make it work for @ethancohen123

README.md

@@ -634,12 +634,10 @@ Alternatively, you can also use <a href="https://github.com/mlfoundations/open_c
 $ pip install open-clip-torch
 ```
 
-Ex. using the <a href="https://laion.ai/blog/large-openclip/">SOTA Open Clip</a> model trained by <a href="https://github.com/rom1504">Romain</a>
-
 ```python
 from dalle2_pytorch import OpenClipAdapter
 
-clip = OpenClipAdapter('ViT-H/14')
+clip = OpenClipAdapter()
 ```
 
 Now you'll just have to worry about training the Prior and the Decoder!
@@ -1068,7 +1066,7 @@ dataloader = create_image_embedding_dataloader(
 )
 for img, emb in dataloader:
     print(img.shape)  # torch.Size([32, 3, 256, 256])
-    print(emb["img"].shape)  # torch.Size([32, 512])
+    print(emb.shape)  # torch.Size([32, 512])
     # Train decoder only as shown above
 
 # Or create a dataset without a loader so you can configure it manually
@@ -1128,7 +1126,6 @@ For detailed information on training the diffusion prior, please refer to the [d
 - [x] add inpainting ability using resampler from repaint paper https://arxiv.org/abs/2201.09865
 - [x] add the final combination of upsample feature maps, used in unet squared, seems to have an effect in local experiments
 - [ ] consider elucidated dalle2 https://arxiv.org/abs/2206.00364
-- [ ] add simple outpainting, text-guided 2x size the image for starters
 - [ ] interface out the vqgan-vae so a pretrained one can be pulled off the shelf to validate latent diffusion + DALL-E2
 
 ## Citations
@@ -1288,24 +1285,4 @@ For detailed information on training the diffusion prior, please refer to the [d
 }
 ```
 
-```bibtex
-@article{Sunkara2022NoMS,
-    title   = {No More Strided Convolutions or Pooling: A New CNN Building Block for Low-Resolution Images and Small Objects},
-    author  = {Raja Sunkara and Tie Luo},
-    journal = {ArXiv},
-    year    = {2022},
-    volume  = {abs/2208.03641}
-}
-```
-
-```bibtex
-@article{Salimans2022ProgressiveDF,
-    title   = {Progressive Distillation for Fast Sampling of Diffusion Models},
-    author  = {Tim Salimans and Jonathan Ho},
-    journal = {ArXiv},
-    year    = {2022},
-    volume  = {abs/2202.00512}
-}
-```
-
 *Creating noise from data is easy; creating data from noise is generative modeling.* - <a href="https://arxiv.org/abs/2011.13456">Yang Song's paper</a>

dalle2_pytorch/__init__.py

@@ -1,6 +1,6 @@
 from dalle2_pytorch.version import __version__
 from dalle2_pytorch.dalle2_pytorch import DALLE2, DiffusionPriorNetwork, DiffusionPrior, Unet, Decoder
-from dalle2_pytorch.dalle2_pytorch import OpenAIClipAdapter, OpenClipAdapter
+from dalle2_pytorch.dalle2_pytorch import OpenAIClipAdapter
 from dalle2_pytorch.trainer import DecoderTrainer, DiffusionPriorTrainer
 
 from dalle2_pytorch.vqgan_vae import VQGanVAE

dalle2_pytorch/dalle2_pytorch.py

@@ -100,9 +100,6 @@ def eval_decorator(fn):
         return out
     return inner
 
-def is_float_dtype(dtype):
-    return any([dtype == float_dtype for float_dtype in (torch.float64, torch.float32, torch.float16, torch.bfloat16)])
-
 def is_list_str(x):
     if not isinstance(x, (list, tuple)):
         return False
@@ -254,9 +251,7 @@ class XClipAdapter(BaseClipAdapter):
         text_mask = text != 0
         encoder_output = self.clip.text_transformer(text)
 
-        encoder_output_is_cls = encoder_output.ndim == 3
-
-        text_cls, text_encodings = (encoder_output[:, 0], encoder_output[:, 1:]) if encoder_output_is_cls else (encoder_output, None)
+        text_cls, text_encodings = (encoder_output[:, 0], encoder_output[:, 1:]) if encoder_output.ndim == 3 else (encoder_output, None)
         text_embed = self.clip.to_text_latent(text_cls)
 
         if exists(text_encodings):
@@ -317,10 +312,7 @@ class OpenAIClipAdapter(BaseClipAdapter):
         self.eos_id = 49407 # for handling 0 being also '!'
 
         text_attention_final = self.find_layer('ln_final')
-
-        self.dim_latent_ = text_attention_final.weight.shape[0]
         self.handle = text_attention_final.register_forward_hook(self._hook)
-
         self.clip_normalize = preprocess.transforms[-1]
         self.cleared = False
 
@@ -339,7 +331,7 @@ class OpenAIClipAdapter(BaseClipAdapter):
 
     @property
     def dim_latent(self):
-        return self.dim_latent_
+        return 512
 
     @property
     def image_size(self):
@@ -360,7 +352,6 @@ class OpenAIClipAdapter(BaseClipAdapter):
         is_eos_id = (text == self.eos_id)
         text_mask_excluding_eos = is_eos_id.cumsum(dim = -1) == 0
         text_mask = F.pad(text_mask_excluding_eos, (1, -1), value = True)
-        text_mask = text_mask & (text != 0)
         assert not self.cleared
 
         text_embed = self.clip.encode_text(text)
@@ -390,8 +381,6 @@ class OpenClipAdapter(BaseClipAdapter):
         self.eos_id = 49407
 
         text_attention_final = self.find_layer('ln_final')
-        self._dim_latent = text_attention_final.weight.shape[0]
-
         self.handle = text_attention_final.register_forward_hook(self._hook)
         self.clip_normalize = preprocess.transforms[-1]
         self.cleared = False
@@ -411,14 +400,11 @@ class OpenClipAdapter(BaseClipAdapter):
 
