fix readme and a small bug in DALLE2 class

This allows pylance to do proper type hinting and makes developing
extensions to the package much easier

README.md

@@ -371,6 +371,7 @@ loss.backward()
 unet1 = Unet(
     dim = 128,
     image_embed_dim = 512,
+    text_embed_dim = 512,
     cond_dim = 128,
     channels = 3,
     dim_mults=(1, 2, 4, 8),
@@ -1112,7 +1113,8 @@ For detailed information on training the diffusion prior, please refer to the [d
 - [x] allow for unet to be able to condition non-cross attention style as well
 - [x] speed up inference, read up on papers (ddim)
 - [x] add inpainting ability using resampler from repaint paper https://arxiv.org/abs/2201.09865
-- [ ] try out the nested unet from https://arxiv.org/abs/2005.09007 after hearing several positive testimonies from researchers, for segmentation anyhow
+- [x] add the final combination of upsample feature maps, used in unet squared, seems to have an effect in local experiments
+- [ ] consider elucidated dalle2 https://arxiv.org/abs/2206.00364
 - [ ] interface out the vqgan-vae so a pretrained one can be pulled off the shelf to validate latent diffusion + DALL-E2
 
 ## Citations

dalle2_pytorch/dalle2_pytorch.py

@@ -516,6 +516,17 @@ class NoiseScheduler(nn.Module):
             extract(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise
         )
 
+    def q_sample_from_to(self, x_from, from_t, to_t, noise = None):
+        shape = x_from.shape
+        noise = default(noise, lambda: torch.randn_like(x_from))
+
+        alpha = extract(self.sqrt_alphas_cumprod, from_t, shape)
+        sigma = extract(self.sqrt_one_minus_alphas_cumprod, from_t, shape)
+        alpha_next = extract(self.sqrt_alphas_cumprod, to_t, shape)
+        sigma_next = extract(self.sqrt_one_minus_alphas_cumprod, to_t, shape)
+
+        return x_from * (alpha_next / alpha) + noise * (sigma_next * alpha - sigma * alpha_next) / alpha
+
     def predict_start_from_noise(self, x_t, t, noise):
         return (
             extract(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
@@ -1492,6 +1503,7 @@ class LinearAttention(nn.Module):
         k = k.softmax(dim = -2)
 
         q = q * self.scale
+        v = v / (x * y)
 
         context = einsum('b n d, b n e -> b d e', k, v)
         out = einsum('b n d, b d e -> b n e', q, context)
@@ -1527,6 +1539,38 @@ class CrossEmbedLayer(nn.Module):
         fmaps = tuple(map(lambda conv: conv(x), self.convs))
         return torch.cat(fmaps, dim = 1)
 
+class UpsampleCombiner(nn.Module):
+    def __init__(
+        self,
+        dim,
+        *,
+        enabled = False,
+        dim_ins = tuple(),
+        dim_outs = tuple()
+    ):
+        super().__init__()
+        assert len(dim_ins) == len(dim_outs)
+        self.enabled = enabled
+
+        if not self.enabled:
+            self.dim_out = dim
+            return
+
+        self.fmap_convs = nn.ModuleList([Block(dim_in, dim_out) for dim_in, dim_out in zip(dim_ins, dim_outs)])
+        self.dim_out = dim + (sum(dim_outs) if len(dim_outs) > 0 else 0)
+
+    def forward(self, x, fmaps = None):
+        target_size = x.shape[-1]
+
+        fmaps = default(fmaps, tuple())
+
+        if not self.enabled or len(fmaps) == 0 or len(self.fmap_convs) == 0:
+            return x
+
+        fmaps = [resize_image_to(fmap, target_size) for fmap in fmaps]
+        outs = [conv(fmap) for fmap, conv in zip(fmaps, self.fmap_convs)]
+        return torch.cat((x, *outs), dim = 1)
+
 class Unet(nn.Module):
     def __init__(
         self,
@@ -1564,6 +1608,7 @@ class Unet(nn.Module):
         scale_skip_connection = False,
         pixel_shuffle_upsample = True,
         final_conv_kernel_size = 1,
+        combine_upsample_fmaps = False, # whether to combine the outputs of all upsample blocks, as in unet squared paper
         **kwargs
     ):
         super().__init__()
@@ -1699,7 +1744,8 @@ class Unet(nn.Module):
         self.ups = nn.ModuleList([])
         num_resolutions = len(in_out)
 
