[Training] Refactor and improve validation datasets (#539)

Author: SolitaryThinker

examples/training/finetune/wan_i2v_14b_480p/crush_smol/validation.json

@@ -0,0 +1,31 @@
+{
+  "data": [
+    {
+      "caption": "A large metal cylinder is seen pressing down on a pile of Oreo cookies, flattening them as if they were under a hydraulic press.",
+      "image_path": null,
+      "video_path": "examples/training/finetune/wan_i2v_14b_480p/crush_smol/validation_dataset/yYcK4nANZz4-Scene-034.mp4",
+      "num_inference_steps": 50,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "A large metal cylinder is seen compressing colorful clay into a compact shape, demonstrating the power of a hydraulic press.",
+      "image_path": null,
+      "video_path": "examples/training/finetune/wan_i2v_14b_480p/crush_smol/validation_dataset/yYcK4nANZz4-Scene-027.mp4",
+      "num_inference_steps": 50,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    },
+    {
+      "caption": "A large metal cylinder is seen pressing down on a pile of colorful candies, flattening them as if they were under a hydraulic press. The candies are crushed and broken into small pieces, creating a mess on the table.",
+      "image_path": null,
+      "video_path": "examples/training/finetune/wan_i2v_14b_480p/crush_smol/validation_dataset/yYcK4nANZz4-Scene-030.mp4",
+      "num_inference_steps": 50,
+      "height": 480,
+      "width": 832,
+      "num_frames": 77
+    }
+  ]
+}

examples/training/finetune/wan_i2v_14b_480p/crush_smol/validation_dataset/yYcK4nANZz4-Scene-027.mp4

@@ -1,19 +1,17 @@
-import os
-
+# SPDX-License-Identifier: Apache-2.0
 from torchvision import transforms
 from torchvision.transforms import Lambda
-from transformers import AutoTokenizer
 
-from fastvideo.v1.dataset.t2v_datasets import T2V_dataset
+from fastvideo.v1.dataset.parquet_dataset_map_style import (
+    build_parquet_map_style_dataloader)
+from fastvideo.v1.dataset.preprocessing_datasets import (
+    VideoCaptionMergedDataset)
 from fastvideo.v1.dataset.transform import (CenterCropResizeVideo, Normalize255,
                                             TemporalRandomCrop)
-
-from .parquet_dataset_map_style import build_parquet_map_style_dataloader
-
-__all__ = ["build_parquet_map_style_dataloader"]
+from fastvideo.v1.dataset.validation_dataset import ValidationDataset
 
 
-def getdataset(args, start_idx=0) -> T2V_dataset:
+def getdataset(args) -> VideoCaptionMergedDataset:
     temporal_sample = TemporalRandomCrop(args.num_frames)  # 16 x
     norm_fun = Lambda(lambda x: 2.0 * x - 1.0)
     resize_topcrop = [
@@ -31,15 +29,17 @@ def getdataset(args, start_idx=0) -> T2V_dataset:
         *resize_topcrop,
         norm_fun,
     ])
-    tokenizer_path = os.path.join(args.model_path, "tokenizer")
-    tokenizer = AutoTokenizer.from_pretrained(tokenizer_path,
-                                              cache_dir=args.cache_dir)
     if args.dataset == "t2v":
-        return T2V_dataset(args,
-                           transform=transform,
-                           temporal_sample=temporal_sample,
-                           tokenizer=tokenizer,
-                           transform_topcrop=transform_topcrop,
-                           start_idx=start_idx)
+        return VideoCaptionMergedDataset(data_merge_path=args.data_merge_path,
+                                         args=args,
+                                         transform=transform,
+                                         temporal_sample=temporal_sample,
+                                         transform_topcrop=transform_topcrop)
 
     raise NotImplementedError(args.dataset)
+
+
+__all__ = [
+    "build_parquet_map_style_dataloader", "ValidationDataset",
+    "VideoCaptionMergedDataset"
+]

examples/training/finetune/wan_i2v_14b_480p/crush_smol/validation_dataset/yYcK4nANZz4-Scene-030.mp4

@@ -48,6 +48,48 @@
     pa.field("fps", pa.float64()),
 ])
 
