[Examples] Add streaming support to the ControlNet training example in JAX (#2859)

* improve stable unclip doc.

* feat: add streaming support to controlnet flax training script.

* fix: CLI arg.

* fix: torch dataloader shuffle setting.

* fix: dataset length.

* fix: wandb config.

* fix: steps_per_epoch in the training loop.

* add: entry about streaming in the readme

* get column names from iterable dataset + fix final logging

---------

Co-authored-by: yiyixuxu <[email protected]>

Author: sayakpaul

examples/controlnet/README.md

@@ -335,15 +335,15 @@ huggingface-cli login
 
 Make sure you have the `MODEL_DIR`,`OUTPUT_DIR` and `HUB_MODEL_ID` environment variables set. The `OUTPUT_DIR` and `HUB_MODEL_ID` variables specify where to save the model to on the Hub:
 
-```
+```bash
 export MODEL_DIR="runwayml/stable-diffusion-v1-5"
 export OUTPUT_DIR="control_out"
 export HUB_MODEL_ID="fill-circle-controlnet"
 ```
 
 And finally start the training
 
-```
+```bash
 python3 train_controlnet_flax.py \
  --pretrained_model_name_or_path=$MODEL_DIR \
  --output_dir=$OUTPUT_DIR \
@@ -363,3 +363,30 @@ python3 train_controlnet_flax.py \
  ```
 
 Since we passed the `--push_to_hub` flag, it will automatically create a model repo under your huggingface account based on `$HUB_MODEL_ID`. By the end of training, the final checkpoint will be automatically stored on the hub. You can find an example model repo [here](https://huggingface.co/YiYiXu/fill-circle-controlnet).
+
+Our training script also provides limited support for streaming large datasets from the Hugging Face Hub. In order to enable streaming, one must also set `--max_train_samples`.  Here is an example command:
+
+```bash
+python3 train_controlnet_flax.py \
+	--pretrained_model_name_or_path=$MODEL_DIR \
+	--output_dir=$OUTPUT_DIR \
+	--dataset_name=multimodalart/facesyntheticsspigacaptioned \
+	--streaming \
+	--conditioning_image_column=spiga_seg \
+	--image_column=image \
+	--caption_column=image_caption \
+	--resolution=512 \
+	--max_train_samples 50 \
+	--max_train_steps 5 \
+	--learning_rate=1e-5 \
+	--validation_steps=2 \
+	--train_batch_size=1 \
+	--revision="flax" \
+	--report_to="wandb"
+```
+
+Note, however, that the performance of the TPUs might get bottlenecked as streaming with `datasets` is not optimized for images. For ensuring maximum throughput, we encourage you to explore the following options:
+
+* [Webdataset](https://webdataset.github.io/webdataset/)
+* [TorchData](https://github.com/pytorch/data)
+* [TensorFlow Datasets](https://www.tensorflow.org/datasets/tfless_tfds)

examples/controlnet/train_controlnet_flax.py

@@ -35,6 +35,7 @@
 from flax.training.common_utils import shard
 from huggingface_hub import HfFolder, Repository, create_repo, whoami
 from PIL import Image
+from torch.utils.data import IterableDataset
 from torchvision import transforms
 from tqdm.auto import tqdm
 from transformers import CLIPTokenizer, FlaxCLIPTextModel, set_seed
@@ -206,7 +207,7 @@ def parse_args():
     parser.add_argument(
         "--from_pt",
         action="store_true",
-        help="Load the pretrained model from a pytorch checkpoint.",
+        help="Load the pretrained model from a PyTorch checkpoint.",
     )
     parser.add_argument(
         "--tokenizer_name",
@@ -332,6 +333,7 @@ def parse_args():
             " or to a folder containing files that 🤗 Datasets can understand."
         ),
     )
+    parser.add_argument("--streaming", action="store_true", help="To stream a large dataset from Hub.")
     parser.add_argument(
         "--dataset_config_name",
         type=str,
@@ -369,7 +371,7 @@ def parse_args():
         default=None,
         help=(
             "For debugging purposes or quicker training, truncate the number of training examples to this "
-            "value if set."
+            "value if set. Needed if `streaming` is set to True."
         ),
     )
     parser.add_argument(
@@ -453,10 +455,15 @@ def parse_args():
             " or the same number of `--validation_prompt`s and `--validation_image`s"
         )
 
+    # This idea comes from
+    # https://github.com/borisdayma/dalle-mini/blob/d2be512d4a6a9cda2d63ba04afc33038f98f705f/src/dalle_mini/data.py#L370
+    if args.streaming and args.max_train_samples is None:
+        raise ValueError("You must specify `max_train_samples` when using dataset streaming.")
+
     return args
 
 
-def make_train_dataset(args, tokenizer):
+def make_train_dataset(args, tokenizer, batch_size=None):
     # Get the datasets: you can either provide your own training and evaluation files (see below)
     # or specify a Dataset from the hub (the dataset will be downloaded automatically from the datasets Hub).
 