     @property
     def dim_latent(self):
-        return self._dim_latent
+        return 512
 
     @property
     def image_size(self):
-        image_size = self.clip.visual.image_size
-        if isinstance(image_size, tuple):
-            return max(image_size)
-        return image_size
+        return self.clip.visual.image_size
 
     @property
     def image_channels(self):
@@ -435,7 +421,6 @@ class OpenClipAdapter(BaseClipAdapter):
         is_eos_id = (text == self.eos_id)
         text_mask_excluding_eos = is_eos_id.cumsum(dim = -1) == 0
         text_mask = F.pad(text_mask_excluding_eos, (1, -1), value = True)
-        text_mask = text_mask & (text != 0)
         assert not self.cleared
 
         text_embed = self.clip.encode_text(text)
@@ -621,7 +606,7 @@ class NoiseScheduler(nn.Module):
         posterior_log_variance_clipped = extract(self.posterior_log_variance_clipped, t, x_t.shape)
         return posterior_mean, posterior_variance, posterior_log_variance_clipped
 
-    def q_sample(self, x_start, t, noise = None):
+    def q_sample(self, x_start, t, noise=None):
         noise = default(noise, lambda: torch.randn_like(x_start))
 
         return (
@@ -629,12 +614,6 @@ class NoiseScheduler(nn.Module):
             extract(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise
         )
 
-    def calculate_v(self, x_start, t, noise = None):
-        return (
-            extract(self.sqrt_alphas_cumprod, t, x_start.shape) * noise -
-            extract(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * x_start
-        )
-
     def q_sample_from_to(self, x_from, from_t, to_t, noise = None):
         shape = x_from.shape
         noise = default(noise, lambda: torch.randn_like(x_from))
@@ -646,12 +625,6 @@ class NoiseScheduler(nn.Module):
 
         return x_from * (alpha_next / alpha) + noise * (sigma_next * alpha - sigma * alpha_next) / alpha
 
-    def predict_start_from_v(self, x_t, t, v):
-        return (
-            extract(self.sqrt_alphas_cumprod, t, x_t.shape) * x_t -
-            extract(self.sqrt_one_minus_alphas_cumprod, t, x_t.shape) * v
-        )
-
     def predict_start_from_noise(self, x_t, t, noise):
         return (
             extract(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
@@ -904,8 +877,6 @@ class Attention(nn.Module):
         # attention
 
         attn = sim.softmax(dim = -1, dtype = torch.float32)
-        attn = attn.type(sim.dtype)
-
         attn = self.dropout(attn)
 
         # aggregate values
@@ -987,8 +958,6 @@ class DiffusionPriorNetwork(nn.Module):
             Rearrange('b (n d) -> b n d', n = num_text_embeds)
         )
 
-        self.continuous_embedded_time = not exists(num_timesteps)
-
         self.to_time_embeds = nn.Sequential(
             nn.Embedding(num_timesteps, dim * num_time_embeds) if exists(num_timesteps) else nn.Sequential(SinusoidalPosEmb(dim), MLP(dim, dim * num_time_embeds)), # also offer a continuous version of timestep embeddings, with a 2 layer MLP
             Rearrange('b (n d) -> b n d', n = num_time_embeds)
@@ -1005,10 +974,7 @@ class DiffusionPriorNetwork(nn.Module):
         # dalle1 learned padding strategy
 
         self.max_text_len = max_text_len
-
-        self.null_text_encodings = nn.Parameter(torch.randn(1, max_text_len, dim))
-        self.null_text_embeds = nn.Parameter(torch.randn(1, num_text_embeds, dim))
-        self.null_image_embed = nn.Parameter(torch.randn(1, dim))
+        self.null_text_embed = nn.Parameter(torch.randn(1, max_text_len, dim))
 
         # whether to use self conditioning, Hinton's group's new ddpm technique
 
@@ -1025,7 +991,7 @@ class DiffusionPriorNetwork(nn.Module):
         if cond_scale == 1:
             return logits
 
-        null_logits = self.forward(*args, text_cond_drop_prob = 1., image_cond_drop_prob = 1, **kwargs)
+        null_logits = self.forward(*args, cond_drop_prob = 1., **kwargs)
         return null_logits + (logits - null_logits) * cond_scale
 
     def forward(
@@ -1036,8 +1002,7 @@ class DiffusionPriorNetwork(nn.Module):
         text_embed,
         text_encodings = None,
         self_cond = None,
-        text_cond_drop_prob = 0.,
-        image_cond_drop_prob = 0.
+        cond_drop_prob = 0.
     ):
         batch, dim, device, dtype = *image_embed.shape, image_embed.device, image_embed.dtype
 
@@ -1055,14 +1020,6 @@ class DiffusionPriorNetwork(nn.Module):
         text_embed = self.to_text_embeds(text_embed)
         image_embed = self.to_image_embeds(image_embed)
 
-        # classifier free guidance masks
-
-        text_keep_mask = prob_mask_like((batch,), 1 - text_cond_drop_prob, device = device)
-        text_keep_mask = rearrange(text_keep_mask, 'b -> b 1 1')
-
-        image_keep_mask = prob_mask_like((batch,), 1 - image_cond_drop_prob, device = device)
-        image_keep_mask = rearrange(image_keep_mask, 'b -> b 1 1')
-
         # make text encodings optional
         # although the paper seems to suggest it is present <--
 
@@ -1083,48 +1040,38 @@ class DiffusionPriorNetwork(nn.Module):
             text_encodings = F.pad(text_encodings, (0, 0, 0, remainder), value = 0.)
             mask = F.pad(mask, (0, remainder), value = False)
 
