Spaces:
Running
on
Zero
Running
on
Zero
File size: 4,103 Bytes
a858bb2 |
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 |
import os
import glob
from functools import partial
from tqdm import tqdm, trange
from multiprocessing import Pool
from PIL import Image
import cv2
import mlxu
from natsort import natsorted
import numpy as np
import einops
import torch
from vqlm_demo.inference import MultiProcessInferenceModel
from vqlm_demo.utils import (
is_video, random_square_crop,
read_frames_from_dir, read_frames_from_video
)
FLAGS, _ = mlxu.define_flags_with_default(
checkpoint='',
input_files='',
frame_input=False,
read_file_list='',
output_dir='',
center_crop=1.0,
n_context_frames=12,
n_new_frames=4,
n_candidates=8,
temperature=1.0,
top_p=1.0,
n_workers=8,
stride=8,
batch_size=32,
torch_devices='',
shuffle=False,
max_examples=0,
)
def save_image(args):
image, filename = args
base = FLAGS.input_files.split('*')[0]
filename = filename[len(base):].replace('/', '_') + '.png'
Image.fromarray(image).save(os.path.join(FLAGS.output_dir, filename))
class VideoDataset(torch.utils.data.Dataset):
def __init__(self, videos, frame_input=False, n_frames=8, stride=1):
self.videos = videos
self.frame_input = frame_input
self.n_frames = n_frames
self.stride = stride
def __getitem__(self, index):
if self.frame_input:
frames = read_frames_from_dir(
self.videos[index], self.n_frames, self.stride,
center_crop=FLAGS.center_crop,
)
else:
frames = read_frames_from_video(
self.videos[index], self.n_frames, self.stride,
center_crop=FLAGS.center_crop,
)
if frames is None:
return self[np.random.randint(0, len(self))]
return frames, self.videos[index]
def __len__(self):
return len(self.videos)
def main(_):
assert FLAGS.checkpoint != '' and FLAGS.output_dir != ''
assert FLAGS.read_file_list != '' or FLAGS.input_files != ''
os.makedirs(FLAGS.output_dir, exist_ok=True)
if FLAGS.read_file_list != '':
with open(FLAGS.read_file_list, 'r') as f:
videos = [x.strip() for x in f.readlines()]
else:
videos = glob.glob(FLAGS.input_files)
if FLAGS.frame_input:
videos = [x for x in videos if os.path.isdir(x)]
else:
videos = [x for x in videos if is_video(x)]
if FLAGS.shuffle:
np.random.shuffle(videos)
if FLAGS.max_examples > 0:
videos = videos[:FLAGS.max_examples]
dataset = VideoDataset(
videos,
frame_input=FLAGS.frame_input,
n_frames=FLAGS.n_context_frames,
stride=FLAGS.stride
)
dataloader = torch.utils.data.DataLoader(
dataset,
batch_size=FLAGS.batch_size,
shuffle=False,
num_workers=FLAGS.n_workers,
prefetch_factor=4,
drop_last=True,
)
if FLAGS.torch_devices == '':
torch_devices = None
else:
torch_devices = [f'cuda:{x}' for x in FLAGS.torch_devices.split(',')]
model = MultiProcessInferenceModel(
checkpoint=FLAGS.checkpoint, torch_devices=torch_devices,
)
save_img_pool = Pool(FLAGS.n_workers)
for batch, filenames in tqdm(dataloader, ncols=0):
batch = batch.numpy()
generated = model(
batch,
n_new_frames=FLAGS.n_new_frames,
n_candidates=FLAGS.n_candidates,
temperature=FLAGS.temperature,
top_p=FLAGS.top_p,
)
generated = np.array(generated)
output_batch = einops.repeat(
batch,
'b s h w c -> b n s h w c',
n=FLAGS.n_candidates,
)
combined = einops.rearrange(
np.concatenate([output_batch, generated], axis=2),
'b n s h w c -> b (n h) (s w) c'
)
combined = (np.clip(combined, 0, 1) * 255).astype(np.uint8)
save_img_pool.imap(save_image, zip(combined, filenames))
if __name__ == '__main__':
mlxu.run(main) |