default=100000,
help='Max rank of the embedding')
parser.add_argument('--classname', help='Name of class in model.py')
parser.add_argument(
'--like',
help='train a new model like this one, pick parameters from the name')
parser.add_argument('--model-file', help='file-name-prefix to save to')
args = parser.parse_args()
word_embeddings = data.read_embeddings(args.embeddings, args.maxrank_emb)
with open(args.train_file) as f:
train_conllu = data.read_conll(f)
inputs_train_dict, outputs_train_dict, output_features = data.prep_data(
train_conllu,
args.dicts_file,
word_embeddings.vocab,
word_seq_len=None,
shuffle=True)
word_seq_len = inputs_train_dict["inp_char_seq"].shape[1]
with open(args.devel_file) as f:
devel_conllu = data.read_conll(f)
inputs_devel_dict, outputs_devel_dict, output_features_dev = data.prep_data(
devel_conllu,
args.dicts_file,
word_embeddings.vocab,
word_seq_len=word_seq_len,
shuffle=False)
assert output_features == output_features_dev
if args.like:
params = params_from_name(args.like)
def train_model(epochs, net, criterion, optimizer, early_stop):
"""Perform the training."""
# DataLoaders
train_loader, val_loader = prep_data(BATCH_SIZE, THREADS, USE_CUDA)
net.train()
loss_list = []
for epoch in tqdm(range(epochs), total=epochs):
running_loss = MetricMeter()
running_accuracy = MetricMeter()
val_loss_meter = MetricMeter()
val_acc_meter = MetricMeter()
for i, dict_ in enumerate(train_loader):
images = dict_['img']
target = dict_['target'].type(torch.FloatTensor) #.long()
if USE_CUDA:
images = images.cuda()
target = target.cuda()
images = Variable(images)
target = Variable(target)
output = net(images).squeeze()
loss = criterion(output, target)
acc = accuracy(target.data, output.data)
running_loss.update(loss.data[0])
running_accuracy.update(acc)
optimizer.zero_grad()
loss.backward()
optimizer.step()
for i, dict_ in enumerate(val_loader):
images = dict_['img']
target = dict_['target'].type(torch.FloatTensor) #.long()
if USE_CUDA:
images = images.cuda()
target = target.cuda()
images = Variable(images)
target = Variable(target)
output = net(images).squeeze()
val_loss = criterion(output, target)
val_acc = accuracy(target.data, output.data)
val_loss_meter.update(val_loss.data[0])
val_acc_meter.update(val_acc)
# hacky early stopping...
loss_list.append(val_loss_meter.avg)
if (mean(loss_list[-12:-4]) < mean(
loss_list[-4:])) and (len(loss_list) > 12) and early_stop:
print('Validation loss no longer decreasing... stopping training.')
break
print("[ loss: {:.4f} | acc: {:.4f} | vloss: {:.4f} | vacc: {:.4f} ]".
format(running_loss.avg, running_accuracy.avg,
val_loss_meter.avg, val_acc_meter.avg))
ts = time.strftime("%Y-%m-%d_%H-%M")
vl = '_{:.5f}'.format(val_loss_meter.avg)
fname = 'model_' + ts + vl + '.pth'
torch.save(net, os.path.join(THIS_DIR, '..', 'weights', fname))