-        skip_connect_dims = [] # keeping track of skip connection dimensions
+        skip_connect_dims = []          # keeping track of skip connection dimensions
+        upsample_combiner_dims = []     # keeping track of dimensions for final upsample feature map combiner
 
         for ind, ((dim_in, dim_out), groups, layer_num_resnet_blocks, layer_self_attn) in enumerate(zip(in_out, resnet_groups, num_resnet_blocks, self_attn)):
             is_first = ind == 0
@@ -1741,6 +1787,8 @@ class Unet(nn.Module):
             elif sparse_attn:
                 attention = Residual(LinearAttention(dim_out, **attn_kwargs))
 
+            upsample_combiner_dims.append(dim_out)
+
             self.ups.append(nn.ModuleList([
                 ResnetBlock(dim_out + skip_connect_dim, dim_out, cond_dim = layer_cond_dim, time_cond_dim = time_cond_dim, groups = groups),
                 nn.ModuleList([ResnetBlock(dim_out + skip_connect_dim, dim_out, cond_dim = layer_cond_dim, time_cond_dim = time_cond_dim, groups = groups)  for _ in range(layer_num_resnet_blocks)]),
@@ -1748,7 +1796,18 @@ class Unet(nn.Module):
                 upsample_klass(dim_out, dim_in) if not is_last or memory_efficient else nn.Identity()
             ]))
 
-        self.final_resnet_block = ResnetBlock(dim * 2, dim, time_cond_dim = time_cond_dim, groups = top_level_resnet_group)
+        # whether to combine outputs from all upsample blocks for final resnet block
+
+        self.upsample_combiner = UpsampleCombiner(
+            dim = dim,
+            enabled = combine_upsample_fmaps,
+            dim_ins = upsample_combiner_dims,
+            dim_outs = (dim,) * len(upsample_combiner_dims)
+        )
+
+        # a final resnet block
+
+        self.final_resnet_block = ResnetBlock(self.upsample_combiner.dim_out + dim, dim, time_cond_dim = time_cond_dim, groups = top_level_resnet_group)
 
         out_dim_in = dim + (channels if lowres_cond else 0)
 
@@ -1942,7 +2001,8 @@ class Unet(nn.Module):
 
         # go through the layers of the unet, down and up
 
-        hiddens = []
+        down_hiddens = []
+        up_hiddens = []
 
         for pre_downsample, init_block, resnet_blocks, attn, post_downsample in self.downs:
             if exists(pre_downsample):
@@ -1952,10 +2012,10 @@ class Unet(nn.Module):
 
             for resnet_block in resnet_blocks:
                 x = resnet_block(x, t, c)
-                hiddens.append(x)
+                down_hiddens.append(x.contiguous())
 
             x = attn(x)
-            hiddens.append(x.contiguous())
+            down_hiddens.append(x.contiguous())
 
             if exists(post_downsample):
                 x = post_downsample(x)
@@ -1967,7 +2027,7 @@ class Unet(nn.Module):
 
         x = self.mid_block2(x, t, mid_c)
 