@@ -468,6 +475,7 @@ def make_train_dataset(args, tokenizer):
             args.dataset_name,
             args.dataset_config_name,
             cache_dir=args.cache_dir,
+            streaming=args.streaming,
         )
     else:
         data_files = {}
@@ -483,7 +491,10 @@ def make_train_dataset(args, tokenizer):
 
     # Preprocessing the datasets.
     # We need to tokenize inputs and targets.
-    column_names = dataset["train"].column_names
+    if isinstance(dataset["train"], IterableDataset):
+        column_names = next(iter(dataset["train"])).keys()
+    else:
+        column_names = dataset["train"].column_names
 
     # 6. Get the column names for input/target.
     if args.image_column is None:
@@ -565,9 +576,20 @@ def preprocess_train(examples):
 
     if jax.process_index() == 0:
         if args.max_train_samples is not None:
-            dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples))
+            if args.streaming:
+                dataset["train"] = dataset["train"].shuffle(seed=args.seed).take(args.max_train_samples)
+            else:
+                dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples))
         # Set the training transforms
-        train_dataset = dataset["train"].with_transform(preprocess_train)
+        if args.streaming:
+            train_dataset = dataset["train"].map(
+                preprocess_train,
+                batched=True,
+                batch_size=batch_size,
+                remove_columns=list(dataset["train"].features.keys()),
+            )
+        else:
+            train_dataset = dataset["train"].with_transform(preprocess_train)
 
     return train_dataset
 
@@ -661,12 +683,12 @@ def main():
         raise NotImplementedError("No tokenizer specified!")
 
     # Get the datasets: you can either provide your own training and evaluation files (see below)
-    train_dataset = make_train_dataset(args, tokenizer)
     total_train_batch_size = args.train_batch_size * jax.local_device_count() * args.gradient_accumulation_steps
+    train_dataset = make_train_dataset(args, tokenizer, batch_size=total_train_batch_size)
 
     train_dataloader = torch.utils.data.DataLoader(
         train_dataset,
-        shuffle=True,
+        shuffle=not args.streaming,
         collate_fn=collate_fn,
         batch_size=total_train_batch_size,
         num_workers=args.dataloader_num_workers,
@@ -897,7 +919,11 @@ def cumul_grad_step(grad_idx, loss_grad_rng):
     vae_params = jax_utils.replicate(vae_params)
 
     # Train!
-    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    if args.streaming:
+        dataset_length = args.max_train_samples
+    else:
+        dataset_length = len(train_dataloader)
+    num_update_steps_per_epoch = math.ceil(dataset_length / args.gradient_accumulation_steps)
 
     # Scheduler and math around the number of training steps.
     if args.max_train_steps is None:
@@ -906,7 +932,7 @@ def cumul_grad_step(grad_idx, loss_grad_rng):
     args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
 
     logger.info("***** Running training *****")
-    logger.info(f"  Num examples = {len(train_dataset)}")
+    logger.info(f"  Num examples = {args.max_train_samples if args.streaming else len(train_dataset)}")
     logger.info(f"  Num Epochs = {args.num_train_epochs}")
     logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
     logger.info(f"  Total train batch size (w. parallel & distributed) = {total_train_batch_size}")
@@ -916,7 +942,7 @@ def cumul_grad_step(grad_idx, loss_grad_rng):
         wandb.define_metric("*", step_metric="train/step")
         wandb.config.update(
             {
-                "num_train_examples": len(train_dataset),
+                "num_train_examples": args.max_train_samples if args.streaming else len(train_dataset),
                 "total_train_batch_size": total_train_batch_size,
                 "total_optimization_step": args.num_train_epochs * num_update_steps_per_epoch,
                 "num_devices": jax.device_count(),
@@ -935,7 +961,11 @@ def cumul_grad_step(grad_idx, loss_grad_rng):
 
         train_metrics = []
 
-        steps_per_epoch = len(train_dataset) // total_train_batch_size
+        steps_per_epoch = (
+            args.max_train_samples // total_train_batch_size
+            if args.streaming
+            else len(train_dataset) // total_train_batch_size
+        )
         train_step_progress_bar = tqdm(
             total=steps_per_epoch,
             desc="Training...",
@@ -980,7 +1010,8 @@ def cumul_grad_step(grad_idx, loss_grad_rng):
 
     # Create the pipeline using using the trained modules and save it.
     if jax.process_index() == 0:
-        image_logs = log_validation(controlnet, state.params, tokenizer, args, validation_rng, weight_dtype)
+        if args.validation_prompt is not None:
+            image_logs = log_validation(controlnet, state.params, tokenizer, args, validation_rng, weight_dtype)
 
         controlnet.save_pretrained(
             args.output_dir,