only pass text encodings conditioning in diffusion prior if specified on initialization

README.md

@@ -499,10 +499,12 @@ loss.backward()
 
 ### DALL-E2 with Latent Diffusion
 
-This repository decides to take the next step and offer DALL-E2 combined with <a href="https://huggingface.co/spaces/multimodalart/latentdiffusion">latent diffusion</a>, from Rombach et al.
+This repository decides to take the next step and offer DALL-E v2 combined with <a href="https://huggingface.co/spaces/multimodalart/latentdiffusion">latent diffusion</a>, from Rombach et al.
 
 You can use it as follows. Latent diffusion can be limited to just the first U-Net in the cascade, or to any number you wish.
 
+The repository also comes equipped with all the necessary settings to recreate `ViT-VQGan` from the <a href="https://arxiv.org/abs/2110.04627">Improved VQGans</a> paper. Furthermore, the <a href="https://github.com/lucidrains/vector-quantize-pytorch">vector quantization</a> library also comes equipped to do <a href="https://arxiv.org/abs/2203.01941">residual or multi-headed quantization</a>, which I believe will give an even further boost in performance to the autoencoder.
+
 ```python
 import torch
 from dalle2_pytorch import Unet, Decoder, CLIP, VQGanVAE
@@ -645,11 +647,12 @@ Once built, images will be saved to the same directory the command is invoked
 - [x] use attention-based upsampling https://arxiv.org/abs/2112.11435
 - [x] use inheritance just this once for sharing logic between decoder and prior network ddpms
 - [x] bring in vit-vqgan https://arxiv.org/abs/2110.04627 for the latent diffusion
-- [ ] abstract interface for CLIP adapter class, so other CLIPs can be brought in
+- [x] abstract interface for CLIP adapter class, so other CLIPs can be brought in
 - [ ] become an expert with unets, cleanup unet code, make it fully configurable, port all learnings over to https://github.com/lucidrains/x-unet
 - [ ] copy the cascading ddpm code to a separate repo (perhaps https://github.com/lucidrains/denoising-diffusion-pytorch) as the main contribution of dalle2 really is just the prior network
 - [ ] transcribe code to Jax, which lowers the activation energy for distributed training, given access to TPUs
 - [ ] train on a toy task, offer in colab
+- [ ] think about how best to design a declarative training config that handles preencoding for prior and training of multiple networks in decoder
 - [ ] extend diffusion head to use diffusion-gan (potentially using lightweight-gan) to speed up inference
 - [ ] bring in tools to train vqgan-vae
 
@@ -697,16 +700,6 @@ Once built, images will be saved to the same directory the command is invoked
 }
 ```
 
