def main(args):
input_size = 3
output_size = 1
# Make the training and testing set to be less bias
fire_dataset = FireDataset(args.csv_path[0])
fire_train, fire_test = random_split(
fire_dataset,
(round(0.7 * len(fire_dataset)), round(0.3 * len(fire_dataset))))
trainloader = DataLoader(fire_train,
batch_size=4096,
shuffle=True,
num_workers=2)
testloader = DataLoader(fire_test,
batch_size=512,
shuffle=False,
num_workers=2)
save_weights_pth = args.weights_path[0]
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = MLP(input_size=input_size, output_size=output_size)
model.to(device)
criterion = nn.BCELoss()
optimizer = optim.Adam(model.parameters(), lr=5e-7)
epochs = 30
if args.eval_only:
do_test(model, device, testloader, save_weights_pth)
else:
do_train(model, device, trainloader, criterion, optimizer, epochs,
save_weights_pth)
do_test(model, device, testloader, save_weights_pth)
def main():
# get unity environment
env, brain = get_unity_envs()
# get arguments
args = get_arguments()
print(args)
# set gpu environment
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
cudnn.enabled = True
cudnn.benchmark = True
cuda = torch.cuda.is_available()
# set random seed
rn = set_seeds(args.random_seed, cuda)
# make directory
os.makedirs(args.snapshot_dir, exist_ok=True)
# get validation dataset
val_set = get_validation_dataset(args)
print("len of test set: ", len(val_set))
val_loader = data.DataLoader(val_set,
batch_size=args.real_batch_size,
shuffle=False,
num_workers=args.num_workers,
pin_memory=True)
# generate training list
with open(args.syn_list_path, "w") as fp:
for i in range(args.syn_img_num):
if i % 10 != 0:
fp.write(str(i + 1) + '\n')
# get main model
main_model = MLP(args.num_inputs, args.num_outputs, args.hidden_size)
if args.resume != "":
main_model.load_state_dict(torch.load(args.resume))
# get task model
if args.task_model_name == "FCN8s":
task_model = FCN8s_sourceonly(n_class=args.num_classes)
vgg16 = VGG16(pretrained=True)
task_model.copy_params_from_vgg16(vgg16)
else:
raise ValueError("Specified model name: FCN8s")
# save initial task model
torch.save(task_model.state_dict(),
os.path.join(args.snapshot_dir, "task_model_init.pth"))
if cuda:
main_model = main_model.cuda()
task_model = task_model.cuda()
# get optimizer
main_optimizer = optim.Adam(main_model.parameters(), lr=args.main_lr)
task_optimizer = optim.SGD(task_model.parameters(),
lr=args.task_lr,
momentum=0.9,
weight_decay=1e-4)
frame_idx = 0
whole_start_time = time.time()
while frame_idx < args.max_frames:
log_probs = []
rewards = []
start_time = time.time()
for i_step in range(1, args.step_each_frame + 1):
# get initial attribute list
state = np.random.rand(1, args.num_inputs)
state = torch.from_numpy(state).float()
if cuda:
state = state.cuda()
# get modified attribute list
dist = main_model(state)
action = dist.sample()
action_actual = action.float() / 10.0 # [0, 0.9]
# generate images by attribute list
print("action: " + str(action_actual.cpu().numpy()))
get_images_by_attributes(args, i_step, env, brain,
action_actual[0].cpu().numpy())
train_set = get_training_dataset(args, i_step)
train_loader = data.DataLoader(train_set,
batch_size=args.syn_batch_size,
shuffle=True,
num_workers=args.num_workers,
pin_memory=True)
# train the task model using synthetic dataset
task_model.load_state_dict(
torch.load(
os.path.join(args.snapshot_dir, "task_model_init.pth")))
reward = train_task_model(train_loader, val_loader, task_model,
task_optimizer, args, cuda)
log_prob = dist.log_prob(action)[0]
log_probs.append(log_prob)
rewards.append(torch.FloatTensor([reward]))
frame_idx += 1
if frame_idx == 1:
moving_start = torch.FloatTensor([reward])
baseline = compute_returns(rewards, moving_start)
moving_start = baseline[-1]
log_probs = torch.cat(log_probs)
baseline = torch.cat(baseline).detach()
rewards = torch.cat(rewards).detach()
advantage = rewards - baseline
if cuda:
advantage = advantage.cuda()
loss = -(log_probs * advantage.detach()).mean()
with open(os.path.join(args.snapshot_dir, "logs.txt"), 'a') as fp:
fp.write(
"frame idx: {0:4d}, state: {1:s}, action: {2:s}, reward: {3:s}, baseline: {4:s}, loss: {5:.2f} \n"
.format(frame_idx, str(state.cpu()[0].numpy()),
str(action.cpu()[0].numpy()), str(rewards.numpy()),
str(baseline.numpy()), loss.item()))
print("optimize the main model parameters")
main_optimizer.zero_grad()
loss.backward()
main_optimizer.step()
elapsed_time = time.time() - start_time
print("[frame: {0:3d}], [loss: {1:.2f}], [time: {2:.1f}]".format(
frame_idx, loss.item(), elapsed_time))
torch.save(
main_model.state_dict(),
os.path.join(args.snapshot_dir, "main_model_%d.pth" % frame_idx))
elapsed_time = time.time() - whole_start_time
print("whole time: {0:.1f}".format(elapsed_time))
env.close()
def train_model(data,
model_name="mlp",
log_interval=10,
loss="mse",
optim="adam",
store=False,
visual=False,
verbose=False,
log=False):
if verbose and log:
print("\nTraining model", model_name)
if store:
# Create directory
checkpoint_path = os.path.join("checkpoints/", model_name)
if not os.path.isdir(checkpoint_path):
try:
os.makedirs(checkpoint_path)
except OSError:
sys.exit("Creation of checkpoint directory failed.")
