add learned padding tokens, same strategy as dalle1, for diffusion prior, and get rid of masking in causal transformer

.github/workflows/ci.yml

@@ -0,0 +1,33 @@
+name: Continuous integration
+
+on:
+  push:
+    branches:
+    - main
+  pull_request:
+    branches:
+    - main
+
+jobs:
+  tests:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: [3.8]
+
+    steps:
+    - uses: actions/checkout@v2
+    - name: Set up Python ${{ matrix.python-version }}
+      uses: actions/setup-python@v2
+      with:
+        python-version: ${{ matrix.python-version }}
+    - name: Install
+      run: |
+        python3 -m venv .env
+        source .env/bin/activate
+        make install
+    - name: Tests
+      run: |
+        source .env/bin/activate
+        make test
+

.gitignore

@@ -136,3 +136,5 @@ dmypy.json
 
 # Pyre type checker
 .pyre/
+.tracker_data
+*.pth

Makefile

@@ -0,0 +1,6 @@
+install:
+	pip install -U pip
+	pip install -e .
+
+test:
+	CUDA_VISIBLE_DEVICES= python train_decoder.py --config_file configs/train_decoder_config.test.json

README.md

@@ -45,6 +45,7 @@ This library would not have gotten to this working state without the help of
 - <a href="https://github.com/rom1504">Romain</a> for the pull request reviews and project management
 - <a href="https://github.com/Ciaohe">He Cao</a> and <a href="https://github.com/xiankgx">xiankgx</a> for the Q&A and for identifying of critical bugs
 - <a href="https://github.com/marunine">Marunine</a> for identifying issues with resizing of the low resolution conditioner, when training the upsampler, in addition to various other bug fixes
+- <a href="https://github.com/malumadev">MalumaDev</a> for proposing the use of pixel shuffle upsampler for fixing checkboard artifacts
 - <a href="https://github.com/crowsonkb">Katherine</a> for her advice
 - <a href="https://stability.ai/">Stability AI</a> for the generous sponsorship
 - <a href="https://huggingface.co">🤗 Huggingface</a> and in particular <a href="https://github.com/sgugger">Sylvain</a> for the <a href="https://github.com/huggingface/accelerate">Accelerate</a> library
@@ -355,7 +356,8 @@ prior_network = DiffusionPriorNetwork(
 diffusion_prior = DiffusionPrior(
     net = prior_network,
     clip = clip,
-    timesteps = 100,
+    timesteps = 1000,
+    sample_timesteps = 64,
     cond_drop_prob = 0.2
 ).cuda()
 
@@ -419,7 +421,7 @@ For the layperson, no worries, training will all be automated into a CLI tool, a
 
 ## Training on Preprocessed CLIP Embeddings
 
-It is likely, when scaling up, that you would first preprocess your images and text into corresponding embeddings before training the prior network. You can do so easily by simply passing in `image_embed`, `text_embed`, and optionally `text_encodings` and `text_mask`
+It is likely, when scaling up, that you would first preprocess your images and text into corresponding embeddings before training the prior network. You can do so easily by simply passing in `image_embed`, `text_embed`, and optionally `text_encodings`
 
 Working example below
 
@@ -583,6 +585,7 @@ unet1 = Unet(
     cond_dim = 128,
     channels = 3,
     dim_mults=(1, 2, 4, 8),
+    text_embed_dim = 512,
     cond_on_text_encodings = True  # set to True for any unets that need to be conditioned on text encodings (ex. first unet in cascade)
 ).cuda()
 
@@ -598,7 +601,8 @@ decoder = Decoder(
     unet = (unet1, unet2),
     image_sizes = (128, 256),
     clip = clip,
-    timesteps = 100,
+    timesteps = 1000,
+    sample_timesteps = (250, 27),
     image_cond_drop_prob = 0.1,
     text_cond_drop_prob = 0.5
 ).cuda()

configs/train_decoder_config.test.json

@@ -0,0 +1,102 @@
+{
+    "decoder": {
+        "unets": [
+            {
+                "dim": 16,
+                "image_embed_dim": 768,
+                "cond_dim": 16,
+                "channels": 3,
+                "dim_mults": [1, 2, 4, 8],
+                "attn_dim_head": 16,
+                "attn_heads": 4,
+		"self_attn": [false, true, true, true]
+            }
+        ],
+        "clip": {
+            "make": "openai",
+            "model": "ViT-L/14"
+        },
+
+	"timesteps": 10,
+        "image_sizes": [64],
+        "channels": 3,
+        "loss_type": "l2",
+        "beta_schedule": ["cosine"],
+        "learned_variance": true
+    },
+    "data": {
+        "webdataset_base_url": "test_data/{}.tar",
+        "num_workers": 4,
+        "batch_size": 4,
+        "start_shard": 0,
+        "end_shard": 9,
+        "shard_width": 1,
+        "index_width": 1,
+        "splits": {
+            "train": 0.75,
+            "val": 0.15,
+            "test": 0.1
+        },
+        "shuffle_train": false,
+        "resample_train": true,
+        "preprocessing": {
+            "RandomResizedCrop": {
+                "size": [224, 224],
+                "scale": [0.75, 1.0],
+                "ratio": [1.0, 1.0]
+            },
+            "ToTensor": true
+        }
+    },
+    "train": {
+        "epochs": 1,
+        "lr": 1e-16,
+        "wd": 0.01,
+        "max_grad_norm": 0.5,
+        "save_every_n_samples": 100,
+        "n_sample_images": 1,
+        "device": "cpu",
+        "epoch_samples": 50,
+        "validation_samples": 5,
+        "use_ema": true,
+        "ema_beta": 0.99,
+        "amp": false,
+        "save_all": false,
+        "save_latest": true,
+        "save_best": true,
+        "unet_training_mask": [true]
+    },
+    "evaluate": {
+        "n_evaluation_samples": 2,
+        "FID": {
+            "feature": 64
+        },
+        "IS": {
+            "feature": 64,
+            "splits": 10
+        },
+        "KID": {
+            "feature": 64,
+            "subset_size": 2
+        },
+        "LPIPS": {
+            "net_type": "vgg",
+            "reduction": "mean"
+        }
+    },
+    "tracker": {
+        "overwrite_data_path": true,
+
+	"log": {
+            "log_type": "console"
+	},
+
+        "load": {
+            "load_from": null
+        },
+
+       "save": [{
+            "save_to": "local"
+        }]
+    }
+}