-        connect_skip = lambda fmap: torch.cat((fmap, hiddens.pop() * self.skip_connect_scale), dim = 1)
+        connect_skip = lambda fmap: torch.cat((fmap, down_hiddens.pop() * self.skip_connect_scale), dim = 1)
 
         for init_block, resnet_blocks, attn, upsample in self.ups:
             x = connect_skip(x)
@@ -1978,8 +2038,12 @@ class Unet(nn.Module):
                 x = resnet_block(x, t, c)
 
             x = attn(x)
+
+            up_hiddens.append(x.contiguous())
             x = upsample(x)
 
+        x = self.upsample_combiner(x, up_hiddens)
+
         x = torch.cat((x, r), dim = 1)
 
         x = self.final_resnet_block(x, t)
@@ -2432,14 +2496,18 @@ class Decoder(nn.Module):
         is_latent_diffusion = False,
         lowres_noise_level = None,
         inpaint_image = None,
-        inpaint_mask = None
+        inpaint_mask = None,
+        inpaint_resample_times = 5
     ):
         device = self.device
 
         b = shape[0]
         img = torch.randn(shape, device = device)
 
-        if exists(inpaint_image):
+        is_inpaint = exists(inpaint_image)
+        resample_times = inpaint_resample_times if is_inpaint else 1
+
+        if is_inpaint:
             inpaint_image = self.normalize_img(inpaint_image)
             inpaint_image = resize_image_to(inpaint_image, shape[-1], nearest = True)
             inpaint_mask = rearrange(inpaint_mask, 'b h w -> b 1 h w').float()
@@ -2449,31 +2517,40 @@ class Decoder(nn.Module):
         if not is_latent_diffusion:
             lowres_cond_img = maybe(self.normalize_img)(lowres_cond_img)
 
-        for i in tqdm(reversed(range(0, noise_scheduler.num_timesteps)), desc = 'sampling loop time step', total = noise_scheduler.num_timesteps):
-            times = torch.full((b,), i, device = device, dtype = torch.long)
+        for time in tqdm(reversed(range(0, noise_scheduler.num_timesteps)), desc = 'sampling loop time step', total = noise_scheduler.num_timesteps):
+            is_last_timestep = time == 0
 
-            if exists(inpaint_image):
-                # following the repaint paper
-                # https://arxiv.org/abs/2201.09865
-                noised_inpaint_image = noise_scheduler.q_sample(inpaint_image, t = times)
-                img = (img * ~inpaint_mask) + (noised_inpaint_image * inpaint_mask)
+            for r in reversed(range(0, resample_times)):
+                is_last_resample_step = r == 0
 
-            img = self.p_sample(
-                unet,
-                img,
-                times,
-                image_embed = image_embed,
-                text_encodings = text_encodings,
-                cond_scale = cond_scale,
-                lowres_cond_img = lowres_cond_img,
-                lowres_noise_level = lowres_noise_level,
-                predict_x_start = predict_x_start,
-                noise_scheduler = noise_scheduler,
-                learned_variance = learned_variance,
-                clip_denoised = clip_denoised
-            )
+                times = torch.full((b,), time, device = device, dtype = torch.long)
 
-        if exists(inpaint_image):
+                if is_inpaint:
+                    # following the repaint paper
+                    # https://arxiv.org/abs/2201.09865
+                    noised_inpaint_image = noise_scheduler.q_sample(inpaint_image, t = times)
+                    img = (img * ~inpaint_mask) + (noised_inpaint_image * inpaint_mask)
+
+                img = self.p_sample(
+                    unet,
+                    img,
+                    times,
+                    image_embed = image_embed,
+                    text_encodings = text_encodings,
+                    cond_scale = cond_scale,
+                    lowres_cond_img = lowres_cond_img,
+                    lowres_noise_level = lowres_noise_level,
+                    predict_x_start = predict_x_start,
+                    noise_scheduler = noise_scheduler,
+                    learned_variance = learned_variance,
+                    clip_denoised = clip_denoised
+                )
+
+                if is_inpaint and not (is_last_timestep or is_last_resample_step):
+                    # in repaint, you renoise and resample up to 10 times every step
+                    img = noise_scheduler.q_sample_from_to(img, times - 1, times)
+
+        if is_inpaint:
             img = (img * ~inpaint_mask) + (inpaint_image * inpaint_mask)
 