-        # mask out text encodings with null encodings
-
-        null_text_encodings = self.null_text_encodings.to(text_encodings.dtype)
+        null_text_embeds = self.null_text_embed.to(text_encodings.dtype)
 
         text_encodings = torch.where(
-            rearrange(mask, 'b n -> b n 1').clone() & text_keep_mask,
+            rearrange(mask, 'b n -> b n 1').clone(),
             text_encodings,
-            null_text_encodings
-        )
-
-        # mask out text embeddings with null text embeddings
-
-        null_text_embeds = self.null_text_embeds.to(text_embed.dtype)
-
-        text_embed = torch.where(
-            text_keep_mask,
-            text_embed,
             null_text_embeds
         )
 
-        # mask out image embeddings with null image embeddings
+        # classifier free guidance
 
-        null_image_embed = self.null_image_embed.to(image_embed.dtype)
+        keep_mask = prob_mask_like((batch,), 1 - cond_drop_prob, device = device)
+        keep_mask = rearrange(keep_mask, 'b -> b 1')
 
-        image_embed = torch.where(
-            image_keep_mask,
-            image_embed,
-            null_image_embed
-        )
+        mask &= keep_mask
+
+        # whether text embedding is masked or not depends on the classifier free guidance conditional masking
+
+        keep_mask = repeat(keep_mask, 'b 1 -> b n', n = num_text_embeds)
+        mask = torch.cat((mask, keep_mask), dim = 1)
 
         # 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 self.continuous_embedded_time:
-            diffusion_timesteps = diffusion_timesteps.type(dtype)
+        attend_padding = 1 + num_time_embeds + num_image_embeds + int(self.self_cond) # 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)
 
         learned_queries = repeat(self.learned_query, 'd -> b 1 d', b = batch)
 
         if self.self_cond:
-            learned_queries = torch.cat((self_cond, learned_queries), dim = -2)
+            learned_queries = torch.cat((image_embed, self_cond), dim = -2)
 
         tokens = torch.cat((
             text_encodings,
@@ -1156,11 +1103,8 @@ class DiffusionPrior(nn.Module):
         timesteps = 1000,
         sample_timesteps = None,
         cond_drop_prob = 0.,
-        text_cond_drop_prob = None,
-        image_cond_drop_prob = None,
         loss_type = "l2",
         predict_x_start = True,
-        predict_v = False,
         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,       # whether to l2norm clamp the image embed at each denoising iteration (analogous to -1 to 1 clipping for usual DDPMs)
@@ -1197,22 +1141,15 @@ class DiffusionPrior(nn.Module):
 
         self.net = net
         self.image_embed_dim = default(image_embed_dim, lambda: clip.dim_latent)
-
-        assert net.dim == self.image_embed_dim, f'your diffusion prior network has a dimension of {net.dim}, but you set your image embedding dimension (keyword image_embed_dim) on DiffusionPrior to {self.image_embed_dim}'
-        assert not exists(clip) or clip.dim_latent == self.image_embed_dim, f'you passed in a CLIP to the diffusion prior with latent dimensions of {clip.dim_latent}, but your image embedding dimension (keyword image_embed_dim) for the DiffusionPrior was set to {self.image_embed_dim}'
-
         self.channels = default(image_channels, lambda: clip.image_channels)
 
-        self.text_cond_drop_prob = default(text_cond_drop_prob, cond_drop_prob)
-        self.image_cond_drop_prob = default(image_cond_drop_prob, cond_drop_prob)
-
-        self.can_classifier_guidance = self.text_cond_drop_prob > 0. and self.image_cond_drop_prob > 0.
+        self.cond_drop_prob = cond_drop_prob
+        self.can_classifier_guidance = cond_drop_prob > 0.
         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
-        self.predict_v = predict_v # takes precedence over predict_x_start
 
         # @crowsonkb 's suggestion - https://github.com/lucidrains/DALLE2-pytorch/issues/60#issue-1226116132
 
@@ -1242,9 +1179,7 @@ class DiffusionPrior(nn.Module):
 
         pred = self.net.forward_with_cond_scale(x, t, cond_scale = cond_scale, self_cond = self_cond, **text_cond)
 
-        if self.predict_v:
-            x_start = self.noise_scheduler.predict_start_from_v(x, t = t, v = pred)
-        elif self.predict_x_start:
+        if self.predict_x_start:
             x_start = pred
         else:
             x_start = self.noise_scheduler.predict_start_from_noise(x, t = t, noise = pred)
@@ -1293,7 +1228,7 @@ class DiffusionPrior(nn.Module):
     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(-1., total_timesteps, steps = timesteps + 1)[:-1]
+        times = torch.linspace(0., total_timesteps, steps = timesteps + 2)[:-1]
 
         times = list(reversed(times.int().tolist()))
         time_pairs = list(zip(times[:-1], times[1:]))
@@ -1315,16 +1250,12 @@ class DiffusionPrior(nn.Module):
 
             pred = self.net.forward_with_cond_scale(image_embed, time_cond, self_cond = self_cond, cond_scale = cond_scale, **text_cond)
 
-            # derive x0
-
-            if self.predict_v:
-                x_start = self.noise_scheduler.predict_start_from_v(image_embed, t = time_cond, v = pred)
-            elif self.predict_x_start:
+            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)
-
-            # clip x0 before maybe predicting noise
+                pred_noise = pred
 
             if not self.predict_x_start:
                 x_start.clamp_(-1., 1.)
@@ -1332,17 +1263,6 @@ class DiffusionPrior(nn.Module):
             if self.predict_x_start and self.sampling_clamp_l2norm:
                 x_start = self.l2norm_clamp_embed(x_start)
 