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
@@ -160,7 +165,7 @@ def unnormalize_zero_to_one(normed_img):
 
 # clip related adapters
 
-EmbeddedText = namedtuple('EmbedTextReturn', ['text_embed', 'text_encodings', 'text_mask'])
+EmbeddedText = namedtuple('EmbedTextReturn', ['text_embed', 'text_encodings'])
 EmbeddedImage = namedtuple('EmbedImageReturn', ['image_embed', 'image_encodings'])
 
 class BaseClipAdapter(nn.Module):
@@ -169,6 +174,11 @@ class BaseClipAdapter(nn.Module):
         self.clip = clip
         self.overrides = kwargs
 
+    def validate_and_resize_image(self, image):
+        image_size = image.shape[-1]
+        assert image_size >= self.image_size, f'you are passing in an image of size {image_size} but CLIP requires the image size to be at least {self.image_size}'
+        return resize_image_to(image, self.image_size)
+
     @property
     def dim_latent(self):
         raise NotImplementedError
@@ -215,11 +225,12 @@ 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)
-        return EmbeddedText(l2norm(text_embed), text_encodings, text_mask)
+        text_encodings = text_encodings.masked_fill(~text_mask[..., None], 0.)
+        return EmbeddedText(l2norm(text_embed), text_encodings)
 
     @torch.no_grad()
     def embed_image(self, image):
-        image = resize_image_to(image, self.image_size)
+        image = self.validate_and_resize_image(image)
         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)
@@ -250,11 +261,12 @@ class CoCaAdapter(BaseClipAdapter):
         text = text[..., :self.max_text_len]
         text_mask = text != 0
         text_embed, text_encodings = self.clip.embed_text(text)
-        return EmbeddedText(text_embed, text_encodings, text_mask)
+        text_encodings = text_encodings.masked_fill(~text_mask[..., None], 0.)
+        return EmbeddedText(text_embed, text_encodings)
 
     @torch.no_grad()
     def embed_image(self, image):
-        image = resize_image_to(image, self.image_size)
+        image = self.validate_and_resize_image(image)
         image_embed, image_encodings = self.clip.embed_image(image)
         return EmbeddedImage(image_embed, image_encodings)
 
@@ -309,13 +321,14 @@ 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)
+        return EmbeddedText(l2norm(text_embed.float()), text_encodings.float())
 
     @torch.no_grad()
     def embed_image(self, image):
         assert not self.cleared
-        image = resize_image_to(image, self.image_size)
+        image = self.validate_and_resize_image(image)
         image = self.clip_normalize(image)
         image_embed = self.clip.encode_image(image)
         return EmbeddedImage(l2norm(image_embed.float()), None)
@@ -500,6 +513,12 @@ class NoiseScheduler(nn.Module):
             extract(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
         )
 
+    def predict_noise_from_start(self, x_t, t, x0):
+        return (
+            (x0 - extract(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t) / \
+            extract(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
+        )
+
     def p2_reweigh_loss(self, loss, times):
         if not self.has_p2_loss_reweighting:
             return loss
@@ -763,17 +782,13 @@ class CausalTransformer(nn.Module):
         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.project_out = nn.Linear(dim, dim, bias = False) if final_proj else nn.Identity()
 
-    def forward(
-        self,
-        x,
-        mask = None    # we will need a mask here, due to variable length of the text encodings - also offer dalle1 strategy with padding token embeddings
-    ):
+    def forward(self, x):
         n, device = x.shape[1], x.device
 
         attn_bias = self.rel_pos_bias(n, n + 1, device = device)
 
         for attn, ff in self.layers:
-            x = attn(x, mask = mask, attn_bias = attn_bias) + x
+            x = attn(x, attn_bias = attn_bias) + x
             x = ff(x) + x
 
         out = self.norm(x)
@@ -787,6 +802,7 @@ class DiffusionPriorNetwork(nn.Module):
         num_time_embeds = 1,
         num_image_embeds = 1,
         num_text_embeds = 1,
+        max_text_len = 256,
         **kwargs
     ):
         super().__init__()
@@ -812,6 +828,11 @@ class DiffusionPriorNetwork(nn.Module):
         self.learned_query = nn.Parameter(torch.randn(dim))
         self.causal_transformer = CausalTransformer(dim = dim, **kwargs)
 
+        # dalle1 learned padding strategy
+
+        self.max_text_len = max_text_len
+        self.null_text_embed = nn.Parameter(torch.randn(1, max_text_len, dim))
+
     def forward_with_cond_scale(
         self,
         *args,
@@ -833,7 +854,6 @@ class DiffusionPriorNetwork(nn.Module):
         *,
         text_embed,
         text_encodings = None,
-        mask = None,
         cond_drop_prob = 0.
     ):
         batch, dim, device, dtype = *image_embed.shape, image_embed.device, image_embed.dtype
@@ -851,9 +871,28 @@ class DiffusionPriorNetwork(nn.Module):
 
         if not exists(text_encodings):
             text_encodings = torch.empty((batch, 0, dim), device = device, dtype = dtype)
+    
+        mask = torch.any(text_encodings != 0., dim = -1)
 
-        if not exists(mask):
-            mask = torch.ones((batch, text_encodings.shape[-2]), device = device, dtype = torch.bool)
+        # replace any padding in the text encodings with learned padding tokens unique across position
+
+        text_encodings = text_encodings[:, :self.max_text_len]
+        mask = mask[:, :self.max_text_len]
+
+        text_len = text_encodings.shape[-2]
+        remainder = self.max_text_len - text_len
+
+        if remainder > 0:
+            text_encodings = F.pad(text_encodings, (0, 0, 0, remainder), value = 0.)
+            mask = F.pad(mask, (0, remainder), value = False)
+
+        null_text_embeds = self.null_text_embed.to(text_encodings.dtype)
+
+        text_encodings = torch.where(
+            rearrange(mask, 'b n -> b n 1'),
+            text_encodings,
+            null_text_embeds
+        )
 
