fix: bug fudan-generative-vision#124 (fudan-generative-vision#133)

add temp unet in log validation
add face emb validation in extract meta info function

scripts/extract_meta_info_stage1.py

@@ -21,6 +21,8 @@
 import os
 from pathlib import Path
 
+import torch
+
 
 def collect_video_folder_paths(root_path: Path) -> list:
     """
@@ -52,6 +54,10 @@ def construct_meta_info(frames_dir_path: Path) -> dict:
         print(f"Mask path not found: {mask_path}")
         return None
 
+    if torch.load(face_emb_path) is None:
+        print(f"Face emb is None: {face_emb_path}")
+        return None
+
     return {
         "image_path": str(frames_dir_path),
         "mask_path": mask_path,

scripts/train_stage1.py

@@ -16,6 +16,7 @@
 """
 
 import argparse
+import copy
 import logging
 import math
 import os
@@ -211,6 +212,7 @@ def log_validation(
     logger.info("Running validation... ")
 
     ori_net = accelerator.unwrap_model(net)
+    ori_net = copy.deepcopy(ori_net)
     reference_unet = ori_net.reference_unet
     denoising_unet = ori_net.denoising_unet
     face_locator = ori_net.face_locator
@@ -278,6 +280,7 @@ def log_validation(
         canvas.save(out_file)
 
     del pipe
+    del ori_net
     torch.cuda.empty_cache()
 
     return pil_images