIP works with DALL-E2 from x-clip https://arxiv.org/abs/2111.07783
- [ ] make sure resnet hyperparameters can be configurable across unet depth (groups and expansion factor)
+- [ ] offer save / load methods on the trainer classes to automatically take care of state dicts for scalers / optimizers / saving versions and checking for breaking changes
+- [ ] bring in skip-layer excitatons (from lightweight gan paper) to see if it helps for either decoder of unet or vqgan-vae training
+- [ ] decoder needs one day worth of refactor for tech debt
+- [ ] allow for unet to be able to condition non-cross attention style as well

 ## Citations

@@ -985,4 +1088,25 @@ Once built, images will be saved to the same directory the command is invoked
 }
 ```

+```bibtex
+@article{Yu2022CoCaCC,
+    title   = {CoCa: Contrastive Captioners are Image-Text Foundation Models},
+    author  = {Jiahui Yu and Zirui Wang and Vijay Vasudevan and Legg Yeung and Mojtaba Seyedhosseini and Yonghui Wu},
+    journal = {ArXiv},
+    year    = {2022},
+    volume  = {abs/2205.01917}
+}
+```
+
+```bibtex
+@misc{wang2021crossformer,
+    title   = {CrossFormer: A Versatile Vision Transformer Hinging on Cross-scale Attention},
+    author  = {Wenxiao Wang and Lu Yao and Long Chen and Binbin Lin and Deng Cai and Xiaofei He and Wei Liu},
+    year    = {2021},
+    eprint  = {2108.00154},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.CV}
+}
+```
+
 *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>
--- a/dalle2_pytorch/init.py
+++ b/dalle2_pytorch/init.py
@@ -1,6 +1,6 @@
 from dalle2_pytorch.dalle2_pytorch import DALLE2, DiffusionPriorNetwork, DiffusionPrior, Unet, Decoder
 from dalle2_pytorch.dalle2_pytorch import OpenAIClipAdapter
-from dalle2_pytorch.train import DecoderTrainer
+from dalle2_pytorch.train import DecoderTrainer, DiffusionPriorTrainer

 from dalle2_pytorch.vqgan_vae import VQGanVAE
 from x_clip import CLIP
--- a/dalle2_pytorch/dalle2_pytorch.py
+++ b/dalle2_pytorch/dalle2_pytorch.py
@@ -1,9 +1,10 @@
 import math
 from tqdm import tqdm
 from inspect import isfunction
-from functools import partial
+from functools import partial, wraps
 from contextlib import contextmanager
 from collections import namedtuple
+from pathlib import Path

 import torch
 import torch.nn.functional as F
@@ -23,9 +24,18 @@ from dalle2_pytorch.vqgan_vae import NullVQGanVAE, VQGanVAE

 from resize_right import resize

+# rotary embeddings
+
+from rotary_embedding_torch import RotaryEmbedding
+
 # use x-clip

 from x_clip import CLIP
+from coca_pytorch import CoCa
+
+# constants
+
+NAT = 1. / math.log(2.)

 # helper functions

@@ -35,6 +45,14 @@ def exists(val):
 def identity(t, *args, **kwargs):
    return t

+def maybe(fn):
+    @wraps(fn)
+    def inner(x):
+        if not exists(x):
+            return x
+        return fn(x)
+    return inner
+
 def default(val, d):
    if exists(val):
        return val
@@ -85,6 +103,9 @@ def freeze_model_and_make_eval_(model):

 # tensor helpers

+def log(t, eps = 1e-12):
+    return torch.log(t.clamp(min = eps))
+
 def l2norm(t):
    return F.normalize(t, dim = -1)

@@ -101,10 +122,10 @@ def resize_image_to(image, target_image_size):
 # ddpms expect images to be in the range of -1 to 1
 # but CLIP may otherwise

-def normalize_img(img):
+def normalize_neg_one_to_one(img):
    return img * 2 - 1

-def unnormalize_img(normed_img):
+def unnormalize_zero_to_one(normed_img):
    return (normed_img + 1) * 0.5

 # clip related adapters
@@ -113,9 +134,10 @@ EmbeddedText = namedtuple('EmbedTextReturn', ['text_embed', 'text_encodings', 't
 EmbeddedImage = namedtuple('EmbedImageReturn', ['image_embed', 'image_encodings'])

 class BaseClipAdapter(nn.Module):
-    def __init__(self, clip):
+    def __init__(self, clip, **kwargs):
        super().__init__()
        self.clip = clip
+        self.overrides = kwargs

    @property
    def dim_latent(self):
@@ -173,6 +195,39 @@ class XClipAdapter(BaseClipAdapter):
        image_embed = self.clip.to_visual_latent(image_cls)
        return EmbeddedImage(l2norm(image_embed), image_encodings)

+class CoCaAdapter(BaseClipAdapter):
+    @property
+    def dim_latent(self):
+        return self.clip.dim
+
+    @property
+    def image_size(self):
+        assert 'image_size' in self.overrides
+        return self.overrides['image_size']
+
+    @property
+    def image_channels(self):
+        assert 'image_channels' in self.overrides
+        return self.overrides['image_channels']
+
+    @property
+    def max_text_len(self):
+        assert 'max_text_len' in self.overrides
+        return self.overrides['max_text_len']
+
+    @torch.no_grad()
+    def embed_text(self, text):
+        text = text[..., :self.max_text_len]
+        text_mask = text != 0
+        text_embed, text_encodings = self.clip.embed_text(text)
+        return EmbeddedText(text_embed, text_encodings, text_mask)
+
+    @torch.no_grad()
+    def embed_image(self, image):
+        image = resize_image_to(image, self.image_size)
+        image_embed, image_encodings = self.clip.embed_image(image)
+        return EmbeddedImage(image_embed, image_encodings)
+
 class OpenAIClipAdapter(BaseClipAdapter):
    def __init__(
        self,
@@ -225,15 +280,15 @@ class OpenAIClipAdapter(BaseClipAdapter):
        text_embed = self.clip.encode_text(text)
        text_encodings = self.text_encodings
        del self.text_encodings
-        return EmbeddedText(text_embed.float(), text_encodings.float(), text_mask)
+        return EmbeddedText(l2norm(text_embed.float()), text_encodings.float(), text_mask)

    @torch.no_grad()
    def embed_image(self, image):
        assert not self.cleared
        image = resize_image_to(image, self.image_size)
-        image = self.clip_normalize(unnormalize_img(image))
+        image = self.clip_normalize(image)
        image_embed = self.clip.encode_image(image)
-        return EmbeddedImage(image_embed.float(), None)
+        return EmbeddedImage(l2norm(image_embed.float()), None)

 # classifier free guidance functions

@@ -257,13 +312,43 @@ def noise_like(shape, device, repeat=False):
    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))))
+
+def normal_kl(mean1, logvar1, mean2, logvar2):
+    return 0.5 * (-1.0 + logvar2 - logvar1 + torch.exp(logvar1 - logvar2) + ((mean1 - mean2) ** 2) * torch.exp(-logvar2))
+
+def approx_standard_normal_cdf(x):
+    return 0.5 * (1.0 + torch.tanh(((2.0 / math.pi) ** 0.5) * (x + 0.044715 * (x ** 3))))
+
+def discretized_gaussian_log_likelihood(x, *, means, log_scales, thres = 0.999):
+    assert x.shape == means.shape == log_scales.shape
+
+    centered_x = x - means
+    inv_stdv = torch.exp(-log_scales)
+    plus_in = inv_stdv * (centered_x + 1. / 255.)
+    cdf_plus = approx_standard_normal_cdf(plus_in)
+    min_in = inv_stdv * (centered_x - 1. / 255.)
