def restore(self, ckpt_path):
logger.info(f'Restoring model from {ckpt_path}')
checkpoint = torch.load(ckpt_path)
self.model.load_state_dict(checkpoint['model'])
self.optimizer.load_state_dict(checkpoint['optimizer'])
epoch = checkpoint['epoch']
logger.info(f'Successfully restored at epoch={epoch}')
return epoch
def save(self, ckpt_path, epoch):
state = {
'epoch': epoch,
'model': self.model.state_dict(),
'optimizer': self.optimizer.state_dict()
}
logger.info(f'Saving checkpoint to {ckpt_path}')
torch.save(state, ckpt_path)
def restore_latest(self, ckpt_dir):
files = ckpt_dir.glob('*.pt')
try:
latest = max(files, key=os.path.getctime)
except ValueError:
logger.info(f'No checkpoints found in {ckpt_dir}')
return 0
return self.restore(latest)
def __init__(self, epochs, logdir):
pl.style.use('ggplot')
self.loaders = self.get_loaders()
self.device = torch.device(
'cuda:0' if torch.cuda.is_available() else 'cpu')
logger.info(f'Using pytorch device={self.device}')
self.model = self.get_model().to(self.device)
self.optimizer = torch.optim.Adam(self.model.parameters(),
lr=self.learning_rate *
self.batch_size)
Path(logdir).mkdir(exist_ok=True)
self.epochs = epochs
self.logdir = logdir
self.checkpoint_dir = Path(logdir) / 'checkpoints'
self.checkpoint_dir.mkdir(exist_ok=True)
def get_loaders(traindir, batch_size, time_steps):
training_videos = [
'/home/users/daniel/data/shanghaitech/training/nvvl_videos/01_001.mp4'
]
logger.info(f'Found {len(training_videos)} video files in training set.')
testing_videos = [
'/home/users/daniel/data/shanghaitech/testing/videos/01_0014.mp4'
]
logger.info(f'Found {len(testing_videos)} video files in test set.')
index_map = np.arange(16)
processing = {
'input':
nvvl.ProcessDesc(
# scale_width=width, scale_height=height,
normalized=True,
dimension_order='cfhw',
index_map=list(index_map))
}
frame_mask_dataset = FrameMaskDataset(
'/home/users/daniel/data/shanghaitech/testing/test_frame_mask',
index_map=index_map,
video_paths=testing_videos)
datasets = novelly.datasets.Split(
nvvl.VideoDataset(training_videos, time_steps, processing=processing),
nvvl.VideoDataset(testing_videos,
time_steps,
processing=processing,
get_label=frame_mask_dataset.get_label))
loaders = novelly.datasets.Split(
nvvl.VideoLoader(datasets.train,
batch_size=batch_size,
shuffle=True,
buffer_length=3),
nvvl.VideoLoader(datasets.test, batch_size=batch_size,
buffer_length=3))
return loaders, frame_mask_dataset