-            # predict noise
-
-            if self.predict_x_start or self.predict_v:
-                pred_noise = self.noise_scheduler.predict_noise_from_start(image_embed, t = time_cond, x0 = x_start)
-            else:
-                pred_noise = pred
-
-            if time_next < 0:
-                image_embed = x_start
-                continue
-
             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.
@@ -1384,20 +1304,14 @@ class DiffusionPrior(nn.Module):
             image_embed_noisy,
             times,
             self_cond = self_cond,
-            text_cond_drop_prob = self.text_cond_drop_prob,
-            image_cond_drop_prob = self.image_cond_drop_prob,
+            cond_drop_prob = self.cond_drop_prob,
             **text_cond
         )
 
         if self.predict_x_start and self.training_clamp_l2norm:
             pred = self.l2norm_clamp_embed(pred)
 
-        if self.predict_v:
-            target = self.noise_scheduler.calculate_v(image_embed, times, noise)
-        elif self.predict_x_start:
-            target = image_embed
-        else:
-            target = noise
+        target = noise if not self.predict_x_start else image_embed
 
         loss = self.noise_scheduler.loss_fn(pred, target)
         return loss
@@ -1467,7 +1381,7 @@ class DiffusionPrior(nn.Module):
         **kwargs
     ):
         assert exists(text) ^ exists(text_embed), 'either text or text embedding must be supplied'
-        assert exists(image) ^ exists(image_embed), 'either image or image embedding must be supplied'
+        assert exists(image) ^ exists(image_embed), 'either text or text embedding must be supplied'
         assert not (self.condition_on_text_encodings and (not exists(text_encodings) and not exists(text))), 'text encodings must be present if you specified you wish to condition on it on initialization'
 
         if exists(image):
@@ -1537,14 +1451,9 @@ class PixelShuffleUpsample(nn.Module):
     def forward(self, x):
         return self.net(x)
 
-def Downsample(dim, dim_out = None):
-    # https://arxiv.org/abs/2208.03641 shows this is the most optimal way to downsample
-    # named SP-conv in the paper, but basically a pixel unshuffle
+def Downsample(dim, *, dim_out = None):
     dim_out = default(dim_out, dim)
-    return nn.Sequential(
-        Rearrange('b c (h s1) (w s2) -> b (c s1 s2) h w', s1 = 2, s2 = 2),
-        nn.Conv2d(dim * 4, dim_out, 1)
-    )
+    return nn.Conv2d(dim, dim_out, 4, 2, 1)
 
 class WeightStandardizedConv2d(nn.Conv2d):
     """
@@ -1573,8 +1482,6 @@ class SinusoidalPosEmb(nn.Module):
 
     def forward(self, x):
         dtype, device = x.dtype, x.device
-        assert is_float_dtype(dtype), 'input to sinusoidal pos emb must be a float type'
-
         half_dim = self.dim // 2
         emb = math.log(10000) / (half_dim - 1)
         emb = torch.exp(torch.arange(half_dim, device = device, dtype = dtype) * -emb)
@@ -1728,7 +1635,6 @@ class CrossAttention(nn.Module):
             sim = sim.masked_fill(~mask, max_neg_value)
 
         attn = sim.softmax(dim = -1, dtype = torch.float32)
-        attn = attn.type(sim.dtype)
 
         out = einsum('b h i j, b h j d -> b h i d', attn, v)
         out = rearrange(out, 'b h n d -> b n (h d)')
@@ -2473,7 +2379,6 @@ class Decoder(nn.Module):
         loss_type = 'l2',
         beta_schedule = None,
         predict_x_start = False,
-        predict_v = False,
         predict_x_start_for_latent_diffusion = False,
         image_sizes = None,                         # for cascading ddpm, image size at each stage
         random_crop_sizes = None,                   # whether to random crop the image at that stage in the cascade (super resoluting convolutions at the end may be able to generalize on smaller crops)
@@ -2496,7 +2401,7 @@ class Decoder(nn.Module):
         dynamic_thres_percentile = 0.95,
         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,
-        ddim_sampling_eta = 0.                      # can be set to 0. for deterministic sampling afaict
+        ddim_sampling_eta = 1.                      # can be set to 0. for deterministic sampling afaict
     ):
         super().__init__()
 
@@ -2646,10 +2551,6 @@ class Decoder(nn.Module):
 
         self.predict_x_start = cast_tuple(predict_x_start, len(unets)) if not predict_x_start_for_latent_diffusion else tuple(map(lambda t: isinstance(t, VQGanVAE), self.vaes))
 
-        # predict v
-
-        self.predict_v = cast_tuple(predict_v, len(unets))
-
         # input image range
 
         self.input_image_range = (-1. if not auto_normalize_img else 0., 1.)
@@ -2761,16 +2662,14 @@ class Decoder(nn.Module):
         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, self_cond = None, clip_denoised = True, predict_x_start = False, predict_v = False, learned_variance = False, cond_scale = 1., model_output = None, lowres_noise_level = None):
+    def p_mean_variance(self, unet, x, t, image_embed, noise_scheduler, text_encodings = None, lowres_cond_img = None, self_cond = None, clip_denoised = True, predict_x_start = False, learned_variance = False, cond_scale = 1., model_output = None, lowres_noise_level = 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)'
 
         model_output = 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, self_cond = self_cond, lowres_noise_level = lowres_noise_level))
 
         pred, var_interp_frac_unnormalized = self.parse_unet_output(learned_variance, model_output)
 
-        if predict_v:
-            x_start = noise_scheduler.predict_start_from_v(x, t = t, v = pred)
-        elif predict_x_start:
+        if predict_x_start:
             x_start = pred
         else:
             x_start = noise_scheduler.predict_start_from_noise(x, t = t, noise = pred)
@@ -2797,9 +2696,9 @@ class Decoder(nn.Module):
         return model_mean, posterior_variance, posterior_log_variance, x_start
 
