def prepare():
logging.getLogger().setLevel('INFO')
logging.info("Loading pytorch models")
net_rand = basic.SpeechImage(basic.DEFAULT_CONFIG).to(_device)
net_rand.eval()
net_train = basic.SpeechImage(basic.DEFAULT_CONFIG)
net_train.load_state_dict(torch.load("net.20.pt").state_dict())
net_train.to(_device)
net_train.eval()
nets = [('trained', net_train), ('random', net_rand)]
with torch.no_grad():
save_data(nets, directory='.', batch_size=32)
shuffle=True),
val=D.NewmanRatner_loader(split='val',
register="ADS",
root=root,
batch_size=16,
shuffle=False))
config = dict(SpeechEncoder=dict(conv=dict(in_channels=39,
out_channels=64,
kernel_size=6,
stride=2,
padding=0,
bias=False),
rnn=dict(input_size=64,
hidden_size=1024,
num_layers=4,
bidirectional=True,
dropout=0),
att=dict(in_size=2048, hidden_size=128)),
ImageEncoder=dict(linear=dict(in_size=768, out_size=2 * 1024),
norm=True),
margin_size=0.2)
logging.info('Building model')
net = M.SpeechImage(config)
run_config = dict(max_lr=2 * 1e-4, epochs=50, seed=seed)
logging.info('Training')
folder = os.path.dirname(os.path.abspath(__file__))
M.experiment(net, data, run_config, folder=folder)
def train(args):
# Setting general configuration
torch.manual_seed(args.seed)
random.seed(args.seed)
# Logging the arguments
logging.info('Arguments: {}'.format(args))
logging.info('Loading data')
if args.dataset_name == 'flickr8k':
data = dict(train=D.flickr8k_loader(
args.flickr8k_root,
args.flickr8k_meta,
args.flickr8k_language,
args.audio_features_fn,
split='train',
batch_size=32,
shuffle=True,
downsampling_factor=args.downsampling_factor),
val=D.flickr8k_loader(args.flickr8k_root,
args.flickr8k_meta,
args.flickr8k_language,
args.audio_features_fn,
split='val',
batch_size=32,
shuffle=False))
elif args.dataset_name == "spokencoco":
data = dict(train=D.spokencoco_loader(
args.spokencoco_root,
args.spokencoco_meta,
args.audio_features_fn,
split='train',
batch_size=32,
shuffle=True,
downsampling_factor=args.downsampling_factor,
debug=args.debug),
val=D.spokencoco_loader(args.spokencoco_root,
args.spokencoco_meta,
args.audio_features_fn,
split='val',
batch_size=32,
shuffle=False,
debug=args.debug))
else:
raise ValueError(
"dataset_name should be in ['flickr8k', 'spokencoco']")
config = dict(
SpeechEncoder=dict(conv=dict(in_channels=39,
out_channels=64,
kernel_size=6,
stride=2,
padding=0,
bias=False),
rnn=dict(input_size=64,
hidden_size=args.hidden_size_factor,
num_layers=4,
bidirectional=True,
dropout=0),
att=dict(in_size=2 * args.hidden_size_factor,
hidden_size=128)),
ImageEncoder=dict(linear=dict(in_size=2048,
out_size=2 * args.hidden_size_factor),
norm=True),
margin_size=0.2)
logging.info('Building model')
net = M.SpeechImage(config)
run_config = dict(
max_lr=args.cyclic_lr_max,
min_lr=args.cyclic_lr_min,
epochs=args.epochs,
l2_regularization=args.l2_regularization,
)
logging.info('Training')
old_time = datetime.datetime.now()
logging.info(f'Start of training - {old_time}')
M.experiment(net, data, run_config, wandb_mode='disabled')
new_time = datetime.datetime.now()
logging.info(f'End of training - {new_time}')
diff_time = new_time - old_time
logging.info(f'Total duration: {diff_time}')
