File size: 8,102 Bytes
b14067d |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 |
import os
import random
import json
import torch
import torch.nn as nn
import torchvision.transforms as transforms
import torchvision.transforms.functional as F
import numpy as np
from decord import VideoReader
from torch.utils.data.dataset import Dataset
from packaging import version as pver
class RandomHorizontalFlipWithPose(nn.Module):
def __init__(self, p=0.5):
super(RandomHorizontalFlipWithPose, self).__init__()
self.p = p
def get_flip_flag(self, n_image):
return torch.rand(n_image) < self.p
def forward(self, image, flip_flag=None):
n_image = image.shape[0]
if flip_flag is not None:
assert n_image == flip_flag.shape[0]
else:
flip_flag = self.get_flip_flag(n_image)
ret_images = []
for fflag, img in zip(flip_flag, image):
if fflag:
ret_images.append(F.hflip(img))
else:
ret_images.append(img)
return torch.stack(ret_images, dim=0)
class RealEstate10KPCDRenderDataset(Dataset):
def __init__(
self,
video_root_dir,
sample_n_frames=49,
image_size=[480, 720],
shuffle_frames=False,
hflip_p=0.0,
):
if hflip_p != 0.0:
use_flip = True
else:
use_flip = False
root_path = video_root_dir
self.root_path = root_path
self.sample_n_frames = sample_n_frames
self.source_video_root = os.path.join(self.root_path, 'videos')
self.mask_video_root = os.path.join(self.root_path, 'masked_videos')
self.captions_root = os.path.join(self.root_path, 'captions')
self.dataset = sorted([n.replace('.mp4','') for n in os.listdir(self.source_video_root)])
self.length = len(self.dataset)
sample_size = image_size
sample_size = tuple(sample_size) if not isinstance(sample_size, int) else (sample_size, sample_size)
self.sample_size = sample_size
if use_flip:
pixel_transforms = [transforms.Resize(sample_size),
RandomHorizontalFlipWithPose(hflip_p),
transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True)]
else:
pixel_transforms = [transforms.Resize(sample_size),
transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True)]
self.sample_wh_ratio = sample_size[1] / sample_size[0]
self.pixel_transforms = pixel_transforms
self.shuffle_frames = shuffle_frames
self.use_flip = use_flip
def load_video_reader(self, idx):
clip_name = self.dataset[idx]
video_path = os.path.join(self.source_video_root, clip_name + '.mp4')
video_reader = VideoReader(video_path)
mask_video_path = os.path.join(self.mask_video_root, clip_name + '.mp4')
mask_video_reader = VideoReader(mask_video_path)
caption_path = os.path.join(self.captions_root, clip_name + '.txt')
if os.path.exists(caption_path):
caption = open(caption_path, 'r').read().strip()
else:
caption = ''
return clip_name, video_reader, mask_video_reader, caption
def get_batch(self, idx):
clip_name, video_reader, mask_video_reader, video_caption = self.load_video_reader(idx)
if self.use_flip:
flip_flag = self.pixel_transforms[1].get_flip_flag(self.sample_n_frames)
else:
flip_flag = torch.zeros(self.sample_n_frames, dtype=torch.bool)
indices = np.arange(self.sample_n_frames)
pixel_values = torch.from_numpy(video_reader.get_batch(indices).asnumpy()).permute(0, 3, 1, 2).contiguous()
pixel_values = pixel_values / 255.
anchor_pixels = torch.from_numpy(mask_video_reader.get_batch(indices).asnumpy()).permute(0, 3, 1, 2).contiguous()
anchor_pixels = anchor_pixels / 255.
return pixel_values, anchor_pixels, video_caption, flip_flag, clip_name
def __len__(self):
return self.length
def __getitem__(self, idx):
while True:
try:
video, anchor_video, video_caption, flip_flag, clip_name = self.get_batch(idx)
break
except Exception as e:
idx = random.randint(0, self.length - 1)
if self.use_flip:
video = self.pixel_transforms[0](video)
video = self.pixel_transforms[1](video, flip_flag)
video = self.pixel_transforms[2](video)
anchor_video = self.pixel_transforms[0](anchor_video)
anchor_video = self.pixel_transforms[1](anchor_video, flip_flag)
anchor_video = self.pixel_transforms[2](anchor_video)
else:
for transform in self.pixel_transforms:
video = transform(video)
anchor_video = transform(anchor_video)
data = {
'video': video,
'anchor_video': anchor_video,
'caption': video_caption,
}
return data
class RealEstate10KPCDRenderCapEmbDataset(RealEstate10KPCDRenderDataset):
def __init__(
self,
video_root_dir,
text_embedding_path,
sample_n_frames=49,
image_size=[480, 720],
shuffle_frames=False,
hflip_p=0.0,
):
super().__init__(
video_root_dir,
sample_n_frames=sample_n_frames,
image_size=image_size,
shuffle_frames=shuffle_frames,
hflip_p=hflip_p,
)
self.text_embedding_path = text_embedding_path
self.mask_root = os.path.join(self.root_path, 'masks')
def get_batch(self, idx):
clip_name, video_reader, mask_video_reader, video_caption = self.load_video_reader(idx)
cap_emb_path = os.path.join(self.text_embedding_path, clip_name + '.pt')
video_caption_emb = torch.load(cap_emb_path, weights_only=True)
if self.use_flip:
flip_flag = self.pixel_transforms[1].get_flip_flag(self.sample_n_frames)
else:
flip_flag = torch.zeros(self.sample_n_frames, dtype=torch.bool)
indices = np.arange(self.sample_n_frames)
pixel_values = torch.from_numpy(video_reader.get_batch(indices).asnumpy()).permute(0, 3, 1, 2).contiguous()
pixel_values = pixel_values / 255.
anchor_pixels = torch.from_numpy(mask_video_reader.get_batch(indices).asnumpy()).permute(0, 3, 1, 2).contiguous()
anchor_pixels = anchor_pixels / 255.
try:
masks = np.load(os.path.join(self.mask_root, clip_name + '.npz'))['mask']*1.0
masks = torch.from_numpy(masks).unsqueeze(1)
except:
threshold = 0.1 # you can adjust this value
masks = (anchor_pixels.sum(dim=1, keepdim=True) < threshold).float()
return pixel_values, anchor_pixels, masks, video_caption_emb, flip_flag, clip_name
def __getitem__(self, idx):
while True:
try:
video, anchor_video, mask, video_caption_emb, flip_flag, clip_name = self.get_batch(idx)
break
except Exception as e:
idx = random.randint(0, self.length - 1)
if self.use_flip:
video = self.pixel_transforms[0](video)
video = self.pixel_transforms[1](video, flip_flag)
video = self.pixel_transforms[2](video)
anchor_video = self.pixel_transforms[0](anchor_video)
anchor_video = self.pixel_transforms[1](anchor_video, flip_flag)
anchor_video = self.pixel_transforms[2](anchor_video)
else:
for transform in self.pixel_transforms:
video = transform(video)
anchor_video = transform(anchor_video)
data = {
'video': video,
'anchor_video': anchor_video,
'caption_emb': video_caption_emb,
'mask': mask
}
return data |