-```bibtex
-@article{Arar2021LearnedQF,
-    title   = {Learned Queries for Efficient Local Attention},
-    author  = {Moab Arar and Ariel Shamir and Amit H. Bermano},
-    journal = {ArXiv},
-    year    = {2021},
-    volume  = {abs/2112.11435}
-}
-```
-
 ```bibtex
 @article{Yu2021VectorquantizedIM,
     title   = {Vector-quantized Image Modeling with Improved VQGAN},

dalle2_pytorch/attention.py

@@ -1,130 +0,0 @@
-import torch
-from torch import nn, einsum
-import torch.nn.functional as F
-
-from einops import rearrange, repeat
-
-class LayerNormChan(nn.Module):
-    def __init__(
-        self,
-        dim,
-        eps = 1e-5
-    ):
-        super().__init__()
-        self.eps = eps
-        self.gamma = nn.Parameter(torch.ones(1, dim, 1, 1))
-
-    def forward(self, x):
-        var = torch.var(x, dim = 1, unbiased = False, keepdim = True)
-        mean = torch.mean(x, dim = 1, keepdim = True)
-        return (x - mean) / (var + self.eps).sqrt() * self.gamma
-
-# attention-based upsampling
-# from https://arxiv.org/abs/2112.11435
-
-class QueryAndAttend(nn.Module):
-    def __init__(
-        self,
-        *,
-        dim,
-        num_queries = 1,
-        dim_head = 32,
-        heads = 8,
-        window_size = 3
-    ):
-        super().__init__()
-        self.scale = dim_head ** -0.5
-        inner_dim = dim_head * heads
-        self.heads = heads
-        self.dim_head = dim_head
-        self.window_size = window_size
-        self.num_queries = num_queries
-
-        self.rel_pos_bias = nn.Parameter(torch.randn(heads, num_queries, window_size * window_size, 1, 1))
-
-        self.queries = nn.Parameter(torch.randn(heads, num_queries, dim_head))
-        self.to_kv = nn.Conv2d(dim, dim_head * 2, 1, bias = False)
-
-        self.to_out = nn.Sequential(
-            nn.Conv2d(inner_dim, dim * 2, 1, bias = False),
-            nn.Tanh(),
-            nn.Conv2d(dim * 2, dim, 1, bias = False)
-        )
-
-    def forward(self, x):
-        """
-        einstein notation
-        b - batch
-        h - heads
-        l - num queries
-        d - head dimension
-        x - height
-        y - width
-        j - source sequence for attending to (kernel size squared in this case)
-        """
-
-        wsz, heads, dim_head, num_queries = self.window_size, self.heads, self.dim_head, self.num_queries
-        batch, _, height, width = x.shape
-
-        is_one_query = self.num_queries == 1
-
-        # queries, keys, values
-
-        q = self.queries * self.scale
-        k, v = self.to_kv(x).chunk(2, dim = 1)
-
-        # similarities
-
-        sim = einsum('h l d, b d x y -> b h l x y', q, k)
-        sim = rearrange(sim, 'b ... x y -> b (...) x y')
-
-        # unfold the similarity scores, with float(-inf) as padding value
-
-        mask_value = -torch.finfo(sim.dtype).max
-        sim = F.pad(sim, ((wsz // 2,) * 4), value = mask_value)
-        sim = F.unfold(sim, kernel_size = wsz)
-        sim = rearrange(sim, 'b (h l j) (x y) -> b h l j x y', h = heads, l = num_queries, x = height, y = width)
-
-        # rel pos bias
-
-        sim = sim + self.rel_pos_bias
-
-        # numerically stable attention
-
-        sim = sim - sim.amax(dim = -3, keepdim = True).detach()
-        attn = sim.softmax(dim = -3)
-
-        # unfold values
-
-        v = F.pad(v, ((wsz // 2,) * 4), value = 0.)
-        v = F.unfold(v, kernel_size = wsz)
-        v = rearrange(v, 'b (d j) (x y) -> b d j x y', d = dim_head, x = height, y = width)
-
-        # aggregate values
-
-        out = einsum('b h l j x y, b d j x y -> b l h d x y', attn, v)
-
-        # combine heads
-
-        out = rearrange(out, 'b l h d x y -> (b l) (h d) x y')
-        out = self.to_out(out)
-        out = rearrange(out, '(b l) d x y -> b l d x y', b = batch)
-
-        # return original input if one query
-
-        if is_one_query:
-            out = rearrange(out, 'b 1 ... -> b ...')
-
-        return out
-
-class QueryAttnUpsample(nn.Module):
-    def __init__(self, dim, **kwargs):
-        super().__init__()
-        self.norm = LayerNormChan(dim)
-        self.qna = QueryAndAttend(dim = dim, num_queries = 4, **kwargs)
-
-    def forward(self, x):
-        x = self.norm(x)
