fix decoder needing separate conditional dropping probabilities for image embeddings and text encodings, thanks to @xiankgx !

Fix spelling and grammatical errors

README.md

@@ -47,7 +47,7 @@ clip = CLIP(
     use_all_token_embeds = True,            # whether to use fine-grained contrastive learning (FILIP)
     decoupled_contrastive_learning = True,  # use decoupled contrastive learning (DCL) objective function, removing positive pairs from the denominator of the InfoNCE loss (CLOOB + DCL)
     extra_latent_projection = True,         # whether to use separate projections for text-to-image vs image-to-text comparisons (CLOOB)
-    use_visual_ssl = True,                  # whether to do self supervised learning on iages
+    use_visual_ssl = True,                  # whether to do self supervised learning on images
     visual_ssl_type = 'simclr',             # can be either 'simclr' or 'simsiam', depending on using DeCLIP or SLIP
     use_mlm = False,                        # use masked language learning (MLM) on text (DeCLIP)
     text_ssl_loss_weight = 0.05,            # weight for text MLM loss
@@ -110,7 +110,8 @@ decoder = Decoder(
     unet = unet,
     clip = clip,
     timesteps = 100,
-    cond_drop_prob = 0.2
+    image_cond_drop_prob = 0.1,
+    text_cond_drop_prob = 0.5
 ).cuda()
 
 # mock images (get a lot of this)
@@ -229,7 +230,8 @@ decoder = Decoder(
     unet = (unet1, unet2),            # insert both unets in order of low resolution to highest resolution (you can have as many stages as you want here)
     image_sizes = (256, 512),         # resolutions, 256 for first unet, 512 for second. these must be unique and in ascending order (matches with the unets passed in)
     timesteps = 1000,
-    cond_drop_prob = 0.2
+    image_cond_drop_prob = 0.1,
+    text_cond_drop_prob = 0.5
 ).cuda()
 
 # mock images (get a lot of this)
@@ -348,7 +350,8 @@ decoder = Decoder(
     image_sizes = (128, 256),
     clip = clip,
     timesteps = 100,
-    cond_drop_prob = 0.2,
+    image_cond_drop_prob = 0.1,
+    text_cond_drop_prob = 0.5,
     condition_on_text_encodings = False  # set this to True if you wish to condition on text during training and sampling
 ).cuda()
 
@@ -430,8 +433,8 @@ images = torch.randn(4, 3, 256, 256).cuda()
 # precompute the text and image embeddings
 # here using the diffusion prior class, but could be done with CLIP alone
 
-clip_image_embeds = diffusion_prior.get_image_embed(images)
-clip_text_embeds = diffusion_prior.get_text_cond(text).get('text_embed')
+clip_image_embeds = diffusion_prior.clip.embed_image(images).image_embed
+clip_text_embeds = diffusion_prior.clip.embed_text(text).text_embed
 
 # feed text and images into diffusion prior network
 
@@ -495,6 +498,95 @@ loss.backward()
 # now the diffusion prior can generate image embeddings from the text embeddings
 ```
 
+## OpenAI CLIP
+
+Although there is the possibility they are using an unreleased, more powerful CLIP, you can use one of the released ones, if you do not wish to train your own CLIP from scratch. This will also allow the community to more quickly validate the conclusions of the paper.
+
+To use a pretrained OpenAI CLIP, simply import `OpenAIClipAdapter` and pass it into the `DiffusionPrior` or `Decoder` like so
+
+```python
+import torch
+from dalle2_pytorch import DALLE2, DiffusionPriorNetwork, DiffusionPrior, Unet, Decoder, OpenAIClipAdapter
+
+# openai pretrained clip - defaults to ViT/B-32
+
+clip = OpenAIClipAdapter()
+
+# 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()
+
+loss = diffusion_prior(text, images)
+loss.backward()
+
+# do above for many steps ...
