use eval vs training mode to determine whether to call backprop on trainer forward

README.md

@@ -895,14 +895,14 @@ dataset = ImageEmbeddingDataset(
 )
 ```
 
-## Scripts
+### Scripts (wip)
 
-### Using the `train_diffusion_prior.py` script
+#### `train_diffusion_prior.py`
 
 This script allows training the DiffusionPrior on pre-computed text and image embeddings. The working example below elucidates this process.
 Please note that the script internally passes text_embed and image_embed to the DiffusionPrior, unlike the example below.
 
-### Usage 
+#### Usage
 
 ```bash
 $ python train_diffusion_prior.py
@@ -910,58 +910,49 @@ $ python train_diffusion_prior.py
 
 The most significant parameters for the script are as follows:
 
---image-embed-url, default = "https://mystic.the-eye.eu/public/AI/cah/laion5b/embeddings/laion2B-en/img_emb/")
+- `image-embed-url`, default = `"https://mystic.the-eye.eu/public/AI/cah/laion5b/embeddings/laion2B-en/img_emb/"`
 
---text-embed-url, default = "https://mystic.the-eye.eu/public/AI/cah/laion5b/embeddings/laion2B-en/text_emb/")
+- `text-embed-url`, default = `"https://mystic.the-eye.eu/public/AI/cah/laion5b/embeddings/laion2B-en/text_emb/"`
 
---image-embed-dim, default=768 - 768 corresponds to the ViT iL/14 embedding size,change it to what your chosen ViT generates
+- `image-embed-dim`, default = `768` - 768 corresponds to the ViT iL/14 embedding size,change it to what your chosen ViT generates
 
---learning-rate, default=1.1e-4
+- `learning-rate`, default = `1.1e-4`
 
---weight-decay,  default=6.02e-2
+- `weight-decay`,  default = `6.02e-2`
 
---max-grad-norm, default=0.5
+- `max-grad-norm`, default = `0.5`
 
---batch-size, default=10 ** 4
+- `batch-size`, default = `10 ** 4`
 
---num-epochs, default=5
+- `num-epochs`, default = `5`
 
---clip, default=None # Signals the prior to use pre-computed embeddings
+- `clip`, default = `None` # Signals the prior to use pre-computed embeddings
 
-### Sample wandb run log
+#### Loading and Saving the DiffusionPrior model
-
-Please find a sample wandb run log at : https://wandb.ai/laion/diffusion-prior/runs/1blxu24j
-
-### Loading and saving the Diffusion Prior model
 
 Two methods are provided, load_diffusion_model and save_diffusion_model, the names being self-explanatory. 
 
-## from dalle2_pytorch.train import load_diffusion_model, save_diffusion_model
+```python
+from dalle2_pytorch.train import load_diffusion_model, save_diffusion_model
+```
+
+##### Loading
 
     load_diffusion_model(dprior_path, device) 
-
         dprior_path : path to saved model(.pth)
-    
         device      : the cuda device you're running on
     
+##### Saving
+
     save_diffusion_model(save_path, model, optimizer, scaler, config, image_embed_dim)
-    
         save_path : path to save at
-    
         model     : object of Diffusion_Prior
-    
         optimizer : optimizer object - see train_diffusion_prior.py for how to create one. 
-    
             e.g: optimizer = get_optimizer(diffusion_prior.net.parameters(), wd=weight_decay, lr=learning_rate)
-    
         scaler    : a GradScaler object.
-    
             e.g: scaler = GradScaler(enabled=amp)
-    
         config    : config object created in train_diffusion_prior.py - see file for example. 
-    
         image_embed_dim - the dimension of the image_embedding
-    
             e.g: 768
 
 ## CLI (wip)
@@ -1021,6 +1012,7 @@ Once built, images will be saved to the same directory the command is invoked
 - [ ] bring in skip-layer excitatons (from lightweight gan paper) to see if it helps for either decoder of unet or vqgan-vae training
 - [ ] decoder needs one day worth of refactor for tech debt
 - [ ] allow for unet to be able to condition non-cross attention style as well
+- [ ] for all model classes with hyperparameters that changes the network architecture, make it requirement that they must expose a config property, and write a simple function that asserts that it restores the object correctly
 
