updates

training/mochi-1/prepare_dataset.py

@@ -13,7 +13,7 @@ from typing import Any, Dict, List, Optional, Union
 
 import torch
 import torch.distributed as dist
-from diffusers import AutoencoderKL
+from diffusers import AutoencoderKLMochi
 from diffusers.training_utils import set_seed
 from diffusers.utils import export_to_video, get_logger
 from torch.utils.data import DataLoader
@@ -24,7 +24,9 @@ from transformers import T5EncoderModel, T5Tokenizer
 
 import decord  # isort:skip
 
-from ..dataset import BucketSampler, VideoDatasetWithResizing, VideoDatasetWithResizeAndRectangleCrop  # isort:skip
+import sys
+sys.path.append("..")
+from dataset import BucketSampler, VideoDatasetWithResizing, VideoDatasetWithResizeAndRectangleCrop  # isort:skip
 
 
 decord.bridge.set_bridge("torch")
@@ -485,12 +487,12 @@ def main():
         tokenizer = T5Tokenizer.from_pretrained(args.model_id, subfolder="tokenizer")
         text_encoder = T5EncoderModel.from_pretrained(
             args.model_id, subfolder="text_encoder", torch_dtype=weight_dtype
-        )
-        text_encoder = text_encoder.to(device)
-
-        vae = AutoencoderKL.from_pretrained(args.model_id, subfolder="vae", torch_dtype=weight_dtype)
-        vae = vae.to(device)
-
+        ).to(device)
+        
+        vae = AutoencoderKLMochi.from_pretrained(
+            args.model_id, subfolder="vae", torch_dtype=weight_dtype
+        ).to(device)
+        
         if args.use_slicing:
             vae.enable_slicing()
         if args.use_tiling:

training/mochi-1/prepare_dataset.sh

@@ -0,0 +1,48 @@
+#!/bin/bash
+
+MODEL_ID="genmo/mochi-1-preview"
+
+NUM_GPUS=1
+
+# For more details on the expected data format, please refer to the README.
+DATA_ROOT="/home/sayak/cogvideox-factory/video-dataset-disney"  # This needs to be the path to the base directory where your videos are located.
+CAPTION_COLUMN="prompt.txt"
+VIDEO_COLUMN="videos.txt"
+OUTPUT_DIR="/home/sayak/cogvideox-factory/video-dataset-disney/mochi-1/preprocessed-dataset"
+HEIGHT_BUCKETS="480"
+WIDTH_BUCKETS="848"
+FRAME_BUCKETS="84"
+MAX_NUM_FRAMES="84"
+MAX_SEQUENCE_LENGTH=256
+TARGET_FPS=30
+BATCH_SIZE=1
+DTYPE=fp32
+
+# To create a folder-style dataset structure without pre-encoding videos and captions
+# For Image-to-Video finetuning, make sure to pass `--save_image_latents`
+CMD_WITHOUT_PRE_ENCODING="\
+  torchrun --nproc_per_node=$NUM_GPUS \
+    prepare_dataset.py \
+      --model_id $MODEL_ID \
+      --data_root $DATA_ROOT \
+      --caption_column $CAPTION_COLUMN \
+      --video_column $VIDEO_COLUMN \
+      --output_dir $OUTPUT_DIR \
+      --height_buckets $HEIGHT_BUCKETS \
+      --width_buckets $WIDTH_BUCKETS \
+      --frame_buckets $FRAME_BUCKETS \
+      --max_num_frames $MAX_NUM_FRAMES \
+      --max_sequence_length $MAX_SEQUENCE_LENGTH \
+      --target_fps $TARGET_FPS \
+      --batch_size $BATCH_SIZE \
+      --dtype $DTYPE
+"
+
+CMD_WITH_PRE_ENCODING="$CMD_WITHOUT_PRE_ENCODING --save_latents_and_embeddings"
+
+# Select which you'd like to run
+CMD=$CMD_WITH_PRE_ENCODING
+
+echo "===== Running \`$CMD\` ====="
+eval $CMD
+echo -ne "===== Finished running script =====\n"