extra insurance in case eos id is not there

fix the previous commit which assumes open_clip is installed

dalle2_pytorch/dalle2_pytorch.py

@@ -360,6 +360,7 @@ 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)
@@ -434,6 +435,7 @@ 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)
@@ -629,7 +631,7 @@ class NoiseScheduler(nn.Module):
 
     def calculate_v(self, x_start, t, noise = None):
         return (
-            extract(self.sqrt_alphas_cumprod, t, x_start.shape) * noise +
+            extract(self.sqrt_alphas_cumprod, t, x_start.shape) * noise -
             extract(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * x_start
         )
 
@@ -1320,7 +1322,7 @@ class DiffusionPrior(nn.Module):
             elif self.predict_x_start:
                 x_start = pred
             else:
-                x_start = self.noise_scheduler.predict_start_from_noise(image_embed, t = time_cond, noise = pred_noise)
+                x_start = self.noise_scheduler.predict_start_from_noise(image_embed, t = time_cond, noise = pred)
 
             # clip x0 before maybe predicting noise
 

dalle2_pytorch/train_configs.py

@@ -4,11 +4,13 @@ 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,
@@ -117,6 +119,10 @@ 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":

dalle2_pytorch/trainer.py

@@ -236,7 +236,7 @@ class DiffusionPriorTrainer(nn.Module):
         )
 
         if exists(cosine_decay_max_steps):
-            self.scheduler = CosineAnnealingLR(optimizer, T_max = cosine_decay_max_steps)
+            self.scheduler = CosineAnnealingLR(self.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__ = '1.11.0'
+__version__ = '1.11.4'

setup.py

@@ -26,6 +26,7 @@ setup(
   install_requires=[
     'accelerate',
     'click',
+    'open-clip-torch>=2.0.0,<3.0.0',
     'clip-anytorch>=2.4.0',
     'coca-pytorch>=0.0.5',
     'ema-pytorch>=0.0.7',