protect against random cropping for base unet

Accelerate removes the ability to collate strings. Likely since it
cannot gather strings.

.github/FUNDING.yml

@@ -1 +1 @@
-github: [lucidrains]
+github: [nousr, Veldrovive, lucidrains]

dalle2_pytorch/dalle2_pytorch.py

@@ -146,7 +146,7 @@ def resize_image_to(
         scale_factors = target_image_size / orig_image_size
         out = resize(image, scale_factors = scale_factors, **kwargs)
     else:
-        out = F.interpolate(image, target_image_size, mode = 'nearest', align_corners = False)
+        out = F.interpolate(image, target_image_size, mode = 'nearest')
 
     if exists(clamp_range):
         out = out.clamp(*clamp_range)
@@ -278,6 +278,7 @@ class OpenAIClipAdapter(BaseClipAdapter):
         import clip
         openai_clip, preprocess = clip.load(name)
         super().__init__(openai_clip)
+        self.eos_id = 49407 # for handling 0 being also '!'
 
         text_attention_final = self.find_layer('ln_final')
         self.handle = text_attention_final.register_forward_hook(self._hook)
@@ -316,7 +317,10 @@ class OpenAIClipAdapter(BaseClipAdapter):
     @torch.no_grad()
     def embed_text(self, text):
         text = text[..., :self.max_text_len]
-        text_mask = text != 0
+
+        is_eos_id = (text == self.eos_id)
+        text_mask_excluding_eos = is_eos_id.cumsum(dim = -1) == 0
+        text_mask = F.pad(text_mask_excluding_eos, (1, -1), value = True)
         assert not self.cleared
 
         text_embed = self.clip.encode_text(text)
@@ -527,25 +531,31 @@ class NoiseScheduler(nn.Module):
 # diffusion prior
 
 class LayerNorm(nn.Module):
-    def __init__(self, dim, eps = 1e-5):
+    def __init__(self, dim, eps = 1e-5, stable = False):
         super().__init__()
         self.eps = eps
+        self.stable = stable
         self.g = nn.Parameter(torch.ones(dim))
 
     def forward(self, x):
-        x = x / x.amax(dim = -1, keepdim = True).detach()
+        if self.stable:
+            x = x / x.amax(dim = -1, keepdim = True).detach()
+
         var = torch.var(x, dim = -1, unbiased = False, keepdim = True)
         mean = torch.mean(x, dim = -1, keepdim = True)
         return (x - mean) * (var + self.eps).rsqrt() * self.g
 
 class ChanLayerNorm(nn.Module):
-    def __init__(self, dim, eps = 1e-5):
+    def __init__(self, dim, eps = 1e-5, stable = False):
         super().__init__()
         self.eps = eps
+        self.stable = stable
         self.g = nn.Parameter(torch.ones(1, dim, 1, 1))
 
     def forward(self, x):
-        x = x / x.amax(dim = 1, keepdim = True).detach()
+        if self.stable:
+            x = x / x.amax(dim = 1, keepdim = True).detach()
+
         var = torch.var(x, dim = 1, unbiased = False, keepdim = True)
         mean = torch.mean(x, dim = 1, keepdim = True)
         return (x - mean) * (var + self.eps).rsqrt() * self.g
@@ -669,7 +679,7 @@ class Attention(nn.Module):
         dropout = 0.,
         causal = False,
         rotary_emb = None,
-        pb_relax_alpha = 32 ** 2
+        pb_relax_alpha = 128
     ):
         super().__init__()
         self.pb_relax_alpha = pb_relax_alpha
@@ -782,7 +792,7 @@ class CausalTransformer(nn.Module):
                 FeedForward(dim = dim, mult = ff_mult, dropout = ff_dropout, post_activation_norm = normformer)
             ]))
 