     @torch.no_grad()
-    def p_sample(self, unet, x, t, image_embed, noise_scheduler, text_encodings = None, cond_scale = 1., lowres_cond_img = None, self_cond = None, predict_x_start = False, predict_v = False, learned_variance = False, clip_denoised = True, lowres_noise_level = None):
+    def p_sample(self, unet, x, t, image_embed, noise_scheduler, text_encodings = None, cond_scale = 1., lowres_cond_img = None, self_cond = None, predict_x_start = False, learned_variance = False, clip_denoised = True, lowres_noise_level = None):
         b, *_, device = *x.shape, x.device
-        model_mean, _, model_log_variance, x_start = 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, self_cond = self_cond, clip_denoised = clip_denoised, predict_x_start = predict_x_start, predict_v = predict_v, noise_scheduler = noise_scheduler, learned_variance = learned_variance, lowres_noise_level = lowres_noise_level)
+        model_mean, _, model_log_variance, x_start = 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, self_cond = self_cond, clip_denoised = clip_denoised, predict_x_start = predict_x_start, noise_scheduler = noise_scheduler, learned_variance = learned_variance, lowres_noise_level = lowres_noise_level)
         noise = torch.randn_like(x)
         # no noise when t == 0
         nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
@@ -2814,7 +2713,6 @@ class Decoder(nn.Module):
         image_embed,
         noise_scheduler,
         predict_x_start = False,
-        predict_v = False,
         learned_variance = False,
         clip_denoised = True,
         lowres_cond_img = None,
@@ -2873,7 +2771,6 @@ class Decoder(nn.Module):
                     lowres_cond_img = lowres_cond_img,
                     lowres_noise_level = lowres_noise_level,
                     predict_x_start = predict_x_start,
-                    predict_v = predict_v,
                     noise_scheduler = noise_scheduler,
                     learned_variance = learned_variance,
                     clip_denoised = clip_denoised
@@ -2899,7 +2796,6 @@ class Decoder(nn.Module):
         timesteps,
         eta = 1.,
         predict_x_start = False,
-        predict_v = False,
         learned_variance = False,
         clip_denoised = True,
         lowres_cond_img = None,
@@ -2911,13 +2807,12 @@ class Decoder(nn.Module):
         inpaint_mask = None,
         inpaint_resample_times = 5
     ):
-        batch, device, total_timesteps, alphas, eta = shape[0], self.device, noise_scheduler.num_timesteps, noise_scheduler.alphas_cumprod, self.ddim_sampling_eta
+        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:]))
-        time_pairs = list(filter(lambda t: t[0] > t[1], time_pairs))
 
         is_inpaint = exists(inpaint_image)
         resample_times = inpaint_resample_times if is_inpaint else 1
@@ -2959,27 +2854,16 @@ class Decoder(nn.Module):
 
                 pred, _ = self.parse_unet_output(learned_variance, unet_output)
 
-                # predict x0
-
-                if predict_v:
-                    x_start = noise_scheduler.predict_start_from_v(img, t = time_cond, v = pred)
-                elif predict_x_start:
+                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)
-
-                # maybe clip x0
+                    pred_noise = pred
 
                 if clip_denoised:
                     x_start = self.dynamic_threshold(x_start)
 
-                # predict noise
-
-                if predict_x_start or predict_v:
-                    pred_noise = noise_scheduler.predict_noise_from_start(img, t = time_cond, x0 = x_start)
-                else:
-                    pred_noise = pred
-
                 c1 = eta * ((1 - alpha / alpha_next) * (1 - alpha_next) / (1 - alpha)).sqrt()
                 c2 = ((1 - alpha_next) - torch.square(c1)).sqrt()
                 noise = torch.randn_like(img) if not is_last_timestep else 0.
@@ -3012,7 +2896,7 @@ class Decoder(nn.Module):
 
         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, predict_v = False, noise = None, learned_variance = False, clip_denoised = False, is_latent_diffusion = False, lowres_noise_level = None):
+    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, lowres_noise_level = None):
         noise = default(noise, lambda: torch.randn_like(x_start))
 
         # normalize to [-1, 1]
@@ -3057,12 +2941,7 @@ class Decoder(nn.Module):
 
         pred, _ = self.parse_unet_output(learned_variance, unet_output)
 
-        if predict_v:
-            target = noise_scheduler.calculate_v(x_start, times, noise)
-        elif predict_x_start:
-            target = x_start
-        else:
-            target = noise
+        target = noise if not predict_x_start else x_start
 
         loss = noise_scheduler.loss_fn(pred, target, reduction = 'none')
         loss = reduce(loss, 'b ... -> b (...)', 'mean')
@@ -3148,7 +3027,7 @@ class Decoder(nn.Module):
         num_unets = self.num_unets
         cond_scale = cast_tuple(cond_scale, num_unets)
 
-        for unet_number, unet, vae, channel, image_size, predict_x_start, predict_v, learned_variance, noise_scheduler, lowres_cond, sample_timesteps, unet_cond_scale in tqdm(zip(range(1, num_unets + 1), self.unets, self.vaes, self.sample_channels, self.image_sizes, self.predict_x_start, self.predict_v, self.learned_variance, self.noise_schedulers, self.lowres_conds, self.sample_timesteps, cond_scale)):
+        for unet_number, unet, vae, channel, image_size, predict_x_start, learned_variance, noise_scheduler, lowres_cond, sample_timesteps, unet_cond_scale in tqdm(zip(range(1, num_unets + 1), self.unets, self.vaes, self.sample_channels, self.image_sizes, self.predict_x_start, self.learned_variance, self.noise_schedulers, self.lowres_conds, self.sample_timesteps, cond_scale)):
             if unet_number < start_at_unet_number:
                 continue  # It's the easiest way to do it
 