         # classifier free guidance
 
@@ -870,9 +909,8 @@ class DiffusionPriorNetwork(nn.Module):
         # whether text embedding is used for conditioning depends on whether text encodings are available for attention (for classifier free guidance, even though it seems from the paper it was not used in the prior ddpm, as the objective is different)
         # but let's just do it right
 
-        if exists(mask):
-            attend_padding = 1 + num_time_embeds + num_image_embeds # 1 for learned queries + number of image embeds + time embeds
-            mask = F.pad(mask, (0, attend_padding), value = True) # extend mask for text embedding, noised image embedding, time step embedding, and learned query
+        attend_padding = 1 + num_time_embeds + num_image_embeds # 1 for learned queries + number of image embeds + time embeds
+        mask = F.pad(mask, (0, attend_padding), value = True) # extend mask for text embedding, noised image embedding, time step embedding, and learned query
 
         time_embed = self.to_time_embeds(diffusion_timesteps)
 
@@ -888,7 +926,7 @@ class DiffusionPriorNetwork(nn.Module):
 
         # attend
 
-        tokens = self.causal_transformer(tokens, mask = mask)
+        tokens = self.causal_transformer(tokens)
 
         # get learned query, which should predict the image embedding (per DDPM timestep)
 
@@ -906,19 +944,23 @@ class DiffusionPrior(nn.Module):
         image_size = None,
         image_channels = 3,
         timesteps = 1000,
+        sample_timesteps = None,
         cond_drop_prob = 0.,
         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__()
 
+        self.sample_timesteps = sample_timesteps
+
         self.noise_scheduler = NoiseScheduler(
             beta_schedule = beta_schedule,
             timesteps = timesteps,
@@ -949,23 +991,32 @@ 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)'
 
@@ -973,8 +1024,6 @@ class DiffusionPrior(nn.Module):
 
         if self.predict_x_start:
             x_recon = pred
-            # not 100% sure of this above line - for any spectators, let me know in the github issues (or through a pull request) if you know how to correctly do this
-            # i'll be rereading https://arxiv.org/abs/2111.14822, where i think a similar approach is taken
         else:
             x_recon = self.noise_scheduler.predict_start_from_noise(x, t = t, noise = pred)
 
@@ -997,21 +1046,81 @@ class DiffusionPrior(nn.Module):
         return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
 
     @torch.no_grad()
-    def p_sample_loop(self, shape, text_cond, cond_scale = 1.):
-        device = self.device
-
-        b = shape[0]
-        image_embed = torch.randn(shape, device=device)
+    def p_sample_loop_ddpm(self, shape, text_cond, cond_scale = 1.):
+        batch, device = shape[0], self.device
+        image_embed = torch.randn(shape, device = device)
 
         if self.init_image_embed_l2norm:
             image_embed = l2norm(image_embed) * self.image_embed_scale
 
         for i in tqdm(reversed(range(0, self.noise_scheduler.num_timesteps)), desc='sampling loop time step', total=self.noise_scheduler.num_timesteps):
-            times = torch.full((b,), i, device = device, dtype = torch.long)
+            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()
+    def p_sample_loop_ddim(self, shape, text_cond, *, timesteps, eta = 1., cond_scale = 1.):
+        batch, device, alphas, total_timesteps = shape[0], self.device, self.noise_scheduler.alphas_cumprod_prev, self.noise_scheduler.num_timesteps
+
+        times = torch.linspace(0., total_timesteps, steps = timesteps + 2)[:-1]
+
+        times = list(reversed(times.int().tolist()))
+        time_pairs = list(zip(times[:-1], times[1:]))
+
+        image_embed = torch.randn(shape, device = device)
+
+        if self.init_image_embed_l2norm:
+            image_embed = l2norm(image_embed) * self.image_embed_scale
+
+        for time, time_next in tqdm(time_pairs, desc = 'sampling loop time step'):
+            alpha = alphas[time]
+            alpha_next = alphas[time_next]
+
+            time_cond = torch.full((batch,), time, device = device, dtype = torch.long)
+
+            pred = self.net.forward_with_cond_scale(image_embed, time_cond, cond_scale = cond_scale, **text_cond)
+
+            if self.predict_x_start:
+                x_start = pred
+                pred_noise = self.noise_scheduler.predict_noise_from_start(image_embed, t = time_cond, x0 = pred)
+            else:
+                x_start = self.noise_scheduler.predict_start_from_noise(image_embed, t = time_cond, noise = pred)
+                pred_noise = pred
+
+            if not self.predict_x_start:
+                x_start.clamp_(-1., 1.)
+
+            if self.predict_x_start and self.sampling_clamp_l2norm:
+                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.
+
+            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
+
+    @torch.no_grad()
+    def p_sample_loop(self, *args, timesteps = None, **kwargs):
+        timesteps = default(timesteps, self.noise_scheduler.num_timesteps)
+        assert timesteps <= self.noise_scheduler.num_timesteps
+        is_ddim = timesteps < self.noise_scheduler.num_timesteps
+
+        if not is_ddim:
+            return self.p_sample_loop_ddpm(*args, **kwargs)
+
+        return self.p_sample_loop_ddim(*args, **kwargs, timesteps = timesteps)
+
     def p_losses(self, image_embed, times, text_cond, noise = None):
         noise = default(noise, lambda: torch.randn_like(image_embed))
 
@@ -1025,7 +1134,7 @@ 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
 
@@ -1046,7 +1155,15 @@ class DiffusionPrior(nn.Module):
 