-        self.norm = LayerNorm(dim) if norm_out else nn.Identity()  # unclear in paper whether they projected after the classic layer norm for the final denoised image embedding, or just had the transformer output it directly: plan on offering both options
+        self.norm = LayerNorm(dim, stable = True) if norm_out else nn.Identity()  # unclear in paper whether they projected after the classic layer norm for the final denoised image embedding, or just had the transformer output it directly: plan on offering both options
         self.project_out = nn.Linear(dim, dim, bias = False) if final_proj else nn.Identity()
 
     def forward(self, x):
@@ -894,7 +904,7 @@ class DiffusionPriorNetwork(nn.Module):
         null_text_embeds = self.null_text_embed.to(text_encodings.dtype)
 
         text_encodings = torch.where(
-            rearrange(mask, 'b n -> b n 1'),
+            rearrange(mask, 'b n -> b n 1').clone(),
             text_encodings,
             null_text_embeds
         )
@@ -1245,6 +1255,14 @@ class DiffusionPrior(nn.Module):
 
 # decoder
 
+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)
+    )
+
 class PixelShuffleUpsample(nn.Module):
     """
     code shared by @MalumaDev at DALLE2-pytorch for addressing checkboard artifacts
@@ -1651,7 +1669,7 @@ class Unet(nn.Module):
 
         # upsample klass
 
-        upsample_klass = ConvTransposeUpsample if not pixel_shuffle_upsample else PixelShuffleUpsample
+        upsample_klass = NearestUpsample if not pixel_shuffle_upsample else PixelShuffleUpsample
 
         # give memory efficient unet an initial resnet block
 
@@ -1713,7 +1731,10 @@ 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)
+
+        out_dim_in = dim + (channels if lowres_cond else 0)
+
+        self.to_out = nn.Conv2d(out_dim_in, 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
 
@@ -1933,23 +1954,26 @@ class Unet(nn.Module):
         x = torch.cat((x, r), dim = 1)
 
         x = self.final_resnet_block(x, t)
+
+        if exists(lowres_cond_img):
+            x = torch.cat((x, lowres_cond_img), dim = 1)
+
         return self.to_out(x)
 
 class LowresConditioner(nn.Module):
     def __init__(
         self,
         downsample_first = True,
-        downsample_mode_nearest = False,
+        blur_prob = 0.5,
         blur_sigma = 0.6,
         blur_kernel_size = 3,
         input_image_range = None
     ):
         super().__init__()
         self.downsample_first = downsample_first
-        self.downsample_mode_nearest = downsample_mode_nearest
-
         self.input_image_range = input_image_range
 
+        self.blur_prob = blur_prob
         self.blur_sigma = blur_sigma
         self.blur_kernel_size = blur_kernel_size
 
@@ -1962,20 +1986,27 @@ class LowresConditioner(nn.Module):
         blur_sigma = None,
         blur_kernel_size = None
     ):
-        if self.training and self.downsample_first and exists(downsample_image_size):
-            cond_fmap = resize_image_to(cond_fmap, downsample_image_size, clamp_range = self.input_image_range, nearest = self.downsample_mode_nearest)
+        if self.downsample_first and exists(downsample_image_size):
+            cond_fmap = resize_image_to(cond_fmap, downsample_image_size, clamp_range = self.input_image_range, nearest = True)
+
+        # blur is only applied 50% of the time
+        # section 3.1 in https://arxiv.org/abs/2106.15282
+
+        if random.random() < self.blur_prob:
 
-        if self.training:
             # when training, blur the low resolution conditional image
+
             blur_sigma = default(blur_sigma, self.blur_sigma)
             blur_kernel_size = default(blur_kernel_size, self.blur_kernel_size)
 
             # allow for drawing a random sigma between lo and hi float values
+
             if isinstance(blur_sigma, tuple):
                 blur_sigma = tuple(map(float, blur_sigma))
                 blur_sigma = random.uniform(*blur_sigma)
 
