File size: 5,740 Bytes
9e14838 | 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 | import os
import cv2
import sys
import yaml
import json
import math
import torch
import pickle
import shutil
import logging
import warnings
import argparse
import numpy as np
from os import path
from datetime import datetime
from torchvision.io import VideoReader
from src.utility.builtin import ODTrainer, ODLightningCLI
def parse_args(args=None):
parser = argparse.ArgumentParser()
parser.add_argument("model_cfg_path", type=str)
parser.add_argument("model_ckpt_path", type=str)
parser.add_argument("video_path", type=str)
parser.add_argument("--out_path", type=str, default=None)
parser.add_argument("--threshold", type=float, default=0.5)
parser.add_argument("--precision", type=str, default="16")
parser.add_argument("--batch_size", type=int, default=30)
return parser.parse_args(args=args)
def configure_logging():
logging_fmt = "[%(levelname)s][%(filename)s:%(lineno)d]: %(message)s"
logging.basicConfig(level="INFO", format=logging_fmt)
warnings.filterwarnings(action="ignore")
@torch.inference_mode()
def demo_driver(cli, ckpt_path, video_path, out_path, batch_size, threshold):
# setup model
model = cli.model
try:
model = model.__class__.load_from_checkpoint(ckpt_path)
except Exception as e:
print(f"Unable to load model from checkpoint in strict mode: {e}")
print(f"Loading model from checkpoint in non-strict mode.")
model = model.__class__.load_from_checkpoint(ckpt_path, strict=False)
model.eval()
transforms = model.transform
BATCH = batch_size
stride = 0.333
# load original video
vid_reader = VideoReader(video_path, "video", num_threads=1)
vid_ext = os.path.splitext(video_path)[-1]
vid_name = os.path.split(video_path)[1].replace(vid_ext, "")
fps = vid_reader.get_metadata()["video"]["fps"][0]
frames = []
for frame_data in vid_reader:
frames.append(frame_data["data"])
frames = torch.stack(frames)
del vid_reader
_, H, W = frames[0].shape
# load bboxes of original video
with open(video_path.replace("videos", "frame_data").replace(vid_ext, ".pickle"), "rb") as f:
fdata = pickle.load(f)
bboxes = []
for data in fdata:
data["bboxes"] = [
bbox.reshape(2, -1)
if len(bbox.shape) == 1 else bbox
for bbox in data["bboxes"]
]
face_idx = np.argsort([
np.linalg.norm((bbox[0] - bbox[1])) for bbox in data["bboxes"]
])[-1]
bboxes.append(data["bboxes"][face_idx])
# load face cropped video
vid_reader = VideoReader(
video_path.replace("/videos", "/cropped/videos").replace(vid_ext, ".avi"),
"video",
num_threads=1
)
cropped_frames = []
for frame_data in vid_reader:
cropped_frames.append(frame_data["data"])
cropped_frames = torch.stack(cropped_frames)
del vid_reader
# sample frames and inference
indices = torch.tensor([int(math.floor(i * stride * fps)) for i in range(10)], dtype=torch.long)
probs = []
i = 0
clip_count = len(cropped_frames) - indices[-1]
while (i < clip_count):
batch = min(clip_count - i, BATCH)
clips = torch.stack([
transforms(cropped_frames[indices + i + j]) for j in range(batch)
]).to("cuda")
results = model.evaluate(clips)
probs.extend(results["logits"].softmax(dim=-1)[:, 1].flatten().cpu().tolist())
i += batch
# draw and write to video
bbox_frames = []
for frame, bbox, prob in zip(frames[indices[-1]:], bboxes[indices[-1]:], probs):
frame = frame.permute(1, 2, 0).numpy()
frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
thickness = int(np.linalg.norm(bbox[0] - bbox[1]) * 0.01)
color = (0, 255, 0) if prob < threshold else (0, 0, 255)
category = "REAL" if prob < threshold else "FAKE"
frame = cv2.rectangle(
frame,
bbox[0].astype(int),
bbox[1].astype(int),
color,
thickness
)
frame = cv2.putText(
frame,
f'{round(prob,2)}',
[int(bbox[0][0]), int(bbox[1][1] - thickness)],
cv2.FONT_HERSHEY_SIMPLEX,
1, color, thickness, cv2.LINE_AA
)
frame = cv2.putText(
frame,
category,
[int(bbox[0][0]), int(bbox[0][1] - thickness)],
cv2.FONT_HERSHEY_SIMPLEX,
1, color, thickness, cv2.LINE_AA
)
bbox_frames.append(frame)
out_path = (f'pred_{vid_name}.avi' if out_path is None else out_path)
writer = cv2.VideoWriter(
out_path,
cv2.VideoWriter_fourcc('X', 'V', 'I', 'D'),
fps,
(W, H)
)
for frame in bbox_frames:
writer.write(frame)
writer.release()
if __name__ == "__main__":
configure_logging()
params = parse_args()
cli = ODLightningCLI(
run=False,
trainer_class=ODTrainer,
save_config_callback=None,
parser_kwargs={
"parser_mode": "omegaconf"
},
auto_configure_optimizers=False,
seed_everything_default=1019,
args=[
'-c', params.model_cfg_path,
'--trainer.logger=null',
f'--trainer.devices=1',
f'--trainer.precision={params.precision}',
],
)
ckpt_path = params.model_ckpt_path
video_path = params.video_path
demo_driver(
cli=cli,
ckpt_path=ckpt_path,
video_path=video_path,
batch_size=params.batch_size,
threshold=params.threshold,
out_path=params.out_path
)
|