fix all clipping / clamping issues

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()
 
@@ -558,7 +561,8 @@ 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()
 
@@ -616,7 +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,
@@ -669,7 +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)

dalle2_pytorch/dalle2_pytorch.py

@@ -173,12 +173,12 @@ class OpenAIClipAdapter(BaseClipAdapter):
         name = 'ViT-B/32'
     ):
         import clip
-        openai_clip, _ = clip.load(name)
+        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 = T.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711))
+        self.clip_normalize = preprocess.transforms[-1]
         self.cleared = False
 
     def find_layer(self,  layer):
@@ -736,6 +736,7 @@ class DiffusionPrior(BaseGaussianDiffusion):
         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
     ):
         super().__init__(
             beta_schedule = beta_schedule,
@@ -764,6 +765,9 @@ class DiffusionPrior(BaseGaussianDiffusion):
         self.predict_x_start = predict_x_start
         # in paper, they do not predict the noise, but predict x0 directly for image embedding, claiming empirically better results. I'll just offer both.
 
+        # whether to force an l2norm, similar to clipping denoised, when sampling
+        self.sampling_clamp_l2norm = sampling_clamp_l2norm
+
     def p_mean_variance(self, x, t, text_cond, clip_denoised: bool):
         pred = self.net(x, t, **text_cond)
 
@@ -777,6 +781,9 @@ class DiffusionPrior(BaseGaussianDiffusion):
         if clip_denoised and not self.predict_x_start:
             x_recon.clamp_(-1., 1.)
 
+        if self.predict_x_start and self.sampling_clamp_l2norm:
+            x_recon = l2norm(x_recon)
+
         model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
         return model_mean, posterior_variance, posterior_log_variance
 
@@ -1101,6 +1108,8 @@ class Unet(nn.Module):
         # for classifier free guidance
 
         self.null_image_embed = nn.Parameter(torch.randn(1, num_image_tokens, cond_dim))
+
+        self.max_text_len = max_text_len
         self.null_text_embed = nn.Parameter(torch.randn(1, max_text_len, cond_dim))
 
         # attention related params
@@ -1174,7 +1183,7 @@ 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(
@@ -1185,7 +1194,9 @@ class Unet(nn.Module):
         image_embed,
         lowres_cond_img = None,
         text_encodings = None,
-        cond_drop_prob = 0.,
+        text_mask = None,
+        image_cond_drop_prob = 0.,
+        text_cond_drop_prob = 0.,
         blur_sigma = None,
         blur_kernel_size = None
     ):
@@ -1204,8 +1215,10 @@ class Unet(nn.Module):
 
         # conditional dropout
 
-        keep_mask = prob_mask_like((batch_size,), 1 - cond_drop_prob, device = device)
-        keep_mask = rearrange(keep_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
@@ -1216,7 +1229,7 @@ class Unet(nn.Module):
             image_tokens = self.image_to_cond(image_embed)
 
             image_tokens = torch.where(
-                keep_mask,
+                image_keep_mask,
                 image_tokens,
                 self.null_image_embed
             )
@@ -1227,10 +1240,25 @@ class Unet(nn.Module):
 
         if exists(text_encodings) and self.cond_on_text_encodings:
             text_tokens = self.text_to_cond(text_encodings)
+            text_tokens = text_tokens[:, :self.max_text_len]
+
+            text_tokens_len = text_tokens.shape[1]
+            remainder = self.max_text_len - text_tokens_len
+
+            if remainder > 0:
+                text_tokens = F.pad(text_tokens, (0, 0, 0, remainder))
+
+            if exists(text_mask):
+                if remainder > 0:
+                    text_mask = F.pad(text_mask, (0, remainder), value = False)
+
+                text_mask = rearrange(text_mask, 'b n -> b n 1')
+                text_keep_mask = text_mask & text_keep_mask
+
             text_tokens = torch.where(
-                keep_mask,
+                text_keep_mask,
                 text_tokens,
-                self.null_text_embed[:, :text_tokens.shape[1]]
+                self.null_text_embed
             )
 