     @torch.no_grad()
     @eval_decorator
-    def sample(self, text, num_samples_per_batch = 2, cond_scale = 1.):
+    def sample(
+        self,
+        text,
+        num_samples_per_batch = 2,
+        cond_scale = 1.,
+        timesteps = None
+    ):
+        timesteps = default(timesteps, self.sample_timesteps)
+
         # in the paper, what they did was
         # sample 2 image embeddings, choose the top 1 similarity, as judged by CLIP
         text = repeat(text, 'b ... -> (b r) ...', r = num_samples_per_batch)
@@ -1054,14 +1171,14 @@ class DiffusionPrior(nn.Module):
         batch_size = text.shape[0]
         image_embed_dim = self.image_embed_dim
 
-        text_embed, text_encodings, text_mask = self.clip.embed_text(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}
+            text_cond = {**text_cond, 'text_encodings': text_encodings}
 
-        image_embeds = self.p_sample_loop((batch_size, image_embed_dim), text_cond = text_cond, cond_scale = cond_scale)
+        image_embeds = self.p_sample_loop((batch_size, image_embed_dim), text_cond = text_cond, cond_scale = cond_scale, timesteps = timesteps)
 
         # retrieve original unscaled image embed
 
@@ -1087,7 +1204,6 @@ class DiffusionPrior(nn.Module):
         text_embed = None,      # allow for training on preprocessed CLIP text and image embeddings
         image_embed = None,
         text_encodings = None,  # as well as CLIP text encodings
-        text_mask = None,       # text mask <- may eventually opt for the learned padding tokens technique from DALL-E1 to reduce complexity
         *args,
         **kwargs
     ):
@@ -1101,13 +1217,13 @@ class DiffusionPrior(nn.Module):
         # calculate text conditionings, based on what is passed in
 
         if exists(text):
-            text_embed, text_encodings, text_mask = self.clip.embed_text(text)
+            text_embed, text_encodings = self.clip.embed_text(text)
 
         text_cond = dict(text_embed = text_embed)
 
         if self.condition_on_text_encodings:
             assert exists(text_encodings), 'text encodings must be present for diffusion prior if specified'
-            text_cond = {**text_cond, 'text_encodings': text_encodings, 'mask': text_mask}
+            text_cond = {**text_cond, 'text_encodings': text_encodings}
 
         # timestep conditioning from ddpm
 
@@ -1124,16 +1240,35 @@ 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)
@@ -1397,7 +1532,7 @@ 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
     ):
@@ -1463,10 +1598,12 @@ class Unet(nn.Module):
         # text encoding conditioning (optional)
 
         self.text_to_cond = None
+        self.text_embed_dim = None
 
         if cond_on_text_encodings:
             assert exists(text_embed_dim), 'text_embed_dim must be given to the unet if cond_on_text_encodings is True'
             self.text_to_cond = nn.Linear(text_embed_dim, cond_dim)
+            self.text_embed_dim = text_embed_dim
 
         # finer control over whether to condition on image embeddings and text encodings
         # so one can have the latter unets in the cascading DDPMs only focus on super-resoluting
@@ -1509,7 +1646,7 @@ 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
 
@@ -1573,6 +1710,8 @@ 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(
@@ -1623,7 +1762,6 @@ class Unet(nn.Module):
         image_embed,
         lowres_cond_img = None,
         text_encodings = None,
-        text_mask = None,
         image_cond_drop_prob = 0.,
         text_cond_drop_prob = 0.,
         blur_sigma = None,
@@ -1695,21 +1833,27 @@ class Unet(nn.Module):
         text_tokens = None
 
         if exists(text_encodings) and self.cond_on_text_encodings:
+            assert text_encodings.shape[0] == batch_size, f'the text encodings being passed into the unet does not have the proper batch size - text encoding shape {text_encodings.shape} - required batch size is {batch_size}'
+            assert self.text_embed_dim == text_encodings.shape[-1], f'the text encodings you are passing in have a dimension of {text_encodings.shape[-1]}, but the unet was created with text_embed_dim of {self.text_embed_dim}.'
+
+            text_mask = torch.any(text_encodings != 0., dim = -1)
+
             text_tokens = self.text_to_cond(text_encodings)
+
             text_tokens = text_tokens[:, :self.max_text_len]
+            text_mask = text_mask[:, :self.max_text_len]
 
             text_tokens_len = text_tokens.shape[1]
             remainder = self.max_text_len - text_tokens_len
 
             if remainder > 0:
                 text_tokens = F.pad(text_tokens, (0, 0, 0, remainder))
+                text_mask = F.pad(text_mask, (0, remainder), value = False)
 
-            if exists(text_mask):
-                if remainder > 0:
-                    text_mask = F.pad(text_mask, (0, remainder), value = False)
+            text_mask = rearrange(text_mask, 'b n -> b n 1')
 
-                text_mask = rearrange(text_mask, 'b n -> b n 1')
-                text_keep_mask = text_mask & text_keep_mask
+            assert text_mask.shape[0] == text_keep_mask.shape[0], f'text_mask has shape of {text_mask.shape} while text_keep_mask has shape {text_keep_mask.shape}. text encoding is of shape {text_encodings.shape}'
+            text_keep_mask = text_mask & text_keep_mask
 
             null_text_embed = self.null_text_embed.to(text_tokens.dtype) # for some reason pytorch AMP not working
 
@@ -1848,6 +1992,7 @@ class Decoder(nn.Module):
         channels = 3,
         vae = tuple(),
         timesteps = 1000,
+        sample_timesteps = None,
         image_cond_drop_prob = 0.1,
         text_cond_drop_prob = 0.5,
         loss_type = 'l2',
@@ -1871,7 +2016,8 @@ class Decoder(nn.Module):
         use_dynamic_thres = False,                  # from the Imagen paper
         dynamic_thres_percentile = 0.9,
         p2_loss_weight_gamma = 0.,                  # p2 loss weight, from https://arxiv.org/abs/2204.00227 - 0 is equivalent to weight of 1 across time - 1. is recommended
-        p2_loss_weight_k = 1
+        p2_loss_weight_k = 1,
+        ddim_sampling_eta = 1.                      # can be set to 0. for deterministic sampling afaict
     ):
         super().__init__()
 
@@ -1951,6 +2097,11 @@ class Decoder(nn.Module):
             self.unets.append(one_unet)
             self.vaes.append(one_vae.copy_for_eval())
 
