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

README.md

@@ -355,7 +355,8 @@ prior_network = DiffusionPriorNetwork(
 diffusion_prior = DiffusionPrior(
     net = prior_network,
     clip = clip,
-    timesteps = 100,
+    timesteps = 1000,
+    sample_timesteps = 64,
     cond_drop_prob = 0.2
 ).cuda()
 
@@ -583,6 +584,7 @@ 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 +600,8 @@ 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

@@ -41,7 +41,7 @@
         "resample_train": true,
         "preprocessing": {
             "RandomResizedCrop": {
-                "size": [64, 64],
+                "size": [224, 224],
                 "scale": [0.75, 1.0],
                 "ratio": [1.0, 1.0]
             },

dalle2_pytorch/dalle2_pytorch.py

@@ -77,6 +77,11 @@ def cast_tuple(val, length = None):
 def module_device(module):
     return next(module.parameters()).device
 
+def zero_init_(m):
+    nn.init.zeros_(m.weight)
+    if exists(m.bias):
+        nn.init.zeros_(m.bias)
+
 @contextmanager
 def null_context(*args, **kwargs):
     yield
@@ -220,6 +225,7 @@ class XClipAdapter(BaseClipAdapter):
         encoder_output = self.clip.text_transformer(text)
         text_cls, text_encodings = encoder_output[:, 0], encoder_output[:, 1:]
         text_embed = self.clip.to_text_latent(text_cls)
+        text_encodings = text_encodings.masked_fill(~text_mask[..., None], 0.)
         return EmbeddedText(l2norm(text_embed), text_encodings, text_mask)
 
     @torch.no_grad()
@@ -255,6 +261,7 @@ class CoCaAdapter(BaseClipAdapter):
         text = text[..., :self.max_text_len]
         text_mask = text != 0
         text_embed, text_encodings = self.clip.embed_text(text)
+        text_encodings = text_encodings.masked_fill(~text_mask[..., None], 0.)
         return EmbeddedText(text_embed, text_encodings, text_mask)
 
     @torch.no_grad()
@@ -314,6 +321,7 @@ class OpenAIClipAdapter(BaseClipAdapter):
 
         text_embed = self.clip.encode_text(text)
         text_encodings = self.text_encodings
+        text_encodings = text_encodings.masked_fill(~text_mask[..., None], 0.)
         del self.text_encodings
         return EmbeddedText(l2norm(text_embed.float()), text_encodings.float(), text_mask)
 
@@ -505,6 +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
@@ -911,19 +925,23 @@ 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,
@@ -954,23 +972,32 @@ 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)'
 
@@ -978,8 +1005,6 @@ 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)
 
@@ -1002,21 +1027,81 @@ 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))
 
@@ -1030,7 +1115,7 @@ 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
 
@@ -1051,7 +1136,15 @@ 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)
@@ -1066,7 +1159,7 @@ class DiffusionPrior(nn.Module):
         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, 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
 
@@ -1112,6 +1205,7 @@ class DiffusionPrior(nn.Module):
 
         if self.condition_on_text_encodings:
             assert exists(text_encodings), 'text encodings must be present for diffusion prior if specified'
+            text_mask = default(text_mask, lambda: torch.any(text_encodings != 0., dim = -1))
             text_cond = {**text_cond, 'text_encodings': text_encodings, 'mask': text_mask}
 
         # timestep conditioning from ddpm
@@ -1129,16 +1223,35 @@ class DiffusionPrior(nn.Module):
 
 # decoder
 
-def ConvTransposeUpsample(dim, dim_out = None):
-    dim_out = default(dim_out, dim)
-    return nn.ConvTranspose2d(dim, dim_out, 4, 2, 1)
+class PixelShuffleUpsample(nn.Module):
+    """
+    code shared by @MalumaDev at DALLE2-pytorch for addressing checkboard artifacts
+    https://arxiv.org/ftp/arxiv/papers/1707/1707.02937.pdf
+    """
+    def __init__(self, dim, dim_out = None):
+        super().__init__()
+        dim_out = default(dim_out, dim)
+        conv = nn.Conv2d(dim, dim_out * 4, 1)
 
-def NearestUpsample(dim, dim_out = None):
-    dim_out = default(dim_out, dim)
-    return nn.Sequential(
-        nn.Upsample(scale_factor = 2, mode = 'nearest'),
-        nn.Conv2d(dim, dim_out, 3, padding = 1)
-    )
+        self.net = nn.Sequential(
+            conv,
+            nn.SiLU(),
+            nn.PixelShuffle(2)
+        )
+
+        self.init_conv_(conv)
+
+    def init_conv_(self, conv):
+        o, i, h, w = conv.weight.shape
+        conv_weight = torch.empty(o // 4, i, h, w)
+        nn.init.kaiming_uniform_(conv_weight)
+        conv_weight = repeat(conv_weight, 'o ... -> (o 4) ...')
+
+        conv.weight.data.copy_(conv_weight)
+        nn.init.zeros_(conv.bias.data)
+
+    def forward(self, x):
+        return self.net(x)
 
