Spaces:
Build error
Build error
File size: 6,557 Bytes
b6af722 |
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 |
# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Run this command to interactively debug:
PYTHONPATH=. python cosmos_predict1/tokenizer/training/datasets/video_dataset.py
Adapted from:
https://github.com/bytedance/IRASim/blob/main/dataset/dataset_3D.py
"""
import traceback
import warnings
from concurrent.futures import ThreadPoolExecutor, as_completed
from glob import glob
import numpy as np
import torch
from decord import VideoReader, cpu
from torch.utils.data import Dataset
from torchvision import transforms as T
from tqdm import tqdm
from cosmos_predict1.diffusion.training.datasets.dataset_utils import ToTensorVideo
class Dataset(Dataset):
def __init__(
self,
video_pattern,
sequence_interval=1,
start_frame_interval=1,
num_video_frames=25,
):
"""Dataset class for loading image-text-to-video generation data.
Args:
video_pattern (str): path/to/videos/*.mp4
sequence_interval (int): Interval between sampled frames in a sequence
num_frames (int): Number of frames to load per sequence
video_size (list): Target size [H,W] for video frames
Returns dict with:
- video: RGB frames tensor [T,C,H,W]
- video_name: Dict with episode/frame metadata
"""
super().__init__()
self.video_directory_or_pattern = video_pattern
self.start_frame_interval = start_frame_interval
self.sequence_interval = sequence_interval
self.sequence_length = num_video_frames
self.video_paths = sorted(glob(str(video_pattern)))
print(f"{len(self.video_paths)} videos in total")
self.samples = self._init_samples(self.video_paths)
self.samples = sorted(self.samples, key=lambda x: (x["video_path"], x["frame_ids"][0]))
print(f"{len(self.samples)} samples in total")
self.wrong_number = 0
self.preprocess = T.Compose(
[
ToTensorVideo(),
]
)
def __str__(self):
return f"{len(self.video_paths)} samples from {self.video_directory_or_pattern}"
def _init_samples(self, video_paths):
samples = []
with ThreadPoolExecutor(32) as executor:
future_to_video_path = {
executor.submit(self._load_and_process_video_path, video_path): video_path for video_path in video_paths
}
for future in tqdm(as_completed(future_to_video_path), total=len(video_paths)):
samples.extend(future.result())
return samples
def _load_and_process_video_path(self, video_path):
vr = VideoReader(video_path, ctx=cpu(0), num_threads=2)
n_frames = len(vr)
samples = []
for frame_i in range(0, n_frames, self.start_frame_interval):
sample = dict()
sample["video_path"] = video_path
sample["frame_ids"] = []
curr_frame_i = frame_i
while True:
if curr_frame_i > (n_frames - 1):
break
sample["frame_ids"].append(curr_frame_i)
if len(sample["frame_ids"]) == self.sequence_length:
break
curr_frame_i += self.sequence_interval
# make sure there are sequence_length number of frames
if len(sample["frame_ids"]) == self.sequence_length:
samples.append(sample)
return samples
def __len__(self):
return len(self.samples)
def _load_video(self, video_path, frame_ids):
vr = VideoReader(video_path, ctx=cpu(0), num_threads=2)
assert (np.array(frame_ids) < len(vr)).all()
assert (np.array(frame_ids) >= 0).all()
vr.seek(0)
frame_data = vr.get_batch(frame_ids).asnumpy()
return frame_data
def _get_frames(self, video_path, frame_ids):
frames = self._load_video(video_path, frame_ids)
frames = frames.astype(np.uint8)
frames = torch.from_numpy(frames).permute(0, 3, 1, 2) # (l, c, h, data)
frames = self.preprocess(frames)
frames = torch.clamp(frames * 255.0, 0, 255).to(torch.uint8)
return frames
def __getitem__(self, index):
try:
sample = self.samples[index]
video_path = sample["video_path"]
frame_ids = sample["frame_ids"]
data = dict()
video = self._get_frames(video_path, frame_ids)
video = video.permute(1, 0, 2, 3) # Rearrange from [T, C, H, W] to [C, T, H, W]
data["video"] = video
data["video_name"] = {
"video_path": video_path,
"start_frame_id": str(frame_ids[0]),
}
data["fps"] = 24
data["image_size"] = torch.tensor([704, 1280, 704, 1280]) # .cuda() # TODO: Does this matter?
data["num_frames"] = self.sequence_length
data["padding_mask"] = torch.zeros(1, 704, 1280) # .cuda()
return data
except Exception:
warnings.warn(
f"Invalid data encountered: {self.samples[index]['video_path']}. Skipped "
f"(by randomly sampling another sample in the same dataset)."
)
warnings.warn("FULL TRACEBACK:")
warnings.warn(traceback.format_exc())
self.wrong_number += 1
print(self.wrong_number)
return self[np.random.randint(len(self.samples))]
if __name__ == "__main__":
dataset = Dataset(
video_directory_or_pattern="assets/example_training_data/videos/*.mp4",
sequence_interval=1,
num_frames=57,
video_size=[240, 360],
)
indices = [0, 13, 200, -1]
for idx in indices:
data = dataset[idx]
print((f"{idx=} " f"{data['video'].sum()=}\n" f"{data['video'].shape=}\n" f"{data['video_name']=}\n" "---"))
|