make it so even if text mask is omitted, it will be derived based on whether text encodings are all 0s or not, simplify dataloading

README.md

@@ -355,7 +355,8 @@ prior_network = DiffusionPriorNetwork(
 diffusion_prior = DiffusionPrior(
     net = prior_network,
     clip = clip,
-    timesteps = 100,
+    timesteps = 1000,
+    sample_timesteps = 64,
     cond_drop_prob = 0.2
 ).cuda()
 

dalle2_pytorch/dalle2_pytorch.py

@@ -220,6 +220,7 @@ class XClipAdapter(BaseClipAdapter):
         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)
+        text_encodings = text_encodings.masked_fill(~text_mask[..., None], 0.)
         return EmbeddedText(l2norm(text_embed), text_encodings, text_mask)
 
     @torch.no_grad()
@@ -255,6 +256,7 @@ class CoCaAdapter(BaseClipAdapter):
         text = text[..., :self.max_text_len]
         text_mask = text != 0
         text_embed, text_encodings = self.clip.embed_text(text)
+        text_encodings = text_encodings.masked_fill(~text_mask[..., None], 0.)
         return EmbeddedText(text_embed, text_encodings, text_mask)
 
     @torch.no_grad()
@@ -314,6 +316,7 @@ class OpenAIClipAdapter(BaseClipAdapter):
 
         text_embed = self.clip.encode_text(text)
         text_encodings = self.text_encodings
+        text_encodings = text_encodings.masked_fill(~text_mask[..., None], 0.)
         del self.text_encodings
         return EmbeddedText(l2norm(text_embed.float()), text_encodings.float(), text_mask)
 
@@ -922,11 +925,12 @@ class DiffusionPrior(nn.Module):
         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,
+        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,       # whether to l2norm clamp the image embed at each denoising iteration (analogous to -1 to 1 clipping for usual DDPMs)
+        sampling_final_clamp_l2norm = False, # whether to l2norm the final image embedding output (this is also done for images in ddpm)
         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
+        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__()
@@ -963,23 +967,32 @@ class DiffusionPrior(nn.Module):
         self.condition_on_text_encodings = condition_on_text_encodings
 
         # in paper, they do not predict the noise, but predict x0 directly for image embedding, claiming empirically better results. I'll just offer both.
+
         self.predict_x_start = predict_x_start
 
         # @crowsonkb 's suggestion - https://github.com/lucidrains/DALLE2-pytorch/issues/60#issue-1226116132
+
         self.image_embed_scale = default(image_embed_scale, self.image_embed_dim ** 0.5)
 
         # whether to force an l2norm, similar to clipping denoised, when sampling
+
         self.sampling_clamp_l2norm = sampling_clamp_l2norm
+        self.sampling_final_clamp_l2norm = sampling_final_clamp_l2norm
+
         self.training_clamp_l2norm = training_clamp_l2norm
         self.init_image_embed_l2norm = init_image_embed_l2norm
 
         # device tracker
+
         self.register_buffer('_dummy', torch.tensor([True]), persistent = False)
 
     @property
     def device(self):
         return self._dummy.device
 
+    def l2norm_clamp_embed(self, image_embed):
+        return l2norm(image_embed) * self.image_embed_scale
+
     def p_mean_variance(self, x, t, text_cond, clip_denoised = False, cond_scale = 1.):
         assert not (cond_scale != 1. and not self.can_classifier_guidance), 'the model was not trained with conditional dropout, and thus one cannot use classifier free guidance (cond_scale anything other than 1)'
 
@@ -1020,6 +1033,9 @@ class DiffusionPrior(nn.Module):
             times = torch.full((batch,), i, device = device, dtype = torch.long)
             image_embed = self.p_sample(image_embed, times, text_cond = text_cond, cond_scale = cond_scale)
 
+        if self.sampling_final_clamp_l2norm and self.predict_x_start:
+            image_embed = self.l2norm_clamp_embed(image_embed)
+
         return image_embed
 
     @torch.no_grad()
@@ -1055,15 +1071,18 @@ class DiffusionPrior(nn.Module):
                 x_start.clamp_(-1., 1.)
 
             if self.predict_x_start and self.sampling_clamp_l2norm:
-                x_start = l2norm(x_start) * self.image_embed_scale
+                x_start = self.l2norm_clamp_embed(x_start)
 
             c1 = eta * ((1 - alpha / alpha_next) * (1 - alpha_next) / (1 - alpha)).sqrt()
             c2 = ((1 - alpha_next) - torch.square(c1)).sqrt()
             noise = torch.randn_like(image_embed) if time_next > 0 else 0.
 
