ReLaX-VQA / src /feature_fragment_pool.py
Xinyi Wang
first commit
211b431
raw
history blame
16.6 kB
from PIL import Image
import pandas as pd
import numpy as np
import os
from pathlib import Path
import scipy.io
import shutil
import torch
import time
import cv2
from torchvision import models, transforms
from utils.logger_setup import logger
from extractor import visualise_vgg_layer, visualise_resnet_layer, visualise_vit_layer, vf_extract
def load_metadata(video_type):
print(f'video_type: {video_type}\n')
# Test
if video_type == 'test':
return pd.read_csv("../metadata/test_videos.csv")
# NR:
elif video_type == 'resolution_ugc':
resolution = '360P'
return pd.read_csv(f"../metadata/YOUTUBE_UGC_{resolution}_metadata.csv")
else:
return pd.read_csv(f'../metadata/{video_type.upper()}_metadata.csv')
def get_video_paths(network_name, video_type, videodata, i):
video_name = videodata['vid'][i]
video_width = videodata['width'][i]
video_height = videodata['height'][i]
pixfmt = videodata['pixfmt'][i]
framerate = videodata['framerate'][i]
common_path = os.path.join('..', 'video_sampled_frame')
# Test
if video_type == 'test':
video_path = f"../ugc_original_videos/{video_name}.mp4"
# NR:
elif video_type == 'konvid_1k':
video_path = Path("D:/video_dataset/KoNViD_1k/KoNViD_1k_videos") / f"{video_name}.mp4"
elif video_type == 'lsvq_train' or video_type == 'lsvq_test' or video_type == 'lsvq_test_1080P':
print(f'video_name: {video_name}')
video_path = Path("D:/video_dataset/LSVQ") / f"{video_name}.mp4"
print(f'video_path: {video_path}')
video_name = os.path.splitext(os.path.basename(video_path))[0]
elif video_type == 'live_vqc':
video_path = Path("D:/video_dataset/LIVE-VQC/video") / f"{video_name}.mp4"
elif video_type == 'live_qualcomm':
video_path = Path("D:/video_dataset/LIVE-Qualcomm") / f"{video_name}.yuv"
video_name = os.path.splitext(os.path.basename(video_path))[0]
elif video_type == 'cvd_2014':
video_path = Path("D:/video_dataset/CVD2014") / f"{video_name}.avi"
video_name = os.path.splitext(os.path.basename(video_path))[0]
elif video_type == 'youtube_ugc':
video_path = Path("D:/video_dataset/ugc-dataset/youtube_ugc/") / f"{video_name}.mkv"
video_name = os.path.splitext(os.path.basename(video_path))[0]
sampled_frame_path = os.path.join(common_path, f'fragment_layerstack', f'video_{str(i + 1)}')
feature_name = f"{network_name}_feature_map"
if video_type == 'resolution_ugc':
resolution = '360P'
# video_path = f'/user/work/um20242/dataset/ugc-dataset/{resolution}/{video_name}.mkv'
video_path = Path(f"D:/video_dataset/ugc-dataset/youtube_ugc/original_videos/{resolution}") / f"{video_name}.mkv"
sampled_frame_path = os.path.join(common_path, f'ytugc_sampled_frame_{resolution}', f'video_{str(i + 1)}')
feature_name = f"{network_name}_feature_map_{resolution}"
return video_name, video_path, sampled_frame_path, feature_name, video_width, video_height, pixfmt, framerate
def get_deep_feature(network_name, video_name, frame, frame_number, model, device, layer_name):
png_path = f'../visualisation/{network_name}_{layer_name}/{video_name}/'
os.makedirs(png_path, exist_ok=True)
if network_name == 'resnet50':
if layer_name == 'pool':
visual_layer = 'resnet50.avgpool' # before avg_pool
resnet50 = model
activations_dict, _, total_flops, total_params = visualise_resnet_layer.process_video_frame(video_name, frame, frame_number, visual_layer, resnet50, device)
elif network_name == 'vgg16':
if layer_name == 'pool':
# visual_layer = 'fc1'
visual_layer = 'fc2' # fc1 = vgg16.classifier[0], fc2 = vgg16.classifier[3]
vgg16 = model