+    cdf_min = approx_standard_normal_cdf(min_in)
+    log_cdf_plus = log(cdf_plus)
+    log_one_minus_cdf_min = log(1. - cdf_min)
+    cdf_delta = cdf_plus - cdf_min
+
+    log_probs = torch.where(x < -thres,
+        log_cdf_plus,
+        torch.where(x > thres,
+            log_one_minus_cdf_min,
+            log(cdf_delta)))
+
+    return log_probs
+
 def cosine_beta_schedule(timesteps, s = 0.008):
    """
    cosine schedule
    as proposed in https://openreview.net/forum?id=-NEXDKk8gZ
    """
    steps = timesteps + 1
-    x = torch.linspace(0, timesteps, steps)
+    x = torch.linspace(0, timesteps, steps, dtype = torch.float64)
    alphas_cumprod = torch.cos(((x / timesteps) + s) / (1 + s) * torch.pi * 0.5) ** 2
    alphas_cumprod = alphas_cumprod / alphas_cumprod[0]
    betas = 1 - (alphas_cumprod[1:] / alphas_cumprod[:-1])
@@ -274,21 +359,21 @@ def linear_beta_schedule(timesteps):
    scale = 1000 / timesteps
    beta_start = scale * 0.0001
    beta_end = scale * 0.02
-    return torch.linspace(beta_start, beta_end, timesteps)
+    return torch.linspace(beta_start, beta_end, timesteps, dtype = torch.float64)


 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) ** 2
+    return torch.linspace(beta_start**2, beta_end**2, timesteps, dtype = torch.float64) ** 2


 def sigmoid_beta_schedule(timesteps):
    scale = 1000 / timesteps
    beta_start = scale * 0.0001
    beta_end = scale * 0.02
-    betas = torch.linspace(-6, 6, timesteps)
+    betas = torch.linspace(-6, 6, timesteps, dtype = torch.float64)
    return torch.sigmoid(betas) * (beta_end - beta_start) + beta_start


@@ -328,17 +413,21 @@ class BaseGaussianDiffusion(nn.Module):
        self.loss_type = loss_type
        self.loss_fn = loss_fn

-        self.register_buffer('betas', betas)
-        self.register_buffer('alphas_cumprod', alphas_cumprod)
-        self.register_buffer('alphas_cumprod_prev', alphas_cumprod_prev)
+        # register buffer helper function to cast double back to float
+
+        register_buffer = lambda name, val: self.register_buffer(name, val.to(torch.float32))
+
+        register_buffer('betas', betas)
+        register_buffer('alphas_cumprod', alphas_cumprod)
+        register_buffer('alphas_cumprod_prev', alphas_cumprod_prev)

        # calculations for diffusion q(x_t | x_{t-1}) and others

-        self.register_buffer('sqrt_alphas_cumprod', torch.sqrt(alphas_cumprod))
-        self.register_buffer('sqrt_one_minus_alphas_cumprod', torch.sqrt(1. - alphas_cumprod))
-        self.register_buffer('log_one_minus_alphas_cumprod', torch.log(1. - alphas_cumprod))
-        self.register_buffer('sqrt_recip_alphas_cumprod', torch.sqrt(1. / alphas_cumprod))
-        self.register_buffer('sqrt_recipm1_alphas_cumprod', torch.sqrt(1. / alphas_cumprod - 1))
+        register_buffer('sqrt_alphas_cumprod', torch.sqrt(alphas_cumprod))
+        register_buffer('sqrt_one_minus_alphas_cumprod', torch.sqrt(1. - alphas_cumprod))
+        register_buffer('log_one_minus_alphas_cumprod', torch.log(1. - alphas_cumprod))
+        register_buffer('sqrt_recip_alphas_cumprod', torch.sqrt(1. / alphas_cumprod))
+        register_buffer('sqrt_recipm1_alphas_cumprod', torch.sqrt(1. / alphas_cumprod - 1))

        # calculations for posterior q(x_{t-1} | x_t, x_0)

@@ -346,19 +435,13 @@ class BaseGaussianDiffusion(nn.Module):

        # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)

-        self.register_buffer('posterior_variance', posterior_variance)
+        register_buffer('posterior_variance', posterior_variance)

        # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain

-        self.register_buffer('posterior_log_variance_clipped', torch.log(posterior_variance.clamp(min =1e-20)))
-        self.register_buffer('posterior_mean_coef1', betas * torch.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod))
-        self.register_buffer('posterior_mean_coef2', (1. - alphas_cumprod_prev) * torch.sqrt(alphas) / (1. - alphas_cumprod))
-
-    def q_mean_variance(self, x_start, t):
-        mean = extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start
-        variance = extract(1. - self.alphas_cumprod, t, x_start.shape)
-        log_variance = extract(self.log_one_minus_alphas_cumprod, t, x_start.shape)
-        return mean, variance, log_variance
+        register_buffer('posterior_log_variance_clipped', torch.log(posterior_variance.clamp(min =1e-20)))
+        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))

    def q_posterior(self, x_start, x_t, t):
        posterior_mean = (
@@ -531,7 +614,7 @@ class Attention(nn.Module):
        heads = 8,
        dropout = 0.,
        causal = False,
-        post_norm = False
+        rotary_emb = None
    ):
        super().__init__()
        self.scale = dim_head ** -0.5
@@ -540,16 +623,17 @@ class Attention(nn.Module):

        self.causal = causal
        self.norm = LayerNorm(dim)
-        self.post_norm = LayerNorm(dim)     # sandwich norm from Coqview paper + Normformer
        self.dropout = nn.Dropout(dropout)

        self.null_kv = nn.Parameter(torch.randn(2, dim_head))
        self.to_q = nn.Linear(dim, inner_dim, bias = False)
        self.to_kv = nn.Linear(dim, dim_head * 2, bias = False)

+        self.rotary_emb = rotary_emb
+
        self.to_out = nn.Sequential(
            nn.Linear(inner_dim, dim, bias = False),
-            LayerNorm(dim) if post_norm else nn.Identity()
+            LayerNorm(dim)
        )

    def forward(self, x, mask = None, attn_bias = None):
@@ -559,6 +643,12 @@ class Attention(nn.Module):
        q, k, v = (self.to_q(x), *self.to_kv(x).chunk(2, dim = -1))

        q = rearrange(q, 'b n (h d) -> b h n d', h = self.heads)
+        q = q * self.scale
+
+        # rotary embeddings
+
+        if exists(self.rotary_emb):
+            q, k = map(self.rotary_emb.rotate_queries_or_keys, (q, k))

        # add null key / value for classifier free guidance in prior net

@@ -566,7 +656,7 @@ class Attention(nn.Module):
        k = torch.cat((nk, k), dim = -2)
        v = torch.cat((nv, v), dim = -2)

-        q = q * self.scale
+        # calculate query / key similarities

        sim = einsum('b h i d, b j d -> b h i j', q, k)

@@ -591,7 +681,7 @@ class Attention(nn.Module):

        # attention

-        sim = sim - sim.amax(dim = -1, keepdim = True)
+        sim = sim - sim.amax(dim = -1, keepdim = True).detach()
        attn = sim.softmax(dim = -1)
        attn = self.dropout(attn)

@@ -600,8 +690,7 @@ class Attention(nn.Module):
        out = einsum('b h i j, b j d -> b h i d', attn, v)

        out = rearrange(out, 'b h n d -> b n (h d)')
-        out = self.to_out(out)
-        return self.post_norm(out)
+        return self.to_out(out)

 class CausalTransformer(nn.Module):
    def __init__(
@@ -616,15 +705,18 @@ class CausalTransformer(nn.Module):
        attn_dropout = 0.,
        ff_dropout = 0.,
        final_proj = True,
-        normformer = False
+        normformer = False,
+        rotary_emb = True
    ):
        super().__init__()
        self.rel_pos_bias = RelPosBias(heads = heads)

+        rotary_emb = RotaryEmbedding(dim = min(32, dim_head)) if rotary_emb else None
+
        self.layers = nn.ModuleList([])
        for _ in range(depth):
            self.layers.append(nn.ModuleList([
-                Attention(dim = dim, causal = True, dim_head = dim_head, heads = heads, dropout = attn_dropout, post_norm = normformer),
+                Attention(dim = dim, causal = True, dim_head = dim_head, heads = heads, dropout = attn_dropout, rotary_emb = rotary_emb),
                FeedForward(dim = dim, mult = ff_mult, dropout = ff_dropout, post_activation_norm = normformer)
            ]))

@@ -652,10 +744,31 @@ class DiffusionPriorNetwork(nn.Module):
        self,
        dim,
        num_timesteps = None,
+        num_time_embeds = 1,
+        num_image_embeds = 1,
+        num_text_embeds = 1,
        **kwargs
    ):
        super().__init__()
-        self.time_embeddings = nn.Embedding(num_timesteps, dim) if exists(num_timesteps) else nn.Sequential(Rearrange('b -> b 1'), MLP(1, dim)) # also offer a continuous version of timestep embeddings, with a 2 layer MLP
+        self.num_time_embeds = num_time_embeds
+        self.num_image_embeds = num_image_embeds
+        self.num_text_embeds = num_text_embeds
+
+        self.to_text_embeds = nn.Sequential(
+            nn.Linear(dim, dim * num_text_embeds) if num_text_embeds > 1 else nn.Identity(),
+            Rearrange('b (n d) -> b n d', n = num_text_embeds)
+        )
+
+        self.to_time_embeds = nn.Sequential(
+            nn.Embedding(num_timesteps, dim * num_time_embeds) if exists(num_timesteps) else nn.Sequential(SinusoidalPosEmb(dim), MLP(dim, dim * num_time_embeds)), # also offer a continuous version of timestep embeddings, with a 2 layer MLP
+            Rearrange('b (n d) -> b n d', n = num_time_embeds)
+        )
+
+        self.to_image_embeds = nn.Sequential(
+            nn.Linear(dim, dim * num_image_embeds) if num_image_embeds > 1 else nn.Identity(),
+            Rearrange('b (n d) -> b n d', n = num_image_embeds)
+        )
+
        self.learned_query = nn.Parameter(torch.randn(dim))
        self.causal_transformer = CausalTransformer(dim = dim, **kwargs)

@@ -685,10 +798,13 @@ class DiffusionPriorNetwork(nn.Module):
    ):
        batch, dim, device, dtype = *image_embed.shape, image_embed.device, image_embed.dtype

+        num_time_embeds, num_image_embeds, num_text_embeds = self.num_time_embeds, self.num_image_embeds, self.num_text_embeds