             # allow for drawing a random kernel size between lo and hi int values
+
             if isinstance(blur_kernel_size, tuple):
                 blur_kernel_size = tuple(map(int, blur_kernel_size))
                 kernel_size_lo, kernel_size_hi = blur_kernel_size
@@ -1983,8 +2014,7 @@ class LowresConditioner(nn.Module):
 
             cond_fmap = gaussian_blur2d(cond_fmap, cast_tuple(blur_kernel_size, 2), cast_tuple(blur_sigma, 2))
 
-        cond_fmap = resize_image_to(cond_fmap, target_image_size, clamp_range = self.input_image_range)
-
+        cond_fmap = resize_image_to(cond_fmap, target_image_size, clamp_range = self.input_image_range, nearest = True)
         return cond_fmap
 
 class Decoder(nn.Module):
@@ -2007,7 +2037,7 @@ class Decoder(nn.Module):
         image_sizes = None,                         # for cascading ddpm, image size at each stage
         random_crop_sizes = None,                   # whether to random crop the image at that stage in the cascade (super resoluting convolutions at the end may be able to generalize on smaller crops)
         lowres_downsample_first = True,             # cascading ddpm - resizes to lower resolution, then to next conditional resolution + blur
-        lowres_downsample_mode_nearest = False,     # cascading ddpm - whether to use nearest mode downsampling for lower resolution
+        blur_prob = 0.5,                            # cascading ddpm - when training, the gaussian blur is only applied 50% of the time
         blur_sigma = 0.6,                           # cascading ddpm - blur sigma
         blur_kernel_size = 3,                       # cascading ddpm - blur kernel size
         clip_denoised = True,
@@ -2142,6 +2172,7 @@ class Decoder(nn.Module):
         # random crop sizes (for super-resoluting unets at the end of cascade?)
 
         self.random_crop_sizes = cast_tuple(random_crop_sizes, len(image_sizes))
+        assert not exists(self.random_crop_sizes[0]), 'you would not need to randomly crop the image for the base unet'
 
         # predict x0 config
 
@@ -2158,7 +2189,7 @@ class Decoder(nn.Module):
 
         self.to_lowres_cond = LowresConditioner(
             downsample_first = lowres_downsample_first,
-            downsample_mode_nearest = lowres_downsample_mode_nearest,
+            blur_prob = blur_prob,
             blur_sigma = blur_sigma,
             blur_kernel_size = blur_kernel_size,
             input_image_range = self.input_image_range
@@ -2322,6 +2353,9 @@ class Decoder(nn.Module):
 
         img = torch.randn(shape, device = device)
 
+        if not is_latent_diffusion:
+            lowres_cond_img = maybe(self.normalize_img)(lowres_cond_img)
+
         for time, time_next in tqdm(time_pairs, desc = 'sampling loop time step'):
             alpha = alphas[time]
             alpha_next = alphas[time_next]
@@ -2465,7 +2499,10 @@ class Decoder(nn.Module):
         img = None
         is_cuda = next(self.parameters()).is_cuda
 
-        for unet_number, unet, vae, channel, image_size, predict_x_start, learned_variance, noise_scheduler, sample_timesteps in tqdm(zip(range(1, len(self.unets) + 1), self.unets, self.vaes, self.sample_channels, self.image_sizes, self.predict_x_start, self.learned_variance, self.noise_schedulers, self.sample_timesteps)):
+        num_unets = len(self.unets)
+        cond_scale = cast_tuple(cond_scale, num_unets)
+
+        for unet_number, unet, vae, channel, image_size, predict_x_start, learned_variance, noise_scheduler, sample_timesteps, unet_cond_scale in tqdm(zip(range(1, num_unets + 1), self.unets, self.vaes, self.sample_channels, self.image_sizes, self.predict_x_start, self.learned_variance, self.noise_schedulers, self.sample_timesteps, cond_scale)):
 
             context = self.one_unet_in_gpu(unet = unet) if is_cuda and not distributed else null_context()
 