+        # sampling timesteps, defaults to non-ddim with full timesteps sampling
+
+        self.sample_timesteps = cast_tuple(sample_timesteps, num_unets)
+        self.ddim_sampling_eta = ddim_sampling_eta
+
         # create noise schedulers per unet
 
         if not exists(beta_schedule):
@@ -1961,7 +2112,9 @@ class Decoder(nn.Module):
 
         self.noise_schedulers = nn.ModuleList([])
 
-        for unet_beta_schedule, unet_p2_loss_weight_gamma in zip(beta_schedule, p2_loss_weight_gamma):
+        for ind, (unet_beta_schedule, unet_p2_loss_weight_gamma, sample_timesteps) in enumerate(zip(beta_schedule, p2_loss_weight_gamma, self.sample_timesteps)):
+            assert not exists(sample_timesteps) or sample_timesteps <= timesteps, f'sampling timesteps {sample_timesteps} must be less than or equal to the number of training timesteps {timesteps} for unet {ind + 1}'
+
             noise_scheduler = NoiseScheduler(
                 beta_schedule = unet_beta_schedule,
                 timesteps = timesteps,
@@ -2062,10 +2215,30 @@ class Decoder(nn.Module):
         for unet, device in zip(self.unets, devices):
             unet.to(device)
 
-    def p_mean_variance(self, unet, x, t, image_embed, noise_scheduler, text_encodings = None, text_mask = None, lowres_cond_img = None, clip_denoised = True, predict_x_start = False, learned_variance = False, cond_scale = 1., model_output = None):
+    def dynamic_threshold(self, x):
+        """ proposed in https://arxiv.org/abs/2205.11487 as an improved clamping in the setting of classifier free guidance """
+        
+        # s is the threshold amount
+        # static thresholding would just be s = 1
+        s = 1.
+        if self.use_dynamic_thres:
+            s = torch.quantile(
+                rearrange(x, 'b ... -> b (...)').abs(),
+                self.dynamic_thres_percentile,
+                dim = -1
+            )
+
+            s.clamp_(min = 1.)
+            s = s.view(-1, *((1,) * (x.ndim - 1)))
+
+        # clip by threshold, depending on whether static or dynamic
+        x = x.clamp(-s, s) / s
+        return x
+
+    def p_mean_variance(self, unet, x, t, image_embed, noise_scheduler, text_encodings = None, lowres_cond_img = None, clip_denoised = True, predict_x_start = False, learned_variance = False, cond_scale = 1., model_output = None):
         assert not (cond_scale != 1. and not self.can_classifier_guidance), 'the decoder was not trained with conditional dropout, and thus one cannot use classifier free guidance (cond_scale anything other than 1)'
 
-        pred = default(model_output, lambda: unet.forward_with_cond_scale(x, t, image_embed = image_embed, text_encodings = text_encodings, text_mask = text_mask, cond_scale = cond_scale, lowres_cond_img = lowres_cond_img))
+        pred = default(model_output, lambda: unet.forward_with_cond_scale(x, t, image_embed = image_embed, text_encodings = text_encodings, cond_scale = cond_scale, lowres_cond_img = lowres_cond_img))
 
         if learned_variance:
             pred, var_interp_frac_unnormalized = pred.chunk(2, dim = 1)
@@ -2076,21 +2249,7 @@ class Decoder(nn.Module):
             x_recon = noise_scheduler.predict_start_from_noise(x, t = t, noise = pred)
 
         if clip_denoised:
-            # s is the threshold amount
-            # static thresholding would just be s = 1
-            s = 1.
-            if self.use_dynamic_thres:
-                s = torch.quantile(
-                    rearrange(x_recon, 'b ... -> b (...)').abs(),
-                    self.dynamic_thres_percentile,
-                    dim = -1
-                )
-
-                s.clamp_(min = 1.)
-                s = s.view(-1, *((1,) * (x_recon.ndim - 1)))
-
-            # clip by threshold, depending on whether static or dynamic
-            x_recon = x_recon.clamp(-s, s) / s
+            x_recon = self.dynamic_threshold(x_recon)
 
         model_mean, posterior_variance, posterior_log_variance = noise_scheduler.q_posterior(x_start=x_recon, x_t=x, t=t)
 
@@ -2111,16 +2270,16 @@ class Decoder(nn.Module):
         return model_mean, posterior_variance, posterior_log_variance
 
     @torch.no_grad()
-    def p_sample(self, unet, x, t, image_embed, noise_scheduler, text_encodings = None, text_mask = None, cond_scale = 1., lowres_cond_img = None, predict_x_start = False, learned_variance = False, clip_denoised = True):
+    def p_sample(self, unet, x, t, image_embed, noise_scheduler, text_encodings = None, cond_scale = 1., lowres_cond_img = None, predict_x_start = False, learned_variance = False, clip_denoised = True):
         b, *_, device = *x.shape, x.device
-        model_mean, _, model_log_variance = self.p_mean_variance(unet, x = x, t = t, image_embed = image_embed, text_encodings = text_encodings, text_mask = text_mask, cond_scale = cond_scale, lowres_cond_img = lowres_cond_img, clip_denoised = clip_denoised, predict_x_start = predict_x_start, noise_scheduler = noise_scheduler, learned_variance = learned_variance)
+        model_mean, _, model_log_variance = self.p_mean_variance(unet, x = x, t = t, image_embed = image_embed, text_encodings = text_encodings, cond_scale = cond_scale, lowres_cond_img = lowres_cond_img, clip_denoised = clip_denoised, predict_x_start = predict_x_start, noise_scheduler = noise_scheduler, learned_variance = learned_variance)
         noise = torch.randn_like(x)
         # no noise when t == 0
         nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
         return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
 