-            img = x_start * alpha_next.sqrt() + \
-                  c1 * noise + \
-                  c2 * pred_noise
+            image_embed = x_start * alpha_next.sqrt() + \
+                          c1 * noise + \
+                          c2 * pred_noise
+
+        if self.predict_x_start and self.sampling_final_clamp_l2norm:
+            image_embed = self.l2norm_clamp_embed(image_embed)
 
         return image_embed
 
@@ -1091,7 +1110,7 @@ class DiffusionPrior(nn.Module):
         )
 
         if self.predict_x_start and self.training_clamp_l2norm:
-            pred = l2norm(pred) * self.image_embed_scale
+            pred = self.l2norm_clamp_embed(pred)
 
         target = noise if not self.predict_x_start else image_embed
 
@@ -1181,6 +1200,7 @@ class DiffusionPrior(nn.Module):
 
         if self.condition_on_text_encodings:
             assert exists(text_encodings), 'text encodings must be present for diffusion prior if specified'
+            text_mask = default(text_mask, lambda: torch.any(text_encodings != 0., dim = -1))
             text_cond = {**text_cond, 'text_encodings': text_encodings, 'mask': text_mask}
 
         # timestep conditioning from ddpm
@@ -2394,6 +2414,9 @@ class Decoder(nn.Module):
         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'
 
+        if self.condition_on_text_encodings:
+            text_mask = default(text_mask, lambda: torch.any(text_encodings != 0., dim = -1))
+
         img = None
         is_cuda = next(self.parameters()).is_cuda
 
@@ -2477,6 +2500,9 @@ class Decoder(nn.Module):
         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'
 
+        if self.condition_on_text_encodings:
+            text_mask = default(text_mask, lambda: torch.any(text_encodings != 0., dim = -1))
+
         lowres_cond_img = self.to_lowres_cond(image, target_image_size = target_image_size, downsample_image_size = self.image_sizes[unet_index - 1]) if unet_number > 1 else None
         image = resize_image_to(image, target_image_size)
 

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '0.19.4'
+__version__ = '0.20.0'

train_decoder.py

@@ -557,7 +557,7 @@ def initialize_training(config: TrainDecoderConfig, config_path):
 
     # Create the decoder model and print basic info
     decoder = config.decoder.create()
-    num_parameters = sum(p.numel() for p in decoder.parameters())
+    get_num_parameters = lambda model, only_training=False: sum(p.numel() for p in model.parameters() if (p.requires_grad or not only_training))
 
     # Create and initialize the tracker if we are the master
     tracker = create_tracker(accelerator, config, config_path, dummy = rank!=0)
@@ -586,7 +586,10 @@ def initialize_training(config: TrainDecoderConfig, config_path):
     accelerator.print(print_ribbon("Loaded Config", repeat=40))
     accelerator.print(f"Running training with {accelerator.num_processes} processes and {accelerator.distributed_type} distributed training")
     accelerator.print(f"Training using {data_source_string}. {'conditioned on text' if conditioning_on_text else 'not conditioned on text'}")
-    accelerator.print(f"Number of parameters: {num_parameters}")
+    accelerator.print(f"Number of parameters: {get_num_parameters(decoder)} total; {get_num_parameters(decoder, only_training=True)} training")
+    for i, unet in enumerate(decoder.unets):
+        accelerator.print(f"Unet {i} has {get_num_parameters(unet)} total; {get_num_parameters(unet, only_training=True)} training")
+
     train(dataloaders, decoder, accelerator,
         tracker=tracker,
         inference_device=accelerator.device,