 ## Citations
 

dalle2_pytorch/dalle2_pytorch.py

@@ -1988,8 +1988,7 @@ class Decoder(BaseGaussianDiffusion):
                 image_size = vae.get_encoded_fmap_size(image_size)
                 shape = (batch_size, vae.encoded_dim, image_size, image_size)
 
-                if exists(lowres_cond_img):
+                lowres_cond_img = maybe(vae.encode)(lowres_cond_img)
-                    lowres_cond_img = vae.encode(lowres_cond_img)
 
                 img = self.p_sample_loop(
                     unet,
@@ -2063,9 +2062,7 @@ class Decoder(BaseGaussianDiffusion):
         vae.eval()
         with torch.no_grad():
             image = vae.encode(image)
-
+            lowres_cond_img = maybe(vae.encode)(lowres_cond_img)
-            if exists(lowres_cond_img):
-                lowres_cond_img = vae.encode(lowres_cond_img)
 
         return self.p_losses(unet, image, times, image_embed = image_embed, text_encodings = text_encodings, text_mask = text_mask, lowres_cond_img = lowres_cond_img, predict_x_start = predict_x_start, learned_variance = learned_variance)
 

dalle2_pytorch/train.py

@@ -80,13 +80,13 @@ def split_args_and_kwargs(*args, split_size = None, **kwargs):
 
     batch_size = len(first_tensor)
     split_size = default(split_size, batch_size)
-    chunk_size = ceil(batch_size / split_size)
+    num_chunks = ceil(batch_size / split_size)
 
     dict_len = len(kwargs)
     dict_keys = kwargs.keys()
     split_kwargs_index = len_all_args - dict_len
 
-    split_all_args = [split(arg, split_size = split_size) if exists(arg) and isinstance(arg, (torch.Tensor, Iterable)) else ((arg,) * chunk_size) for arg in all_args]
+    split_all_args = [split(arg, split_size = split_size) if exists(arg) and isinstance(arg, (torch.Tensor, Iterable)) else ((arg,) * num_chunks) for arg in all_args]
     chunk_sizes = tuple(map(len, split_all_args[0]))
 
     for (chunk_size, *chunked_all_args) in tuple(zip(chunk_sizes, *split_all_args)):
@@ -279,7 +279,9 @@ class DiffusionPriorTrainer(nn.Module):
                 loss = loss * chunk_size_frac
 
             total_loss += loss.item()
-            self.scaler.scale(loss).backward()
+
+            if self.training:
+                self.scaler.scale(loss).backward()
 
         return total_loss
 
@@ -406,6 +408,8 @@ class DecoderTrainer(nn.Module):
                 loss = loss * chunk_size_frac
 
             total_loss += loss.item()
-            self.scale(loss, unet_number = unet_number).backward()
+
+            if self.training:
+                self.scale(loss, unet_number = unet_number).backward()
 
         return total_loss

setup.py

@@ -10,7 +10,7 @@ setup(
       'dream = dalle2_pytorch.cli:dream'
     ],
   },
-  version = '0.2.31',
+  version = '0.2.32',
   license='MIT',
   description = 'DALL-E 2',
   author = 'Phil Wang',

train_diffusion_prior.py

@@ -111,37 +111,110 @@ def report_cosine_sims(diffusion_prior,image_reader,text_reader,train_set_size,N
             "CosineSimilarity(text_embed,predicted_unrelated_embed)": np.mean(unrelated_similarity),
             "Cosine similarity difference":np.mean(predicted_similarity - original_similarity)})
 