     @torch.no_grad()
-    def p_sample_loop(self, unet, shape, image_embed, noise_scheduler, predict_x_start = False, learned_variance = False, clip_denoised = True, lowres_cond_img = None, text_encodings = None, text_mask = None, cond_scale = 1, is_latent_diffusion = False):
+    def p_sample_loop_ddpm(self, unet, shape, image_embed, noise_scheduler, predict_x_start = False, learned_variance = False, clip_denoised = True, lowres_cond_img = None, text_encodings = None, cond_scale = 1, is_latent_diffusion = False):
         device = self.device
 
         b = shape[0]
@@ -2136,7 +2295,6 @@ class Decoder(nn.Module):
                 torch.full((b,), i, device = device, dtype = torch.long),
                 image_embed = image_embed,
                 text_encodings = text_encodings,
-                text_mask = text_mask,
                 cond_scale = cond_scale,
                 lowres_cond_img = lowres_cond_img,
                 predict_x_start = predict_x_start,
@@ -2148,7 +2306,63 @@ class Decoder(nn.Module):
         unnormalize_img = self.unnormalize_img(img)
         return unnormalize_img
 
-    def p_losses(self, unet, x_start, times, *, image_embed, noise_scheduler, lowres_cond_img = None, text_encodings = None, text_mask = None, predict_x_start = False, noise = None, learned_variance = False, clip_denoised = False, is_latent_diffusion = False):
+    @torch.no_grad()
+    def p_sample_loop_ddim(self, unet, shape, image_embed, noise_scheduler, timesteps, eta = 1., predict_x_start = False, learned_variance = False, clip_denoised = True, lowres_cond_img = None, text_encodings = None, cond_scale = 1, is_latent_diffusion = False):
+        batch, device, total_timesteps, alphas, eta = shape[0], self.device, noise_scheduler.num_timesteps, noise_scheduler.alphas_cumprod_prev, self.ddim_sampling_eta
+
+        times = torch.linspace(0., total_timesteps, steps = timesteps + 2)[:-1]
+
+        times = list(reversed(times.int().tolist()))
+        time_pairs = list(zip(times[:-1], times[1:]))
+
+        img = torch.randn(shape, device = device)
+
+        for time, time_next in tqdm(time_pairs, desc = 'sampling loop time step'):
+            alpha = alphas[time]
+            alpha_next = alphas[time_next]
+
+            time_cond = torch.full((batch,), time, device = device, dtype = torch.long)
+
+            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)
+
+            if learned_variance:
+                pred, _ = pred.chunk(2, dim = 1)
+
+            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
+
+            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 time_next > 0 else 0.
+
+            img = x_start * alpha_next.sqrt() + \
+                  c1 * noise + \
+                  c2 * pred_noise
+
+        img = self.unnormalize_img(img)
+        return img
+
+    @torch.no_grad()
+    def p_sample_loop(self, *args, noise_scheduler, timesteps = None, **kwargs):
+        num_timesteps = noise_scheduler.num_timesteps
+
+        timesteps = default(timesteps, num_timesteps)
+        assert timesteps <= num_timesteps
+        is_ddim = timesteps < num_timesteps
+
+        if not is_ddim:
+            return self.p_sample_loop_ddpm(*args, noise_scheduler = noise_scheduler, **kwargs)
+
+        return self.p_sample_loop_ddim(*args, noise_scheduler = noise_scheduler, timesteps = timesteps, **kwargs)
+
+    def p_losses(self, unet, x_start, times, *, image_embed, noise_scheduler, lowres_cond_img = None, text_encodings = None, predict_x_start = False, noise = None, learned_variance = False, clip_denoised = False, is_latent_diffusion = False):
         noise = default(noise, lambda: torch.randn_like(x_start))
 
         # normalize to [-1, 1]
@@ -2166,7 +2380,6 @@ class Decoder(nn.Module):
             times,
             image_embed = image_embed,
             text_encodings = text_encodings,
-            text_mask = text_mask,
             lowres_cond_img = lowres_cond_img,
             image_cond_drop_prob = self.image_cond_drop_prob,
             text_cond_drop_prob = self.text_cond_drop_prob,
@@ -2226,7 +2439,6 @@ class Decoder(nn.Module):
         self,
         image_embed = None,
         text = None,
-        text_mask = None,
         text_encodings = None,
         batch_size = 1,
         cond_scale = 1.,
@@ -2240,7 +2452,7 @@ class Decoder(nn.Module):
 
         if exists(text) and not exists(text_encodings) and not self.unconditional:
             assert exists(self.clip)
-            _, text_encodings, text_mask = self.clip.embed_text(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'
         assert not (not self.condition_on_text_encodings and exists(text_encodings)), 'decoder specified not to be conditioned on text, yet it is presented'
@@ -2248,7 +2460,7 @@ 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 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)):
+        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)):
 
             context = self.one_unet_in_gpu(unet = unet) if is_cuda and not distributed else null_context()
 
@@ -2270,14 +2482,14 @@ class Decoder(nn.Module):
                     shape,
                     image_embed = image_embed,
                     text_encodings = text_encodings,
-                    text_mask = text_mask,
                     cond_scale = cond_scale,
                     predict_x_start = predict_x_start,
                     learned_variance = learned_variance,
                     clip_denoised = not is_latent_diffusion,
                     lowres_cond_img = lowres_cond_img,
                     is_latent_diffusion = is_latent_diffusion,
-                    noise_scheduler = noise_scheduler
+                    noise_scheduler = noise_scheduler,
+                    timesteps = sample_timesteps
                 )
 
                 img = vae.decode(img)
@@ -2293,7 +2505,6 @@ class Decoder(nn.Module):
         text = None,
         image_embed = None,
         text_encodings = None,
-        text_mask = None,
         unet_number = None,
         return_lowres_cond_image = False # whether to return the low resolution conditioning images, for debugging upsampler purposes
     ):
@@ -2322,7 +2533,7 @@ class Decoder(nn.Module):
 
         if exists(text) and not exists(text_encodings) and not self.unconditional:
             assert exists(self.clip), 'if you are passing in raw text, you need to supply `clip` to the decoder'
-            _, text_encodings, text_mask = self.clip.embed_text(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'
         assert not (not self.condition_on_text_encodings and exists(text_encodings)), 'decoder specified not to be conditioned on text, yet it is presented'
@@ -2345,7 +2556,7 @@ class Decoder(nn.Module):
             image = vae.encode(image)
             lowres_cond_img = maybe(vae.encode)(lowres_cond_img)
 