activations_dict, _, total_flops, total_params = visualise_vgg_layer.process_video_frame(video_name, frame, frame_number, visual_layer, vgg16, device)
elif network_name == 'vit':
patch_size = 16
activations_dict, _, total_flops, total_params = visualise_vit_layer.process_video_frame(video_name, frame, frame_number, model, patch_size, device)
return png_path, activations_dict, total_flops, total_params
def process_video_feature(video_feature, network_name, layer_name):
# print(f'video frame number: {len(video_feature)}')
# initialize an empty list to store processed frames
averaged_frames = []
# iterate through each frame in the video_feature
for frame in video_feature:
frame_features = []
if layer_name == 'pool':
if network_name == 'vit':
# global mean and std
global_mean = torch.mean(frame, dim=0)
global_max = torch.max(frame, dim=0)[0]
global_std = torch.std(frame, dim=0)
# concatenate all pooling
combined_features = torch.hstack([global_mean, global_max, global_std])
frame_features.append(combined_features)
elif network_name == 'resnet50':
frame = torch.squeeze(torch.tensor(frame))
# global mean and std
global_mean = torch.mean(frame, dim=0)
global_max = torch.max(frame, dim=0)[0]
global_std = torch.std(frame, dim=0)
# concatenate all pooling
combined_features = torch.hstack([frame, global_mean, global_max, global_std])
frame_features.append(combined_features)
# concatenate the layer means horizontally to form the processed frame
processed_frame = torch.hstack(frame_features)
averaged_frames.append(processed_frame)
averaged_frames = torch.stack(averaged_frames)
# output the shape of the resulting feature vector
logger.debug(f"Shape of feature vector after global pooling: {averaged_frames.shape}")
return averaged_frames
def flow_to_rgb(flow):
mag, ang = cv2.cartToPolar(flow[..., 0], flow[..., 1])
mag = cv2.normalize(mag, None, 0, 255, cv2.NORM_MINMAX)
# convert angle to hue
hue = ang * 180 / np.pi / 2
# create HSV
hsv = np.zeros((flow.shape[0], flow.shape[1], 3), dtype=np.uint8)
hsv[..., 0] = hue
hsv[..., 1] = 255
hsv[..., 2] = cv2.normalize(mag, None, 0, 255, cv2.NORM_MINMAX)
# convert HSV to RGB
rgb = cv2.cvtColor(hsv, cv2.COLOR_HSV2BGR)
return rgb
def get_patch_diff(residual_frame, patch_size):
h, w = residual_frame.shape[2:] # Assuming (1, C, H, W) shape
h_adj = (h // patch_size) * patch_size
w_adj = (w // patch_size) * patch_size
residual_frame_adj = residual_frame[:, :, :h_adj, :w_adj]
# calculate absolute patch difference
diff = torch.zeros((h_adj // patch_size, w_adj // patch_size), device=residual_frame.device)
for i in range(0, h_adj, patch_size):
for j in range(0, w_adj, patch_size):
patch = residual_frame_adj[:, :, i:i + patch_size, j:j + patch_size]
# absolute sum
diff[i // patch_size, j // patch_size] = torch.sum(torch.abs(patch))
return diff
def extract_important_patches(residual_frame, diff, patch_size=16, target_size=224, top_n=196):
# find top n patches indices
patch_idx = torch.argsort(-diff.view(-1))
top_patches = [(idx // diff.shape[1], idx % diff.shape[1]) for idx in patch_idx[:top_n]]
sorted_idx = sorted(top_patches, key=lambda x: (x[0], x[1]))
imp_patches_img = torch.zeros((residual_frame.shape[1], target_size, target_size), dtype=residual_frame.dtype, device=residual_frame.device)
patches_per_row = target_size // patch_size # 14
# order the patch in the original location relation
positions = []
for idx, (y, x) in enumerate(sorted_idx):
patch = residual_frame[:, :, y * patch_size:(y + 1) * patch_size, x * patch_size:(x + 1) * patch_size]
# new patch location
row_idx = idx // patches_per_row
col_idx = idx % patches_per_row
start_y = row_idx * patch_size