+
        # in section 2.2, last paragraph
        # "... consisting of encoded text, CLIP text embedding, diffusion timestep embedding, noised CLIP image embedding, final embedding for prediction"

-        text_embed, image_embed = rearrange_many((text_embed, image_embed), 'b d -> b 1 d')
+        text_embed = self.to_text_embeds(text_embed)
+        image_embed = self.to_image_embeds(image_embed)

        # make text encodings optional
        # although the paper seems to suggest it is present <--
@@ -708,16 +824,17 @@ class DiffusionPriorNetwork(nn.Module):

        # whether text embedding is masked or not depends on the classifier free guidance conditional masking

+        keep_mask = repeat(keep_mask, 'b 1 -> b n', n = num_text_embeds)
        mask = torch.cat((mask, keep_mask), dim = 1)

        # whether text embedding is used for conditioning depends on whether text encodings are available for attention (for classifier free guidance, even though it seems from the paper it was not used in the prior ddpm, as the objective is different)
        # but let's just do it right

        if exists(mask):
-            mask = F.pad(mask, (0, 2), value = True) # extend mask for text embedding, noised image embedding, time step embedding, and learned query
+            attend_padding = 1 + num_time_embeds + num_image_embeds # 1 for learned queries + number of image embeds + time embeds
+            mask = F.pad(mask, (0, attend_padding), value = True) # extend mask for text embedding, noised image embedding, time step embedding, and learned query

-        time_embed = self.time_embeddings(diffusion_timesteps)
-        time_embed = rearrange(time_embed, 'b d -> b 1 d')
+        time_embed = self.to_time_embeds(diffusion_timesteps)

        learned_queries = repeat(self.learned_query, 'd -> b 1 d', b = batch)

@@ -725,6 +842,7 @@ class DiffusionPriorNetwork(nn.Module):
            text_encodings,
            text_embed,
            time_embed,
+            image_embed,
            learned_queries
        ), dim = -2)

@@ -748,13 +866,16 @@ class DiffusionPrior(BaseGaussianDiffusion):
        image_size = None,
        image_channels = 3,
        timesteps = 1000,
-        cond_drop_prob = 0.2,
-        loss_type = "l1",
+        cond_drop_prob = 0.,
+        loss_type = "l2",
        predict_x_start = True,
        beta_schedule = "cosine",
        condition_on_text_encodings = True, # the paper suggests this is needed, but you can turn it off for your CLIP preprocessed text embed -> image embed training
        sampling_clamp_l2norm = False,
+        training_clamp_l2norm = False,
+        init_image_embed_l2norm = False,
        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__(
            beta_schedule = beta_schedule,
@@ -764,7 +885,9 @@ class DiffusionPrior(BaseGaussianDiffusion):

        if exists(clip):
            if isinstance(clip, CLIP):
-                clip = XClipAdapter(clip)
+                clip = XClipAdapter(clip, **clip_adapter_overrides)
+            elif isinstance(clip, CoCa):
+                clip = CoCaAdapter(clip, **clip_adapter_overrides)

            assert isinstance(clip, BaseClipAdapter)
            freeze_model_and_make_eval_(clip)
@@ -788,6 +911,8 @@ class DiffusionPrior(BaseGaussianDiffusion):

        # whether to force an l2norm, similar to clipping denoised, when sampling
        self.sampling_clamp_l2norm = sampling_clamp_l2norm
+        self.training_clamp_l2norm = training_clamp_l2norm
+        self.init_image_embed_l2norm = init_image_embed_l2norm

    def p_mean_variance(self, x, t, text_cond, clip_denoised: bool):
        pred = self.net(x, t, **text_cond)
@@ -822,11 +947,16 @@ class DiffusionPrior(BaseGaussianDiffusion):
        device = self.betas.device

        b = shape[0]
-        img = torch.randn(shape, device=device)
+        image_embed = torch.randn(shape, device=device)
+
+        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):
-            img = self.p_sample(img, torch.full((b,), i, device = device, dtype = torch.long), text_cond = text_cond)
-        return img
+            times = torch.full((b,), i, device = device, dtype = torch.long)
+            image_embed = self.p_sample(image_embed, times, text_cond = text_cond)
+
+        return image_embed

    def p_losses(self, image_embed, times, text_cond, noise = None):
        noise = default(noise, lambda: torch.randn_like(image_embed))
@@ -840,11 +970,26 @@ class DiffusionPrior(BaseGaussianDiffusion):
            **text_cond
        )

+        if self.predict_x_start and self.training_clamp_l2norm:
+            pred = l2norm(pred) * self.image_embed_scale
+
        target = noise if not self.predict_x_start else image_embed

        loss = self.loss_fn(pred, target)
        return loss

+    @torch.inference_mode()
+    @eval_decorator
+    def sample_batch_size(self, batch_size, text_cond):
+        device = self.betas.device
+        shape = (batch_size, self.image_embed_dim)
+
+        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):
+            img = self.p_sample(img, torch.full((batch_size,), i, device = device, dtype = torch.long), text_cond = text_cond)
+        return img
+
    @torch.inference_mode()
    @eval_decorator
    def sample(self, text, num_samples_per_batch = 2):
@@ -1018,6 +1163,7 @@ class CrossAttention(nn.Module):
        dim_head = 64,
        heads = 8,
        dropout = 0.,
+        norm_context = False
    ):
        super().__init__()
        self.scale = dim_head ** -0.5
@@ -1027,13 +1173,17 @@ class CrossAttention(nn.Module):
        context_dim = default(context_dim, dim)

        self.norm = LayerNorm(dim)
-        self.norm_context = LayerNorm(context_dim)
+        self.norm_context = LayerNorm(context_dim) if norm_context else nn.Identity()
        self.dropout = nn.Dropout(dropout)

        self.null_kv = nn.Parameter(torch.randn(2, dim_head))
        self.to_q = nn.Linear(dim, inner_dim, bias = False)
        self.to_kv = nn.Linear(context_dim, inner_dim * 2, bias = False)
-        self.to_out = nn.Linear(inner_dim, dim, bias = False)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, dim, bias = False),
+            LayerNorm(dim)
+        )

    def forward(self, x, context, mask = None):
        b, n, device = *x.shape[:2], x.device
@@ -1062,7 +1212,7 @@ 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)
+        sim = sim - sim.amax(dim = -1, keepdim = True).detach()
        attn = sim.softmax(dim = -1)

        out = einsum('b h i j, b h j d -> b h i d', attn, v)
@@ -1123,6 +1273,33 @@ class LinearAttention(nn.Module):
        out = self.nonlin(out)
        return self.to_out(out)

+class CrossEmbedLayer(nn.Module):
+    def __init__(
+        self,
+        dim_in,
+        kernel_sizes,
+        dim_out = None,
+        stride = 2
+    ):
+        super().__init__()
+        assert all([*map(lambda t: (t % 2) == (stride % 2), kernel_sizes)])
+        dim_out = default(dim_out, dim_in)
+
+        kernel_sizes = sorted(kernel_sizes)
+        num_scales = len(kernel_sizes)
+
+        # calculate the dimension at each scale
+        dim_scales = [int(dim_out / (2 ** i)) for i in range(1, num_scales)]
+        dim_scales = [*dim_scales, dim_out - sum(dim_scales)]
+
+        self.convs = nn.ModuleList([])
+        for kernel, dim_scale in zip(kernel_sizes, dim_scales):
+            self.convs.append(nn.Conv2d(dim_in, dim_scale, kernel, stride = stride, padding = (kernel - stride) // 2))
+
+    def forward(self, x):
+        fmaps = tuple(map(lambda conv: conv(x), self.convs))
+        return torch.cat(fmaps, dim = 1)
+
 class Unet(nn.Module):
    def __init__(
        self,
@@ -1136,6 +1313,7 @@ class Unet(nn.Module):
        out_dim = None,
        dim_mults=(1, 2, 4, 8),
        channels = 3,
+        channels_out = None,
        attn_dim_head = 32,
        attn_heads = 16,
        lowres_cond = False, # for cascading diffusion - https://cascaded-diffusion.github.io/
@@ -1146,8 +1324,10 @@ class Unet(nn.Module):
        cond_on_image_embeds = False,
        init_dim = None,
        init_conv_kernel_size = 7,
-        block_type = 'resnet',
-        block_resnet_groups = 8,
+        resnet_groups = 8,
+        init_cross_embed_kernel_sizes = (3, 7, 15),
+        cross_embed_downsample = False,
+        cross_embed_downsample_kernel_sizes = (2, 4),
        **kwargs
    ):
        super().__init__()
@@ -1164,12 +1344,12 @@ class Unet(nn.Module):
        # determine dimensions

        self.channels = channels
+        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 // 2)
+        init_dim = default(init_dim, dim // 3 * 2)

-        assert (init_conv_kernel_size % 2) == 1
-        self.init_conv = nn.Conv2d(init_channels, init_dim, init_conv_kernel_size, padding = init_conv_kernel_size // 2)
+        self.init_conv = CrossEmbedLayer(init_channels, dim_out = init_dim, kernel_sizes = init_cross_embed_kernel_sizes, stride = 1)

        dims = [init_dim, *map(lambda m: dim * m, dim_mults)]
        in_out = list(zip(dims[:-1], dims[1:]))
@@ -1199,6 +1379,9 @@ class Unet(nn.Module):
            Rearrange('b (n d) -> b n d', n = num_image_tokens)
        ) if image_embed_dim != cond_dim else nn.Identity()

+        self.norm_cond = nn.LayerNorm(cond_dim)
+        self.norm_mid_cond = nn.LayerNorm(cond_dim)
+
        # text encoding conditioning (optional)

        self.text_to_cond = None
@@ -1225,7 +1408,15 @@ class Unet(nn.Module):