@@ -3184,7 +3063,6 @@ class Decoder(nn.Module):
                     text_encodings = text_encodings,
                     cond_scale = unet_cond_scale,
                     predict_x_start = predict_x_start,
-                    predict_v = predict_v,
                     learned_variance = learned_variance,
                     clip_denoised = not is_latent_diffusion,
                     lowres_cond_img = lowres_cond_img,
@@ -3224,7 +3102,6 @@ class Decoder(nn.Module):
         lowres_conditioner  = self.lowres_conds[unet_index]
         target_image_size   = self.image_sizes[unet_index]
         predict_x_start     = self.predict_x_start[unet_index]
-        predict_v           = self.predict_v[unet_index]
         random_crop_size    = self.random_crop_sizes[unet_index]
         learned_variance    = self.learned_variance[unet_index]
         b, c, h, w, device, = *image.shape, image.device
@@ -3263,7 +3140,7 @@ 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, lowres_cond_img = lowres_cond_img, predict_x_start = predict_x_start, predict_v = predict_v, learned_variance = learned_variance, is_latent_diffusion = is_latent_diffusion, noise_scheduler = noise_scheduler, lowres_noise_level = lowres_noise_level)
+        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, lowres_noise_level = lowres_noise_level)
 
         if not return_lowres_cond_image:
             return losses

dalle2_pytorch/train_configs.py

@@ -4,13 +4,11 @@ from pydantic import BaseModel, validator, root_validator
 from typing import List, Optional, Union, Tuple, Dict, Any, TypeVar
 
 from x_clip import CLIP as XCLIP
-from open_clip import list_pretrained
 from coca_pytorch import CoCa
 
 from dalle2_pytorch.dalle2_pytorch import (
     CoCaAdapter,
     OpenAIClipAdapter,
-    OpenClipAdapter,
     Unet,
     Decoder,
     DiffusionPrior,
@@ -119,10 +117,6 @@ class AdapterConfig(BaseModel):
     def create(self):
         if self.make == "openai":
             return OpenAIClipAdapter(self.model)
-        elif self.make == "open_clip":
-            pretrained = dict(list_pretrained())
-            checkpoint = pretrained[self.model]
-            return OpenClipAdapter(name=self.model, pretrained=checkpoint)
         elif self.make == "x-clip":
             return XClipAdapter(XCLIP(**self.base_model_kwargs))
         elif self.make == "coca":
@@ -247,7 +241,7 @@ 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[Optional[int]]] = None
+    sample_timesteps: Optional[SingularOrIterable[int]] = None
     loss_type: str = 'l2'
     beta_schedule: ListOrTuple[str] = None  # None means all cosine
     learned_variance: SingularOrIterable[bool] = True
@@ -313,7 +307,6 @@ class DecoderTrainConfig(BaseModel):
     wd: SingularOrIterable[float] = 0.01
     warmup_steps: Optional[SingularOrIterable[int]] = None
     find_unused_parameters: bool = True
-    static_graph: bool = True
     max_grad_norm: SingularOrIterable[float] = 0.5
     save_every_n_samples: int = 100000
     n_sample_images: int = 6                       # The number of example images to produce when sampling the train and test dataset

dalle2_pytorch/trainer.py

@@ -236,7 +236,7 @@ class DiffusionPriorTrainer(nn.Module):
         )
 
         if exists(cosine_decay_max_steps):
-            self.scheduler = CosineAnnealingLR(self.optimizer, T_max = cosine_decay_max_steps)
+            self.scheduler = CosineAnnealingLR(optimizer, T_max = cosine_decay_max_steps)
         else:
             self.scheduler = LambdaLR(self.optimizer, lr_lambda = lambda _: 1.0)
         

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '1.12.2'
+__version__ = '1.8.3'

setup.py

@@ -26,8 +26,7 @@ setup(
   install_requires=[
     'accelerate',
     'click',
-    'open-clip-torch>=2.0.0,<3.0.0',
-    'clip-anytorch>=2.5.2',
+    'clip-anytorch>=2.4.0',
     'coca-pytorch>=0.0.5',
     'ema-pytorch>=0.0.7',
     'einops>=0.4',

train_decoder.py

@@ -134,7 +134,7 @@ 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, clip=None, start_unet=1, end_unet=None, condition_on_text_encodings=False, cond_scale=1.0, device=None, text_prepend="", match_image_size=True):
+def generate_samples(trainer, example_data, start_unet=1, end_unet=None, condition_on_text_encodings=False, cond_scale=1.0, device=None, text_prepend="", match_image_size=True):
     """
     Takes example data and generates images from the embeddings
     Returns three lists: real images, generated images, and captions
@@ -144,9 +144,7 @@ def generate_samples(trainer, example_data, clip=None, start_unet=1, end_unet=No
     if img_embeddings[0] is None:
         # Generate image embeddings from clip
         imgs_tensor = torch.stack(real_images)
-        assert clip is not None, "clip is None, but img_embeddings is None"
-        imgs_tensor.to(device=device)
-        img_embeddings, img_encoding = clip.embed_image(imgs_tensor)
+        img_embeddings, *_ = trainer.embed_image(imgs_tensor)
         sample_params["image_embed"] = img_embeddings
     else:
         # Then we are using precomputed image embeddings
@@ -155,10 +153,8 @@ def generate_samples(trainer, example_data, clip=None, start_unet=1, end_unet=No
     if condition_on_text_encodings:
         if text_embeddings[0] is None:
             # Generate text embeddings from text
-            assert clip is not None, "clip is None, but text_embeddings is None"
-            tokenized_texts = tokenize(txts, truncate=True).to(device=device)
-            text_embed, text_encodings = clip.embed_text(tokenized_texts)
-            sample_params["text_encodings"] = text_encodings
+            tokenized_texts = tokenize(txts, truncate=True)
+            sample_params["text"] = tokenized_texts
         else:
             # Then we are using precomputed text embeddings
             text_embeddings = torch.stack(text_embeddings)
@@ -170,7 +166,7 @@ def generate_samples(trainer, example_data, clip=None, start_unet=1, end_unet=No
         sample_params["image"] = torch.stack(real_images)
     if device is not None:
         sample_params["_device"] = device
-    samples = trainer.sample(**sample_params, _cast_deepspeed_precision=False)  # At sampling time we don't want to cast to FP16
+    samples = trainer.sample(**sample_params)
     generated_images = list(samples)
     captions = [text_prepend + txt for txt in txts]
     if match_image_size:
@@ -178,15 +174,15 @@ def generate_samples(trainer, example_data, clip=None, start_unet=1, end_unet=No
         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, clip=None, start_unet=1, end_unet=None, condition_on_text_encodings=False, cond_scale=1.0, device=None, text_prepend=""):
+def generate_grid_samples(trainer, examples, start_unet=1, end_unet=None, condition_on_text_encodings=False, cond_scale=1.0, device=None, text_prepend=""):
     """
     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, clip, start_unet, end_unet, condition_on_text_encodings, cond_scale, device, text_prepend)
+    real_images, generated_images, captions = generate_samples(trainer, examples, start_unet, end_unet, condition_on_text_encodings, cond_scale, device, text_prepend)
     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
                     