@@ -2474,7 +2511,7 @@ class Decoder(nn.Module):
                 shape = (batch_size, channel, image_size, image_size)
 
                 if unet.lowres_cond:
-                    lowres_cond_img = self.to_lowres_cond(img, target_image_size = image_size)
+                    lowres_cond_img = resize_image_to(img, target_image_size = image_size, clamp_range = self.input_image_range, nearest = True)
 
                 is_latent_diffusion = isinstance(vae, VQGanVAE)
                 image_size = vae.get_encoded_fmap_size(image_size)
@@ -2487,7 +2524,7 @@ class Decoder(nn.Module):
                     shape,
                     image_embed = image_embed,
                     text_encodings = text_encodings,
-                    cond_scale = cond_scale,
+                    cond_scale = unet_cond_scale,
                     predict_x_start = predict_x_start,
                     learned_variance = learned_variance,
                     clip_denoised = not is_latent_diffusion,
@@ -2544,7 +2581,7 @@ class Decoder(nn.Module):
         assert not (not self.condition_on_text_encodings and exists(text_encodings)), 'decoder specified not to be conditioned on text, yet it is presented'
 
         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)
+        image = resize_image_to(image, target_image_size, nearest = True)
 
         if exists(random_crop_size):
             aug = K.RandomCrop((random_crop_size, random_crop_size), p = 1.)

dalle2_pytorch/trainer.py

@@ -673,8 +673,14 @@ class DecoderTrainer(nn.Module):
     def sample(self, *args, **kwargs):
         distributed = self.accelerator.num_processes > 1
         base_decoder = self.accelerator.unwrap_model(self.decoder)
+
+        was_training = base_decoder.training
+        base_decoder.eval()
+
         if kwargs.pop('use_non_ema', False) or not self.use_ema:
-            return base_decoder.sample(*args, **kwargs, distributed = distributed)
+            out = base_decoder.sample(*args, **kwargs, distributed = distributed)
+            base_decoder.train(was_training)
+            return out
 
         trainable_unets = self.accelerator.unwrap_model(self.decoder).unets
         base_decoder.unets = self.unets                  # swap in exponential moving averaged unets for sampling
@@ -687,6 +693,7 @@ class DecoderTrainer(nn.Module):
         for ema in self.ema_unets:
             ema.restore_ema_model_device()
 
+        base_decoder.train(was_training)
         return output
 
     @torch.no_grad()

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '0.23.2'
+__version__ = '0.24.1'

train_decoder.py

@@ -323,7 +323,7 @@ def train(
         last_snapshot = sample
 
         if next_task == 'train':
-            for i, (img, emb, txt) in enumerate(trainer.train_loader):
+            for i, (img, emb, txt) in enumerate(dataloaders["train"]):
                 # We want to count the total number of samples across all processes
                 sample_length_tensor[0] = len(img)
                 all_samples = accelerator.gather(sample_length_tensor)  # TODO: accelerator.reduce is broken when this was written. If it is fixed replace this.
@@ -358,6 +358,7 @@ def train(
                         else:
                             # Then we need to pass the text instead
                             tokenized_texts = tokenize(txt, truncate=True)
+                            assert tokenized_texts.shape[0] == len(img), f"The number of texts ({tokenized_texts.shape[0]}) should be the same as the number of images ({len(img)})"
                             forward_params['text'] = tokenized_texts
                     loss = trainer.forward(img, **forward_params, unet_number=unet)
                     trainer.update(unet_number=unet)
@@ -416,7 +417,7 @@ def train(
             timer = Timer()
             accelerator.wait_for_everyone()
             i = 0
-            for i, (img, emb, txt) in enumerate(trainer.val_loader):  # Use the accelerate prepared loader
+            for i, (img, emb, txt) in enumerate(dataloaders['val']):  # Use the accelerate prepared loader
                 val_sample_length_tensor[0] = len(img)
                 all_samples = accelerator.gather(val_sample_length_tensor)
                 total_samples = all_samples.sum().item()