# Model
model = None
if model_name == "mlp":
model = MLP(
num_features=data.X.shape[1],
hidden_size=config("{}.hidden_layer".format(model_name)),
)
elif model_name == "gcn":
model = GCN(
num_features=data.X.shape[1],
hidden_size=config("{}.hidden_layer".format(model_name)),
)
elif model_name == "gat":
model = GAT(
num_features=data.X.shape[1],
hidden_size=config("{}.hidden_layer".format(model_name)),
)
# Criterion and Loss Function
criterion = torch.nn.MSELoss()
loss_fn = None
if loss == "mse":
loss_fn = torch.nn.MSELoss()
elif loss == "nll":
loss_fn = NLLLoss()
# Optimizer
optimizer = None
if optim == "adam":
optimizer = Adam(model.parameters(),
lr=config('{}.learning_rate'.format(model_name)))
elif optim == "adagrad":
optimizer = Adagrad(model.parameters(),
lr=config('{}.learning_rate'.format(model_name)))
# Setup training
fig, axes = None, None
if visual:
fig, axes = make_training_plot(model_name)
start_epoch = 0
stats = []
if store:
# Attempts to restore the latest checkpoint if exists
print('Loading {}...'.format(model_name))
model, start_epoch, stats = restore_checkpoint(
model, config('{}.checkpoint'.format(model_name)))
# Evaluate the randomly initialized model
_evaluate_epoch(axes, data, model, criterion, start_epoch, stats,
log_interval, log)
# Loop over the entire dataset multiple times
patience = config("patience")
best_loss = float('inf')
idx = -1
for epoch in range(start_epoch,
config('{}.num_epochs'.format(model_name))):
# Early stop
if patience < 0:
break
# Train model
_train_epoch(data, model, loss_fn, optimizer)
# Evaluate model
if (epoch + 1) % log_interval == 0:
_evaluate_epoch(axes, data, model, criterion, epoch + 1, stats,
log_interval, log)
if store:
# Save model parameters
save_checkpoint(model, epoch + 1,
config('{}.checkpoint'.format(model_name)),
stats)
valid_loss = stats[-1][0]
if valid_loss < best_loss:
patience = config("patience")
best_loss = valid_loss
idx = epoch
patience -= 1
epoch = idx
idx = min(int((idx + 1) / log_interval), len(stats) - 1)
if verbose:
print("The loss on test dataset is:", stats[idx][2],
"obtained in epoch", epoch)
# Save figure and keep plot open
if visual:
save_training_plot(fig, model_name)
hold_training_plot()
return stats[idx][2]
# build network
mnist_mlp = MLP(io_bits=io_bits)
mnist_mlp.cuda()
try:
data = torch.load('mlp_mnist_nobias.t7')
mnist_mlp.load_state_dict(data)
print('Loading weight')
except:
print('Initializing model')
#
##
criterian = nn.CrossEntropyLoss(size_average=False)
optimizer = optim.SGD(mnist_mlp.parameters(), lr=learning_rate)
best_acc = 0
for i in range(epoches):
#training
running_acc = 0.
