lineup with paper

The default config for clip from openai should be ViT-B/32

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()
 

dalle2_pytorch/dalle2_pytorch.py

@@ -1,7 +1,7 @@
 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
@@ -33,6 +33,10 @@ from rotary_embedding_torch import RotaryEmbedding
 from x_clip import CLIP
 from coca_pytorch import CoCa
 
+# constants
+
+NAT = 1. / math.log(2.)
+
 # helper functions
 
 def exists(val):
@@ -41,6 +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
@@ -91,6 +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)
 
@@ -107,10 +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
@@ -271,7 +286,7 @@ class OpenAIClipAdapter(BaseClipAdapter):
     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(l2norm(image_embed.float()), None)
 
@@ -297,13 +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])
@@ -314,21 +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
 
 
@@ -368,17 +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)
 
@@ -386,19 +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 = (
@@ -571,7 +614,6 @@ class Attention(nn.Module):
         heads = 8,
         dropout = 0.,
         causal = False,
-        post_norm = False,
         rotary_emb = None
     ):
         super().__init__()
@@ -581,7 +623,6 @@ 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))
@@ -592,7 +633,7 @@ class Attention(nn.Module):
 
         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):
@@ -649,8 +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__(
@@ -676,7 +716,7 @@ class CausalTransformer(nn.Module):
         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, rotary_emb = rotary_emb),
+                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)
             ]))
 
@@ -827,7 +867,7 @@ class DiffusionPrior(BaseGaussianDiffusion):
         image_channels = 3,
         timesteps = 1000,
         cond_drop_prob = 0.,
-        loss_type = "l1",
+        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
@@ -1138,7 +1178,11 @@ class CrossAttention(nn.Module):
         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
@@ -1268,6 +1312,7 @@ 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/
@@ -1298,6 +1343,7 @@ 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 // 3 * 2)
@@ -1403,11 +1449,9 @@ class Unet(nn.Module):
                 Upsample(dim_in)
             ]))
 
-        out_dim = default(out_dim, channels)
-
         self.final_conv = nn.Sequential(
             ResnetBlock(dim, dim, groups = resnet_groups[0]),
-            nn.Conv2d(dim, out_dim, 1)
+            nn.Conv2d(dim, self.channels_out, 1)
         )
 
     # if the current settings for the unet are not correct
@@ -1417,13 +1461,25 @@ class Unet(nn.Module):
         *,
         lowres_cond,
         channels,
+        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 and cond_on_text_encodings == self.cond_on_text_encodings:
+        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, 'cond_on_text_encodings': cond_on_text_encodings}
+        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(
@@ -1488,11 +1544,12 @@ 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)
@@ -1516,10 +1573,12 @@ 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)
@@ -1607,7 +1666,7 @@ 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,
@@ -1620,6 +1679,8 @@ class Decoder(BaseGaussianDiffusion):
         clip_denoised = True,
         clip_x_start = True,
         clip_adapter_overrides = dict(),
+        learned_variance = True,
+        vb_loss_weight = 0.001,
         unconditional = False
     ):
         super().__init__(
@@ -1658,10 +1719,18 @@ 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))
 
@@ -1669,12 +1738,14 @@ 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 and not unconditional,
                 cond_on_text_encodings = one_unet.cond_on_text_encodings and not unconditional,
-                channels = unet_channels
+                channels = unet_channels,
+                channels_out = unet_channels_out
             )
 
             self.unets.append(one_unet)
@@ -1737,8 +1808,11 @@ 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
@@ -1749,24 +1823,38 @@ 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,
@@ -1778,17 +1866,26 @@ 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,
@@ -1799,10 +1896,48 @@ 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
@@ -1829,7 +1964,7 @@ class Decoder(BaseGaussianDiffusion):
         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 is_cuda else null_context()
 
@@ -1855,6 +1990,7 @@ 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
                 )
@@ -1884,6 +2020,7 @@ 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)
@@ -1921,7 +2058,7 @@ 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
 
@@ -1971,4 +2108,3 @@ class DALLE2(nn.Module):
             return images[0]
 
         return images
-

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)

dalle2_pytorch/train.py

@@ -90,7 +90,7 @@ class EMA(nn.Module):
     def __init__(
         self,
         model,
-        beta = 0.99,
+        beta = 0.9999,
         update_after_step = 1000,
         update_every = 10,
     ):
@@ -105,6 +105,10 @@ 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
 
@@ -143,6 +147,7 @@ class DiffusionPriorTrainer(nn.Module):
         use_ema = True,
         lr = 3e-4,
         wd = 1e-2,
+        eps = 1e-6,
         max_grad_norm = None,
         amp = False,
         **kwargs
@@ -169,6 +174,7 @@ class DiffusionPriorTrainer(nn.Module):
             diffusion_prior.parameters(),
             lr = lr,
             wd = wd,
+            eps = eps,
             **kwargs
         )
 
@@ -217,8 +223,9 @@ class DecoderTrainer(nn.Module):
         self,
         decoder,
         use_ema = True,
-        lr = 3e-4,
+        lr = 2e-5,
         wd = 1e-2,
+        eps = 1e-8,
         max_grad_norm = None,
         amp = False,
         **kwargs
@@ -243,13 +250,14 @@ 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
             )
 
@@ -305,6 +313,11 @@ 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(

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.2.8',
+  version = '0.2.22',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',