-def train(image_embed_dim,
+@click.command()
-          image_embed_url,
+@click.option("--wandb-entity", default="laion")
-          text_embed_url,
+@click.option("--wandb-project", default="diffusion-prior")
-          batch_size,
+@click.option("--wandb-dataset", default="LAION-5B")
-          train_percent,
+@click.option("--wandb-arch", default="DiffusionPrior")
-          val_percent,
+@click.option("--image-embed-url", default="https://mystic.the-eye.eu/public/AI/cah/laion5b/embeddings/laion2B-en/img_emb/")
-          test_percent,
+@click.option("--text-embed-url", default="https://mystic.the-eye.eu/public/AI/cah/laion5b/embeddings/laion2B-en/text_emb/")
-          num_epochs,
+@click.option("--learning-rate", default=1.1e-4)
-          dp_loss_type,
+@click.option("--weight-decay", default=6.02e-2)
-          clip,
+@click.option("--dropout", default=5e-2)
-          dp_condition_on_text_encodings,
+@click.option("--max-grad-norm", default=0.5)
-          dp_timesteps,
+@click.option("--batch-size", default=10**4)
-          dp_normformer,
+@click.option("--num-epochs", default=5)
-          dp_cond_drop_prob,
+@click.option("--image-embed-dim", default=768)
-          dpn_depth,
+@click.option("--train-percent", default=0.7)
-          dpn_dim_head,
+@click.option("--val-percent", default=0.2)
-          dpn_heads,
+@click.option("--test-percent", default=0.1)
-          save_interval,
+@click.option("--dpn-depth", default=6)
-          save_path,
+@click.option("--dpn-dim-head", default=64)
-          device,
+@click.option("--dpn-heads", default=8)
-          RESUME,
+@click.option("--dp-condition-on-text-encodings", default=False)
-          DPRIOR_PATH,
+@click.option("--dp-timesteps", default=100)
-          config,
+@click.option("--dp-normformer", default=False)
-          wandb_entity,
+@click.option("--dp-cond-drop-prob", default=0.1)
-          wandb_project,
+@click.option("--dp-loss-type", default="l2")
-          learning_rate=0.001,
+@click.option("--clip", default=None)
-          max_grad_norm=0.5,
+@click.option("--amp", default=False)
-          weight_decay=0.01,
+@click.option("--save-interval", default=30)
-          dropout=0.05,
+@click.option("--save-path", default="./diffusion_prior_checkpoints")
-          amp=False):
+@click.option("--pretrained-model-path", default=None)
+def train(
+    wandb_entity,
+    wandb_project,
+    wandb_dataset,
+    wandb_arch,
+    image_embed_url,
+    text_embed_url,
+    learning_rate,
+    weight_decay,
+    dropout,
+    max_grad_norm,
+    batch_size,
+    num_epochs,
+    image_embed_dim,
+    train_percent,
+    val_percent,
+    test_percent,
+    dpn_depth,
+    dpn_dim_head,
+    dpn_heads,
+    dp_condition_on_text_encodings,
+    dp_timesteps,
+    dp_normformer,
+    dp_cond_drop_prob,
+    dp_loss_type,
+    clip,
+    amp,
+    save_interval,
+    save_path,
+    pretrained_model_path
+):
+    config = {
+        "learning_rate": learning_rate,
+        "architecture": wandb_arch,
+        "dataset": wandb_dataset,
+        "weight_decay": weight_decay,
+        "max_gradient_clipping_norm": max_grad_norm,
+        "batch_size": batch_size,
+        "epochs": num_epochs,
+        "diffusion_prior_network": {
+            "depth": dpn_depth,
+            "dim_head": dpn_dim_head,
+            "heads": dpn_heads,
+            "normformer": dp_normformer
+        },
+        "diffusion_prior": {
+            "condition_on_text_encodings": dp_condition_on_text_encodings,
+            "timesteps": dp_timesteps,
+            "cond_drop_prob": dp_cond_drop_prob,
+            "loss_type": dp_loss_type,
+            "clip": clip
+        }
+    }
 
+    # Check if DPRIOR_PATH exists(saved model path)
+
+    DPRIOR_PATH = args.pretrained_model_path
+    RESUME = exists(DPRIOR_PATH)
+
+    if not RESUME:
+        tracker.init(
+            entity = wandb_entity,
+            project = wandb_project,
+            config = config
+        )
+
+    # Obtain the utilized device.
+
+    has_cuda = torch.cuda.is_available()
+    if has_cuda:
+        device = torch.device("cuda:0")
+        torch.cuda.set_device(device)
+
+    # Training loop
     # diffusion prior network
 
     prior_network = DiffusionPriorNetwork( 
@@ -269,140 +342,5 @@ def train(image_embed_dim,
     end = num_data_points
     eval_model(diffusion_prior,device,image_reader,text_reader,start,end,batch_size,dp_loss_type,phase="Test")
 
