add PixelShuffleUpsample thanks to @MalumaDev and @marunine for running the experiment and verifyng absence of checkboard artifacts

dalle2_pytorch/dalle2_pytorch.py

@@ -77,6 +77,11 @@ def cast_tuple(val, length = None):
 def module_device(module):
     return next(module.parameters()).device
 
+def zero_init_(m):
+    nn.init.zeros_(m.weight)
+    if exists(m.bias):
+        nn.init.zeros_(m.bias)
+
 @contextmanager
 def null_context(*args, **kwargs):
     yield
@@ -220,6 +225,7 @@ 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 +261,7 @@ 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 +321,7 @@ 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)
 
@@ -1197,6 +1205,7 @@ 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
@@ -1214,16 +1223,35 @@ class DiffusionPrior(nn.Module):
 
 # decoder
 
-def ConvTransposeUpsample(dim, dim_out = None):
-    dim_out = default(dim_out, dim)
-    return nn.ConvTranspose2d(dim, dim_out, 4, 2, 1)
+class PixelShuffleUpsample(nn.Module):
+    """
+    code shared by @MalumaDev at DALLE2-pytorch for addressing checkboard artifacts
+    https://arxiv.org/ftp/arxiv/papers/1707/1707.02937.pdf
+    """
+    def __init__(self, dim, dim_out = None):
+        super().__init__()
+        dim_out = default(dim_out, dim)
+        conv = nn.Conv2d(dim, dim_out * 4, 1)
 
-def NearestUpsample(dim, dim_out = None):
-    dim_out = default(dim_out, dim)
-    return nn.Sequential(
-        nn.Upsample(scale_factor = 2, mode = 'nearest'),
-        nn.Conv2d(dim, dim_out, 3, padding = 1)
-    )
+        self.net = nn.Sequential(
+            conv,
+            nn.SiLU(),
+            nn.PixelShuffle(2)
+        )
+
+        self.init_conv_(conv)
+
+    def init_conv_(self, conv):
+        o, i, h, w = conv.weight.shape
+        conv_weight = torch.empty(o // 4, i, h, w)
+        nn.init.kaiming_uniform_(conv_weight)
+        conv_weight = repeat(conv_weight, 'o ... -> (o 4) ...')
+
+        conv.weight.data.copy_(conv_weight)
+        nn.init.zeros_(conv.bias.data)
+
+    def forward(self, x):
+        return self.net(x)
 
 def Downsample(dim, *, dim_out = None):
     dim_out = default(dim_out, dim)
@@ -1487,7 +1515,7 @@ class Unet(nn.Module):
         cross_embed_downsample_kernel_sizes = (2, 4),
         memory_efficient = False,
         scale_skip_connection = False,
-        nearest_upsample = False,
+        pixel_shuffle_upsample = True,
         final_conv_kernel_size = 1,
         **kwargs
     ):
@@ -1601,7 +1629,7 @@ class Unet(nn.Module):
 
         # upsample klass
 
-        upsample_klass = ConvTransposeUpsample if not nearest_upsample else NearestUpsample
+        upsample_klass = ConvTransposeUpsample if not pixel_shuffle_upsample else PixelShuffleUpsample
 
         # give memory efficient unet an initial resnet block
 
@@ -1665,6 +1693,8 @@ class Unet(nn.Module):
         self.final_resnet_block = ResnetBlock(dim * 2, dim, time_cond_dim = time_cond_dim, groups = top_level_resnet_group)
         self.to_out = nn.Conv2d(dim, self.channels_out, kernel_size = final_conv_kernel_size, padding = final_conv_kernel_size // 2)
 
+        zero_init_(self.to_out) # since both OpenAI and @crowsonkb are doing it
+
     # if the current settings for the unet are not correct
     # for cascading DDPM, then reinit the unet with the right settings
     def cast_model_parameters(
@@ -2410,6 +2440,9 @@ 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
 
@@ -2493,6 +2526,9 @@ 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__ = '0.19.6'
+__version__ = '0.21.0'

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,