 def Downsample(dim, *, dim_out = None):
     dim_out = default(dim_out, dim)
@@ -1402,7 +1515,7 @@ class Unet(nn.Module):
         cross_embed_downsample_kernel_sizes = (2, 4),
         memory_efficient = False,
         scale_skip_connection = False,
-        nearest_upsample = False,
+        pixel_shuffle_upsample = True,
         final_conv_kernel_size = 1,
         **kwargs
     ):
@@ -1468,10 +1581,12 @@ 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
@@ -1514,7 +1629,7 @@ class Unet(nn.Module):
 
         # upsample klass
 
-        upsample_klass = ConvTransposeUpsample if not nearest_upsample else NearestUpsample
+        upsample_klass = ConvTransposeUpsample if not pixel_shuffle_upsample else PixelShuffleUpsample
 
         # give memory efficient unet an initial resnet block
 
@@ -1578,6 +1693,8 @@ class Unet(nn.Module):
         self.final_resnet_block = ResnetBlock(dim * 2, dim, time_cond_dim = time_cond_dim, groups = top_level_resnet_group)
         self.to_out = nn.Conv2d(dim, self.channels_out, kernel_size = final_conv_kernel_size, padding = final_conv_kernel_size // 2)
 
+        zero_init_(self.to_out) # since both OpenAI and @crowsonkb are doing it
+
     # if the current settings for the unet are not correct
     # for cascading DDPM, then reinit the unet with the right settings
     def cast_model_parameters(
@@ -1700,6 +1817,8 @@ class Unet(nn.Module):
         text_tokens = None
 
         if exists(text_encodings) and self.cond_on_text_encodings:
+            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_tokens = self.text_to_cond(text_encodings)
             text_tokens = text_tokens[:, :self.max_text_len]
 
@@ -1853,6 +1972,7 @@ 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',
@@ -1876,7 +1996,8 @@ 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__()
 
@@ -1956,6 +2077,11 @@ 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):
@@ -1966,7 +2092,9 @@ 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,
@@ -2067,6 +2195,26 @@ class Decoder(nn.Module):
         for unet, device in zip(self.unets, devices):
             unet.to(device)
 
+    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, text_mask = 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)'
 
@@ -2081,21 +2229,7 @@ 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)
 
@@ -2125,7 +2259,7 @@ class Decoder(nn.Module):
         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, text_mask = None, cond_scale = 1, is_latent_diffusion = False):
         device = self.device
 
         b = shape[0]
@@ -2153,6 +2287,62 @@ class Decoder(nn.Module):
         unnormalize_img = self.unnormalize_img(img)
         return unnormalize_img
 
+    @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, text_mask = 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, text_mask = text_mask, 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, text_mask = 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))
 
@@ -2250,10 +2440,13 @@ class Decoder(nn.Module):
         assert not (self.condition_on_text_encodings and not exists(text_encodings)), 'text or text encodings must be passed into decoder if specified'
         assert not (not self.condition_on_text_encodings and exists(text_encodings)), 'decoder specified not to be conditioned on text, yet it is presented'
 
+        if self.condition_on_text_encodings:
+            text_mask = default(text_mask, lambda: torch.any(text_encodings != 0., dim = -1))
+
         img = None
         is_cuda = next(self.parameters()).is_cuda
 
-        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()
 
@@ -2282,7 +2475,8 @@ class Decoder(nn.Module):
                     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)
@@ -2332,6 +2526,9 @@ class Decoder(nn.Module):
         assert not (self.condition_on_text_encodings and not exists(text_encodings)), 'text or text encodings must be passed into decoder if specified'
         assert not (not self.condition_on_text_encodings and exists(text_encodings)), 'decoder specified not to be conditioned on text, yet it is presented'
 
+        if self.condition_on_text_encodings:
+            text_mask = default(text_mask, lambda: torch.any(text_encodings != 0., dim = -1))
+
         lowres_cond_img = self.to_lowres_cond(image, target_image_size = target_image_size, downsample_image_size = self.image_sizes[unet_index - 1]) if unet_number > 1 else None
         image = resize_image_to(image, target_image_size)
 

dalle2_pytorch/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

@@ -154,6 +154,7 @@ 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 +234,7 @@ 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__ = '0.17.1'
+__version__ = '0.21.0'

train_decoder.py

@@ -274,6 +274,7 @@ def train(
     trainer = DecoderTrainer(
         decoder=decoder,
         accelerator=accelerator,
+        dataloaders=dataloaders,
         **kwargs
     )
 
@@ -284,7 +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)
@@ -299,6 +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))
@@ -321,7 +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.
@@ -414,7 +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()
@@ -519,6 +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))
@@ -541,7 +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)
@@ -570,7 +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,