-        losses = self.p_losses(unet, image, times, image_embed = image_embed, text_encodings = text_encodings, text_mask = text_mask, lowres_cond_img = lowres_cond_img, predict_x_start = predict_x_start, learned_variance = learned_variance, is_latent_diffusion = is_latent_diffusion, noise_scheduler = noise_scheduler)
+        losses = self.p_losses(unet, image, times, image_embed = image_embed, text_encodings = text_encodings, lowres_cond_img = lowres_cond_img, predict_x_start = predict_x_start, learned_variance = learned_variance, is_latent_diffusion = is_latent_diffusion, noise_scheduler = noise_scheduler)
 
         if not return_lowres_cond_image:
             return losses

dalle2_pytorch/dataloaders/decoder_loader.py

@@ -1,6 +1,7 @@
 import os
 import webdataset as wds
 import torch
+from torch.utils.data import DataLoader
 import numpy as np
 import fsspec
 import shutil
@@ -255,7 +256,7 @@ def create_image_embedding_dataloader(
     )
     if shuffle_num is not None and shuffle_num > 0:
         ds.shuffle(1000)
-    return wds.WebLoader(
+    return DataLoader(
         ds,
         num_workers=num_workers,
         batch_size=batch_size,

dalle2_pytorch/train_configs.py

@@ -129,6 +129,7 @@ class AdapterConfig(BaseModel):
 class DiffusionPriorNetworkConfig(BaseModel):
     dim: int
     depth: int
+    max_text_len: int = None
     num_timesteps: int = None
     num_time_embeds: int = 1
     num_image_embeds: int = 1
@@ -154,6 +155,7 @@ class DiffusionPriorConfig(BaseModel):
     image_size: int
     image_channels: int = 3
     timesteps: int = 1000
+    sample_timesteps: Optional[int] = None
     cond_drop_prob: float = 0.
     loss_type: str = 'l2'
     predict_x_start: bool = True
@@ -233,6 +235,7 @@ class DecoderConfig(BaseModel):
     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
     loss_type: str = 'l2'
     beta_schedule: ListOrTuple(str) = 'cosine'
     learned_variance: bool = True

dalle2_pytorch/trainer.py

@@ -21,7 +21,7 @@ import pytorch_warmup as warmup
 
 from ema_pytorch import EMA
 
-from accelerate import Accelerator
+from accelerate import Accelerator, DistributedType
 
 import numpy as np
 
@@ -76,6 +76,7 @@ def cast_torch_tensor(fn):
     def inner(model, *args, **kwargs):
         device = kwargs.pop('_device', next(model.parameters()).device)
         cast_device = kwargs.pop('_cast_device', True)
+        cast_deepspeed_precision = kwargs.pop('_cast_deepspeed_precision', True)
 
         kwargs_keys = kwargs.keys()
         all_args = (*args, *kwargs.values())
@@ -85,6 +86,21 @@ def cast_torch_tensor(fn):
         if cast_device:
             all_args = tuple(map(lambda t: t.to(device) if exists(t) and isinstance(t, torch.Tensor) else t, all_args))
 
+        if cast_deepspeed_precision:
+            try:
+                accelerator = model.accelerator
+                if accelerator is not None and accelerator.distributed_type == DistributedType.DEEPSPEED:
+                    cast_type_map = {
+                        "fp16": torch.half,
+                        "bf16": torch.bfloat16,
+                        "no": torch.float
+                    }
+                    precision_type = cast_type_map[accelerator.mixed_precision]
+                    all_args = tuple(map(lambda t: t.to(precision_type) if exists(t) and isinstance(t, torch.Tensor) else t, all_args))
+            except AttributeError:
+                # Then this model doesn't have an accelerator
+                pass
+
         args, kwargs_values = all_args[:split_kwargs_index], all_args[split_kwargs_index:]
         kwargs = dict(tuple(zip(kwargs_keys, kwargs_values)))
 
@@ -446,6 +462,7 @@ class DecoderTrainer(nn.Module):
         self,
         decoder,
         accelerator = None,
+        dataloaders = None,
         use_ema = True,
         lr = 1e-4,
         wd = 1e-2,
@@ -508,10 +525,31 @@ class DecoderTrainer(nn.Module):
 
         self.register_buffer('steps', torch.tensor([0] * self.num_unets))
 
+        if self.accelerator.distributed_type == DistributedType.DEEPSPEED and decoder.clip is not None:
+            # Then we need to make sure clip is using the correct precision or else deepspeed will error
+            cast_type_map = {
+                "fp16": torch.half,
+                "bf16": torch.bfloat16,
+                "no": torch.float
+            }
+            precision_type = cast_type_map[accelerator.mixed_precision]
+            assert precision_type == torch.float, "DeepSpeed currently only supports float32 precision when using on the fly embedding generation from clip"
+            clip = decoder.clip
+            clip.to(precision_type)
+
         decoder, *optimizers = list(self.accelerator.prepare(decoder, *optimizers))
 
         self.decoder = decoder
 
+        # prepare dataloaders
+
+        train_loader = val_loader = None
+        if exists(dataloaders):
+            train_loader, val_loader = self.accelerator.prepare(dataloaders["train"], dataloaders["val"])
+
+        self.train_loader = train_loader
+        self.val_loader = val_loader
+
         # store optimizers
 
         for opt_ind, optimizer in zip(range(len(optimizers)), optimizers):
@@ -675,6 +713,9 @@ class DecoderTrainer(nn.Module):
 
         total_loss = 0.
 