for (img, label) in trainloader:
img = torch.autograd.Variable(img).cuda()
label = torch.autograd.Variable(label).cuda()
save_param, st = quantize_weight(mnist_mlp, w_bits, cells)
noise(mnist_mlp, st)
optimizer.zero_grad()
output = mnist_mlp(img)
loss = criterian(output, label)
loss.backward()
optimizer.step()
def main(args):
input_size = 5
output_size = 3
epochs = 100
# Make the training and testing set to be less bias
hlb_dataset = HLBDataset(args.dataset_path[0])
hlb_train, hlb_test = random_split(
hlb_dataset,
(round(0.8 * len(hlb_dataset)), round(0.2 * len(hlb_dataset))))
print(args.weights_path[0])
print(f'Number of training examples: {len(hlb_train)}')
print(f'Number of testing examples: {len(hlb_test)}')
trainloader = DataLoader(hlb_train,
batch_size=2048,
shuffle=True,
num_workers=2)
testloader = DataLoader(hlb_test,
batch_size=1024,
shuffle=False,
num_workers=2)
save_weights_pth = args.weights_path[0]
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = MLP(input_size=input_size, output_size=output_size)
print(model)
model.to(device)
weights = torch.tensor([0.85, 0.25, 1.0])
criterion = nn.CrossEntropyLoss(weight=weights).cuda()
learning_rate = 1e-6
optimizer = optim.Adam(model.parameters(), lr=learning_rate)
train_acc_array = []
train_loss_array = []
test_acc_array = []
test_loss_array = []
for i in range(epochs): # loop over the dataset multiple times
train_loss, train_acc = do_train(model, device, trainloader, criterion,
optimizer)
print('Epoch {} Train loss: {} Train acc: {}'.format(
i, train_loss, train_acc))
train_acc_array.append(train_acc)
train_loss_array.append(train_loss)
test_loss, test_acc = do_test(model, device, testloader, criterion)
test_acc_array.append(test_acc)
test_loss_array.append(test_loss)
print('Test loss: {} Test acc: {}'.format(test_loss, test_acc))
if i % 50 == 0:
learning_rate = learning_rate * 0.99
for param_group in optimizer.param_groups:
param_group['lr'] = learning_rate
torch.save(model.state_dict(), args.weights_path[0])
save_data(train_loss_array, train_acc_array, test_loss_array,
test_acc_array)
plot_graph(train_loss_array, train_acc_array, test_loss_array,
test_acc_array)
def bench_rnn_backward(batch_size=512, num_batch=100, vocab_size=1024, length=30, embed_size=128,\
hidden_size=128, delta=5):
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
rnn = RNNNative(vocab_size, embed_size, hidden_size).to(device)
delta = np.random.randint(-delta, delta, size=num_batch)
input = torch.LongTensor(batch_size).random_(0, vocab_size).to(device)
label = torch.LongTensor(batch_size).random_(0, vocab_size).to(device)
criteria = nn.CrossEntropyLoss()
optimizer = optim.SGD(rnn.parameters(), lr=0.01)
time_elapsed = 0
for i in range(num_batch):
input = torch.LongTensor(batch_size).random_(0, vocab_size).to(device)
hx = torch.randn(batch_size, hidden_size).to(device)
loss = 0
optimizer.zero_grad()
start = time.time()
for j in range(length + delta[i]):
output, hx = rnn(input, hx)
loss += criteria(output, label)
loss.backward()
optimizer.step()
time_elapsed += time.time() - start
print("Elapsed time for RNNNative backward {:.3f}, avg length: {:.3f}".format(time_elapsed, length + np.mean(delta)))
rnn = RNNTorch(vocab_size, embed_size, hidden_size).to(device)
input = torch.LongTensor(length, batch_size).random_(0, vocab_size).to(device)
label = torch.LongTensor(length*batch_size).random_(0, vocab_size).to(device)
start = time.time()
for i in range(num_batch):
loss = 0
optimizer.zero_grad()
output, _ = rnn(input)
loss = criteria(output.view(length*batch_size, -1), label)
loss.backward()
optimizer.step()
end = time.time()
print("Elapsed time for RNNTorch backward {:.3f}, avg length: {:.3f}".format(end - start, length + np.mean(delta)))
input = torch.LongTensor(1, batch_size).random_(0, vocab_size).to(device)
label = torch.LongTensor(batch_size).random_(0, vocab_size).to(device)
hx = torch.randn(1, batch_size, hidden_size).to(device)
time_elapsed = 0
for i in range(num_batch):
loss = 0
input = torch.LongTensor(1, batch_size).random_(0, vocab_size).to(device)
optimizer.zero_grad()
start = time.time()
for j in range(length + delta[i]):
output, hx = rnn(input, hx)
loss += criteria(output.view(batch_size, -1), label)
loss.backward()
optimizer.step()
time_elapsed += time.time() - start
print("Elapsed time for RNNTorch step backward {:.3f}, avg length: {:.3f}".format(time_elapsed, length + np.mean(delta)))
# mlp
label = torch.LongTensor(batch_size).random_(0, vocab_size).to(device)
mlp = MLP(device, vocab_size, embed_size, hidden_size, length).to(device)
input = []
hx = []
for i in range(length):
input.append(torch.LongTensor(batch_size).random_(0, vocab_size).to(device))
hx.append(torch.randn(batch_size, hidden_size).to(device))
optimizer = optim.SGD(mlp.parameters(), lr=0.01)
start = time.time()
for i in range(num_batch):
optimizer.zero_grad()
output = mlp(input, hx)
loss = 0
for i in range(length):
loss += criteria(output[i], label)
loss.backward
optimizer.step()
end = time.time()
print("Elapsed time for MLP backward {:.3f}".format(end - start))
# mlp2
mlp2 = MLP2(device, vocab_size, embed_size, hidden_size, length).to(device)
input = torch.LongTensor(batch_size).random_(0, vocab_size).to(device)
hx = torch.randn(batch_size, hidden_size).to(device)
optimizer = optim.SGD(mlp2.parameters(), lr=0.01)
start = time.time()
for i in range(num_batch):
optimizer.zero_grad()
loss = 0
for i in range(length):
output = mlp2(input, hx)
loss += criteria(output, label)
loss.backward
optimizer.step()
end = time.time()
print("Elapsed time for MLP2 backward {:.3f}".format(end - start))