more changes

README.md

@@ -54,19 +54,23 @@ python src/emma_perception/commands/run_server.py
 
 ### Extracting features
 
+For training things, we need to extract the features for each image.
 
-#### For the pretrained datasets
+Here's the command you can use to extract features from images. Obviously, you can change the paths to the folder of images, and the output dir, and whatever else you want.
 
 ```bash
+python src/emma_perception/commands/extract_visual_features.py --images_dir <path_to_images> --output_dir <path to output dir>
 ```
 
+<details>
+<summary>`argparse` arguments for the command</summary>
 
-#### For the Alexa Arena
 
+</details>
 
-```bash
-```
+#### Extracting features for the Alexa Arena
 
+If you want to use the fine-tuned model to extract features with the model we trained on the Alexa Arena, just add `--is_arena` onto the above command.
 
 ### Developer tooling
 

src/emma_perception/commands/download_checkpoints.py

@@ -19,7 +19,7 @@ def download_arena_checkpoint(
 
 
 def download_vinvl_checkpoint(
-    *, hf_repo_id: str = HF_REPO_ID, file_name: str = ARENA_CHECKPOINT_NAME
+    *, hf_repo_id: str = HF_REPO_ID, file_name: str = VINVL_CHECKPOINT_NAME
 ) -> Path:
     """Download the pre-trained VinVL checkpoint."""
     file_path = download_file(repo_id=hf_repo_id, repo_type="model", filename=file_name)

src/emma_perception/commands/extract_visual_features.py

@@ -1,59 +1,45 @@
 import argparse
-from typing import Union
+from ast import arg
 
 from maskrcnn_benchmark.config import cfg
 from pytorch_lightning import Trainer
 from scene_graph_benchmark.config import sg_cfg
 
 from emma_perception.callbacks.callbacks import VisualExtractionCacheCallback
-from emma_perception.datamodules.visual_extraction_dataset import (
-    ImageDataset,
-    PredictDataModule,
-    VideoFrameDataset,
+from emma_perception.commands.download_checkpoints import (
+    download_arena_checkpoint,
+    download_vinvl_checkpoint,
 )
+from emma_perception.datamodules.visual_extraction_dataset import ImageDataset, PredictDataModule
 from emma_perception.models.vinvl_extractor import VinVLExtractor, VinVLTransform
 
 
 def parse_args() -> argparse.Namespace:
     """Defines arguments."""
-    parser = argparse.ArgumentParser(prog="PROG")
-
+    parser = argparse.ArgumentParser()
     parser = Trainer.add_argparse_args(parser)  # type: ignore[assignment]
-    parser.add_argument("-i", "--input_path", required=True, help="Path to input dataset")
-    parser.add_argument("-b", "--batch_size", type=int, default=2)
-    parser.add_argument("-w", "--num_workers", type=int, default=0)
-    parser.add_argument("-cs", "--cache_suffix", default=".pt", help="Extension of cached files")
-    parser.add_argument("--config_file", metavar="FILE", help="path to VinVL config file")
-    parser.add_argument("--return_predictions", action="store_true")
-
-    parser.add_argument(
-        "--downsample",
-        type=int,
-        default=0,
-        help="Downsampling factor for videos. If 0 then no downsampling is performed",
-    )
-
     parser.add_argument(
-        "-c", "--cache", default="storage/data/cache", help="Path to store visual features"
+        "-i",
+        "--images_dir",
+        required=True,
+        help="Path to a folder of images to extract features from",
     )
-
     parser.add_argument(
-        "-d", "--dataset", required=True, choices=["images", "frames"], help="Dataset type"
+        "--is_arena",
+        action="store_true",
+        help="If we are extracting features from the Arena images, use the Arena checkpoint",
     )
+    parser.add_argument("-b", "--batch_size", type=int, default=2)
+    parser.add_argument("-w", "--num_workers", type=int, default=0)
     parser.add_argument(
-        "-a",
-        "--ann_csv",
-        help="Path to annotation csv file. Used for video datasets to select only the frames that have annotations",
+        "-c", "--output_dir", default="storage/data/cache", help="Path to store visual features"
     )
-
     parser.add_argument(
-        "-at",
-        "--ann_type",
-        choices=["epic_kitchens"],
-        default="epic_kitchens",
-        help="Annotation parser for video datasets",
+        "--num_gpus",
+        type=int,
+        default=None,
+        help="Number of GPUs to use for visual feature extraction",
     )
-
     parser.add_argument(
         "opts",
         default=None,
@@ -75,34 +61,23 @@ def main() -> None:
     cfg.merge_from_list(args.opts)
     cfg.freeze()
 
-    extractor = VinVLExtractor(cfg=cfg)
-    transform = VinVLTransform(cfg=cfg)
-
-    dataset: Union[ImageDataset, VideoFrameDataset]
-    if args.dataset == "images":
-        dataset = ImageDataset(input_path=args.input_path, preprocess_transform=transform)
-    elif args.dataset == "frames":
-        dataset = VideoFrameDataset(
-            input_path=args.input_path,
-            ann_csv=args.ann_csv,
-            ann_type=args.ann_type,
-            preprocess_transform=transform,
-            downsample=args.downsample,
-        )
+    if args.is_arena:
+        cfg.MODEL.WEIGHT = download_arena_checkpoint().as_posix()
     else:
-        raise OSError(f"Unsupported dataset type {args.dataset}")
+        cfg.MODEL.WEIGHT = download_vinvl_checkpoint().as_posix()
 
+    dataset = ImageDataset(
+        input_path=args.images_dir, preprocess_transform=VinVLTransform(cfg=cfg)
+    )
     dm = PredictDataModule(
         dataset=dataset, batch_size=args.batch_size, num_workers=args.num_workers
     )
+    extractor = VinVLExtractor(cfg=cfg)
     trainer = Trainer(
-        gpus=args.gpus,
-        callbacks=[
-            VisualExtractionCacheCallback(cache_dir=args.cache, cache_suffix=args.cache_suffix)
-        ],
-        profiler="advanced",
+        gpus=args.num_gpus,
+        callbacks=[VisualExtractionCacheCallback(cache_dir=args.output_dir, cache_suffix=".pt")],
     )
-    trainer.predict(extractor, dm, return_predictions=args.return_predictions)
+    trainer.predict(extractor, dm)
 
 
 if __name__ == "__main__":

src/emma_perception/constants/vinvl_x152c4.yaml

@@ -17,6 +17,7 @@ MODEL:
     SCORE_THRESH: 0.2 # 0.0001
     DETECTIONS_PER_IMG: 36 # 600
     MIN_DETECTIONS_PER_IMG: 10
+    NMS_FILTER: 1
   ROI_BOX_HEAD:
     NUM_CLASSES: 1595
   ROI_ATTRIBUTE_HEAD:
@@ -52,6 +53,7 @@ TEST:
   TSV_SAVE_SUBSET: ["rect", "class", "conf", "feature"]
   OUTPUT_FEATURE: True
   GATHER_ON_CPU: True
+  IGNORE_BOX_REGRESSION: False
 OUTPUT_DIR: "./output/X152C5_test"
 DATA_DIR: "./datasets"
 DISTRIBUTED_BACKEND: "gloo"