         unnormalize_img = self.unnormalize_img(img)
@@ -2497,7 +2574,8 @@ class Decoder(nn.Module):
         is_latent_diffusion = False,
         lowres_noise_level = None,
         inpaint_image = None,
-        inpaint_mask = None
+        inpaint_mask = None,
+        inpaint_resample_times = 5
     ):
         batch, device, total_timesteps, alphas, eta = shape[0], self.device, noise_scheduler.num_timesteps, noise_scheduler.alphas_cumprod_prev, self.ddim_sampling_eta
 
@@ -2506,7 +2584,10 @@ class Decoder(nn.Module):
         times = list(reversed(times.int().tolist()))
         time_pairs = list(zip(times[:-1], times[1:]))
 
-        if exists(inpaint_image):
+        is_inpaint = exists(inpaint_image)
+        resample_times = inpaint_resample_times if is_inpaint else 1
+
+        if is_inpaint:
             inpaint_image = self.normalize_img(inpaint_image)
             inpaint_image = resize_image_to(inpaint_image, shape[-1], nearest = True)
             inpaint_mask = rearrange(inpaint_mask, 'b h w -> b 1 h w').float()
@@ -2519,39 +2600,49 @@ class Decoder(nn.Module):
             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]
+            is_last_timestep = time_next == 0
 
-            time_cond = torch.full((batch,), time, device = device, dtype = torch.long)
+            for r in reversed(range(0, resample_times)):
+                is_last_resample_step = r == 0
 
-            if exists(inpaint_image):
-                # following the repaint paper
-                # https://arxiv.org/abs/2201.09865
-                noised_inpaint_image = noise_scheduler.q_sample(inpaint_image, t = time_cond)
-                img = (img * ~inpaint_mask) + (noised_inpaint_image * inpaint_mask)
+                alpha = alphas[time]
+                alpha_next = alphas[time_next]
 
-            pred = unet.forward_with_cond_scale(img, time_cond, image_embed = image_embed, text_encodings = text_encodings, cond_scale = cond_scale, lowres_cond_img = lowres_cond_img, lowres_noise_level = lowres_noise_level)
+                time_cond = torch.full((batch,), time, device = device, dtype = torch.long)
 
-            if learned_variance:
-                pred, _ = pred.chunk(2, dim = 1)
+                if is_inpaint:
+                    # following the repaint paper
+                    # https://arxiv.org/abs/2201.09865
+                    noised_inpaint_image = noise_scheduler.q_sample(inpaint_image, t = time_cond)
+                    img = (img * ~inpaint_mask) + (noised_inpaint_image * inpaint_mask)
 
-            if predict_x_start:
-                x_start = pred
-                pred_noise = noise_scheduler.predict_noise_from_start(img, t = time_cond, x0 = pred)
-            else:
-                x_start = noise_scheduler.predict_start_from_noise(img, t = time_cond, noise = pred)
-                pred_noise = pred
+                pred = unet.forward_with_cond_scale(img, time_cond, image_embed = image_embed, text_encodings = text_encodings, cond_scale = cond_scale, lowres_cond_img = lowres_cond_img, lowres_noise_level = lowres_noise_level)
 
-            if clip_denoised:
-                x_start = self.dynamic_threshold(x_start)
+                if learned_variance:
+                    pred, _ = pred.chunk(2, dim = 1)
 
-            c1 = eta * ((1 - alpha / alpha_next) * (1 - alpha_next) / (1 - alpha)).sqrt()
-            c2 = ((1 - alpha_next) - torch.square(c1)).sqrt()
-            noise = torch.randn_like(img) if time_next > 0 else 0.
+                if predict_x_start:
+                    x_start = pred
+                    pred_noise = noise_scheduler.predict_noise_from_start(img, t = time_cond, x0 = pred)
+                else:
+                    x_start = noise_scheduler.predict_start_from_noise(img, t = time_cond, noise = pred)
+                    pred_noise = pred
 