+
+# decoder (with unet)
+
+unet1 = Unet(
+    dim = 128,
+    image_embed_dim = 512,
+    cond_dim = 128,
+    channels = 3,
+    dim_mults=(1, 2, 4, 8)
+).cuda()
+
+unet2 = Unet(
+    dim = 16,
+    image_embed_dim = 512,
+    cond_dim = 128,
+    channels = 3,
+    dim_mults = (1, 2, 4, 8, 16)
+).cuda()
+
+decoder = Decoder(
+    unet = (unet1, unet2),
+    image_sizes = (128, 256),
+    clip = clip,
+    timesteps = 100,
+    image_cond_drop_prob = 0.1,
+    text_cond_drop_prob = 0.5,
+    condition_on_text_encodings = False  # set this to True if you wish to condition on text during training and sampling
+).cuda()
+
+for unet_number in (1, 2):
+    loss = decoder(images, unet_number = unet_number) # this can optionally be decoder(images, text) if you wish to condition on the text encodings as well, though it was hinted in the paper it didn't do much
+    loss.backward()
+
+# do above for many steps
+
+dalle2 = DALLE2(
+    prior = diffusion_prior,
+    decoder = decoder
+)
+
+images = dalle2(
+    ['a butterfly trying to escape a tornado'],
+    cond_scale = 2. # classifier free guidance strength (> 1 would strengthen the condition)
+)
+
+# save your image (in this example, of size 256x256)
+```
+
+Now you'll just have to worry about training the Prior and the Decoder!
+
 ## Experimental
 
 ### DALL-E2 with Latent Diffusion
@@ -528,7 +620,7 @@ clip = CLIP(
 # 3 unets for the decoder (a la cascading DDPM)
 
 # first two unets are doing latent diffusion
-# vqgan-vae must be trained before hand
+# vqgan-vae must be trained beforehand
 
 vae1 = VQGanVAE(
     dim = 32,
@@ -581,7 +673,8 @@ decoder = Decoder(
     unet = (unet1, unet2, unet3),      # insert unets in order of low resolution to highest resolution (you can have as many stages as you want here)
     image_sizes = (256, 512, 1024),    # resolutions, 256 for first unet, 512 for second, 1024 for third
     timesteps = 100,
-    cond_drop_prob = 0.2
+    image_cond_drop_prob = 0.1,
+    text_cond_drop_prob = 0.5
 ).cuda()
 
 # mock images (get a lot of this)
@@ -651,6 +744,7 @@ Once built, images will be saved to the same directory the command is invoked
 - [ ] become an expert with unets, cleanup unet code, make it fully configurable, port all learnings over to https://github.com/lucidrains/x-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
+- [ ] just take care of the training for the decoder in a wrapper class, as each unet in the cascade will need its own optimizer
 - [ ] 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

dalle2_pytorch/__init__.py

@@ -1,4 +1,5 @@
 from dalle2_pytorch.dalle2_pytorch import DALLE2, DiffusionPriorNetwork, DiffusionPrior, Unet, Decoder
+from dalle2_pytorch.dalle2_pytorch import OpenAIClipAdapter
 
 from dalle2_pytorch.vqgan_vae import VQGanVAE
 from x_clip import CLIP

dalle2_pytorch/dalle2_pytorch.py

@@ -3,6 +3,7 @@ from tqdm import tqdm
 from inspect import isfunction
 from functools import partial
 from contextlib import contextmanager
+from collections import namedtuple
 
 import torch
 import torch.nn.functional as F
@@ -90,8 +91,21 @@ def resize_image_to(t, image_size, mode = 'bilinear'): # take a look at https://
 
     return F.interpolate(t, size = shape, mode = mode, align_corners = False)
 
+# image normalization functions
+# ddpms expect images to be in the range of -1 to 1
+# but CLIP may otherwise
+
+def normalize_img(img):
+    return img * 2 - 1
+
+def unnormalize_img(normed_img):
+    return (normed_img + 1) * 0.5
+
 # clip related adapters
 
+EmbeddedText = namedtuple('EmbedTextReturn', ['text_embed', 'text_encodings', 'text_mask'])
+EmbeddedImage = namedtuple('EmbedImageReturn', ['image_embed', 'image_encodings'])
+
 class BaseClipAdapter(nn.Module):
     def __init__(self, clip):
         super().__init__()
@@ -109,6 +123,10 @@ class BaseClipAdapter(nn.Module):
     def image_channels(self):
         raise NotImplementedError
 
+    @property
+    def max_text_len(self):
+        raise NotImplementedError