+        
+        using_amp = self.accelerator.mixed_precision != 'no'
+
         for chunk_size_frac, (chunked_args, chunked_kwargs) in split_args_and_kwargs(*args, split_size = max_batch_size, **kwargs):
             with self.accelerator.autocast():
                 loss = self.decoder(*chunked_args, unet_number = unet_number, **chunked_kwargs)

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '0.17.0'
+__version__ = '0.23.1'

train_decoder.py

@@ -132,7 +132,7 @@ def get_example_data(dataloader, device, n=5):
             break
     return list(zip(images[:n], img_embeddings[:n], text_embeddings[:n], captions[:n]))
 
-def generate_samples(trainer, example_data, condition_on_text_encodings=False, text_prepend=""):
+def generate_samples(trainer, example_data, condition_on_text_encodings=False, text_prepend="", match_image_size=True):
     """
     Takes example data and generates images from the embeddings
     Returns three lists: real images, generated images, and captions
@@ -160,6 +160,9 @@ def generate_samples(trainer, example_data, condition_on_text_encodings=False, t
     samples = trainer.sample(**sample_params)
     generated_images = list(samples)
     captions = [text_prepend + txt for txt in txts]
+    if match_image_size:
+        generated_image_size = generated_images[0].shape[-1]
+        real_images = [resize_image_to(image, generated_image_size, clamp_range=(0, 1)) for image in real_images]
     return real_images, generated_images, captions
 
 def generate_grid_samples(trainer, examples, condition_on_text_encodings=False, text_prepend=""):
@@ -167,14 +170,6 @@ def generate_grid_samples(trainer, examples, condition_on_text_encodings=False,
     Generates samples and uses torchvision to put them in a side by side grid for easy viewing
     """
     real_images, generated_images, captions = generate_samples(trainer, examples, condition_on_text_encodings, text_prepend)
-
-    real_image_size = real_images[0].shape[-1]
-    generated_image_size = generated_images[0].shape[-1]
-
-    # training images may be larger than the generated one
-    if real_image_size > generated_image_size:
-        real_images = [resize_image_to(image, generated_image_size) for image in real_images]
-
     grid_images = [torchvision.utils.make_grid([original_image, generated_image]) for original_image, generated_image in zip(real_images, generated_images)]
     return grid_images, captions
                     
@@ -279,6 +274,7 @@ def train(
     trainer = DecoderTrainer(
         decoder=decoder,
         accelerator=accelerator,
+        dataloaders=dataloaders,
         **kwargs
     )
 
@@ -289,7 +285,6 @@ def train(
     sample = 0
     samples_seen = 0
     val_sample = 0
-    step = lambda: int(trainer.num_steps_taken(unet_number=1))
 
     if tracker.can_recall:
         start_epoch, validation_losses, next_task, recalled_sample, samples_seen = recall_trainer(tracker, trainer)
@@ -304,6 +299,8 @@ def train(
     if not exists(unet_training_mask):
         # Then the unet mask should be true for all unets in the decoder
         unet_training_mask = [True] * trainer.num_unets
+    first_training_unet = min(index for index, mask in enumerate(unet_training_mask) if mask)
+    step = lambda: int(trainer.num_steps_taken(unet_number=first_training_unet+1))
     assert len(unet_training_mask) == trainer.num_unets, f"The unet training mask should be the same length as the number of unets in the decoder. Got {len(unet_training_mask)} and {trainer.num_unets}"
 
     accelerator.print(print_ribbon("Generating Example Data", repeat=40))
@@ -326,7 +323,7 @@ def train(
         last_snapshot = sample
 
         if next_task == 'train':
-            for i, (img, emb, txt) in enumerate(dataloaders["train"]):
+            for i, (img, emb, txt) in enumerate(trainer.train_loader):
                 # 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.
@@ -419,7 +416,7 @@ def train(
             timer = Timer()
             accelerator.wait_for_everyone()
             i = 0
-            for i, (img, emb, txt) in enumerate(dataloaders["val"]):
+            for i, (img, emb, txt) in enumerate(trainer.val_loader):  # 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()
@@ -524,6 +521,20 @@ def initialize_training(config: TrainDecoderConfig, config_path):
     # Set up accelerator for configurable distributed training
     ddp_kwargs = DistributedDataParallelKwargs(find_unused_parameters=config.train.find_unused_parameters)
     accelerator = Accelerator(kwargs_handlers=[ddp_kwargs])
+
+    if accelerator.num_processes > 1:
+        # We are using distributed training and want to immediately ensure all can connect
+        accelerator.print("Waiting for all processes to connect...")
+        accelerator.wait_for_everyone()
+        accelerator.print("All processes online and connected")
+
+    # If we are in deepspeed fp16 mode, we must ensure learned variance is off
+    if accelerator.mixed_precision == "fp16" and accelerator.distributed_type == accelerate_dataclasses.DistributedType.DEEPSPEED and config.decoder.learned_variance:
+        raise ValueError("DeepSpeed fp16 mode does not support learned variance")
+
+    if accelerator.process_index != accelerator.local_process_index and accelerator.distributed_type == accelerate_dataclasses.DistributedType.DEEPSPEED:
+        # This is an invalid configuration until we figure out how to handle this
+        raise ValueError("DeepSpeed does not support multi-node distributed training")
     
     # Set up data
     all_shards = list(range(config.data.start_shard, config.data.end_shard + 1))
@@ -546,7 +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)
@@ -575,7 +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,

train_diffusion_prior.py

@@ -126,9 +126,9 @@ def report_cosine_sims(
 
         # we are text conditioned, we produce an embedding from the tokenized text
         if text_conditioned:
-            text_embedding, text_encodings, text_mask = trainer.embed_text(text_data)
+            text_embedding, text_encodings = trainer.embed_text(text_data)
             text_cond = dict(
-                text_embed=text_embedding, text_encodings=text_encodings, mask=text_mask
+                text_embed=text_embedding, text_encodings=text_encodings
             )
         else:
             text_embedding = text_data
@@ -146,15 +146,12 @@ def report_cosine_sims(
 
         if text_conditioned:
             text_encodings_shuffled = text_encodings[rolled_idx]
-            text_mask_shuffled = text_mask[rolled_idx]
         else:
             text_encodings_shuffled = None
-            text_mask_shuffled = None
 
         text_cond_shuffled = dict(
             text_embed=text_embed_shuffled,
-            text_encodings=text_encodings_shuffled,
-            mask=text_mask_shuffled,
+            text_encodings=text_encodings_shuffled
         )
 
         # prepare the text embedding