Spaces:
Running on Zero
Running on Zero
File size: 11,052 Bytes
796e051 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 | import os, sys, shutil
import random
from typing import List, Union, Optional, Tuple
from pathlib import Path
import numpy as np
import ffmpeg
import torch
import torch.utils.data
import imageio
import torchvision
import torchvision.transforms.functional as F
from torchvision.transforms import InterpolationMode
from torch.utils.data import Dataset
from PIL import Image
PILImage = Image.Image
def _to_4d_video_tensor(frames_np):
"""
Inputs: [N, H, W, 3] in range [0, 255]
Returns: float tensor [N, C, H, W] in [0, 1]
"""
# numpy -> torch, NHWC -> NCHW
frames = torch.from_numpy(frames_np.copy())
# Divide 255 to the range [0, 1]
frames = frames.float().div_(255.0)
# Reshape to [T,C,H,W]
frames = frames.permute(0, 3, 1, 2).contiguous() # [T,C,H,W]
# Return
return frames
def save_video_mp4(
video: torch.Tensor,
path: str = "video.mp4",
fps: int = 24,
mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225),
assume_normalized: bool | None = None, # None=自动判断;True/False=强制
):
"""
Save torch video tensor to mp4.
Accepts:
- [T, C, H, W] float tensor
- [B, T, C, H, W] float tensor (will use first sample)
It will auto unnormalize (ImageNet) if it detects normalized inputs.
Output:
- mp4 with uint8 frames [T, H, W, 3]
"""
if not torch.is_tensor(video):
raise TypeError(f"video must be torch.Tensor, got {type(video)}")
v = video.detach().cpu()
# Handle [B, T, C, H, W]
if v.ndim == 5:
v = v[0]
if v.ndim != 4:
raise ValueError(f"Expected [T,C,H,W] (or [B,T,C,H,W]), got shape {tuple(v.shape)}")
T, C, H, W = v.shape
if C not in (1, 3):
raise ValueError(f"Expected C=1 or 3, got C={C}")
v = v.to(torch.float32)
# ---- Decide whether to unnormalize ----
# Heuristic: normalized ImageNet tensors often have values outside [0,1]
# (e.g., negative or >1). Raw unnormalized typically stays in [0,1].
if assume_normalized is None:
minv = float(v.min())
maxv = float(v.max())
is_normalized = (minv < -0.05) or (maxv > 1.05)
else:
is_normalized = bool(assume_normalized)
if is_normalized:
mean_t = torch.tensor(mean, dtype=v.dtype).view(1, 3, 1, 1)
std_t = torch.tensor(std, dtype=v.dtype).view(1, 3, 1, 1)
if C == 1:
v = v.repeat(1, 3, 1, 1)
C = 3
v = v * std_t + mean_t # unnormalize back to roughly [0,1]
# Clamp to valid range and convert to uint8
v = v.clamp(0.0, 1.0)
v = (v * 255.0).round().to(torch.uint8)
# [T, C, H, W] -> [T, H, W, C]
v = v.permute(0, 2, 3, 1).contiguous().numpy()
imageio.mimsave(path, v, fps=fps)
# print("Save video at", path)
class Video_Augmentation_Transform:
"""
Clip-wise (video-constant) augmentation.
All randomness sampled ONCE per call, applied identically to every frame.
"""
def __init__(
self,
set_type: str = "train",
horizontal_flip_prob: float = 0.5,
vertical_flip_prob: float = 0.1,
jitter_prob: float = 0.5,
jitter_param: Tuple[float, float, float, float] = (0.2, 0.2, 0.2, 0.05), # b,c,s,h
grayscale_prob: float = 0.1,
# ---- add blur ----
blur_prob: float = 0.1,
blur_kernel_size: int = 8,
blur_sigma: Tuple[float, float] = (0.1, 2.0),
# ---- add noise ----
noise_prob: float = 0.2,
noise_sigma: Tuple[float, float] = (0.003, 0.01), # for [0,1] images
noise_clip: Tuple[float, float] = (0.0, 1.0),
# ---- add compression ----
compression_prob = 0.0,
compression_choices = ["jpeg", "webp"],
# ---- Basic Normlization ----
normalize_mean: Tuple[float, float, float] = (0.485, 0.456, 0.406),
normalize_std: Tuple[float, float, float] = (0.229, 0.224, 0.225),
):
self.set_type = set_type
self.horizontal_flip_prob = horizontal_flip_prob
self.vertical_flip_prob = vertical_flip_prob
self.jitter_prob = jitter_prob
self.jitter_param = jitter_param
self.grayscale_prob = grayscale_prob
self.blur_prob = blur_prob
self.blur_kernel_size = blur_kernel_size
self.blur_sigma = blur_sigma
self.noise_prob = noise_prob
self.noise_sigma = noise_sigma
self.noise_clip = noise_clip
self.compression_prob = compression_prob
self.compression_choices = compression_choices
self.mean = normalize_mean
self.std = normalize_std
@staticmethod
def _sample_color_jitter_params(b, c, s, h):
# Sample factors similar to torchvision ColorJitter
def _rand_factor(x):
if x is None or x == 0:
return None
lo = max(0.0, 1.0 - x)
hi = 1.0 + x
return random.uniform(lo, hi)
brightness_factor = _rand_factor(b)
contrast_factor = _rand_factor(c)