+
     def embed_text(self, text):
         raise NotImplementedError
 
@@ -128,12 +146,18 @@ class XClipAdapter(BaseClipAdapter):
     def image_channels(self):
         return self.clip.image_channels
 
+    @property
+    def max_text_len(self):
+        return self.clip.text_seq_len
+
     @torch.no_grad()
     def embed_text(self, text):
+        text = text[..., :self.max_text_len]
+        text_mask = text != 0
         encoder_output = self.clip.text_transformer(text)
         text_cls, text_encodings = encoder_output[:, 0], encoder_output[:, 1:]
         text_embed = self.clip.to_text_latent(text_cls)
-        return l2norm(text_embed), text_encodings
+        return EmbeddedText(l2norm(text_embed), text_encodings, text_mask)
 
     @torch.no_grad()
     def embed_image(self, image):
@@ -141,7 +165,69 @@ class XClipAdapter(BaseClipAdapter):
         encoder_output = self.clip.visual_transformer(image)
         image_cls, image_encodings = encoder_output[:, 0], encoder_output[:, 1:]
         image_embed = self.clip.to_visual_latent(image_cls)
-        return l2norm(image_embed), image_encodings
+        return EmbeddedImage(l2norm(image_embed), image_encodings)
+
+class OpenAIClipAdapter(BaseClipAdapter):
+    def __init__(
+        self,
+        name = 'ViT-B/32'
+    ):
+        import clip
+        openai_clip, preprocess = clip.load(name)
+        super().__init__(openai_clip)
+
+        text_attention_final = self.find_layer('ln_final')
+        self.handle = text_attention_final.register_forward_hook(self._hook)
+        self.clip_normalize = preprocess.transforms[-1]
+        self.cleared = False
+
+    def find_layer(self,  layer):
+        modules = dict([*self.clip.named_modules()])
+        return modules.get(layer, None)
+
+    def clear(self):
+        if self.cleared:
+            return
+
+        self.handle()
+
+    def _hook(self, _, inputs, outputs):
+        self.text_encodings = outputs
+
+    @property
+    def dim_latent(self):
+        return 512
+
+    @property
+    def image_size(self):
+        return self.clip.visual.input_resolution
+
+    @property
+    def image_channels(self):
+        return 3
+
+    @property
+    def max_text_len(self):
+        return self.clip.context_length
+
+    @torch.no_grad()
+    def embed_text(self, text):
+        text = text[..., :self.max_text_len]
+        text_mask = text != 0
+        assert not self.cleared
+
+        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)
+
+    @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_embed = self.clip.encode_image(image)
+        return EmbeddedImage(image_embed.float(), None)
 
 # classifier free guidance functions
 
@@ -223,7 +309,18 @@ class BaseGaussianDiffusion(nn.Module):
 
         timesteps, = betas.shape
         self.num_timesteps = int(timesteps)
+
+        if loss_type == 'l1':
+            loss_fn = F.l1_loss
+        elif loss_type == 'l2':
+            loss_fn = F.mse_loss
+        elif loss_type == 'huber':
+            loss_fn = F.smooth_l1_loss
+        else:
+            raise NotImplementedError()
+
         self.loss_type = loss_type
+        self.loss_fn = loss_fn
 
         self.register_buffer('betas', betas)
         self.register_buffer('alphas_cumprod', alphas_cumprod)
@@ -587,14 +684,14 @@ class DiffusionPriorNetwork(nn.Module):
 
         # classifier free guidance
 
-        cond_prob_mask = prob_mask_like((batch,), cond_drop_prob, device = device)
-        cond_prob_mask = rearrange(cond_prob_mask, 'b -> b 1')
+        keep_mask = prob_mask_like((batch,), 1 - cond_drop_prob, device = device)
+        keep_mask = rearrange(keep_mask, 'b -> b 1')
 
-        mask &= cond_prob_mask
+        mask &= keep_mask
 
         # whether text embedding is masked or not depends on the classifier free guidance conditional masking
 
-        mask = torch.cat((mask, cond_prob_mask), dim = 1)
+        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