-def evaluate_trainer(trainer, dataloader, device, start_unet, end_unet, clip=None, condition_on_text_encodings=False, cond_scale=1.0, inference_device=None, n_evaluation_samples=1000, FID=None, IS=None, KID=None, LPIPS=None):
+def evaluate_trainer(trainer, dataloader, device,  start_unet, end_unet, condition_on_text_encodings=False, cond_scale=1.0, inference_device=None, n_evaluation_samples=1000, FID=None, IS=None, KID=None, LPIPS=None):
     """
     Computes evaluation metrics for the decoder
     """
@@ -196,7 +192,7 @@ def evaluate_trainer(trainer, dataloader, device, start_unet, end_unet, clip=Non
     if len(examples) == 0:
         print("No data to evaluate. Check that your dataloader has shards.")
         return metrics
-    real_images, generated_images, captions = generate_samples(trainer, examples, clip, start_unet, end_unet, condition_on_text_encodings, cond_scale, inference_device)
+    real_images, generated_images, captions = generate_samples(trainer, examples, start_unet, end_unet, condition_on_text_encodings, cond_scale, inference_device)
     real_images = torch.stack(real_images).to(device=device, dtype=torch.float)
     generated_images = torch.stack(generated_images).to(device=device, dtype=torch.float)
     # Convert from [0, 1] to [0, 255] and from torch.float to torch.uint8
@@ -229,8 +225,8 @@ def evaluate_trainer(trainer, dataloader, device, start_unet, end_unet, clip=Non
         metrics["KID_std"] = kid_std.item()
     if exists(LPIPS):
         # Convert from [0, 1] to [-1, 1]
-        renorm_real_images = real_images.mul(2).sub(1).clamp(-1,1)
-        renorm_generated_images = generated_images.mul(2).sub(1).clamp(-1,1)
+        renorm_real_images = real_images.mul(2).sub(1)
+        renorm_generated_images = generated_images.mul(2).sub(1)
         lpips = LearnedPerceptualImagePatchSimilarity(**LPIPS, dist_sync_fn=null_sync)
         lpips.to(device=device)
         lpips.update(renorm_real_images, renorm_generated_images)
@@ -269,7 +265,6 @@ def train(
     accelerator: Accelerator,
     tracker: Tracker,
     inference_device,
-    clip=None,
     evaluate_config=None,
     epoch_samples = None,  # If the training dataset is resampling, we have to manually stop an epoch
     validation_samples = None,
@@ -376,19 +371,15 @@ def train(
                         forward_params['image_embed'] = img_emb
                     else:
                         # Forward pass automatically generates embedding
-                        assert clip is not None
-                        img_embed, img_encoding = clip.embed_image(img)
-                        forward_params['image_embed'] = img_embed
+                        pass
                     if condition_on_text_encodings:
                         if has_text_embedding:
                             forward_params['text_encodings'] = text_emb
                         else:
                             # Then we need to pass the text instead
-                            assert clip is not None
-                            tokenized_texts = tokenize(txt, truncate=True).to(inference_device)
+                            tokenized_texts = tokenize(txt, truncate=True)
                             assert tokenized_texts.shape[0] == len(img), f"The number of texts ({tokenized_texts.shape[0]}) should be the same as the number of images ({len(img)})"
-                            text_embed, text_encodings = clip.embed_text(tokenized_texts)
-                            forward_params['text_encodings'] = text_encodings
+                            forward_params['text'] = tokenized_texts
                     loss = trainer.forward(img, **forward_params, unet_number=unet, _device=inference_device)
                     trainer.update(unet_number=unet)
                     unet_losses_tensor[i % TRAIN_CALC_LOSS_EVERY_ITERS, unet-1] = loss
@@ -428,7 +419,7 @@ def train(
                     save_trainer(tracker, trainer, epoch, sample, next_task, validation_losses, samples_seen)
                     if exists(n_sample_images) and n_sample_images > 0:
                         trainer.eval()
-                        train_images, train_captions = generate_grid_samples(trainer, train_example_data, clip, first_trainable_unet, last_trainable_unet, condition_on_text_encodings, cond_scale, inference_device, "Train: ")
+                        train_images, train_captions = generate_grid_samples(trainer, train_example_data, first_trainable_unet, last_trainable_unet, condition_on_text_encodings, cond_scale, inference_device, "Train: ")
                         tracker.log_images(train_images, captions=train_captions, image_section="Train Samples", step=step())
                 