        # resnet block klass

-        block_klass = partial(ResnetBlock, groups = block_resnet_groups)
+        resnet_groups = cast_tuple(resnet_groups, len(in_out))
+
+        assert len(resnet_groups) == len(in_out)
+
+        # downsample klass
+
+        downsample_klass = Downsample
+        if cross_embed_downsample:
+            downsample_klass = partial(CrossEmbedLayer, kernel_sizes = cross_embed_downsample_kernel_sizes)

        # layers

@@ -1233,39 +1424,38 @@ class Unet(nn.Module):
        self.ups = nn.ModuleList([])
        num_resolutions = len(in_out)

-        for ind, (dim_in, dim_out) in enumerate(in_out):
+        for ind, ((dim_in, dim_out), groups) in enumerate(zip(in_out, resnet_groups)):
            is_first = ind == 0
            is_last = ind >= (num_resolutions - 1)
            layer_cond_dim = cond_dim if not is_first else None

            self.downs.append(nn.ModuleList([
-                block_klass(dim_in, dim_out, time_cond_dim = time_cond_dim),
+                ResnetBlock(dim_in, dim_out, time_cond_dim = time_cond_dim, groups = groups),
                Residual(LinearAttention(dim_out, **attn_kwargs)) if sparse_attn else nn.Identity(),
-                block_klass(dim_out, dim_out, cond_dim = layer_cond_dim, time_cond_dim = time_cond_dim),
-                Downsample(dim_out) if not is_last else nn.Identity()
+                ResnetBlock(dim_out, dim_out, cond_dim = layer_cond_dim, time_cond_dim = time_cond_dim, groups = groups),
+                downsample_klass(dim_out) if not is_last else nn.Identity()
            ]))

        mid_dim = dims[-1]

-        self.mid_block1 = block_klass(mid_dim, mid_dim, cond_dim = cond_dim, time_cond_dim = time_cond_dim)
+        self.mid_block1 = ResnetBlock(mid_dim, mid_dim, cond_dim = cond_dim, time_cond_dim = time_cond_dim, groups = resnet_groups[-1])
        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 = block_klass(mid_dim, mid_dim, cond_dim = cond_dim, time_cond_dim = time_cond_dim)
+        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) in enumerate(reversed(in_out[1:])):
+        for ind, ((dim_in, dim_out), groups) in enumerate(zip(reversed(in_out[1:]), reversed(resnet_groups))):
            is_last = ind >= (num_resolutions - 2)
            layer_cond_dim = cond_dim if not is_last else None

            self.ups.append(nn.ModuleList([
-                block_klass(dim_out * 2, dim_in, cond_dim = layer_cond_dim, time_cond_dim = time_cond_dim),
+                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(),
-                block_klass(dim_in, dim_in, cond_dim = layer_cond_dim, time_cond_dim = time_cond_dim),
+                ResnetBlock(dim_in, dim_in, cond_dim = layer_cond_dim, time_cond_dim = time_cond_dim, groups = groups),
                Upsample(dim_in)
            ]))

-        out_dim = default(out_dim, channels)
        self.final_conv = nn.Sequential(
-            block_klass(dim, dim),
-            nn.Conv2d(dim, out_dim, 1)
+            ResnetBlock(dim, dim, groups = resnet_groups[0]),
+            nn.Conv2d(dim, self.channels_out, 1)
        )

    # if the current settings for the unet are not correct
@@ -1275,12 +1465,25 @@ class Unet(nn.Module):
        *,
        lowres_cond,
        channels,
-        cond_on_image_embeds
+        channels_out,
+        cond_on_image_embeds,
+        cond_on_text_encodings
    ):
-        if lowres_cond == self.lowres_cond and channels == self.channels and cond_on_image_embeds == self.cond_on_image_embeds:
+        if lowres_cond == self.lowres_cond and \
+            channels == self.channels and \
+            cond_on_image_embeds == self.cond_on_image_embeds and \
+            cond_on_text_encodings == self.cond_on_text_encodings and \
+            channels_out == self.channels_out:
            return self

-        updated_kwargs = {'lowres_cond': lowres_cond, 'channels': channels, 'cond_on_image_embeds': cond_on_image_embeds}
+        updated_kwargs = dict(
+            lowres_cond = lowres_cond,
+            channels = channels,
+            channels_out = channels_out,
+            cond_on_image_embeds = cond_on_image_embeds,
+            cond_on_text_encodings = cond_on_text_encodings
+        )
+
        return self.__class__(**{**self._locals, **updated_kwargs})

    def forward_with_cond_scale(
@@ -1345,11 +1548,12 @@ class Unet(nn.Module):

        if self.cond_on_image_embeds:
            image_tokens = self.image_to_cond(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_tokens,
-                self.null_image_embed
+                null_image_embed
            )

        # take care of text encodings (optional)
@@ -1373,10 +1577,12 @@ class Unet(nn.Module):
                text_mask = rearrange(text_mask, 'b n -> b n 1')
                text_keep_mask = text_mask & text_keep_mask

+            null_text_embed = self.null_text_embed.to(text_tokens.dtype) # for some reason pytorch AMP not working
+
            text_tokens = torch.where(
                text_keep_mask,
                text_tokens,
-                self.null_text_embed
+                null_text_embed
            )

        # main conditioning tokens (c)
@@ -1391,6 +1597,11 @@ class Unet(nn.Module):

        mid_c = c if not exists(text_tokens) else torch.cat((c, text_tokens), dim = -2)

+        # normalize conditioning tokens
+
+        c = self.norm_cond(c)
+        mid_c = self.norm_mid_cond(mid_c)
+
        # go through the layers of the unet, down and up

        hiddens = []
@@ -1464,7 +1675,7 @@ class Decoder(BaseGaussianDiffusion):
        timesteps = 1000,
        image_cond_drop_prob = 0.1,
        text_cond_drop_prob = 0.5,
-        loss_type = 'l1',
+        loss_type = 'l2',
        beta_schedule = 'cosine',
        predict_x_start = False,
        predict_x_start_for_latent_diffusion = False,
@@ -1475,7 +1686,11 @@ class Decoder(BaseGaussianDiffusion):
        blur_kernel_size = 3,                       # cascading ddpm - blur kernel size
        condition_on_text_encodings = False,        # the paper suggested that this didn't do much in the decoder, but i'm allowing the option for experimentation
        clip_denoised = True,
-        clip_x_start = True
+        clip_x_start = True,
+        clip_adapter_overrides = dict(),
+        learned_variance = True,
+        vb_loss_weight = 0.001,
+        unconditional = False
    ):
        super().__init__(
            beta_schedule = beta_schedule,
@@ -1483,12 +1698,17 @@ class Decoder(BaseGaussianDiffusion):
            loss_type = loss_type
        )

+        self.unconditional = unconditional
+        assert not (condition_on_text_encodings and unconditional), 'unconditional decoder image generation cannot be set to True if conditioning on text is present'
+
        assert exists(clip) ^ exists(image_size), 'either CLIP is supplied, or you must give the image_size and channels (usually 3 for RGB)'

        self.clip = None
        if exists(clip):
            if isinstance(clip, CLIP):
-                clip = XClipAdapter(clip)
+                clip = XClipAdapter(clip, **clip_adapter_overrides)
+            elif isinstance(clip, CoCa):
+                clip = CoCaAdapter(clip, **clip_adapter_overrides)

            freeze_model_and_make_eval_(clip)
            assert isinstance(clip, BaseClipAdapter)
@@ -1508,10 +1728,18 @@ class Decoder(BaseGaussianDiffusion):
        unets = cast_tuple(unet)
        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
+
+        learned_variance = pad_tuple_to_length(cast_tuple(learned_variance), len(unets), fillvalue = False)
+        self.learned_variance = learned_variance
+        self.vb_loss_weight = vb_loss_weight
+
+        # construct unets and vaes
+
        self.unets = nn.ModuleList([])
        self.vaes = nn.ModuleList([])

-        for ind, (one_unet, one_vae) in enumerate(zip(unets, vaes)):
+        for ind, (one_unet, one_vae, one_unet_learned_var) in enumerate(zip(unets, vaes, learned_variance)):
            assert isinstance(one_unet, Unet)
            assert isinstance(one_vae, (VQGanVAE, NullVQGanVAE))

@@ -1519,11 +1747,14 @@ class Decoder(BaseGaussianDiffusion):
            latent_dim = one_vae.encoded_dim if exists(one_vae) else None

            unet_channels = default(latent_dim, self.channels)
+            unet_channels_out = unet_channels * (1 if not one_unet_learned_var else 2)

            one_unet = one_unet.cast_model_parameters(
                lowres_cond = not is_first,
-                cond_on_image_embeds = is_first,
-                channels = unet_channels
+                cond_on_image_embeds = is_first and not unconditional,
+                cond_on_text_encodings = one_unet.cond_on_text_encodings and not unconditional,
+                channels = unet_channels,
+                channels_out = unet_channels_out
            )

            self.unets.append(one_unet)
@@ -1586,8 +1817,11 @@ class Decoder(BaseGaussianDiffusion):
        yield
        unet.cpu()

-    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, cond_scale = 1.):
-        pred = 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)
+    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):
+        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))
+
+        if learned_variance:
+            pred, var_interp_frac_unnormalized = pred.chunk(2, dim = 1)

        if predict_x_start:
            x_recon = pred
@@ -1598,24 +1832,38 @@ class Decoder(BaseGaussianDiffusion):
            x_recon.clamp_(-1., 1.)