         # main conditioning tokens (c)
@@ -1318,7 +1346,8 @@ 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,
@@ -1329,6 +1358,8 @@ class Decoder(BaseGaussianDiffusion):
         blur_sigma = 0.1,                           # cascading ddpm - blur sigma
         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
     ):
         super().__init__(
             beta_schedule = beta_schedule,
@@ -1402,7 +1433,13 @@ 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
+
+        # whether to clip when sampling
+
+        self.clip_denoised = clip_denoised
+        self.clip_x_start = clip_x_start
 
     def get_unet(self, unet_number):
         assert 0 < unet_number <= len(self.unets)
@@ -1429,31 +1466,31 @@ class Decoder(BaseGaussianDiffusion):
         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)
+    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)
 
         if predict_x_start:
             x_recon = pred
         else:
             x_recon = self.predict_start_from_noise(x, t = t, noise = pred)
 
-        if clip_denoised and not predict_x_start:
+        if clip_denoised:
             x_recon.clamp_(-1., 1.)
 
         model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
         return model_mean, posterior_variance, posterior_log_variance
 
     @torch.no_grad()
-    def p_sample(self, unet, x, t, image_embed, text_encodings = 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, 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, 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)
         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.no_grad()
-    def p_sample_loop(self, unet, shape, image_embed, predict_x_start = False, lowres_cond_img = None, text_encodings = None, cond_scale = 1):
+    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):
         device = self.betas.device
 
         b = shape[0]
@@ -1466,14 +1503,16 @@ class Decoder(BaseGaussianDiffusion):
                 torch.full((b,), i, device = device, dtype = torch.long),
                 image_embed = image_embed,
                 text_encodings = text_encodings,
+                text_mask = text_mask,
                 cond_scale = cond_scale,
                 lowres_cond_img = lowres_cond_img,
-                predict_x_start = predict_x_start
+                predict_x_start = predict_x_start,
+                clip_denoised = clip_denoised
             )
 
         return img
 
-    def p_losses(self, unet, x_start, times, *, 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, text_mask = 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 = times, noise = noise)
@@ -1483,8 +1522,10 @@ class Decoder(BaseGaussianDiffusion):
             times,
             image_embed = image_embed,
             text_encodings = text_encodings,
+            text_mask = text_mask,
             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
@@ -1497,9 +1538,9 @@ class Decoder(BaseGaussianDiffusion):
     def sample(self, image_embed, text = None, cond_scale = 1.):
         batch_size = image_embed.shape[0]
 
-        text_encodings = None
+        text_encodings = text_mask = None
         if exists(text):
-            _, text_encodings, _ = self.clip.embed_text(text)
+            _, text_encodings, text_mask = 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'
@@ -1517,6 +1558,7 @@ class Decoder(BaseGaussianDiffusion):
                 if unet.lowres_cond:
                     lowres_cond_img = self.to_lowres_cond(img, target_image_size = image_size)
 
+                is_latent_diffusion = isinstance(vae, VQGanVAE)
                 image_size = vae.get_encoded_fmap_size(image_size)
                 shape = (batch_size, vae.encoded_dim, image_size, image_size)
 
@@ -1528,8 +1570,10 @@ class Decoder(BaseGaussianDiffusion):
                     shape,
                     image_embed = image_embed,
                     text_encodings = text_encodings,
+                    text_mask = text_mask,
                     cond_scale = cond_scale,
                     predict_x_start = predict_x_start,
+                    clip_denoised = not is_latent_diffusion,
                     lowres_cond_img = lowres_cond_img
                 )
 
@@ -1565,9 +1609,9 @@ class Decoder(BaseGaussianDiffusion):
         if not exists(image_embed):
             image_embed, _ = self.clip.embed_image(image)
 
-        text_encodings = None
+        text_encodings = text_mask = None
         if exists(text) and not exists(text_encodings):
-            _, text_encodings, _ = self.clip.embed_text(text)
+            _, text_encodings, text_mask = 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'
@@ -1582,7 +1626,7 @@ 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, 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)
 
 # main class
 

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.0.73',
+  version = '0.0.76',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',