def main():
transformations = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
train_dataset = CifarDataset(TRAIN_CSV_PATH, TRAIN_IMG_PATH, transformations)
train_loader = CifarDataloader(train_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=2)
test_dataset = CifarDataset(TEST_CSV_PATH, TEST_IMG_PATH, transformations)
test_loader = CifarDataloader(test_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=2)
model = resnet50(pretrained=True, num_classes=10)
criterion = nn.CrossEntropyLoss()
if USE_GPU:
model = model.cuda()
criterion = criterion.cuda()
optimizer = optim.Adam(model.parameters(), lr=LEARNING_RATE)
# load_checkpoint(os.path.join('checkpoint', 'last_checkpoint.pth.tar'), model, optimizer)
for epoch in range(EPOCHS):
train(train_loader, model, criterion, optimizer, epoch+1, USE_GPU)
test(test_loader, model, USE_GPU)
save_checkpoint({
'epoch': epoch+1,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
}, os.path.join('checkpoint'))
def new_training_data():
try:
training_data = request.get_json()
if training_data is None:
raise BadRequest(description=NO_JSON)
train(training_data)
except BadRequest as e:
return json_error_message('Failed to parse JSON data',
error_data=e.description)
return json_success('Training data added successfully')
train_batch_size = 50
test_batch_size = 50
lr = 1e-4
if (__name__ == '__main__'):
# use pre-trained embed if avalible
word_embed = th.Tensor(vocab_size, embed_dim)
label_embed = th.Tensor(label_num, embed_dim)
net = HyperIM(word_num,
word_embed,
label_embed,
hidden_size=embed_dim,
if_gru=if_gru)
net.to(cuda_device)
loss = nn.BCEWithLogitsLoss()
optim = gt.optim.RiemannianAdam(net.parameters(), lr=lr)
train_data_loader, test_data_loader = data.load_data(
data_path, train_batch_size, test_batch_size, word_num)
train.train(epoch,
net,
loss,
optim,
if_neg_samp=False,
train_data_loader=train_data_loader)
evalu.evaluate(net, if_log=if_log, test_data_loader=test_data_loader)
if __name__ == '__main__':
tf.enable_eager_execution()
parser = argparse.ArgumentParser(
description='Train using a specified config')
parser.add_argument('--config', help='config to run')
parser.add_argument('--experiment', help='experiment to run')
parser.add_argument('--experiment_name',
help='name of the experiment for logging')
parser.add_argument('--experiment_index',
help='index of the experiment to run',
type=int)
parser.add_argument(
'--split',
help='cross validation split number to use as validation data',
default=None,
type=int)
args = parser.parse_args()
if args.experiment:
experiments = generate_configs_from_experiment(args.experiment)
conf = copy.deepcopy(experiments[args.experiment_index])
if args.split is not None:
conf['dataset']['cross_validation_split'] = args.split
train(conf, args.experiment_name)
else:
conf = copy.deepcopy(configs[args.config])
if args.split is not None:
conf['dataset']['cross_validation_split'] = args.split
train(conf, args.experiment_name)
import tensorflow as tf
import boto3
import json
import os
import sys
import time
from util.train import train
conn_sqs = boto3.resource('sqs')
if __name__ == '__main__':
tf.enable_eager_execution()
queue = conn_sqs.get_queue_by_name(QueueName='experiment-configs')
messages = queue.receive_messages(
MaxNumberOfMessages=1,
MessageAttributeNames=['experiment_name'],
WaitTimeSeconds=0)
for message in messages:
conf = json.loads(message.body)
experiment_name = message.message_attributes.get(
'experiment_name').get('StringValue')
message.delete()
train(conf, experiment_name)
)
scheduler = ReduceLROnPlateau(optimizer,
mode="min",
factor=0.1,
patience=0)
with open(log_file, "a") as f:
f.write(f"Run: {run + 1}\n")
for epoch in tqdm(range(_epoch_num), desc="[Epoch]"):
start_time = time.time()
train_loss, train_acc, dev_loss, dev_acc = train(
train_data_loader,
probe,
bert,
loss_fct,
optimizer,
dev_data_loader=dev_data_loader,
scheduler=scheduler,
)
secs = int(time.time() - start_time)
mins = secs / 60
secs = secs % 60
if optimizer.param_groups[0][
"lr"] < _stop_lr or epoch == _epoch_num - 1:
test_loss, test_acc = evaluate(test_data_loader, probe, bert,
loss_fct)
with open(log_file, "a") as f:
def main():
#Data preprocessing for Transfrom , and Normalize the data with the imageNet weight
transformations = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
#Get traing data and test data
train_dataset = CifarDataset(TRAIN_CSV_PATH, TRAIN_IMG_PATH,
transformations)
train_loader = CifarDataloader(train_dataset,
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=2)
test_dataset = CifarDataset(TEST_CSV_PATH, TEST_IMG_PATH, transformations)
test_loader = CifarDataloader(test_dataset,
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=2)
#choose the model with resnet50
model = resnet50(pretrained=True, num_classes=10)
criterion = nn.CrossEntropyLoss()
if USE_GPU:
model = model.cuda()
criterion = criterion.cuda()
optimizer = optim.Adam(model.parameters(), lr=LEARNING_RATE)
# load_checkpoint(os.path.join('checkpoint', 'last_checkpoint.pth.tar'), model, optimizer)
# make optimizer into a horovod Version .
hvd.broadcast_parameters(model.state_dict(), root_rank=0)
optimizer = hvd.DistributedOptimizer(
optimizer, named_parameters=model.named_parameters())
# def save_checkpoint(epoch):
# if hvd.rank() == 0:
# filepath = args.checkpoint_format.format(epoch=epoch + 1)
# state = {
# 'model': model.state_dict(),
# 'optimizer': optimizer.state_dict(),
# }
# torch.save(state, filepath)
for epoch in range(EPOCHS):
train(train_loader,
model,
criterion,
optimizer,
epoch + 1,
USE_GPU,
writer=writer)
test(test_loader, model, USE_GPU)
save_checkpoint(epoch)
if hvd.rank() == 0:
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
}, os.path.join('checkpoint'))