        model_mean, posterior_variance, posterior_log_variance = self.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)
+            var_interp_frac = unnormalize_zero_to_one(var_interp_frac_unnormalized)
+
+            posterior_log_variance = var_interp_frac * max_log + (1 - var_interp_frac) * min_log
+            posterior_variance = posterior_log_variance.exp()
+
        return model_mean, posterior_variance, posterior_log_variance

    @torch.inference_mode()
-    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, clip_denoised = True, repeat_noise = False):
+    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):
        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)
+        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)
        # 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.inference_mode()
-    def p_sample_loop(self, unet, shape, image_embed, predict_x_start = False, clip_denoised = True, lowres_cond_img = None, text_encodings = None, text_mask = None, cond_scale = 1):
+    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):
        device = self.betas.device

        b = shape[0]
        img = torch.randn(shape, device = device)

+        lowres_cond_img = maybe(normalize_neg_one_to_one)(lowres_cond_img)
+
        for i in tqdm(reversed(range(0, self.num_timesteps)), desc = 'sampling loop time step', total = self.num_timesteps):
            img = self.p_sample(
                unet,
@@ -1627,17 +1875,26 @@ class Decoder(BaseGaussianDiffusion):
                cond_scale = cond_scale,
                lowres_cond_img = lowres_cond_img,
                predict_x_start = predict_x_start,
+                learned_variance = learned_variance,
                clip_denoised = clip_denoised
            )

-        return img
+        unnormalize_img = unnormalize_zero_to_one(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):
+    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):
        noise = default(noise, lambda: torch.randn_like(x_start))

+        # normalize to [-1, 1]
+
+        x_start = normalize_neg_one_to_one(x_start)
+        lowres_cond_img = maybe(normalize_neg_one_to_one)(lowres_cond_img)
+
+        # get x_t
+
        x_noisy = self.q_sample(x_start = x_start, t = times, noise = noise)

-        pred = unet(
+        model_output = unet(
            x_noisy,
            times,
            image_embed = image_embed,
@@ -1648,21 +1905,63 @@ class Decoder(BaseGaussianDiffusion):
            text_cond_drop_prob = self.text_cond_drop_prob,
        )

+        if learned_variance:
+            pred, _ = model_output.chunk(2, dim = 1)
+        else:
+            pred = model_output
+
        target = noise if not predict_x_start else x_start

        loss = self.loss_fn(pred, target)
-        return loss
+
+        if not learned_variance:
+            # return simple loss if not using learned variance
+            return loss
+
+        # most of the code below is transcribed from
+        # https://github.com/hojonathanho/diffusion/blob/master/diffusion_tf/diffusion_utils_2.py
+        # the Improved DDPM paper then further modified it so that the mean is detached (shown a couple lines before), and weighted to be smaller than the l1 or l2 "simple" loss
+        # it is questionable whether this is really needed, looking at some of the figures in the paper, but may as well stay faithful to their implementation
+
+        # 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)
+
+        # kl loss with detached model predicted mean, for stability reasons as in paper
+
+        detached_model_mean = model_mean.detach()
+
+        kl = normal_kl(true_mean, true_log_variance_clipped, detached_model_mean, model_log_variance)
+        kl = meanflat(kl) * NAT
+
+        decoder_nll = -discretized_gaussian_log_likelihood(x_start, means = detached_model_mean, log_scales = 0.5 * model_log_variance)
+        decoder_nll = meanflat(decoder_nll) * NAT
+
+        # at the first timestep return the decoder NLL, otherwise return KL(q(x_{t-1}|x_t,x_0) || p(x_{t-1}|x_t))
+
+        vb_losses = torch.where(times == 0, decoder_nll, kl)
+
+        # weight the vb loss smaller, for stability, as in the paper (recommended 0.001)
+
+        vb_loss = vb_losses.mean() * self.vb_loss_weight
+
+        return loss + vb_loss

    @torch.inference_mode()
    @eval_decorator
    def sample(
        self,
-        image_embed,
+        image_embed = None,
        text = None,
+        batch_size = 1,
        cond_scale = 1.,
        stop_at_unet_number = None
    ):
-        batch_size = image_embed.shape[0]
+        assert self.unconditional or exists(image_embed), 'image embed must be present on sampling from decoder unless if trained unconditionally'
+
+        if not self.unconditional:
+            batch_size = image_embed.shape[0]

        text_encodings = text_mask = None
        if exists(text):
@@ -1672,10 +1971,11 @@ class Decoder(BaseGaussianDiffusion):
        assert not (not self.condition_on_text_encodings and exists(text_encodings)), 'decoder specified not to be conditioned on text, yet it is presented'

        img = None
+        is_cuda = next(self.parameters()).is_cuda

-        for unet_number, unet, vae, channel, image_size, predict_x_start in tqdm(zip(range(1, len(self.unets) + 1), self.unets, self.vaes, self.sample_channels, self.image_sizes, self.predict_x_start)):
+        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)):

-            context = self.one_unet_in_gpu(unet = unet) if image_embed.is_cuda else null_context()
+            context = self.one_unet_in_gpu(unet = unet) if is_cuda else null_context()

            with context:
                lowres_cond_img = None
@@ -1699,6 +1999,7 @@ class Decoder(BaseGaussianDiffusion):
                    text_mask = text_mask,
                    cond_scale = cond_scale,
                    predict_x_start = predict_x_start,
+                    learned_variance = learned_variance,
                    clip_denoised = not is_latent_diffusion,
                    lowres_cond_img = lowres_cond_img
                )
@@ -1728,6 +2029,7 @@ class Decoder(BaseGaussianDiffusion):
        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]
+        learned_variance    = self.learned_variance[unet_index]
        b, c, h, w, device, = *image.shape, image.device

        check_shape(image, 'b c h w', c = self.channels)
@@ -1765,7 +2067,7 @@ class Decoder(BaseGaussianDiffusion):
            if exists(lowres_cond_img):
                lowres_cond_img = 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)
+        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)

 # main class

--- a/dalle2_pytorch/optimizer.py
+++ b/dalle2_pytorch/optimizer.py
@@ -7,16 +7,17 @@ def separate_weight_decayable_params(params):

 def get_optimizer(
    params,
-    lr = 3e-4,
+    lr = 2e-5,
    wd = 1e-2,
    betas = (0.9, 0.999),
+    eps = 1e-8,
    filter_by_requires_grad = False
 ):
    if filter_by_requires_grad:
        params = list(filter(lambda t: t.requires_grad, params))

    if wd == 0:
-        return Adam(params, lr = lr, betas = betas)
+        return Adam(params, lr = lr, betas = betas, eps = eps)

    params = set(params)
    wd_params, no_wd_params = separate_weight_decayable_params(params)
@@ -26,4 +27,4 @@ def get_optimizer(
        {'params': list(no_wd_params), 'weight_decay': 0},
    ]

-    return AdamW(param_groups, lr = lr, weight_decay = wd, betas = betas)
+    return AdamW(param_groups, lr = lr, weight_decay = wd, betas = betas, eps = eps)
--- a/dalle2_pytorch/train.py
+++ b/dalle2_pytorch/train.py
@@ -1,11 +1,14 @@
+import time
 import copy
+from math import ceil
 from functools import partial
+from collections.abc import Iterable

 import torch
 from torch import nn
 from torch.cuda.amp import autocast, GradScaler

-from dalle2_pytorch.dalle2_pytorch import Decoder
+from dalle2_pytorch.dalle2_pytorch import Decoder, DiffusionPrior
 from dalle2_pytorch.optimizer import get_optimizer

 # helper functions
@@ -13,6 +16,9 @@ from dalle2_pytorch.optimizer import get_optimizer
 def exists(val):
    return val is not None

+def default(val, d):
+    return val if exists(val) else d
+
 def cast_tuple(val, length = 1):
    return val if isinstance(val, tuple) else ((val,) * length)

@@ -39,13 +45,98 @@ def groupby_prefix_and_trim(prefix, d):
    kwargs_without_prefix = dict(map(lambda x: (x[0][len(prefix):], x[1]), tuple(kwargs_with_prefix.items())))
    return kwargs_without_prefix, kwargs

+# gradient accumulation functions
+
+def split_iterable(it, split_size):
+    accum = []
+    for ind in range(ceil(len(it) / split_size)):
+        start_index = ind * split_size
+        accum.append(it[start_index: (start_index + split_size)])
+    return accum
+
+def split(t, split_size = None):
+    if not exists(split_size):
+        return t
+
+    if isinstance(t, torch.Tensor):
+        return t.split(split_size, dim = 0)
+
+    if isinstance(t, Iterable):
+        return split_iterable(t, split_size)
+
+    return TypeError
+
+def split_args_and_kwargs(x, *args, split_size = None, **kwargs):
+    batch_size = len(x)
+    split_size = default(split_size, batch_size)
+    chunk_size = ceil(batch_size / split_size)
+
+    dict_len = len(kwargs)
+    dict_keys = kwargs.keys()
+    all_args = (x, *args, *kwargs.values())
+    len_all_args = len(all_args)
+    split_kwargs_index = len_all_args - dict_len
+
+    split_all_args = [split(arg, split_size = split_size) if exists(arg) and isinstance(arg, (torch.Tensor, Iterable)) else ((arg,) * chunk_size) for arg in all_args]
+    chunk_sizes = tuple(map(len, split_all_args[0]))
+
+    for (chunk_size, *chunked_all_args) in tuple(zip(chunk_sizes, *split_all_args)):
+        chunked_args, chunked_kwargs_values = chunked_all_args[:split_kwargs_index], chunked_all_args[split_kwargs_index:]
+        chunked_kwargs = dict(tuple(zip(dict_keys, chunked_kwargs_values)))
+        chunk_size_frac = chunk_size / batch_size
+        yield chunk_size_frac, (chunked_args, chunked_kwargs)
+
+# print helpers
+
+def print_ribbon(s, symbol = '=', repeat = 40):
+    flank = symbol * repeat
+    return f'{flank} {s} {flank}'
+
+# saving and loading functions
+
+# for diffusion prior
+
+def load_diffusion_model(dprior_path, device):
+    dprior_path = Path(dprior_path)
+    assert dprior_path.exists(), 'Dprior model file does not exist'
+    loaded_obj = torch.load(str(dprior_path), map_location='cpu')
+
+    # Get hyperparameters of loaded model
+    dpn_config = loaded_obj['hparams']['diffusion_prior_network']
+    dp_config = loaded_obj['hparams']['diffusion_prior']
+    image_embed_dim = loaded_obj['image_embed_dim']['image_embed_dim']
+
+    # Create DiffusionPriorNetwork and DiffusionPrior with loaded hyperparameters
+
+    # DiffusionPriorNetwork
+    prior_network = DiffusionPriorNetwork( dim = image_embed_dim, **dpn_config).to(device)
+
+    # DiffusionPrior with text embeddings and image embeddings pre-computed
+    diffusion_prior = DiffusionPrior(net = prior_network, **dp_config, image_embed_dim = image_embed_dim).to(device)
+
+    # Load state dict from saved model
+    diffusion_prior.load_state_dict(loaded_obj['model'])
+
+    return diffusion_prior
+
+def save_diffusion_model(save_path, model, optimizer, scaler, config, image_embed_dim):
+    # Saving State Dict
+    print_ribbon('Saving checkpoint')
+
+    state_dict = dict(model=model.state_dict(),
+                      optimizer=optimizer.state_dict(),
+                      scaler=scaler.state_dict(),
+                      hparams = config,
+                      image_embed_dim = {"image_embed_dim":image_embed_dim})
+    torch.save(state_dict, save_path+'/'+str(time.time())+'_saved_model.pth')
+
 # exponential moving average wrapper

 class EMA(nn.Module):
    def __init__(
        self,
        model,
-        beta = 0.99,
+        beta = 0.9999,
        update_after_step = 1000,
        update_every = 10,
    ):
@@ -60,6 +151,10 @@ class EMA(nn.Module):
        self.register_buffer('initted', torch.Tensor([False]))
        self.register_buffer('step', torch.tensor([0.]))