-@click.command()
-@click.option("--wandb-entity", default="laion")
-@click.option("--wandb-project", default="diffusion-prior")
-@click.option("--wandb-dataset", default="LAION-5B")
-@click.option("--wandb-arch", default="DiffusionPrior")
-@click.option("--image-embed-url", default="https://mystic.the-eye.eu/public/AI/cah/laion5b/embeddings/laion2B-en/img_emb/")
-@click.option("--text-embed-url", default="https://mystic.the-eye.eu/public/AI/cah/laion5b/embeddings/laion2B-en/text_emb/")
-@click.option("--learning-rate", default=1.1e-4)
-@click.option("--weight-decay", default=6.02e-2)
-@click.option("--dropout", default=5e-2)
-@click.option("--max-grad-norm", default=0.5)
-@click.option("--batch-size", default=10**4)
-@click.option("--num-epochs", default=5)
-@click.option("--image-embed-dim", default=768)
-@click.option("--train-percent", default=0.7)
-@click.option("--val-percent", default=0.2)
-@click.option("--test-percent", default=0.1)
-@click.option("--dpn-depth", default=6)
-@click.option("--dpn-dim-head", default=64)
-@click.option("--dpn-heads", default=8)
-@click.option("--dp-condition-on-text-encodings", default=False)
-@click.option("--dp-timesteps", default=100)
-@click.option("--dp-normformer", default=False)
-@click.option("--dp-cond-drop-prob", default=0.1)
-@click.option("--dp-loss-type", default="l2")
-@click.option("--clip", default=None)
-@click.option("--amp", default=False)
-@click.option("--save-interval", default=30)
-@click.option("--save-path", default="./diffusion_prior_checkpoints")
-@click.option("--pretrained-model-path", default=None)
-def main(
-    wandb_entity,
-    wandb_project,
-    wandb_dataset,
-    wandb_arch,
-    image_embed_url,
-    text_embed_url,
-    learning_rate,
-    weight_decay,
-    dropout,
-    max_grad_norm,
-    batch_size,
-    num_epochs,
-    image_embed_dim,
-    train_percent,
-    val_percent,
-    test_percent,
-    dpn_depth,
-    dpn_dim_head,
-    dpn_heads,
-    dp_condition_on_text_encodings,
-    dp_timesteps,
-    dp_normformer,
-    dp_cond_drop_prob,
-    dp_loss_type,
-    clip,
-    amp,
-    save_interval,
-    save_path,
-    pretrained_model_path
-):
-    config = {
-        "learning_rate": learning_rate,
-        "architecture": wandb_arch,
-        "dataset": wandb_dataset,
-        "weight_decay": weight_decay,
-        "max_gradient_clipping_norm": max_grad_norm,
-        "batch_size": batch_size,
-        "epochs": num_epochs,
-        "diffusion_prior_network": {
-            "depth": dpn_depth,
-            "dim_head": dpn_dim_head,
-            "heads": dpn_heads,
-            "normformer": dp_normformer
-        },
-        "diffusion_prior": {
-            "condition_on_text_encodings": dp_condition_on_text_encodings,
-            "timesteps": dp_timesteps,
-            "cond_drop_prob": dp_cond_drop_prob,
-            "loss_type": dp_loss_type,
-            "clip": clip
-        }
-    }
-
-    # Check if DPRIOR_PATH exists(saved model path)
-
-    DPRIOR_PATH = args.pretrained_model_path
-    RESUME = exists(DPRIOR_PATH)
-
-    if not RESUME:
-        tracker.init(
-            entity = wandb_entity,
-            project = wandb_project,
-            config = config
-        )
-
-    # Obtain the utilized device.
-
-    has_cuda = torch.cuda.is_available()
-    if has_cuda:
-        device = torch.device("cuda:0")
-        torch.cuda.set_device(device)
-
-    # Training loop
-    train(image_embed_dim,
-          image_embed_url,
-          text_embed_url,
-          batch_size,
-          train_percent,
-          val_percent,
-          test_percent,
-          num_epochs,
-          dp_loss_type,
-          clip,
-          dp_condition_on_text_encodings,
-          dp_timesteps,
-          dp_normformer,
-          dp_cond_drop_prob,
-          dpn_depth,
-          dpn_dim_head,
-          dpn_heads,
-          save_interval,
-          save_path,
-          device,
-          RESUME,
-          DPRIOR_PATH,
-          config,
-          wandb_entity,
-          wandb_project,
-          learning_rate,
-          max_grad_norm,
-          weight_decay,
-          dropout,
-          amp)
-
 if __name__ == "__main__":
-    main()
+    train()