start_x = col_idx * patch_size
imp_patches_img[:, start_y:start_y + patch_size, start_x:start_x + patch_size] = patch
positions.append((y.item(), x.item()))
return imp_patches_img, positions
def get_frame_patches(frame, positions, patch_size, target_size):
imp_patches_img = torch.zeros((frame.shape[1], target_size, target_size), dtype=frame.dtype, device=frame.device)
patches_per_row = target_size // patch_size
for idx, (y, x) in enumerate(positions):
start_y = y * patch_size
start_x = x * patch_size
end_y = start_y + patch_size
end_x = start_x + patch_size
patch = frame[:, :, start_y:end_y, start_x:end_x]
row_idx = idx // patches_per_row
col_idx = idx % patches_per_row
target_start_y = row_idx * patch_size
target_start_x = col_idx * patch_size
imp_patches_img[:, target_start_y:target_start_y + patch_size,
target_start_x:target_start_x + patch_size] = patch.squeeze(0)
return imp_patches_img
def process_patches(original_path, frag_name, residual, patch_size, target_size, top_n):
diff = get_patch_diff(residual, patch_size)
imp_patches, positions = extract_important_patches(residual, diff, patch_size, target_size, top_n)
if frag_name == 'frame_diff':
frag_path = original_path.replace('.png', '_residual_imp.png')
elif frag_name == 'optical_flow':
frag_path = original_path.replace('.png', '_residual_of_imp.png')
# cv2.imwrite(frag_path, imp_patches)
return frag_path, imp_patches, positions
# Frame Differencing
def compute_frame_difference(frame_tensor, frame_next_tensor, frame_path, patch_size, target_size, top_n):
residual = torch.abs(frame_next_tensor - frame_tensor)
return process_patches(frame_path, 'frame_diff', residual, patch_size, target_size, top_n)
# Optical Flow
def compute_optical_flow(frame, frame_next, frame_path, patch_size, target_size, top_n, device):
flow = cv2.calcOpticalFlowFarneback(cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY),
cv2.cvtColor(frame_next, cv2.COLOR_BGR2GRAY),
None, 0.5, 3, 15, 3, 5, 1.2, 0)
opticalflow_rgb = flow_to_rgb(flow)
opticalflow_rgb_tensor = transforms.ToTensor()(opticalflow_rgb).unsqueeze(0).to(device)
return process_patches(frame_path, 'optical_flow', opticalflow_rgb_tensor, patch_size, target_size, top_n)
def merge_fragments(diff_fragment, flow_fragment):
alpha = 0.5
merged_fragment = diff_fragment * alpha + flow_fragment * (1 - alpha)
return merged_fragment
def concatenate_features(frame_feature, residual_feature):
return torch.cat((frame_feature, residual_feature), dim=-1)
if __name__ == '__main__':
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
if device.type == "cuda":
torch.cuda.set_device(0)
# device = torch.device("cpu")
video_type = 'test' # test
# resolution_ugc/konvid_1k/live_vqc/cvd_2014/live_qualcomm
# lsvq_train/lsvq_test/lsvq_test_1080P/
frag_name = 'framediff_frag' # framediff_frag, opticalflow_frag, sampled_frag, merged_frag
network_name = 'vit'
layer_name = 'pool'
if network_name == 'vit':
model = visualise_vit_layer.VitGenerator('vit_base', 16, device, evaluate=True, random=False, verbose=True)
elif network_name == 'resnet50':
model = models.resnet50(pretrained=True).to(device)
else:
model = models.vgg16(pretrained=True).to(device)
logger.info(f"video type: {video_type}, frag name: {frag_name}, network name: {network_name}, layer name: {layer_name}")
logger.info(f"torch cuda: {torch.cuda.is_available()}")
videodata = load_metadata(video_type)
valid_video_types = ['test',
'resolution_ugc', 'konvid_1k', 'live_vqc', 'cvd_2014', 'live_qualcomm',
'lsvq_train', 'lsvq_test', 'lsvq_test_1080P']
target_size = 224
patch_size = 16
top_n = int((target_size / patch_size) * (target_size / patch_size))