+    def restore_ema_model_device(self):
+        device = self.initted.device
+        self.ema_model.to(device)
+
    def update(self):
        self.step += 1

@@ -89,15 +184,102 @@ class EMA(nn.Module):
    def __call__(self, *args, **kwargs):
        return self.ema_model(*args, **kwargs)

-# trainers
+# diffusion prior trainer
+
+class DiffusionPriorTrainer(nn.Module):
+    def __init__(
+        self,
+        diffusion_prior,
+        use_ema = True,
+        lr = 3e-4,
+        wd = 1e-2,
+        eps = 1e-6,
+        max_grad_norm = None,
+        amp = False,
+        **kwargs
+    ):
+        super().__init__()
+        assert isinstance(diffusion_prior, DiffusionPrior)
+        ema_kwargs, kwargs = groupby_prefix_and_trim('ema_', kwargs)
+
+        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.optimizer = get_optimizer(
+            diffusion_prior.parameters(),
+            lr = lr,
+            wd = wd,
+            eps = eps,
+            **kwargs
+        )
+
+        # gradient clipping if needed
+
+        self.max_grad_norm = max_grad_norm
+
+    def update(self):
+        if exists(self.max_grad_norm):
+            self.scaler.unscale_(self.optimizer)
+            nn.utils.clip_grad_norm_(self.diffusion_prior.parameters(), self.max_grad_norm)
+
+        self.scaler.step(self.optimizer)
+        self.scaler.update()
+        self.optimizer.zero_grad()
+
+        if self.use_ema:
+            self.ema_diffusion_prior.update()
+
+    @torch.inference_mode()
+    def p_sample_loop(self, *args, **kwargs):
+        return self.ema_diffusion_prior.ema_model.p_sample_loop(*args, **kwargs)
+
+    @torch.inference_mode()
+    def sample(self, *args, **kwargs):
+        return self.ema_diffusion_prior.ema_model.sample(*args, **kwargs)
+
+    @torch.inference_mode()
+    def sample_batch_size(self, *args, **kwargs):
+        return self.ema_diffusion_prior.ema_model.sample_batch_size(*args, **kwargs)
+
+    def forward(
+        self,
+        x,
+        *args,
+        max_batch_size = None,
+        **kwargs
+    ):
+        total_loss = 0.
+
+        for chunk_size_frac, (chunked_args, chunked_kwargs) in split_args_and_kwargs(x, *args, split_size = max_batch_size, **kwargs):
+            with autocast(enabled = self.amp):
+                loss = self.diffusion_prior(*chunked_args, **chunked_kwargs)
+
+                total_loss += loss.item() * chunk_size_frac
+                self.scaler.scale(loss * chunk_size_frac).backward()
+
+        return total_loss
+
+# decoder trainer

 class DecoderTrainer(nn.Module):
    def __init__(
        self,
        decoder,
        use_ema = True,
-        lr = 3e-4,
+        lr = 2e-5,
        wd = 1e-2,
+        eps = 1e-8,
        max_grad_norm = None,
        amp = False,
        **kwargs
@@ -122,13 +304,14 @@ class DecoderTrainer(nn.Module):
        # be able to finely customize learning rate, weight decay
        # per unet

-        lr, wd = map(partial(cast_tuple, length = self.num_unets), (lr, wd))
+        lr, wd, eps = map(partial(cast_tuple, length = self.num_unets), (lr, wd, eps))

-        for ind, (unet, unet_lr, unet_wd) in enumerate(zip(self.decoder.unets, lr, wd)):
+        for ind, (unet, unet_lr, unet_wd, unet_eps) in enumerate(zip(self.decoder.unets, lr, wd, eps)):
            optimizer = get_optimizer(
                unet.parameters(),
                lr = unet_lr,
                wd = unet_wd,
+                eps = unet_eps,
                **kwargs
            )

@@ -184,6 +367,11 @@ class DecoderTrainer(nn.Module):

        if self.use_ema:
            self.decoder.unets = trainable_unets             # restore original training unets
+
+        # cast the ema_model unets back to original device
+        for ema in self.ema_unets:
+            ema.restore_ema_model_device()
+
        return output

    def forward(
@@ -191,9 +379,17 @@ class DecoderTrainer(nn.Module):
        x,
        *,
        unet_number,
-        divisor = 1,
+        max_batch_size = None,
        **kwargs
    ):
-        with autocast(enabled = self.amp):
-            loss = self.decoder(x, unet_number = unet_number, **kwargs)
-        return self.scale(loss / divisor, unet_number = unet_number)
+        total_loss = 0.
+
+        for chunk_size_frac, (chunked_args, chunked_kwargs) in split_args_and_kwargs(x, split_size = max_batch_size, **kwargs):
+            with autocast(enabled = self.amp):
+                loss = self.decoder(*chunked_args, unet_number = unet_number, **chunked_kwargs)
+
+                total_loss += loss.item() * chunk_size_frac
+
+                self.scale(loss * chunk_size_frac, unet_number = unet_number).backward()
+
+        return total_loss
--- a/dalle2_pytorch/train_vqgan_vae.py
+++ b/dalle2_pytorch/train_vqgan_vae.py
@@ -3,14 +3,15 @@ import copy
 from random import choice
 from pathlib import Path
 from shutil import rmtree
+from PIL import Image

 import torch
 from torch import nn
-
-from PIL import Image
-from torchvision.datasets import ImageFolder
-import torchvision.transforms as T
+from torch.cuda.amp import autocast, GradScaler
 from torch.utils.data import Dataset, DataLoader, random_split
+
+import torchvision.transforms as T
+from torchvision.datasets import ImageFolder
 from torchvision.utils import make_grid, save_image

 from einops import rearrange
@@ -99,6 +100,7 @@ class VQGanVAETrainer(nn.Module):
        ema_update_after_step = 2000,
        ema_update_every = 10,
        apply_grad_penalty_every = 4,
+        amp = False
    ):
        super().__init__()
        assert isinstance(vae, VQGanVAE), 'vae must be instance of VQGanVAE'
@@ -120,6 +122,10 @@ class VQGanVAETrainer(nn.Module):
        self.optim = get_optimizer(vae_parameters, lr = lr, wd = wd)
        self.discr_optim = get_optimizer(discr_parameters, lr = lr, wd = wd)

+        self.amp = amp
+        self.scaler = GradScaler(enabled = amp)
+        self.discr_scaler = GradScaler(enabled = amp)
+
        # create dataset

        self.ds = ImageDataset(folder, image_size = image_size)
@@ -178,20 +184,22 @@ class VQGanVAETrainer(nn.Module):
            img = next(self.dl)
            img = img.to(device)

-            loss = self.vae(
-                img,
-                return_loss = True,
-                apply_grad_penalty = apply_grad_penalty
-            )
+            with autocast(enabled = self.amp):
+                loss = self.vae(
+                    img,
+                    return_loss = True,
+                    apply_grad_penalty = apply_grad_penalty
+                )
+
+
+                self.scaler.scale(loss / self.grad_accum_every).backward()

            accum_log(logs, {'loss': loss.item() / self.grad_accum_every})

-            (loss / self.grad_accum_every).backward()
-
-        self.optim.step()
+        self.scaler.step(self.optim)
+        self.scaler.update()
        self.optim.zero_grad()

-
        # update discriminator

        if exists(self.vae.discr):
@@ -200,12 +208,15 @@ class VQGanVAETrainer(nn.Module):
                img = next(self.dl)
                img = img.to(device)

-                loss = self.vae(img, return_discr_loss = True)
+                with autocast(enabled = self.amp):
+                    loss = self.vae(img, return_discr_loss = True)
+
+                    self.discr_scaler.scale(loss / self.grad_accum_every).backward()
+
                accum_log(logs, {'discr_loss': loss.item() / self.grad_accum_every})

-                (loss / self.grad_accum_every).backward()
-
-            self.discr_optim.step()
+            self.discr_scaler.step(self.discr_optim)
+            self.discr_scaler.update()
            self.discr_optim.zero_grad()

            # log
--- a/setup.py
+++ b/setup.py
@@ -10,11 +10,12 @@ setup(
      'dream = dalle2_pytorch.cli:dream'
    ],
  },
-  version = '0.0.107',
+  version = '0.2.27',
  license='MIT',
  description = 'DALL-E 2',
  author = 'Phil Wang',
  author_email = 'lucidrains@gmail.com',
+  long_description_content_type = 'text/markdown',
  url = 'https://github.com/lucidrains/dalle2-pytorch',
  keywords = [
    'artificial intelligence',
@@ -24,12 +25,14 @@ setup(
  install_requires=[
    'click',
    'clip-anytorch',
+    'coca-pytorch>=0.0.5',
    'einops>=0.4',
    'einops-exts>=0.0.3',
    'embedding-reader',
    'kornia>=0.5.4',
    'pillow',
    'resize-right>=0.0.2',
+    'rotary-embedding-torch',
    'torch>=1.10',
    'torchvision',
    'tqdm',
--- a/train_diffusion_prior.py
+++ b/train_diffusion_prior.py
@@ -7,6 +7,7 @@ import torch
 from torch import nn
 from embedding_reader import EmbeddingReader
 from dalle2_pytorch import DiffusionPrior, DiffusionPriorNetwork
+from dalle2_pytorch.train import load_diffusion_model, save_diffusion_model, print_ribbon
 from dalle2_pytorch.optimizer import get_optimizer
 from torch.cuda.amp import autocast,GradScaler

@@ -41,37 +42,56 @@ def eval_model(model,device,image_reader,text_reader,start,end,batch_size,loss_t
        avg_loss = (total_loss / total_samples)
        wandb.log({f'{phase} {loss_type}': avg_loss})

-def save_model(save_path, state_dict):
-    # Saving State Dict
-    print("====================================== Saving checkpoint ======================================")
-    torch.save(state_dict, save_path+'/'+str(time.time())+'_saved_model.pth')
+def report_cosine_sims(diffusion_prior,image_reader,text_reader,train_set_size,NUM_TEST_EMBEDDINGS,device):
+    diffusion_prior.eval()

-def report_cosine_sims(diffusion_prior,image_reader,text_reader,train_set_size,val_set_size,NUM_TEST_EMBEDDINGS,device):