@@ -703,29 +800,21 @@ class DiffusionPrior(BaseGaussianDiffusion):
             img = self.p_sample(img, torch.full((b,), i, device = device, dtype = torch.long), text_cond = text_cond)
         return img
 
-    def p_losses(self, image_embed, t, text_cond, noise = None):
+    def p_losses(self, image_embed, times, text_cond, noise = None):
         noise = default(noise, lambda: torch.randn_like(image_embed))
 
-        image_embed_noisy = self.q_sample(x_start = image_embed, t = t, noise = noise)
+        image_embed_noisy = self.q_sample(x_start = image_embed, t = times, noise = noise)
 
-        x_recon = self.net(
+        pred = self.net(
             image_embed_noisy,
-            t,
+            times,
             cond_drop_prob = self.cond_drop_prob,
             **text_cond
         )
 
-        to_predict = noise if not self.predict_x_start else image_embed
-
-        if self.loss_type == 'l1':
-            loss = F.l1_loss(to_predict, x_recon)
-        elif self.loss_type == 'l2':
-            loss = F.mse_loss(to_predict, x_recon)
-        elif self.loss_type == "huber":
-            loss = F.smooth_l1_loss(to_predict, x_recon)
-        else:
-            raise NotImplementedError()
+        target = noise if not self.predict_x_start else image_embed
 
+        loss = self.loss_fn(pred, target)
         return loss
 
     @torch.no_grad()
@@ -738,12 +827,12 @@ class DiffusionPrior(BaseGaussianDiffusion):
         batch_size = text.shape[0]
         image_embed_dim = self.image_embed_dim
 
-        text_embed, text_encodings = self.clip.embed_text(text)
+        text_embed, text_encodings, text_mask = self.clip.embed_text(text)
 
         text_cond = dict(text_embed = text_embed)
 
         if self.condition_on_text_encodings:
-            text_cond = {**text_cond, 'text_encodings': text_encodings, 'mask': text != 0}
+            text_cond = {**text_cond, 'text_encodings': text_encodings, 'mask': text_mask}
 
         image_embeds = self.p_sample_loop((batch_size, image_embed_dim), text_cond = text_cond)
         text_embeds = text_cond['text_embed']
@@ -780,8 +869,7 @@ class DiffusionPrior(BaseGaussianDiffusion):
         # calculate text conditionings, based on what is passed in
 
         if exists(text):
-            text_embed, text_encodings = self.clip.embed_text(text)
-            text_mask = text != 0
+            text_embed, text_encodings, text_mask = self.clip.embed_text(text)
 
         text_cond = dict(text_embed = text_embed)
 
@@ -1066,13 +1154,14 @@ class Unet(nn.Module):
         self,
         *,
         lowres_cond,
-        channels
+        channels,
+        cond_on_image_embeds
     ):
-        if lowres_cond == self.lowres_cond and channels == self.channels:
+        if lowres_cond == self.lowres_cond and channels == self.channels and cond_on_image_embeds == self.cond_on_image_embeds:
             return self
 
-        updated_kwargs = {**self._locals, 'lowres_cond': lowres_cond, 'channels': channels}
-        return self.__class__(**updated_kwargs)
+        updated_kwargs = {'lowres_cond': lowres_cond, 'channels': channels, 'cond_on_image_embeds': cond_on_image_embeds}
+        return self.__class__(**{**self._locals, **updated_kwargs})
 
     def forward_with_cond_scale(
         self,
@@ -1085,7 +1174,7 @@ class Unet(nn.Module):
         if cond_scale == 1:
             return logits
 
-        null_logits = self.forward(*args, cond_drop_prob = 1., **kwargs)
+        null_logits = self.forward(*args, text_cond_drop_prob = 1., image_cond_drop_prob = 1., **kwargs)
         return null_logits + (logits - null_logits) * cond_scale
 
     def forward(
@@ -1096,7 +1185,8 @@ class Unet(nn.Module):
         image_embed,
         lowres_cond_img = None,
         text_encodings = None,
-        cond_drop_prob = 0.,
+        image_cond_drop_prob = 0.,
+        text_cond_drop_prob = 0.,
         blur_sigma = None,
         blur_kernel_size = None
     ):
@@ -1115,8 +1205,10 @@ class Unet(nn.Module):
 
