implemented RWTH Phoenix data loading script..?

rwth_phoenix_weather_2014.py

@@ -1,12 +1,5 @@
-import queue
-from concurrent import futures
-from functools import wraps
-from typing import Generator
-
-import cv2
 import datasets
-import numpy as np
-import scipy.io
+import pandas as pd
 
 _CITATION = """\
 @article{koller2015continuous,
@@ -55,6 +48,7 @@ class RWTHPhoenixWeather2014Config(datasets.BuilderConfig):
         super(RWTHPhoenixWeather2014Config, self).__init__(version=datasets.Version("1.0.0"), **kwargs)
         self.main_data_folder = main_data_folder
 
+
 class RWTHPhoenixWeather2014(datasets.GeneratorBasedBuilder):
     """RWTH-PHOENIX-Weather 2014: Continuous Sign Language Recognition Dataset."""
 
@@ -79,8 +73,7 @@ class RWTHPhoenixWeather2014(datasets.GeneratorBasedBuilder):
             description=_DESCRIPTION + self.config.description,
             features=datasets.Features(
                 {
-                    "label": datasets.ClassLabel(names=_LABELS),
-                    "frames_interval": datasets.Sequence(feature=datasets.Value("uint32"), length=2),
+                    "tokens": datasets.Sequence(feature=str),
                     "frames": datasets.Sequence(feature=datasets.Image()),
                 }
             ),
@@ -92,8 +85,9 @@ class RWTHPhoenixWeather2014(datasets.GeneratorBasedBuilder):
         )
 
     def _split_generators(self, dl_manager: datasets.DownloadManager):
-        videos = {}
-        data = {}
+        split_frames = {}
+        split_tokens = {}
+        split_ids = {}
 
         dataDirMapper = {
             datasets.Split.TRAIN: "train",
@@ -106,31 +100,34 @@ class RWTHPhoenixWeather2014(datasets.GeneratorBasedBuilder):
             datasets.Split.VALIDATION,
             datasets.Split.TEST,
         ]:
-            base_urls = [
-                f"data/{self.config.main_data_folder}/features/fullFrame-210x260px/{dataDirMapper[split]}"
-                for idx in split_ids[split]
-            ]
+            base_url = f"data/{self.config.main_data_folder}"
 
-            video_urls = [
-                f"{base_url}_color.mp4"
-                for base_url in base_urls
-            ]
+            data_csv = dl_manager.download(f"{base_url}/annotations/manual/{dataDirMapper[split]}.corpus.csv")
+            df = pd.read_csv(data_csv, sep='|')
 
-            data_urls = [
-                f"{base_url}_data.mat"
-                for base_url in base_urls
-            ]
+            split_tokens[split] = df.annotation.map(
+                lambda x: x.strip().split()
+            )
+
+            split_ids[split] = df.id
+
+            frame_archive_urls = dl_manager.download([
+                f"{base_url}/features/fullFrame-210x260px/{dataDirMapper[split]}/{id}.tar"
+                for id in df.id
+            ])
 
-            videos[split] = dl_manager.download(video_urls)
-            data[split] = dl_manager.download(data_urls)
+            split_frames[split] = [
+                dl_manager.iter_archive(url)
+                for url in frame_archive_urls
+            ]
 
         return [
             datasets.SplitGenerator(
                 name=split,
                 gen_kwargs={
-                    "ids": split_ids[split],
-                    "video_files": videos[split],
-                    "data_files": data[split],
+                    "split_ids": split_ids[split],
+                    "split_frames": split_frames[split],
+                    "split_tokens": split_tokens[split],
                 },
             )
             for split in [
@@ -140,94 +137,9 @@ class RWTHPhoenixWeather2014(datasets.GeneratorBasedBuilder):
             ]
         ]
 
-    def _generate_examples(self, ids, video_files, data_files):
-        with futures.ThreadPoolExecutor() as executor:
-            video_processor = VideoProcessor(executor)
-
-            for idx, video_file, data_file in zip(ids, video_files, data_files):
-                mat = scipy.io.loadmat(data_file)
-                data = mat['Video'][0, 0][4][0]
-
-                frame_queue = video_processor.open(video_file)
-                frames = []
-
-                last_end_frame = 0
-                sub_sample_idx = 0
-                curr_frame_idx = 0
-
-                while True:
-                    try:
-                        [label], [[start_frame]], [[end_frame]] = data[sub_sample_idx]
-                    except Exception as e:
-                        print(f"\nskipping this example: something weird happened at idx = {idx}, sub_sample_idx = {sub_sample_idx}: {e}")
-                        break
-
-                    image = next(frame_queue, None)
-
-                    if image is None:
-                        # no more images
-                        break
-
-                    # last read was successful
-                    curr_frame_idx += 1
-
-                    if last_end_frame > start_frame:
-                        # sanity check
-                        raise RuntimeError(
-                            f"example frames are not monotonically increasing: last_end_frame = {last_end_frame} > start_frame = {start_frame}")
-
-                    if curr_frame_idx > end_frame:
-                        # sanity check
-                        raise RuntimeError(f"count = {curr_frame_idx} was greater than end_frame = {start_frame}")
-
-                    if curr_frame_idx >= start_frame:
-                        # save image if we are at the right frame index
-                        frames.append(image)
-
-                    if curr_frame_idx == end_frame:
-                        # we found our end frame and can yield a result
-                        yield f"Sample{idx}/{sub_sample_idx}", {
-                            "label": label,
-                            "frames_interval": (start_frame, end_frame),
-                            "frames": frames,
-                        }
-
-                        frames = []
-                        last_end_frame = end_frame
-                        sub_sample_idx += 1
-
-                        if sub_sample_idx == len(data):
-                            # no more samples to generate
-                            break
-
-
-class VideoProcessor:
-    def __init__(self, executor: futures.Executor):
-        super().__init__()
-        self.executor = executor
-        self.feature = datasets.Image()
-
-    def open(self, video_file) -> Generator[dict, None, None]:
-        video_capture = cv2.VideoCapture(video_file)
-        futures_queue = queue.Queue[futures.Future[dict]](maxsize=30)
-
-        self.executor.submit(self._push_frames_to_queue, video_capture, futures_queue)
-
-        while True:
-            value = futures_queue.get()
-
-            if value is None:
-                break
-
-            yield value.result()
-
-    def _push_frames_to_queue(self, capture: cv2.VideoCapture, futures_queue: queue.Queue):
-        ok = True
-        while ok:
-            ok, image = capture.read()
-
-            if not ok:
-                futures_queue.put(None)
-            else:
-                future = self.executor.submit(self.feature.encode_example, image)
-                futures_queue.put(future)
+    def _generate_examples(self, split_ids, split_frames, split_tokens):
+        for id, frames, tokens in zip(split_ids, split_frames, split_tokens):
+            yield id, {
+                "tokens": tokens,
+                "frames": frames,
+            }