    cos = nn.CosineSimilarity(dim=1, eps=1e-6)

-    tstart = train_set_size+val_set_size
-    tend = train_set_size+val_set_size+NUM_TEST_EMBEDDINGS
-
-    for embt, embi in zip(text_reader(batch_size = NUM_TEST_EMBEDDINGS, start=tstart, end = tend),image_reader(batch_size = NUM_TEST_EMBEDDINGS, start=tstart, end = tend)):
-        text_embed = torch.tensor(embt[0]).to(device)
-        text_embed = text_embed /  text_embed.norm(dim=1, keepdim=True)
-        test_text_cond = dict(text_embed = text_embed)
-
-        test_image_embeddings = torch.tensor(embi[0]).to(device)
-        test_image_embeddings = test_image_embeddings /  test_image_embeddings.norm(dim=1, keepdim=True)
-
-        predicted_image_embeddings = diffusion_prior.p_sample_loop((NUM_TEST_EMBEDDINGS, 768), text_cond = test_text_cond)
-        predicted_image_embeddings = predicted_image_embeddings / predicted_image_embeddings.norm(dim=1, keepdim=True)
-
-        original_similarity = cos(text_embed,test_image_embeddings).cpu().numpy()
-        predicted_similarity = cos(text_embed,predicted_image_embeddings).cpu().numpy()
-
-        wandb.log({"CosineSimilarity(text_embed,image_embed)": np.mean(original_similarity)})
-        wandb.log({"CosineSimilarity(text_embed,predicted_image_embed)":np.mean(predicted_similarity)})
-
-    return np.mean(predicted_similarity - original_similarity)
-
+    tstart = train_set_size
+    tend = train_set_size+NUM_TEST_EMBEDDINGS

+    for embt, embi in zip(text_reader(batch_size=NUM_TEST_EMBEDDINGS, start=tstart, end=tend), 
+            image_reader(batch_size=NUM_TEST_EMBEDDINGS, start=tstart, end=tend)):
+       # make a copy of the text embeddings for shuffling
+       text_embed = torch.tensor(embt[0]).to(device)
+       text_embed_shuffled = text_embed.clone()
+        # roll the text embeddings to simulate "unrelated" captions
+       rolled_idx = torch.roll(torch.arange(NUM_TEST_EMBEDDINGS), 1)
+       text_embed_shuffled = text_embed_shuffled[rolled_idx]
+       text_embed_shuffled = text_embed_shuffled / \
+           text_embed_shuffled.norm(dim=1, keepdim=True)
+       test_text_shuffled_cond = dict(text_embed=text_embed_shuffled)
+        # prepare the text embedding
+       text_embed = text_embed / text_embed.norm(dim=1, keepdim=True)
+       test_text_cond = dict(text_embed=text_embed)
+        # prepare image embeddings
+       test_image_embeddings = torch.tensor(embi[0]).to(device)
+       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(
+           (NUM_TEST_EMBEDDINGS, 768), text_cond=test_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(
+           (NUM_TEST_EMBEDDINGS, 768), text_cond=test_text_shuffled_cond)
+       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()
+       wandb.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)})

 def train(image_embed_dim,
          image_embed_url,
@@ -93,9 +113,15 @@ def train(image_embed_dim,
          save_interval,
          save_path,
          device,
+          RESUME,
+          DPRIOR_PATH,
+          config,
+          wandb_entity,
+          wandb_project,
          learning_rate=0.001,
          max_grad_norm=0.5,
          weight_decay=0.01,
+          dropout=0.05,
          amp=False):

    # DiffusionPriorNetwork 
@@ -104,6 +130,8 @@ def train(image_embed_dim,
            depth = dpn_depth, 
            dim_head = dpn_dim_head, 
            heads = dpn_heads,
+            attn_dropout = dropout,
+            ff_dropout = dropout,
            normformer = dp_normformer).to(device)
    
    # DiffusionPrior with text embeddings and image embeddings pre-computed
@@ -116,16 +144,21 @@ def train(image_embed_dim,
            loss_type = dp_loss_type, 
            condition_on_text_encodings = dp_condition_on_text_encodings).to(device)

-    # Get image and text embeddings from the servers
-    print("==============Downloading embeddings - image and text====================")
-    image_reader = EmbeddingReader(embeddings_folder=image_embed_url, file_format="npy")
-    text_reader  = EmbeddingReader(embeddings_folder=text_embed_url, file_format="npy")
-    num_data_points = text_reader.count
+    # Load pre-trained model from DPRIOR_PATH
+    if RESUME:
+        diffusion_prior=load_diffusion_model(DPRIOR_PATH,device)   
+        wandb.init( entity=wandb_entity, project=wandb_project, config=config) 

    # Create save_path if it doesn't exist
    if not os.path.exists(save_path):
        os.makedirs(save_path)

+    # Get image and text embeddings from the servers
+    print_ribbon("Downloading embeddings - image and text")
+    image_reader = EmbeddingReader(embeddings_folder=image_embed_url, file_format="npy")
+    text_reader  = EmbeddingReader(embeddings_folder=text_embed_url, file_format="npy")
+    num_data_points = text_reader.count
+
    ### Training code ###
    scaler = GradScaler(enabled=amp)
    optimizer = get_optimizer(diffusion_prior.net.parameters(), wd=weight_decay, lr=learning_rate)
@@ -136,12 +169,15 @@ def train(image_embed_dim,

    train_set_size = int(train_percent*num_data_points)
    val_set_size = int(val_percent*num_data_points)
+    eval_start = train_set_size

    for _ in range(epochs):
-        diffusion_prior.train()

        for emb_images,emb_text in zip(image_reader(batch_size=batch_size, start=0, end=train_set_size),
                text_reader(batch_size=batch_size, start=0, end=train_set_size)):
+
+            diffusion_prior.train()
+            
            emb_images_tensor = torch.tensor(emb_images[0]).to(device)
            emb_text_tensor = torch.tensor(emb_text[0]).to(device)

@@ -156,9 +192,13 @@ def train(image_embed_dim,
            if(int(time.time()-t) >= 60*save_interval):
                t = time.time()

-                save_model(
+                save_diffusion_model(
                    save_path,
-                    dict(model=diffusion_prior.state_dict(), optimizer=optimizer.state_dict(), scaler=scaler.state_dict()))
+                    diffusion_prior,
+                    optimizer,
+                    scaler,
+                    config,
+                    image_embed_dim)

            # Log to wandb
            wandb.log({"Training loss": loss.item(),
@@ -168,14 +208,22 @@ def train(image_embed_dim,
            # 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:
-                diff_cosine_sim = report_cosine_sims(diffusion_prior,
+                report_cosine_sims(diffusion_prior,
                        image_reader,
                        text_reader,
                        train_set_size,
-                        val_set_size,
                        NUM_TEST_EMBEDDINGS,
                        device)
-                wandb.log({"Cosine similarity difference": diff_cosine_sim})
+                ### Evaluate model(validation run) ###
+                eval_model(diffusion_prior,
+                        device,
+                        image_reader,
+                        text_reader,
+                        eval_start,
+                        eval_start+NUM_TEST_EMBEDDINGS,
+                        NUM_TEST_EMBEDDINGS,
+                        dp_loss_type,
+                        phase="Validation")

            scaler.unscale_(optimizer)
            nn.utils.clip_grad_norm_(diffusion_prior.parameters(), max_grad_norm)
@@ -184,11 +232,6 @@ def train(image_embed_dim,
            scaler.update()
            optimizer.zero_grad()

-        ### Evaluate model(validation run) ###
-        start = train_set_size
-        end=start+val_set_size
-        eval_model(diffusion_prior,device,image_reader,text_reader,start,end,batch_size,dp_loss_type,phase="Validation")
-
    ### Test run ###
    test_set_size = int(test_percent*train_set_size) 
    start=train_set_size+val_set_size
@@ -200,7 +243,6 @@ def main():
    # Logging
    parser.add_argument("--wandb-entity", type=str, default="laion")
    parser.add_argument("--wandb-project", type=str, default="diffusion-prior")
-    parser.add_argument("--wandb-name", type=str, default="laion-dprior")
    parser.add_argument("--wandb-dataset", type=str, default="LAION-5B")
    parser.add_argument("--wandb-arch", type=str, default="DiffusionPrior")
    # URLs for embeddings 
@@ -209,6 +251,7 @@ def main():
    # Hyperparameters
    parser.add_argument("--learning-rate", type=float, default=1.1e-4)
    parser.add_argument("--weight-decay", type=float, default=6.02e-2)
+    parser.add_argument("--dropout", type=float, default=5e-2)
    parser.add_argument("--max-grad-norm", type=float, default=0.5)
    parser.add_argument("--batch-size", type=int, default=10**4)
    parser.add_argument("--num-epochs", type=int, default=5)
@@ -226,7 +269,6 @@ def main():
    # DiffusionPrior(dp) parameters
    parser.add_argument("--dp-condition-on-text-encodings", type=bool, default=False)
    parser.add_argument("--dp-timesteps", type=int, default=100)
-    parser.add_argument("--dp-l2norm-output", type=bool, default=False)
    parser.add_argument("--dp-normformer", type=bool, default=False)
    parser.add_argument("--dp-cond-drop-prob", type=float, default=0.1)
    parser.add_argument("--dp-loss-type", type=str, default="l2")
@@ -235,22 +277,40 @@ def main():
    # Model checkpointing interval(minutes)
    parser.add_argument("--save-interval", type=int, default=30)
    parser.add_argument("--save-path", type=str, default="./diffusion_prior_checkpoints")
+    # Saved model path 
+    parser.add_argument("--pretrained-model-path", type=str, default=None)

    args = parser.parse_args()

-    print("Setting up wandb logging... Please wait...")