         # conditional dropout
 
-        cond_prob_mask = prob_mask_like((batch_size,), cond_drop_prob, device = device)
-        cond_prob_mask = rearrange(cond_prob_mask, 'b -> b 1 1')
+        image_keep_mask = prob_mask_like((batch_size,), 1 - image_cond_drop_prob, device = device)
+        text_keep_mask = prob_mask_like((batch_size,), 1 - text_cond_drop_prob, device = device)
+
+        image_keep_mask, text_keep_mask = rearrange_many((image_keep_mask, text_keep_mask), 'b -> b 1 1')
 
         # mask out image embedding depending on condition dropout
         # for classifier free guidance
@@ -1127,7 +1219,7 @@ class Unet(nn.Module):
             image_tokens = self.image_to_cond(image_embed)
 
             image_tokens = torch.where(
-                cond_prob_mask,
+                image_keep_mask,
                 image_tokens,
                 self.null_image_embed
             )
@@ -1139,7 +1231,7 @@ class Unet(nn.Module):
         if exists(text_encodings) and self.cond_on_text_encodings:
             text_tokens = self.text_to_cond(text_encodings)
             text_tokens = torch.where(
-                cond_prob_mask,
+                text_keep_mask,
                 text_tokens,
                 self.null_text_embed[:, :text_tokens.shape[1]]
             )
@@ -1209,7 +1301,7 @@ class LowresConditioner(nn.Module):
         target_image_size = cast_tuple(target_image_size, 2)
 
         if self.training and self.downsample_first and exists(downsample_image_size):
-            cond_fmap = resize_image_to(cond_fmap, target_image_size, mode = self.cond_upsample_mode)
+            cond_fmap = resize_image_to(cond_fmap, downsample_image_size, mode = self.cond_upsample_mode)
 
         if self.training:
             # when training, blur the low resolution conditional image
@@ -1229,7 +1321,8 @@ class Decoder(BaseGaussianDiffusion):
         clip,
         vae = tuple(),
         timesteps = 1000,
-        cond_drop_prob = 0.2,
+        image_cond_drop_prob = 0.1,
+        text_cond_drop_prob = 0.5,
         loss_type = 'l1',
         beta_schedule = 'cosine',
         predict_x_start = False,
@@ -1279,6 +1372,7 @@ class Decoder(BaseGaussianDiffusion):
 
             one_unet = one_unet.cast_model_parameters(
                 lowres_cond = not is_first,
+                cond_on_image_embeds = is_first,
                 channels = unet_channels
             )
 
@@ -1312,7 +1406,8 @@ class Decoder(BaseGaussianDiffusion):
 
         # classifier free guidance
 
-        self.cond_drop_prob = cond_drop_prob
+        self.image_cond_drop_prob = image_cond_drop_prob
+        self.text_cond_drop_prob = text_cond_drop_prob
 
     def get_unet(self, unet_number):
         assert 0 < unet_number <= len(self.unets)
@@ -1336,11 +1431,8 @@ class Decoder(BaseGaussianDiffusion):
 
     @torch.no_grad()
     def get_image_embed(self, image):
-        image = resize_image_to(image, self.clip_image_size)
-        image_encoding = self.clip.visual_transformer(image)
-        image_cls = image_encoding[:, 0]
-        image_embed = self.clip.to_visual_latent(image_cls)
-        return l2norm(image_embed)
+        image_embed, _ = self.clip.embed_image(image)
+        return image_embed
 
     def p_mean_variance(self, unet, x, t, image_embed, text_encodings = 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, cond_scale = cond_scale, lowres_cond_img = lowres_cond_img)
@@ -1386,31 +1478,24 @@ class Decoder(BaseGaussianDiffusion):
 
         return img
 
-    def p_losses(self, unet, x_start, t, *, image_embed, lowres_cond_img = None, text_encodings = None, predict_x_start = False, noise = None):
+    def p_losses(self, unet, x_start, times, *, image_embed, lowres_cond_img = None, text_encodings = None, predict_x_start = False, noise = None):
         noise = default(noise, lambda: torch.randn_like(x_start))
 
-        x_noisy = self.q_sample(x_start = x_start, t = t, noise = noise)