begin_time = time.time()
if video_type in valid_video_types:
for i in range(len(videodata)):
start_time = time.time()
video_name, video_path, sampled_frame_path, feature_name, video_width, video_height, pixfmt, framerate = get_video_paths(network_name, video_type, videodata, i)
frames, frames_next = vf_extract.process_video_residual(video_type, video_name, framerate, video_path, sampled_frame_path)
logger.info(f'{video_name}')
all_frame_activations_feats = []
for j, (frame, frame_next) in enumerate(zip(frames, frames_next)):
frame_number = j + 1
frame_path = os.path.join(sampled_frame_path, f'{video_name}_{frame_number}.png')
# compute residual
frame_tensor = transforms.ToTensor()(frame).unsqueeze(0).to(device)
frame_next_tensor = transforms.ToTensor()(frame_next).unsqueeze(0).to(device)
# DNN feature extraction
if frag_name in ['framediff_frag', 'sampled_frag', 'merged_frag']:
residual_frag_path, diff_frag, positions = compute_frame_difference(frame_tensor, frame_next_tensor, frame_path, patch_size, target_size, top_n)
png_path, frag_activations, total_flops, total_params = get_deep_feature(network_name, video_name, diff_frag, frame_number, model, device, layer_name)
if frag_name == 'sampled_frag':
frame_patches = get_frame_patches(frame_tensor, positions, patch_size, target_size)
png_path, frag_activations, total_flops, total_params = get_deep_feature(network_name, video_name, frame_patches, frame_number, model, device, layer_name)
elif frag_name == 'merged_frag':
of_frag_path, flow_frag, _ = compute_optical_flow(frame, frame_next, frame_path, patch_size, target_size, top_n, device)
merged_frag = merge_fragments(diff_frag, flow_frag)
png_path, frag_activations, total_flops, total_params = get_deep_feature(network_name, video_name, merged_frag, frame_number, model, device, layer_name)
elif frag_name == 'opticalflow_frag':
of_frag_path, flow_frag, _ = compute_optical_flow(frame, frame_next, frame_path, patch_size, target_size, top_n, device)
png_path, frag_activations, total_flops, total_params = get_deep_feature(network_name, video_name, flow_frag, frame_number, model, device, layer_name)
# feature combined
all_frame_activations_feats.append(frag_activations)
averaged_frames_feats = process_video_feature(all_frame_activations_feats, network_name, layer_name)
print("Features shape:", averaged_frames_feats.shape)
# remove tmp folders
shutil.rmtree(png_path)
shutil.rmtree(sampled_frame_path)
averaged_npy = averaged_frames_feats.cpu().numpy()
# save the processed data as numpy file
output_npy_path = f'../features/{video_type}/{frag_name}_{network_name}_{layer_name}/'
os.makedirs(output_npy_path, exist_ok=True)
# output_npy_name = f'{output_npy_path}video_{str(i + 1)}_{feature_name}.npy'
# np.save(output_npy_name, averaged_npy)
# print(f'Processed file saved to: {output_npy_name}')
run_time = time.time() - start_time
print(f"Execution time for {video_name} feature extraction: {run_time:.4f} seconds\n")
# save feature mat file
average_data = np.mean(averaged_npy, axis=0)
if i == 0:
feats_matrix = np.zeros((len(videodata),) + average_data.shape)
feats_matrix[i] = average_data
print((f'All features shape: {feats_matrix.shape}'))
logger.debug(f'\n All features shape: {feats_matrix.shape}')
mat_file_path = f'../features/{video_type}/'
mat_file_name = f'{mat_file_path}{video_type}_{frag_name}_{network_name}_{layer_name}_feats.mat'
scipy.io.savemat(mat_file_name, {video_type: feats_matrix})
logger.debug(f'Successfully created {mat_file_name}')
logger.debug(f"Execution time for all feature extraction: {time.time() - begin_time:.4f} seconds\n")