[Research Projects] ORPO diffusion for alignment (#7423)

* barebones orpo

* remove reference model.

* full implementation

* change default of beta_orpo

* add a training command.

* fix: dataloading issues.

* interpreting the formulation.

* revert styling

* add: wds full blown version

* fix: per_gpu_batch_siz

* start debuggin

* debugging

* remove print

* fix

* remove filter keys.

* turn on non-blocking calls.

* device_placement

* let's see.

* add bigger training run command

* reinitialize generator for fair repro

* add: detailed readme and requirements

---------

Co-authored-by: Sayak Paul <sayakpaul@Sayaks-MacBook-Pro-2.local>

examples/research_projects/diffusion_orpo/README.md

@@ -0,0 +1,121 @@
+This project is an attempt to check if it's possible to apply to [ORPO](https://arxiv.org/abs/2403.07691) on a text-conditioned diffusion model to align it on preference data WITHOUT a reference model. The implementation is based on https://github.com/huggingface/trl/pull/1435/. 
+
+> [!WARNING] 
+> We assume that MSE in the diffusion formulation approximates the log-probs as required by ORPO (hat-tip to [@kashif](https://github.com/kashif) for the idea). So, please consider this to be extremely experimental.
+
+## Training
+
+Here's training command you can use on a 40GB A100 to validate things on a [small preference
+dataset](https://hf.co/datasets/kashif/pickascore): 
+
+```bash
+accelerate launch train_diffusion_orpo_sdxl_lora.py \
+  --pretrained_model_name_or_path=stabilityai/stable-diffusion-xl-base-1.0  \
+  --pretrained_vae_model_name_or_path=madebyollin/sdxl-vae-fp16-fix \
+  --output_dir="diffusion-sdxl-orpo" \
+  --mixed_precision="fp16" \
+  --dataset_name=kashif/pickascore \
+  --train_batch_size=8 \
+  --gradient_accumulation_steps=2 \
+  --gradient_checkpointing \
+  --use_8bit_adam \
+  --rank=8 \
+  --learning_rate=1e-5 \
+  --report_to="wandb" \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --max_train_steps=2000 \
+  --checkpointing_steps=500 \
+  --run_validation --validation_steps=50 \
+  --seed="0" \
+  --report_to="wandb" \
+  --push_to_hub
+```
+
+We also provide a simple script to scale up the training on the [yuvalkirstain/pickapic_v2](https://huggingface.co/datasets/yuvalkirstain/pickapic_v2) dataset:
+
+```bash
+accelerate launch --multi_gpu train_diffusion_orpo_sdxl_lora_wds.py \
+  --pretrained_model_name_or_path=stabilityai/stable-diffusion-xl-base-1.0  \
+  --pretrained_vae_model_name_or_path=madebyollin/sdxl-vae-fp16-fix \
+  --dataset_path="pipe:aws s3 cp s3://diffusion-preference-opt/{00000..00644}.tar -" \
+  --output_dir="diffusion-sdxl-orpo-wds" \
+  --mixed_precision="fp16" \
+  --gradient_accumulation_steps=1 \
+  --gradient_checkpointing \
+  --use_8bit_adam \
+  --rank=8 \
+  --dataloader_num_workers=8 \
+  --learning_rate=3e-5 \
+  --report_to="wandb" \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --max_train_steps=50000 \
+  --checkpointing_steps=2000 \
+  --run_validation --validation_steps=500 \
+  --seed="0" \
+  --report_to="wandb" \
+  --push_to_hub
+```
+
+We tested the above on a node of 8 H100s but it should also work on A100s. It requires the `webdataset` library for faster dataloading. Note that we kept the dataset shards on an S3 bucket but it should be also possible to have them stored locally. 
+
+You can use the code below to convert the original dataset into `webdataset` shards:
+
+```python
+import os
+import io
+import ray
+import webdataset as wds
+from datasets import Dataset
+from PIL import Image
+
+ray.init(num_cpus=8)
+
+
+def convert_to_image(im_bytes):
+    return Image.open(io.BytesIO(im_bytes)).convert("RGB")
+
+def main():
+    dataset_path = "/pickapic_v2/data"
+    wds_shards_path = "/pickapic_v2_webdataset"
+    # get all .parquet files in the dataset path
+    dataset_files = [
+        os.path.join(dataset_path, f)
+        for f in os.listdir(dataset_path)
+        if f.endswith(".parquet")
+    ]
+
+    @ray.remote
+    def create_shard(path):
+        # get basename of the file
+        basename = os.path.basename(path)
+        # get the shard number data-00123-of-01034.parquet -> 00123
+        shard_num = basename.split("-")[1]
+        dataset = Dataset.from_parquet(path)
+        # create a webdataset shard
+        shard = wds.TarWriter(os.path.join(wds_shards_path, f"{shard_num}.tar"))
+        
+        for i, example in enumerate(dataset):
+            wds_example = {
+                "__key__": str(i),
+                "original_prompt.txt": example["caption"],
+                "jpg_0.jpg": convert_to_image(example["jpg_0"]),
+                "jpg_1.jpg": convert_to_image(example["jpg_1"]),
+                "label_0.txt": str(example["label_0"]),
+                "label_1.txt": str(example["label_1"])
+            }
+            shard.write(wds_example)
+        shard.close()
+
+    futures = [create_shard.remote(path) for path in dataset_files]
+    ray.get(futures)
+
+
+if __name__ == "__main__":
+    main()
+```
+
+## Inference
+
+Refer to [sayakpaul/diffusion-sdxl-orpo](https://huggingface.co/sayakpaul/diffusion-sdxl-orpo) for an experimental checkpoint.

examples/research_projects/diffusion_orpo/requirements.txt

@@ -0,0 +1,7 @@
+datasets
+accelerate
+transformers
+torchvision
+wandb
+peft
+webdataset