File size: 6,652 Bytes
5b9e601
 
 
 
 
 
 
 
55d0a5e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5b9e601
55d0a5e
 
 
5b9e601
55d0a5e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5b9e601
55d0a5e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5b9e601
55d0a5e
 
 
 
 
 
 
 
5b9e601
 
55d0a5e
 
5b9e601
55d0a5e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
import io
import json
import numpy as np
import torch
import webdataset as wds
from PIL import Image
from concurrent.futures import ThreadPoolExecutor
import os
from huggingface_hub import HfApi

# 🚀 Speed up HF transfers
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"

class CARLAVideoLoader:
    def __init__(
        self,
        hf_repo="mkxdxd/carla-dataset",
        town="Town01",
        actor_type="pedestrian",
        frames_per_scene=93,
        batch_size=1,
        num_workers=4,
        decode_workers=32,
        shuffle=True
    ):
        """
        Custom video loader for CARLA Stage2 dataset in WebDataset format.
        Optimized for 1-tar-N-scenes format with parallel decoding.
        """
        self.hf_repo = hf_repo
        self.town = town
        self.actor_type = actor_type
        self.frames_per_scene = frames_per_scene
        self.batch_size = batch_size
        self.num_workers = num_workers
        self.decode_workers = decode_workers
        self.shuffle = shuffle
        
        # 1. Get tar URLs from HuggingFace Hub
        api = HfApi()
        all_files = api.list_repo_files(repo_id=hf_repo, repo_type="dataset")
        
        self.tar_urls = []
        prefix = f"{town}/{actor_type}/"
        for f in all_files:
            if f.startswith(prefix) and f.endswith(".tar"):
                self.tar_urls.append(f"https://huggingface.co/datasets/{hf_repo}/resolve/main/{f}")
        
        if not self.tar_urls:
            raise ValueError(f"No tar files found for {town}/{actor_type} in {hf_repo}")
        
        print(f"✅ Found {len(self.tar_urls)} shards for {town}/{actor_type}")
        
        # 2. Build WebDataset pipeline
        self.dataset = (
            wds.WebDataset(self.tar_urls, shardshuffle=self.shuffle, handler=wds.warn_and_continue)
            .to_tuple("__key__", "rgb.png", "depth.npy", "camera.json", "metadata.json", handler=wds.warn_and_continue)
            .map(lambda x: (
                x[0],
                {
                    "rgb.png": x[1],
                    "depth.npy": x[2],
                    "camera.json": x[3],
                    "metadata.json": x[4],
                }
            ), handler=wds.warn_and_continue)
            .compose(self._frames_to_video_sequence)
        )
        
        # 3. Create DataLoader
        self.dataloader = torch.utils.data.DataLoader(
            self.dataset,
            batch_size=self.batch_size,
            num_workers=self.num_workers,
            prefetch_factor=2,
            persistent_workers=True,
            pin_memory=True,
            # Since our compose function yields dictionaries, we handle collate carefully
            # or just yield them one by one if batch_size=1
            collate_fn=self._collate_fn
        )

    def _decode_frame(self, sample):
        """Decodes raw bytes into tensors."""
        # RGB: png -> PIL -> Tensor (3, H, W)
        rgb = Image.open(io.BytesIO(sample["rgb.png"])).convert("RGB")
        rgb = torch.from_numpy(np.array(rgb)).permute(2, 0, 1)

        # Depth: npy -> Tensor (H, W)
        depth = np.load(io.BytesIO(sample["depth.npy"]))
        depth = torch.from_numpy(depth)

        # JSONs
        camera = json.loads(sample["camera.json"])
        metadata = json.loads(sample["metadata.json"])

        return rgb, depth, camera, metadata

    def _frames_to_video_sequence(self, it):
        """WebDataset 'compose' function to group individual frames into full videos."""
        buffer = []
        
        for key, sample in it:
            buffer.append(sample)
            
            if len(buffer) == self.frames_per_scene:
                # 🚀 Parallel decoding for high performance
                with ThreadPoolExecutor(max_workers=self.decode_workers) as executor:
                    decoded = list(executor.map(self._decode_frame, buffer))
                
                # Unpack and stack
                rgbs = torch.stack([d[0] for d in decoded])      # (T, 3, H, W)
                depths = torch.stack([d[1] for d in decoded])    # (T, H, W)
                cameras = [d[2] for d in decoded]
                metas = [d[3] for d in decoded]
                
                # key example: Town01/pedestrian/scene_0001_000
                # base_key will be scene_0001
                base_key = key.rsplit("_", 1)[0]
                
                yield {
                    "video": rgbs,          # (93, 3, 704, 1280)
                    "depth": depths,        # (93, 704, 1280)
                    "camera": cameras,      # list of 93 dicts
                    "metadata": metas,      # list of 93 dicts
                    "scene_id": base_key
                }
                buffer = []

    def _collate_fn(self, batch):
        """
        Handle batching. 
        If batch_size > 1, tensors are stacked, metadata remains as lists.
        """
        if len(batch) == 0:
            return {}
        
        res = {
            "video": torch.stack([b["video"] for b in batch]),     # (B, T, 3, H, W)
            "depth": torch.stack([b["depth"] for b in batch]),     # (B, T, H, W)
            "scene_id": [b["scene_id"] for b in batch],
            "camera": [b["camera"] for b in batch],                # (B, T, dict)
            "metadata": [b["metadata"] for b in batch],            # (B, T, dict)
        }
        return res

    def __iter__(self):
        return iter(self.dataloader)

# --- Example Usage ---
if __name__ == "__main__":
    import time

    print("🚀 Initializing CARLA Video Loader...")
    loader = CARLAVideoLoader(
        hf_repo="mkxdxd/carla-dataset",
        town="Town01",
        actor_type="pedestrian",
        batch_size=2,      # Loading 2 videos at once
        num_workers=4,     # 4 DataLoader workers
        decode_workers=16  # 16 threads for PNG decoding per scene
    )

    print("🎬 Starting stream...")
    start_time = time.time()
    
    for i, batch in enumerate(loader):
        print(f"\n📦 Batch {i+1}")
        print(f"   - Video Shape: {batch['video'].shape}") # (B, 93, 3, 704, 1280)
        print(f"   - Depth Shape: {batch['depth'].shape}") # (B, 93, 704, 1280)
        print(f"   - Scenes: {batch['scene_id']}")
        
        # Access first frame of first video in batch
        first_frame_meta = batch['metadata'][0][0]
        print(f"   - Meta (B0, T0): {first_frame_meta['town']}, frame {first_frame_meta['frame_id']}")
        
        if i >= 2: # Just test a few batches
            break
            
    total_time = time.time() - start_time
    print(f"\n✅ Test finished in {total_time:.2f}s")