                 if epoch_samples is not None and sample >= epoch_samples:
@@ -471,19 +462,15 @@ def train(
                         forward_params['image_embed'] = img_emb.float()
                     else:
                         # Forward pass automatically generates embedding
-                        assert clip is not None
-                        img_embed, img_encoding = clip.embed_image(img)
-                        forward_params['image_embed'] = img_embed
+                        pass
                     if condition_on_text_encodings:
                         if has_text_embedding:
                             forward_params['text_encodings'] = text_emb.float()
                         else:
                             # Then we need to pass the text instead
-                            assert clip is not None
-                            tokenized_texts = tokenize(txt, truncate=True).to(device=inference_device)
+                            tokenized_texts = tokenize(txt, truncate=True)
                             assert tokenized_texts.shape[0] == len(img), f"The number of texts ({tokenized_texts.shape[0]}) should be the same as the number of images ({len(img)})"
-                            text_embed, text_encodings = clip.embed_text(tokenized_texts)
-                            forward_params['text_encodings'] = text_encodings
+                            forward_params['text'] = tokenized_texts
                     loss = trainer.forward(img.float(), **forward_params, unet_number=unet, _device=inference_device)
                     average_val_loss_tensor[0, unet-1] += loss
 
@@ -511,7 +498,7 @@ def train(
         if next_task == 'eval':
             if exists(evaluate_config):
                 accelerator.print(print_ribbon(f"Starting Evaluation {epoch}", repeat=40))
-                evaluation = evaluate_trainer(trainer, dataloaders["val"], inference_device, first_trainable_unet, last_trainable_unet, clip=clip, inference_device=inference_device, **evaluate_config.dict(), condition_on_text_encodings=condition_on_text_encodings, cond_scale=cond_scale)
+                evaluation = evaluate_trainer(trainer, dataloaders["val"], inference_device, first_trainable_unet, last_trainable_unet, inference_device=inference_device, **evaluate_config.dict(), condition_on_text_encodings=condition_on_text_encodings, cond_scale=cond_scale)
                 if is_master:
                     tracker.log(evaluation, step=step())
             next_task = 'sample'
@@ -522,8 +509,8 @@ def train(
                 # Generate examples and save the model if we are the master
                 # Generate sample images
                 print(print_ribbon(f"Sampling Set {epoch}", repeat=40))
-                test_images, test_captions = generate_grid_samples(trainer, test_example_data, clip, first_trainable_unet, last_trainable_unet, condition_on_text_encodings, cond_scale, inference_device, "Test: ")
-                train_images, train_captions = generate_grid_samples(trainer, train_example_data, clip, first_trainable_unet, last_trainable_unet, condition_on_text_encodings, cond_scale, inference_device, "Train: ")
+                test_images, test_captions = generate_grid_samples(trainer, test_example_data,  first_trainable_unet, last_trainable_unet, condition_on_text_encodings, cond_scale, inference_device, "Test: ")
+                train_images, train_captions = generate_grid_samples(trainer, train_example_data,  first_trainable_unet, last_trainable_unet, condition_on_text_encodings, cond_scale, inference_device, "Train: ")
                 tracker.log_images(test_images, captions=test_captions, image_section="Test Samples", step=step())
                 tracker.log_images(train_images, captions=train_captions, image_section="Train Samples", step=step())
 
@@ -545,7 +532,6 @@ def create_tracker(accelerator: Accelerator, config: TrainDecoderConfig, config_
         "NumProcesses": accelerator.num_processes,
         "MixedPrecision": accelerator.mixed_precision
     }
-    accelerator.wait_for_everyone()  # If nodes arrive at this point at different times they might try to autoresume the current run which makes no sense and will cause errors
     tracker: Tracker = tracker_config.create(config, accelerator_config, dummy_mode=dummy)
     tracker.save_config(config_path, config_name='decoder_config.json')
     tracker.add_save_metadata(state_dict_key='config', metadata=config.dict())
@@ -556,7 +542,7 @@ def initialize_training(config: TrainDecoderConfig, config_path):
     torch.manual_seed(config.seed)
 
     # Set up accelerator for configurable distributed training
-    ddp_kwargs = DistributedDataParallelKwargs(find_unused_parameters=config.train.find_unused_parameters, static_graph=config.train.static_graph)
+    ddp_kwargs = DistributedDataParallelKwargs(find_unused_parameters=config.train.find_unused_parameters)
     init_kwargs = InitProcessGroupKwargs(timeout=timedelta(seconds=60*60))
     accelerator = Accelerator(kwargs_handlers=[ddp_kwargs, init_kwargs])
 
@@ -569,6 +555,10 @@ def initialize_training(config: TrainDecoderConfig, config_path):
     # 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))
@@ -589,11 +579,6 @@ def initialize_training(config: TrainDecoderConfig, config_path):
         seed = config.seed,
     )
 
-    # If clip is in the model, we need to remove it for compatibility with deepspeed
-    clip = None
-    if config.decoder.clip is not None:
-        clip = config.decoder.clip.create()  # Of course we keep it to use it during training, just not in the decoder as that causes issues
-        config.decoder.clip = None
     # Create the decoder model and print basic info
     decoder = config.decoder.create()
     get_num_parameters = lambda model, only_training=False: sum(p.numel() for p in model.parameters() if (p.requires_grad or not only_training))
@@ -605,7 +590,7 @@ def initialize_training(config: TrainDecoderConfig, config_path):
     has_text_embeddings = config.data.text_embeddings_url is not None
     conditioning_on_text = any([unet.cond_on_text_encodings for unet in config.decoder.unets])
 
-    has_clip_model = clip is not None
+    has_clip_model = config.decoder.clip is not None
     data_source_string = ""
 
     if has_img_embeddings:
@@ -630,7 +615,6 @@ def initialize_training(config: TrainDecoderConfig, config_path):
         accelerator.print(f"Unet {i} has {get_num_parameters(unet)} total; {get_num_parameters(unet, only_training=True)} training")
 
     train(dataloaders, decoder, accelerator,
-        clip=clip,
         tracker=tracker,
         inference_device=accelerator.device,
         evaluate_config=config.evaluate,