-        out = self.qna(x)
-        out = rearrange(out, 'b (w1 w2) c h w -> b c (h w1) (w w2)', w1 = 2, w2 = 2)
-        return out

dalle2_pytorch/dalle2_pytorch.py

@@ -17,7 +17,6 @@ from kornia.filters import gaussian_blur2d
 
 from dalle2_pytorch.tokenizer import tokenizer
 from dalle2_pytorch.vqgan_vae import NullVQGanVAE, VQGanVAE
-from dalle2_pytorch.attention import QueryAttnUpsample
 
 # use x-clip
 
@@ -36,6 +35,10 @@ def default(val, d):
 def cast_tuple(val, length = 1):
     return val if isinstance(val, tuple) else ((val,) * length)
 
+@contextmanager
+def null_context(*args, **kwargs):
+    yield
+
 def eval_decorator(fn):
     def inner(model, *args, **kwargs):
         was_training = model.training
@@ -86,6 +89,59 @@ def resize_image_to(t, image_size, mode = 'bilinear'): # take a look at https://
 
     return F.interpolate(t, size = shape, mode = mode, align_corners = False)
 
+# clip related adapters
+
+class BaseClipAdapter(nn.Module):
+    def __init__(self, clip):
+        super().__init__()
+        self.clip = clip
+
+    @property
+    def dim_latent(self):
+        raise NotImplementedError
+
+    @property
+    def image_size(self):
+        raise NotImplementedError
+
+    @property
+    def image_channels(self):
+        raise NotImplementedError
+
+    def embed_text(self, text):
+        raise NotImplementedError
+
+    def embed_image(self, image):
+        raise NotImplementedError
+
+class XClipAdapter(BaseClipAdapter):
+    @property
+    def dim_latent(self):
+        return self.clip.dim_latent
+
+    @property
+    def image_size(self):
+        return self.clip.image_size
+
+    @property
+    def image_channels(self):
+        return self.clip.image_channels
+
+    @torch.no_grad()
+    def embed_text(self, text):
+        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)
+        return l2norm(text_embed), text_encodings
+
+    @torch.no_grad()
+    def embed_image(self, image):
+        image = resize_image_to(image, self.image_size)
+        encoder_output = self.clip.visual_transformer(image)
+        image_cls, image_encodings = encoder_output[:, 0], encoder_output[:, 1:]
+        image_embed = self.clip.to_visual_latent(image_cls)
+        return l2norm(image_embed), image_encodings
+
 # classifier free guidance functions
 
 def prob_mask_like(shape, prob, device):
@@ -592,7 +648,7 @@ class DiffusionPrior(BaseGaussianDiffusion):
         if exists(clip):
             assert isinstance(clip, CLIP)
             freeze_model_and_make_eval_(clip)
-            self.clip = clip
+            self.clip = XClipAdapter(clip)
         else:
             assert exists(image_embed_dim), 'latent dimension must be given, if training prior network without CLIP given'
             self.clip = None
@@ -607,29 +663,6 @@ 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.
 
-    @torch.no_grad()
-    def get_image_embed(self, image):
-        assert exists(self.clip)
-
-        image_encoding = self.clip.visual_transformer(image)
-        image_cls = image_encoding[:, 0]
-        image_embed = self.clip.to_visual_latent(image_cls)
-        return l2norm(image_embed)
-
-    @torch.no_grad()
-    def get_text_cond(self, text):
-        assert exists(self.clip)
-
-        text_encodings = self.clip.text_transformer(text)
-        text_cls, text_encodings = text_encodings[:, 0], text_encodings[:, 1:]
-        text_embed = self.clip.to_text_latent(text_cls)
-        text_embed = l2norm(text_embed)
-
-        if not self.condition_on_text_encodings:            
-            return dict(text_embed = text_embed)
-
-        return dict(text_encodings = text_encodings, text_embed = text_embed, mask = text != 0)
-
     def p_mean_variance(self, x, t, text_cond, clip_denoised: bool):
         pred = self.net(x, t, **text_cond)
 
@@ -701,7 +734,12 @@ class DiffusionPrior(BaseGaussianDiffusion):
         batch_size = text.shape[0]
         image_embed_dim = self.image_embed_dim
 
-        text_cond = self.get_text_cond(text)
+        text_embed, text_encodings = self.clip.embed_text(text)
+
+        text_cond = dict(text_embed = text_embed)
+
+        if self.condition_on_text_encodings:
