just take care of the logic for AdamW and transformers

README.md

@@ -495,6 +495,96 @@ 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.
+
+First you'll need to install <a href="https://github.com/openai/CLIP#usage">the prerequisites</a>
+
+Then 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,
+    cond_drop_prob = 0.2,
+    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

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

@@ -90,6 +90,16 @@ 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
 
 class BaseClipAdapter(nn.Module):
@@ -109,6 +119,10 @@ 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 +142,18 @@ 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 l2norm(text_embed), text_encodings, text_mask
 
     @torch.no_grad()
     def embed_image(self, image):
@@ -143,6 +163,72 @@ class XClipAdapter(BaseClipAdapter):
         image_embed = self.clip.to_visual_latent(image_cls)
         return l2norm(image_embed), image_encodings
 
+class OpenAIClipAdapter(BaseClipAdapter):
+    def __init__(
+        self,
+        name = 'ViT-B/32'
+    ):
+        try:
+            import clip
+        except ImportError:
+            print('you must install openai clip in order to use this adapter - `pip install git+https://github.com/openai/CLIP.git` - more instructions at https://github.com/openai/CLIP#usage')
+
+        openai_clip, _ = 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 = T.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711))
+        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 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 image_embed.float(), None
+
 # classifier free guidance functions
 
 def prob_mask_like(shape, prob, device):
@@ -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)
@@ -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,
@@ -1209,7 +1298,7 @@ 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
@@ -1279,6 +1368,7 @@ class Decoder(BaseGaussianDiffusion):
 
             one_unet = one_unet.cast_model_parameters(
                 lowres_cond = not is_first,
+                cond_on_image_embeds = is_first,
                 channels = unet_channels
             )
 
@@ -1336,11 +1426,8 @@ 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)
@@ -1386,14 +1473,14 @@ 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,
@@ -1402,15 +1489,7 @@ class Decoder(BaseGaussianDiffusion):
 
         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 +1499,7 @@ 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 +1567,7 @@ 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'

dalle2_pytorch/openai_clip.py

@@ -0,0 +1,84 @@
+import torch
+from PIL import Image
+
+from dalle2_pytorch.dalle2_pytorch import BaseClipAdapter
+import torchvision.transforms as T
+
+def find_layer(model, layer):
+    modules = dict([*model.named_modules()])
+    return modules.get(layer, None)
+
+def hook(_, input, output):
+    print(output.shape)
+
+import clip
+# image = preprocess(Image.open("CLIP.png")).unsqueeze(0).to(device)
+text = clip.tokenize(["a diagram", "a dog", "a cat"]).cuda()
+image = torch.randn(1, 3, 224, 224).cuda()
+
+
+class OpenAIClipAdapter(BaseClipAdapter):
+    def __init__(self, name = 'ViT-B/32'):
+        try:
+            import clip
+        except ImportError:
+            print('you must install openai clip in order to use this adapter - `pip install git+https://github.com/openai/CLIP.git` - more instructions at https://github.com/openai/CLIP#usage')
+
+        openai_clip, _ = clip.load(name)
+        super().__init__(openai_clip)
+
+        text_attention_final = self.find_layer(self.clip, 'ln_final')
+        self.handle = text_attention_final.register_forward_hook(self._hook)
+        self.clip_normalize = T.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711))
+        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
+
+    @torch.no_grad()
+    def embed_text(self, text):
+        assert not self.cleared
+
+        text_embed = self.clip.encode_text(text)
+        text_encodings = self.text_encodings
+        del self.text_encodings
+        return text_embed, text_encodings
+
+    @torch.no_grad()
+    def embed_image(self, image):
+        assert not self.cleared
+
+        image = self.clip_normalize(image)
+        image_embed = self.clip.encode_image(image)
+        return image_embed, None
+
+clip_adapter = OpenAIClipAdapter().cuda()
+
+# print(model)
+with torch.no_grad():
+    image_features, _ = clip_adapter.embed_image(image)
+    text_features, text_encodings = clip_adapter.embed_text(text)
+    print(text_features.shape, image_features.shape)
+    print(text_encodings.shape)

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 = '0.0.70',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',
@@ -31,7 +31,7 @@ setup(
     'torchvision',
     'tqdm',
     'vector-quantize-pytorch',
-    'x-clip>=0.4.4',
+    'x-clip>=0.5.1',
     'youtokentome'
   ],
   classifiers=[