-            img = x_start * alpha_next.sqrt() + \
-                  c1 * noise + \
-                  c2 * pred_noise
+                if clip_denoised:
+                    x_start = self.dynamic_threshold(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(img) if not is_last_timestep else 0.
+
+                img = x_start * alpha_next.sqrt() + \
+                      c1 * noise + \
+                      c2 * pred_noise
+
+                if is_inpaint and not (is_last_timestep or is_last_resample_step):
+                    # in repaint, you renoise and resample up to 10 times every step
+                    time_next_cond = torch.full((batch,), time_next, device = device, dtype = torch.long)
+                    img = noise_scheduler.q_sample_from_to(img, time_next_cond, time_cond)
 
         if exists(inpaint_image):
             img = (img * ~inpaint_mask) + (inpaint_image * inpaint_mask)
@@ -2658,7 +2749,8 @@ class Decoder(nn.Module):
         stop_at_unet_number = None,
         distributed = False,
         inpaint_image = None,
-        inpaint_mask = None
+        inpaint_mask = None,
+        inpaint_resample_times = 5
     ):
         assert self.unconditional or exists(image_embed), 'image embed must be present on sampling from decoder unless if trained unconditionally'
 
@@ -2730,7 +2822,8 @@ class Decoder(nn.Module):
                     noise_scheduler = noise_scheduler,
                     timesteps = sample_timesteps,
                     inpaint_image = inpaint_image,
-                    inpaint_mask = inpaint_mask
+                    inpaint_mask = inpaint_mask,
+                    inpaint_resample_times = inpaint_resample_times
                 )
 
                 img = vae.decode(img)
@@ -2845,7 +2938,7 @@ class DALLE2(nn.Module):
         image_embed = self.prior.sample(text, num_samples_per_batch = self.prior_num_samples, cond_scale = prior_cond_scale)
 
         text_cond = text if self.decoder_need_text_cond else None
-        images = self.decoder.sample(image_embed, text = text_cond, cond_scale = cond_scale)
+        images = self.decoder.sample(image_embed = image_embed, text = text_cond, cond_scale = cond_scale)
 
         if return_pil_images:
             images = list(map(self.to_pil, images.unbind(dim = 0)))

dalle2_pytorch/trackers.py

@@ -528,8 +528,12 @@ class Tracker:
         elif save_type == 'model':
             if isinstance(trainer, DiffusionPriorTrainer):
                 prior = trainer.ema_diffusion_prior.ema_model if trainer.use_ema else trainer.diffusion_prior
-                state_dict = trainer.accelerator.unwrap_model(prior).state_dict()
-                torch.save(state_dict, file_path)
+                prior: DiffusionPrior = trainer.accelerator.unwrap_model(prior)
+                # Remove CLIP if it is part of the model
+                original_clip = prior.clip
+                prior.clip = None
+                model_state_dict = prior.state_dict()
+                prior.clip = original_clip
             elif isinstance(trainer, DecoderTrainer):
                 decoder: Decoder = trainer.accelerator.unwrap_model(trainer.decoder)
                 # Remove CLIP if it is part of the model

dalle2_pytorch/train_configs.py

@@ -1,7 +1,7 @@
 import json
 from torchvision import transforms as T
 from pydantic import BaseModel, validator, root_validator
-from typing import List, Iterable, Optional, Union, Tuple, Dict, Any
+from typing import List, Optional, Union, Tuple, Dict, Any, TypeVar
 
 from x_clip import CLIP as XCLIP
 from coca_pytorch import CoCa
@@ -25,11 +25,9 @@ def exists(val):
 def default(val, d):
     return val if exists(val) else d
 