+        x_noisy = self.q_sample(x_start = x_start, t = times, noise = noise)
 
-        x_recon = unet(
+        pred = unet(
             x_noisy,
-            t,
+            times,
             image_embed = image_embed,
             text_encodings = text_encodings,
             lowres_cond_img = lowres_cond_img,
-            cond_drop_prob = self.cond_drop_prob
+            image_cond_drop_prob = self.image_cond_drop_prob,
+            text_cond_drop_prob = self.text_cond_drop_prob,
         )
 
         target = noise if not predict_x_start else x_start
 
-        if self.loss_type == 'l1':
-            loss = F.l1_loss(target, x_recon)
-        elif self.loss_type == 'l2':
-            loss = F.mse_loss(target, x_recon)
-        elif self.loss_type == "huber":
-            loss = F.smooth_l1_loss(target, x_recon)
-        else:
-            raise NotImplementedError()
-
+        loss = self.loss_fn(pred, target)
         return loss
 
     @torch.no_grad()
@@ -1420,7 +1505,7 @@ class Decoder(BaseGaussianDiffusion):
 
         text_encodings = None
         if exists(text):
-            _, text_encodings = self.clip.embed_text(text)
+            _, text_encodings, _ = self.clip.embed_text(text)
 
         assert not (self.condition_on_text_encodings and not exists(text_encodings)), 'text or text encodings must be passed into decoder if specified'
         assert not (not self.condition_on_text_encodings and exists(text_encodings)), 'decoder specified not to be conditioned on text, yet it is presented'
@@ -1488,7 +1573,7 @@ class Decoder(BaseGaussianDiffusion):
 
         text_encodings = None
         if exists(text) and not exists(text_encodings):
-            _, text_encodings = self.clip.embed_text(text)
+            _, text_encodings, _ = self.clip.embed_text(text)
 
         assert not (self.condition_on_text_encodings and not exists(text_encodings)), 'text or text encodings must be passed into decoder if specified'
         assert not (not self.condition_on_text_encodings and exists(text_encodings)), 'decoder specified not to be conditioned on text, yet it is presented'
@@ -1553,4 +1638,3 @@ class DALLE2(nn.Module):
             return images[0]
 
         return images
-

dalle2_pytorch/optimizer.py

@@ -0,0 +1,29 @@
+from torch.optim import AdamW, Adam
+
+def separate_weight_decayable_params(params):
+    no_wd_params = set([param for param in params if param.ndim < 2])
+    wd_params = set(params) - no_wd_params
+    return wd_params, no_wd_params
+
+def get_optimizer(
+    params,
+    lr = 3e-4,
+    wd = 1e-2,
+    betas = (0.9, 0.999),
+    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)
+
+    params = set(params)
+    wd_params, no_wd_params = separate_weight_decayable_params(params)
+
+    param_groups = [
+        {'params': list(wd_params)},
+        {'params': list(no_wd_params), 'weight_decay': 0},
+    ]
+
+    return AdamW(param_groups, lr = lr, weight_decay = wd, betas = betas)

dalle2_pytorch/vqgan_vae.py

@@ -545,6 +545,7 @@ class VQGanVAE(nn.Module):
         l2_recon_loss = False,
         use_hinge_loss = True,
         vgg = None,
+        vq_codebook_dim = 256,
         vq_codebook_size = 512,
         vq_decay = 0.8,
         vq_commitment_weight = 1.,
@@ -579,6 +580,7 @@ class VQGanVAE(nn.Module):
 
         self.vq = VQ(
             dim = self.enc_dec.encoded_dim,
+            codebook_dim = vq_codebook_dim,
             codebook_size = vq_codebook_size,
             decay = vq_decay,
             commitment_weight = vq_commitment_weight,

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.0.62',
+  version = '0.0.74',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',
@@ -23,6 +23,7 @@ setup(
   ],
   install_requires=[
     'click',
+    'clip-anytorch',
     'einops>=0.4',
     'einops-exts>=0.0.3',
     'kornia>=0.5.4',
@@ -31,7 +32,7 @@ setup(
     'torchvision',
     'tqdm',
     'vector-quantize-pytorch',
-    'x-clip>=0.4.4',
+    'x-clip>=0.5.1',
     'youtokentome'
   ],
   classifiers=[