+pyarrow_schema_i2v_validation = pa.schema([
+    pa.field("id", pa.string()),
+    # --- Image/Video VAE latents ---
+    # Tensors are stored as raw bytes with shape and dtype info for loading
+    pa.field("vae_latent_bytes", pa.binary()),
+    # e.g., [C, T, H, W] or [C, H, W]
+    pa.field("vae_latent_shape", pa.list_(pa.int64())),
+    # e.g., 'float32'
+    pa.field("vae_latent_dtype", pa.string()),
+    # --- Text encoder output tensor ---
+    # Tensors are stored as raw bytes with shape and dtype info for loading
+    pa.field("text_embedding_bytes", pa.binary()),
+    # e.g., [SeqLen, Dim]
+    pa.field("text_embedding_shape", pa.list_(pa.int64())),
+    # e.g., 'bfloat16' or 'float32'
+    pa.field("text_embedding_dtype", pa.string()),
+    pa.field("text_attention_mask_bytes", pa.binary()),
+    # e.g., [SeqLen]
+    pa.field("text_attention_mask_shape", pa.list_(pa.int64())),
+    # e.g., 'bool' or 'int8'
+    pa.field("text_attention_mask_dtype", pa.string()),
+    #I2V
+    pa.field("clip_feature_bytes", pa.binary()),
+    pa.field("clip_feature_shape", pa.list_(pa.int64())),
+    pa.field("clip_feature_dtype", pa.string()),
+    # I2V Validation
+    pa.field("pil_image_bytes", pa.binary()),
+    pa.field("pil_image_shape", pa.list_(pa.int64())),
+    pa.field("pil_image_dtype", pa.string()),
+    # --- Metadata ---
+    pa.field("file_name", pa.string()),
+    pa.field("caption", pa.string()),
+    pa.field("media_type", pa.string()),  # 'image' or 'video'
+    pa.field("width", pa.int64()),
+    pa.field("height", pa.int64()),
+    # -- Video-specific (can be null/default for images) ---
+    # Number of frames processed (e.g., 1 for image, N for video)
+    pa.field("num_frames", pa.int64()),
+    pa.field("duration_sec", pa.float64()),
+    pa.field("fps", pa.float64()),
+])
+
 pyarrow_schema_t2v = pa.schema([
     pa.field("id", pa.string()),
     # --- Image/Video VAE latents ---
@@ -80,4 +122,38 @@
     pa.field("num_frames", pa.int64()),
     pa.field("duration_sec", pa.float64()),
     pa.field("fps", pa.float64()),
-])
+])
+
+pyarrow_schema_t2v_validation = pa.schema([
+    pa.field("id", pa.string()),
+    # --- Image/Video VAE latents ---
+    # Tensors are stored as raw bytes with shape and dtype info for loading
+    pa.field("vae_latent_bytes", pa.binary()),
+    # e.g., [C, T, H, W] or [C, H, W]
+    pa.field("vae_latent_shape", pa.list_(pa.int64())),
+    # e.g., 'float32'
+    pa.field("vae_latent_dtype", pa.string()),
+    # --- Text encoder output tensor ---
+    # Tensors are stored as raw bytes with shape and dtype info for loading
+    pa.field("text_embedding_bytes", pa.binary()),
+    # e.g., [SeqLen, Dim]
+    pa.field("text_embedding_shape", pa.list_(pa.int64())),
+    # e.g., 'bfloat16' or 'float32'
+    pa.field("text_embedding_dtype", pa.string()),
+    pa.field("text_attention_mask_bytes", pa.binary()),
+    # e.g., [SeqLen]
+    pa.field("text_attention_mask_shape", pa.list_(pa.int64())),
+    # e.g., 'bool' or 'int8'
+    pa.field("text_attention_mask_dtype", pa.string()),
+    # --- Metadata ---
+    pa.field("file_name", pa.string()),
+    pa.field("caption", pa.string()),
+    pa.field("media_type", pa.string()),  # 'image' or 'video'
+    pa.field("width", pa.int64()),
+    pa.field("height", pa.int64()),
+    # -- Video-specific (can be null/default for images) ---
+    # Number of frames processed (e.g., 1 for image, N for video)
+    pa.field("num_frames", pa.int64()),
+    pa.field("duration_sec", pa.float64()),
+    pa.field("fps", pa.float64()),
+])