saturation_factor = _rand_factor(s)
hue_factor = None
if h is not None and h != 0:
hue_factor = random.uniform(-h, h)
# Randomize application order (same as torchvision)
order = ["brightness", "contrast", "saturation", "hue"]
random.shuffle(order)
return {
"brightness": brightness_factor,
"contrast": contrast_factor,
"saturation": saturation_factor,
"hue": hue_factor,
"order": order,
}
@staticmethod
def _fix_blur_kernel(k: int) -> int:
# torchvision GaussianBlur typically expects odd kernel
k = int(k)
if k <= 0:
raise ValueError("blur_kernel_size must be > 0")
if k % 2 == 0:
k = k + 1
return k
@staticmethod
def _add_gaussian_noise(img: torch.Tensor, sigma: float, clip_min: float = 0.0, clip_max: float = 1.0) -> torch.Tensor:
"""
img: [C,H,W] float, assumed in [0,1] (or at least bounded)
sigma: std of Gaussian noise in same scale as img
"""
if sigma <= 0:
return img
noise = torch.randn_like(img) * float(sigma)
img = img + noise
return img.clamp_(clip_min, clip_max)
def _add_compression(self, img, compression_choice):
if compression_choice == "jpeg":
from datasets.compression_utils import jpeg_compress_tensor
# compress
jpeg_compress_tensor(img)
compressed_img = jpeg_compress_tensor(img)
elif compression_choice == "webp":
from datasets.compression_utils import webp_compress_tensor
# compress
compressed_img = webp_compress_tensor(img)
else:
raise NotImplementedError("We do not support comrpession type of", compression_choice)
return compressed_img
def __call__(self, frames: Union[List[PILImage], torch.Tensor], idx = 0) -> torch.Tensor:
"""
Returns normalized float tensor [T, C, H, W]
"""
# Convert to Tesnor
x = _to_4d_video_tensor(frames) # [T, C, H, W] in float
Tt, C, H, W = x.shape
# Decide the prob
if self.set_type == "train":
## Flip
do_horizontal_flip = random.random() < self.horizontal_flip_prob
do_vertical_flip = random.random() < self.vertical_flip_prob
## Color Jitter
do_jitter = random.random() < self.jitter_prob
b, c, s, h = self.jitter_param
jitter_params = self._sample_color_jitter_params(b, c, s, h)
## GaryScale
do_gray = random.random() < self.grayscale_prob
## Blur
do_blur = random.random() < self.blur_prob
blur_kernel = self._fix_blur_kernel(self.blur_kernel_size)
blur_sigma = random.uniform(self.blur_sigma[0], self.blur_sigma[1])
## Gaussian Noise
do_noise = random.random() < self.noise_prob
noise_sigma = random.uniform(self.noise_sigma[0], self.noise_sigma[1]) if do_noise else 0.0
clip_min, clip_max = self.noise_clip
## Compression
do_compression = random.random() < self.compression_prob
compression_choice = random.choice(self.compression_choices)
else: # For Testing, No Augmentation at all
do_horizontal_flip = False
do_vertical_flip = False
do_jitter = False
do_gray = False
do_blur = False
do_noise = False
do_compression = False
# Apply per-frame with shared params
out = []
for t in range(Tt): # Iterate each Frame
img = x[t] # torch shape is [C, H, W]
## Horizontal Flipping
if do_horizontal_flip:
img = F.hflip(img)
if do_vertical_flip:
img = F.vflip(img)
## Color Jitter (shared factors + shared order)
if do_jitter:
for op in jitter_params["order"]:
if op == "brightness" and jitter_params["brightness"] is not None:
img = F.adjust_brightness(img, jitter_params["brightness"])
elif op == "contrast" and jitter_params["contrast"] is not None:
img = F.adjust_contrast(img, jitter_params["contrast"])
elif op == "saturation" and jitter_params["saturation"] is not None:
img = F.adjust_saturation(img, jitter_params["saturation"])
elif op == "hue" and jitter_params["hue"] is not None:
img = F.adjust_hue(img, jitter_params["hue"])
## Gray Scale
if do_gray:
img = F.rgb_to_grayscale(img, num_output_channels=3)
# Gaussian Blur
if do_blur:
img = F.gaussian_blur(img, kernel_size=[blur_kernel, blur_kernel], sigma=[blur_sigma, blur_sigma])
# Gaussian Noise (small, clip-wise sigma)
if do_noise:
img = self._add_gaussian_noise(img, sigma=noise_sigma, clip_min=clip_min, clip_max=clip_max)
# Image compression
if do_compression:
img = self._add_compression(img, compression_choice)
# ImageNet normalize per-frame (Must Do)
img = F.normalize(img, mean = self.mean, std = self.std)
# Append
out.append(img)
# Stack
out = torch.stack(out, dim=0) # Output Shape is [T, C, H, W]
# Save the video (Comment Out Later)
# write_path = "augmented"+str(idx)+".mp4"
# save_video_mp4(out, write_path)
# Return
return out
|