def main():
## params:
config = {'learning_rate': 0.0001,
'nr_epochs': 500,
'batch_size': 8
}
batch_size = 1
model_name = stringify(["model_"]+
[str(k)+'_'+str(config[k]) for k in config \
if type(config[k])==int or type(config[k])==float])+'.pt'
model_path = ''
data_path = ''
train_dataset = dataset()
train_loader =DataLoader(train_dataset,batch_size=batch_size,shuffle = True)
test_dataset = dataset()
val_loader = DataLoader(test_dataset, batch_size = batch_size,shuffle=False)
model_inference = model_def()
model_inference.load_state_dict(torch.load(
os.path.join(model_path, model_name), map_location = device))
model_inference.eval()
labels = np.array([])
outputs = np.array([])
for batch_idx, (data, target) in enumerate(val_loader):
prob = model_inference.forward_inference(data)
labels = np.append(labels, target.data.to('cpu'))
outputs = np.append(outputs, prob.data.to('cpu'))
print(labels)
print(outputs)
def train(config, model_def, dataset, experiment):
print("Device selected: ", device)
data_path = ''
model_path = ''
learning_rate = config['learning_rate']
nr_epochs = config['nr_epochs']
batch_size = config['batch_size']
print(stringify([str(k) + '_' + str(config[k]) for k in config]))
train_dataset = dataset()
train_loader = DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True)
test_dataset = dataset()
test_loader = DataLoader(test_dataset,
batch_size=batch_size,
shuffle=False)
model_inference = model_def()
model_inference = model_inference.to(device)
criterion = F.binary_cross_entropy_with_logits
optimizer = optim.Adam(model_inference.parameters(), lr=learning_rate)
for epoch in range(nr_epochs + 1):
epoch_loss = train_epoch(epoch,
model_inference,
train_loader,
optimizer,
criterion,
verbose=False)
if experiment != None:
experiment.log_metric(stringify(["train_loss_"]+
[str(config[k]) for k in config if type(config[k])==int \
or type(config[k])==float]), epoch_loss, epoch = epoch)
else:
print(
stringify(["train_loss_"] +
[str(k) + '_' + str(config[k]) for k in config]),
epoch_loss)
epoch_loss = evaluate_epoch(epoch,
model_inference,
test_loader,
criterion,
verbose=False)
if experiment != None:
experiment.log_metric(stringify(["val_loss_"]+
[str(config[k]) for k in config if type(config[k])==int \
or type(config[k])==float]), epoch_loss)
else:
print(
stringify(["test_loss_"] +
[str(k) + '_' + str(config[k]) for k in config]),
epoch_loss)
accuracy = calculate_accuracy(model_inference, test_loader)
if experiment != None:
experiment.log_metric(stringify(["accuracy"]+
[str(config[k]) for k in config if type(config[k])==int \
or type(config[k])==float]), accuracy)
else:
print(
stringify(["accuracy"] +
[str(k) + '_' + str(config[k]) for k in config]),
accuracy)
torch.save(model_inference.state_dict(), os.path.join(model_path,stringify(
["model_"]+ [str(k)+'_'+str(config[k]) for k in config \
if type(config[k])==int or type(config[k])==float])+'.pt'))
print("Model written to disc")
def train(config, model_def, dataset, experiment):
print("Device selected: ", device)
data_path = 'temp/dataset'
data_path = '/mnt/nfs/scratch1/vkeshav/vision/dataset'
model_path = 'temp/saved_models'
model_path = '/mnt/nfs/scratch1/vkeshav/vision/saved_models'
learning_rate = config['learning_rate']
nr_epochs = config['nr_epochs']
batch_size = config['batch_size']
print(stringify([str(k) + '_' + str(config[k]) for k in config]))
train_dataset = dataset(data_path, split='train')
train_loader = DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True)
test_dataset = dataset(data_path, split='test')
test_loader = DataLoader(test_dataset,
batch_size=batch_size,
shuffle=False)
model_inference = model_def()
model_inference = model_inference.to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model_inference.parameters(), lr=learning_rate)
for epoch in range(nr_epochs + 1):
epoch_loss = train_epoch(epoch,
model_inference,
train_loader,
optimizer,
criterion,
verbose=False)
if experiment is not None:
experiment.log_metric(stringify(["train_loss_"] + [
str(config[k]) for k in config
if type(config[k]) == int or type(config[k]) == float
]),
epoch_loss,
epoch=epoch)
else:
print(
stringify(["train_loss_"] +
[str(k) + '_' + str(config[k]) for k in config]),
epoch_loss)
epoch_loss = evaluate_epoch(epoch,
model_inference,
test_loader,
criterion,
verbose=False)
if experiment is not None:
experiment.log_metric(
stringify(["val_loss_"] + [
str(config[k]) for k in config
if type(config[k]) == int or type(config[k]) == float
]), epoch_loss)
else:
print(
stringify(["test_loss_"] +
[str(k) + '_' + str(config[k]) for k in config]),
epoch_loss)
accuracies = calculate_accuracy(model_inference, test_loader)
if experiment is not None:
experiment.log_metric(
stringify(["accuracy_path1"] + [
str(config[k]) for k in config
if type(config[k]) == int or type(config[k]) == float
]), accuracies[0])
experiment.log_metric(
stringify(["accuracy_path2"] + [
str(config[k]) for k in config
if type(config[k]) == int or type(config[k]) == float
]), accuracies[1])
experiment.log_metric(
stringify(["accuracy_path3"] + [
str(config[k]) for k in config
if type(config[k]) == int or type(config[k]) == float
]), accuracies[2])
experiment.log_metric(
stringify(["accuracy_path4"] + [
str(config[k]) for k in config
if type(config[k]) == int or type(config[k]) == float
]), accuracies[3])
else:
print(
stringify(["accuracy"] +
[str(k) + '_' + str(config[k]) for k in config]),
np.mean(accuracies))
torch.save(
model_inference.state_dict(),
os.path.join(
model_path,
stringify(["model_"] + [
str(k) + '_' + str(config[k]) for k in config
if type(config[k]) == int or type(config[k]) == float
]) + '.pt'))
print("Model written to disc")