-def ListOrTuple(inner_type):
-    return Union[List[inner_type], Tuple[inner_type]]
-
-def SingularOrIterable(inner_type):
-    return Union[inner_type, ListOrTuple(inner_type)]
+InnerType = TypeVar('InnerType')
+ListOrTuple = Union[List[InnerType], Tuple[InnerType]]
+SingularOrIterable = Union[InnerType, ListOrTuple[InnerType]]
 
 # general pydantic classes
 
@@ -222,13 +220,13 @@ class TrainDiffusionPriorConfig(BaseModel):
 
 class UnetConfig(BaseModel):
     dim: int
-    dim_mults: ListOrTuple(int)
+    dim_mults: ListOrTuple[int]
     image_embed_dim: int = None
     text_embed_dim: int = None
     cond_on_text_encodings: bool = None
     cond_dim: int = None
     channels: int = 3
-    self_attn: ListOrTuple(int)
+    self_attn: ListOrTuple[int]
     attn_dim_head: int = 32
     attn_heads: int = 16
     init_cross_embed: bool = True
@@ -237,16 +235,16 @@ class UnetConfig(BaseModel):
         extra = "allow"
 
 class DecoderConfig(BaseModel):
-    unets: ListOrTuple(UnetConfig)
+    unets: ListOrTuple[UnetConfig]
     image_size: int = None
-    image_sizes: ListOrTuple(int) = None
+    image_sizes: ListOrTuple[int] = None
     clip: Optional[AdapterConfig]   # The clip model to use if embeddings are not provided
     channels: int = 3
     timesteps: int = 1000
-    sample_timesteps: Optional[SingularOrIterable(int)] = None
+    sample_timesteps: Optional[SingularOrIterable[int]] = None
     loss_type: str = 'l2'
-    beta_schedule: ListOrTuple(str) = 'cosine'
-    learned_variance: bool = True
+    beta_schedule: ListOrTuple[str] = None  # None means all cosine
+    learned_variance: SingularOrIterable[bool] = True
     image_cond_drop_prob: float = 0.1
     text_cond_drop_prob: float = 0.5
 
@@ -305,11 +303,11 @@ class DecoderDataConfig(BaseModel):
 
 class DecoderTrainConfig(BaseModel):
     epochs: int = 20
-    lr: SingularOrIterable(float) = 1e-4
-    wd: SingularOrIterable(float) = 0.01
-    warmup_steps: Optional[SingularOrIterable(int)] = None
+    lr: SingularOrIterable[float] = 1e-4
+    wd: SingularOrIterable[float] = 0.01
+    warmup_steps: Optional[SingularOrIterable[int]] = None
     find_unused_parameters: bool = True
-    max_grad_norm: SingularOrIterable(float) = 0.5
+    max_grad_norm: SingularOrIterable[float] = 0.5
     save_every_n_samples: int = 100000
     n_sample_images: int = 6                       # The number of example images to produce when sampling the train and test dataset
     cond_scale: Union[float, List[float]] = 1.0
@@ -320,7 +318,7 @@ class DecoderTrainConfig(BaseModel):
     use_ema: bool = True
     ema_beta: float = 0.999
     amp: bool = False
-    unet_training_mask: ListOrTuple(bool) = None   # If None, use all unets
+    unet_training_mask: ListOrTuple[bool] = None   # If None, use all unets
 
 class DecoderEvaluateConfig(BaseModel):
     n_evaluation_samples: int = 1000

dalle2_pytorch/trainer.py

@@ -300,7 +300,7 @@ class DiffusionPriorTrainer(nn.Module):
 
         # all processes need to load checkpoint. no restriction here
         if isinstance(path_or_state, str):
-            path = Path(path)
+            path = Path(path_or_state)
             assert path.exists()
             loaded_obj = torch.load(str(path), map_location=self.device)
 

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '1.0.1'
+__version__ = '1.2.1'