import av
|
import os
|
import pims
|
import numpy as np
|
from torch.utils.data import Dataset
|
from torchvision.transforms.functional import to_pil_image
|
from PIL import Image
|
|
|
class VideoReader(Dataset):
|
def __init__(self, path, transform=None):
|
self.video = pims.PyAVVideoReader(path)
|
self.rate = self.video.frame_rate
|
self.transform = transform
|
|
@property
|
def frame_rate(self):
|
return self.rate
|
|
def __len__(self):
|
return len(self.video)
|
|
def __getitem__(self, idx):
|
frame = self.video[idx]
|
frame = Image.fromarray(np.asarray(frame))
|
if self.transform is not None:
|
frame = self.transform(frame)
|
return frame
|
|
|
class VideoWriter:
|
def __init__(self, path, frame_rate, bit_rate=1000000,audio_source_path=None):
|
"""
|
:param path: 输出文件路径
|
:param frame_rate: 视频帧率
|
:param bit_rate: 视频码率
|
:param audio_source_path: 原视频文件路径（用于拷贝音频）
|
"""
|
self.container = av.open(path, mode='w')
|
self.v_stream = self.container.add_stream('h264', rate=round(frame_rate))
|
self.v_stream.pix_fmt = 'yuv420p'
|
self.v_stream.bit_rate = bit_rate
|
self.v_stream.codec_context.options = {
|
'crf': '18', # 控制画质，越小越清晰（18是高质量）
|
'preset': 'veryfast', # 影响编码速度，medium 是1x，fast是2x faster是3x veryfast是4x superfast是6x ultrafast是10x，文件体积大约会增加(以medium为基准) 10%，15%，20% 30% 50%
|
'tune': 'film',
|
}
|
self.audio_source_path =audio_source_path
|
self.audio_in_container = None
|
self.audio_stream_in = None
|
self.audio_out_stream = None
|
|
if self.audio_source_path:
|
try:
|
self.audio_in_container = av.open(audio_source_path)
|
if self.audio_in_container.streams.audio:
|
self.audio_stream_in = self.audio_in_container.streams.audio[0]
|
except:
|
pass
|
|
if self.audio_stream_in:
|
self.audio_out_stream = self.container.add_stream(self.audio_stream_in.codec.name)
|
else:
|
self.audio_out_stream = None
|
|
|
def write(self, frames):
|
# frames: [T, C, H, W]
|
self.v_stream.width = frames.size(3)
|
self.v_stream.height = frames.size(2)
|
if frames.size(1) == 1:
|
frames = frames.repeat(1, 3, 1, 1) # convert grayscale to RGB
|
frames = frames.mul(255).byte().cpu().permute(0, 2, 3, 1).numpy()
|
for t in range(frames.shape[0]):
|
frame = frames[t]
|
frame = av.VideoFrame.from_ndarray(frame, format='rgb24')
|
# old version
|
#self.container.mux(self.v_stream.encode(frame))
|
for packet in self.v_stream.encode(frame):
|
self.container.mux(packet)
|
|
def close(self):
|
# old version
|
#self.container.mux(self.v_stream.encode())
|
for packet in self.v_stream.encode():
|
self.container.mux(packet)
|
|
if self.audio_stream_in and self.audio_out_stream:
|
for packet in self.audio_in_container.demux(self.audio_stream_in):
|
if packet.dts is None:
|
continue
|
packet.stream = self.audio_out_stream # 直接改，不用 copy()
|
self.container.mux(packet)
|
|
self.container.close()
|
if self.audio_in_container:
|
self.audio_in_container.close()
|
|
|
class ImageSequenceReader(Dataset):
|
def __init__(self, path, transform=None):
|
self.path = path
|
self.files = sorted(os.listdir(path))
|
self.transform = transform
|
|
def __len__(self):
|
return len(self.files)
|
|
def __getitem__(self, idx):
|
with Image.open(os.path.join(self.path, self.files[idx])) as img:
|
img.load()
|
if self.transform is not None:
|
return self.transform(img)
|
return img
|
|
|
class ImageSequenceWriter:
|
def __init__(self, path, extension='jpg'):
|
self.path = path
|
self.extension = extension
|
self.counter = 0
|
os.makedirs(path, exist_ok=True)
|
|
def write(self, frames):
|
# frames: [T, C, H, W]
|
for t in range(frames.shape[0]):
|
to_pil_image(frames[t]).save(os.path.join(
|
self.path, str(self.counter).zfill(4) + '.' + self.extension))
|
self.counter += 1
|
|
def close(self):
|
pass
|
|
