update scoring

tools/scoring/README.md

@@ -1,8 +1,38 @@
-## Data Scoring and Filtering
+# Data Scoring and Filtering
+Important!!! All scoring jobs require these columns in meta files:
+- `path`: absolute path to a sample
 
-### Aesthetic Score
+## Aesthetic Score
+First prepare the environment and pretrained models.
+```bash
+# install clip
+pip install git+https://github.com/openai/CLIP.git
+pip install decord
 
-### Optical Flow Score
-`python tools/scoring/optical_flow/inference.py --meta_path ./data/Panda-70M/processed/meta/test_intact_cut_head-100.csv`
+# get pretrained model
+wget https://github.com/christophschuhmann/improved-aesthetic-predictor/raw/main/sac+logos+ava1-l14-linearMSE.pth -O pretrained_models/aesthetic.pth
+```
 
-### Matching Score
+Then run:
+```bash
+# output: DATA_aes.csv
+python -m tools.scoring.aesthetic.inference /path/to/meta.csv
+```
+The output should be `/path/to/meta_aes.csv` with column `aes`. Aesthetic scores range from 1 to 10, with 10 being the best quality.
+
+## Optical Flow Score
+First get the pretrained model.
+```bash
+wget https://s3.eu-central-1.amazonaws.com/avg-projects/unimatch/pretrained/gmflow-scale2-regrefine6-mixdata-train320x576-4e7b215d.pth -P pretrained_models/unimatch
+```
+
+Then run:
+```
+python tools/scoring/optical_flow/inference.py /path/to/meta.csv
+```
+The output should be `/path/to/meta_flow.csv` with column `flow`. Higher optical flow scores indicate larger movement.
+
+## Matching Score
+Require column `text` in meta files, which is the caption of the sample.
+
+TODO.

tools/scoring/aesthetic/inference.py

@@ -116,7 +116,7 @@ def main(args):
     )
 
     # compute aesthetic scores
-    dataset.data["aesthetic_score"] = np.nan
+    dataset.data["aes"] = np.nan
     index = 0
     for batch in tqdm(dataloader):
         images = batch["image"].to(device)
@@ -127,10 +127,10 @@ def main(args):
         scores = rearrange(scores, "(b p) 1 -> b p", b=B)
         scores = scores.mean(dim=1)
         scores_np = scores.cpu().numpy()
-        dataset.data.loc[index : index + len(scores_np) - 1, "aesthetic_score"] = scores_np
+        dataset.data.loc[index : index + len(scores_np) - 1, "aes"] = scores_np
         index += len(images)
     dataset.data.to_csv(output_file, index=False)
-    print(f"Saved aesthetic scores to {output_file}.")
+    print(f"New meta with aesthetic scores saved to \'{output_file}\'.")
 
 
 if __name__ == "__main__":

tools/scoring/optical_flow/inference.py

@@ -71,7 +71,7 @@ class VideoTextDataset(torch.utils.data.Dataset):
 
 def main():
     parser = argparse.ArgumentParser()
-    parser.add_argument("--meta_path", type=str, help="Path to the input CSV file")
+    parser.add_argument("meta_path", type=str, help="Path to the input CSV file")
     parser.add_argument("--bs", type=int, default=4, help="Batch size")
     parser.add_argument("--num_workers", type=int, default=16, help="Number of workers")
     args = parser.parse_args()
@@ -92,10 +92,10 @@ def main():
         reg_refine=True,
         task='flow',
     )
-    # ckpt = torch.load(
-    #     './checkpoints/pretrained_models/unimatch/gmflow-scale2-regrefine6-mixdata-train320x576-4e7b215d.pth'
-    # )
-    # model.load_state_dict(ckpt['model'])
+    ckpt = torch.load(
+        './pretrained_models/unimatch/gmflow-scale2-regrefine6-mixdata-train320x576-4e7b215d.pth'
+    )
+    model.load_state_dict(ckpt['model'])
     model = model.to(device)
     model = torch.nn.DataParallel(model)