+    config = ({"learning_rate": args.learning_rate,
+        "architecture": args.wandb_arch,
+        "dataset": args.wandb_dataset,
+        "weight_decay":args.weight_decay,
+        "max_gradient_clipping_norm":args.max_grad_norm,
+        "batch_size":args.batch_size,
+        "epochs": args.num_epochs,
+        "diffusion_prior_network":{"depth":args.dpn_depth,
+        "dim_head":args.dpn_dim_head,
+        "heads":args.dpn_heads,
+        "normformer":args.dp_normformer},
+        "diffusion_prior":{"condition_on_text_encodings": args.dp_condition_on_text_encodings,
+        "timesteps": args.dp_timesteps,
+        "cond_drop_prob":args.dp_cond_drop_prob,
+        "loss_type":args.dp_loss_type,
+        "clip":args.clip}
+        })

-    wandb.init(
-      entity=args.wandb_entity,
-      project=args.wandb_project,
-      config={
-      "learning_rate": args.learning_rate,
-      "architecture": args.wandb_arch,
-      "dataset": args.wandb_dataset,
-      "epochs": args.num_epochs,
-      })
+    RESUME = False
+    # Check if DPRIOR_PATH exists(saved model path)
+    DPRIOR_PATH = args.pretrained_model_path
+    if(DPRIOR_PATH is not None):
+        RESUME = True
+    else:
+        wandb.init(
+          entity=args.wandb_entity,
+          project=args.wandb_project,
+          config=config)

-    print("wandb logging setup done!")
    # Obtain the utilized device.

    has_cuda = torch.cuda.is_available()
@@ -279,9 +339,15 @@ def main():
          args.save_interval,
          args.save_path,
          device,
+          RESUME,
+          DPRIOR_PATH,
+          config,
+          args.wandb_entity,
+          args.wandb_project,
          args.learning_rate,
          args.max_grad_norm,
          args.weight_decay,
+          args.dropout,
          args.amp)

 if __name__ == "__main__":
Author	SHA1	Message	Date
Phil Wang	f1739267e4	simplify more	2022-05-14 17:13:13 -07:00
Phil Wang	b0cd5f24b6	take care of gradient accumulation automatically for researchers, by passing in a `max_batch_size` on the decoder or diffusion prior trainer forward	2022-05-14 17:04:09 -07:00
Phil Wang	b494ed81d4	take care of backwards within trainer classes for diffusion prior and decoder, readying to take care of gradient accumulation as well (plus, unsure if loss should be backwards within autocast block)	2022-05-14 15:49:24 -07:00
Phil Wang	ff3474f05c	normalize conditioning tokens outside of cross attention blocks	2022-05-14 14:23:52 -07:00
Phil Wang	d5293f19f1	lineup with paper	2022-05-14 13:57:00 -07:00
Phil Wang	e697183849	be able to customize adam eps	2022-05-14 13:55:04 -07:00
Phil Wang	591d37e266	lower default initial learning rate to what Jonathan Ho had in his original repo	2022-05-14 13:22:43 -07:00
Phil Wang	d1f02e8f49	always use sandwich norm for attention layer	2022-05-14 12:13:41 -07:00
Phil Wang	9faab59b23	use post-attn-branch layernorm in attempt to stabilize cross attention conditioning in decoder	2022-05-14 11:58:09 -07:00
Phil Wang	5d27029e98	make sure lowres conditioning image is properly normalized to -1 to 1 for cascading ddpm	2022-05-14 01:23:54 -07:00
Phil Wang	3115fa17b3	fix everything around normalizing images to -1 to 1 for ddpm training automatically	2022-05-14 01:17:11 -07:00
Phil Wang	124d8577c8	move the inverse normalization function called before image embeddings are derived from clip to within the diffusion prior and decoder classes	2022-05-14 00:37:52 -07:00
Phil Wang	2db0c9794c	comments	2022-05-12 14:25:20 -07:00
Phil Wang	2277b47ffd	make sure learned variance can work for any number of unets in the decoder, defaults to first unet, as suggested was used in the paper	2022-05-12 14:18:15 -07:00
Phil Wang	28b58e568c	cleanup in preparation of option for learned variance	2022-05-12 12:04:52 -07:00
Phil Wang	924455d97d	align the ema model device back after sampling from the cascading ddpm in the decoder	2022-05-11 19:56:54 -07:00
Phil Wang	6021945fc8	default to l2 loss	2022-05-11 19:24:51 -07:00
Light-V	6f76652d11	fix typo in README.md (#85 ) The default config for clip from openai should be ViT-B/32	2022-05-11 13:38:16 -07:00
Phil Wang	3dda2570ed	fix amp issue for https://github.com/lucidrains/DALLE2-pytorch/issues/82	2022-05-11 08:21:39 -07:00
Phil Wang	2f3c02dba8	numerical accuracy for noise schedule parameters	2022-05-10 15:28:46 -07:00
Phil Wang	908088cfea	wrap up cross embed layer feature	2022-05-10 12:19:34 -07:00
Phil Wang	8dc8a3de0d	product management	2022-05-10 11:51:38 -07:00
Phil Wang	35f89556ba	bring in the cross embed layer from Crossformer paper for initial convolution in unet	2022-05-10 11:50:38 -07:00
Phil Wang	2b55f753b9	fix new issue with github actions and auto pypi package uploading	2022-05-10 10:51:15 -07:00
Phil Wang	fc8fce38fb	make sure cascading DDPM can be trained unconditionally, to ready for CLI one command training for the public	2022-05-10 10:48:10 -07:00
Phil Wang	a1bfb03ba4	project management	2022-05-10 10:13:51 -07:00
Phil Wang	b1e7b5f6bb	make sure resnet groups in unet is finely customizable	2022-05-10 10:12:50 -07:00
z	10b905b445	smol typo (#81 )	2022-05-10 09:52:50 -07:00
Phil Wang	9b322ea634	patch	2022-05-09 19:46:19 -07:00
Phil Wang	ba64ea45cc	0.2.3	2022-05-09 16:50:31 -07:00
Phil Wang	64f7be1926	some cleanup	2022-05-09 16:50:21 -07:00
Phil Wang	db805e73e1	fix a bug with numerical stability in attention, sorry! 🐛	2022-05-09 16:23:37 -07:00
z	cb07b37970	Ensure Eval Mode In Metric Functions (#79 ) * add eval/train toggles * train/eval flags * shift train toggle Co-authored-by: nousr <z@localhost.com>	2022-05-09 16:05:40 -07:00
Phil Wang	a774bfefe2	add attention and feedforward dropouts to train_diffusion_prior script	2022-05-09 13:57:15 -07:00
Phil Wang	2ae57f0cf5	cleanup	2022-05-09 13:51:26 -07:00
Phil Wang	e46eaec817	deal the diffusion prior problem yet another blow	2022-05-09 11:08:52 -07:00
Kumar R	8647cb5e76	Val loss changes, with quite a few other changes. This is in place of the earlier PR(https://github.com/lucidrains/DALLE2-pytorch/pull/67 ) (#77 ) * Val_loss changes - no rebased with lucidrains' master. * Val Loss changes - now rebased with lucidrains' master * train_diffusion_prior.py updates * dalle2_pytorch.py updates * __init__.py changes * Update train_diffusion_prior.py * Update dalle2_pytorch.py * Update train_diffusion_prior.py * Update train_diffusion_prior.py * Update dalle2_pytorch.py * Update train_diffusion_prior.py * Update train_diffusion_prior.py * Update train_diffusion_prior.py * Update train_diffusion_prior.py * Update README.md * Update README.md * Update README.md * Update README.md * Update README.md * Update README.md * Update README.md * Update README.md * Update README.md	2022-05-09 08:53:29 -07:00
Phil Wang	53c189e46a	give more surface area for attention in diffusion prior	2022-05-09 08:08:11 -07:00
Phil Wang	dde51fd362	revert restriction for classifier free guidance for diffusion prior, given @crowsonkb advice	2022-05-07 20:55:41 -07:00
Nasir Khalid	2eac7996fa	Additional image_embed metric (#75 ) Added metric to track image_embed vs predicted_image_embed	2022-05-07 14:32:33 -07:00
Phil Wang	4010aec033	turn off classifier free guidance if predicting x_start for diffusion prior	2022-05-07 09:38:17 -07:00
Phil Wang	c87b84a259	todo	2022-05-07 09:21:08 -07:00
Phil Wang	8b05468653	todo	2022-05-07 08:33:45 -07:00
Phil Wang	830afd3c15	sinusoidal embed time embeddings for diffusion prior as well, for continuous version	2022-05-07 08:32:43 -07:00
Phil Wang	8f93729d19	when in doubt, make it a hyperparameter	2022-05-07 07:52:17 -07:00
z	cd5f2c1de4	simulate unrelated captions as a training metric (#66 ) * add unrelated embedding metric * change to torch.roll Co-authored-by: nousr <z@localhost.com> Co-authored-by: nousr <>	2022-05-07 05:34:59 -07:00
Phil Wang	85ed77d512	fix a potentially huge bug thanks to @CiaoHe https://github.com/lucidrains/DALLE2-pytorch/issues/71	2022-05-07 05:05:54 -07:00
Piero Rolando	fd53fa17db	Fix a typo in README (#70 ) Change "pyhon" for "python" (correct)	2022-05-06 16:53:36 -07:00
Phil Wang	3676ef4d49	make sure vqgan-vae trainer supports mixed precision	2022-05-06 10:44:16 -07:00
Phil Wang	28e944f328	make sure openai clip adapter outputs l2normed embeddings	2022-05-06 10:12:03 -07:00
Phil Wang	14e63a3f67	also offer l2norm clamping in diffusion prior during training, if one were using predict x0 objective	2022-05-06 10:05:14 -07:00
Phil Wang	09e9eaa5a6	project management	2022-05-06 09:00:22 -07:00
Phil Wang	e6d752cf4a	reprioritize	2022-05-06 08:55:26 -07:00
Phil Wang	ad20a14a4d	bring in rotary embeddings for diffusion prior causal transformer (the most powerful relative positional encoding, used in PaLM) - 0.1.0 because of breaking change	2022-05-06 08:45:30 -07:00
Phil Wang	0be1e0d64c	support CoCa, which seems to be better than CLIP (has an autoregressive text encoder) https://arxiv.org/abs/2205.01917	2022-05-06 08:27:12 -07:00
Phil Wang	98df1ba51e	add diffusion prior trainer, which automatically takes care of the exponential moving average (training and sampling), as well as mixed precision, gradient clipping	2022-05-06 08:11:09 -07:00