+            text_cond = {**text_cond, 'text_encodings': text_encodings, 'mask': text_mask}
 
         image_embeds = self.p_sample_loop((batch_size, image_embed_dim), text_cond = text_cond)
         text_embeds = text_cond['text_embed']
@@ -733,18 +771,18 @@ class DiffusionPrior(BaseGaussianDiffusion):
         assert not (self.condition_on_text_encodings and (not exists(text_encodings) and not exists(text))), 'text encodings must be present if you specified you wish to condition on it on initialization'
 
         if exists(image):
-            image_embed = self.get_image_embed(image)
+            image_embed, _ = self.clip.embed_image(image)
 
         # calculate text conditionings, based on what is passed in
 
         if exists(text):
-            text_cond = self.get_text_cond(text)
-        else:
-            text_cond = dict(
-                text_embed = text_embed,
-                text_encodings = text_encodings,
-                mask = text_mask
-            )
+            text_embed, text_encodings = self.clip.embed_text(text)
+            text_mask = text != 0
+
+        text_cond = dict(text_embed = text_embed)
+
+        if self.condition_on_text_encodings:
+            text_cond = {**text_cond, 'text_encodings': text_encodings, 'mask': text_mask}
 
         # timestep conditioning from ddpm
 
@@ -1205,7 +1243,9 @@ class Decoder(BaseGaussianDiffusion):
             loss_type = loss_type
         )
 
-        assert isinstance(clip, CLIP)
+        if isinstance(clip, CLIP):
+            clip = XClipAdapter(clip)
+
         freeze_model_and_make_eval_(clip)
         self.clip = clip
         self.clip_image_size = clip.image_size
@@ -1287,10 +1327,6 @@ class Decoder(BaseGaussianDiffusion):
         yield
         unet.cpu()
 
-    @torch.no_grad()
-    def get_text_encodings(self, text):
-        text_encodings = self.clip.text_transformer(text)
-        return text_encodings[:, 1:]
 
     @torch.no_grad()
     def get_image_embed(self, image):
@@ -1376,14 +1412,19 @@ class Decoder(BaseGaussianDiffusion):
     def sample(self, image_embed, text = None, cond_scale = 1.):
         batch_size = image_embed.shape[0]
 
-        text_encodings = self.get_text_encodings(text) if exists(text) else None
+        text_encodings = None
+        if exists(text):
+            _, text_encodings = 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'
 
         img = None
 
         for unet, vae, channel, image_size, predict_x_start in tqdm(zip(self.unets, self.vaes, self.sample_channels, self.image_sizes, self.predict_x_start)):
-            with self.one_unet_in_gpu(unet = unet):
+
+            context = self.one_unet_in_gpu(unet = unet) if image_embed.is_cuda else null_context()
+
+            with context:
                 lowres_cond_img = None
                 shape = (batch_size, channel, image_size, image_size)
 
@@ -1436,9 +1477,11 @@ class Decoder(BaseGaussianDiffusion):
         times = torch.randint(0, self.num_timesteps, (b,), device = device, dtype = torch.long)
 
         if not exists(image_embed):
-            image_embed = self.get_image_embed(image)
+            image_embed, _ = self.clip.embed_image(image)
 
-        text_encodings = self.get_text_encodings(text) if exists(text) and not exists(text_encodings) else None
+        text_encodings = None
+        if exists(text) and not exists(text_encodings):
+            _, text_encodings = 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'
 

dalle2_pytorch/train.py

@@ -35,7 +35,7 @@ class EMA(nn.Module):
 
         self.update_moving_average(self.ema_model, self.online_model)
 
-    def update_moving_average(ma_model, current_model):
+    def update_moving_average(self, ma_model, current_model):
         def calculate_ema(beta, old, new):
             if not exists(old):
                 return new

dalle2_pytorch/vqgan_vae.py

@@ -15,8 +15,6 @@ from einops import rearrange, reduce, repeat
 from einops_exts import rearrange_many
 from einops.layers.torch import Rearrange
 
-from dalle2_pytorch.attention import QueryAttnUpsample
-
 # constants
 
 MList = nn.ModuleList

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.0.54',
+  version = '0.0.58',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',