def train(config, data_import):
class_num = config["network"]["params"]["class_num"]
loss_params = config["loss"]
class_criterion = nn.CrossEntropyLoss()
transfer_criterion = loss.PADA
center_criterion = config["loss"]["discriminant_loss"](
num_classes=class_num,
feat_dim=config["network"]["params"]["bottleneck_dim"])
## prepare data
pristine_x, pristine_y, noisy_x, noisy_y = data_import
dsets = {}
dset_loaders = {}
#sampling WOR
pristine_indices = torch.randperm(len(pristine_x))
pristine_x_train = pristine_x[pristine_indices]
pristine_y_train = pristine_y[pristine_indices]
noisy_indices = torch.randperm(len(noisy_x))
noisy_x_train = noisy_x[noisy_indices]
noisy_y_train = noisy_y[noisy_indices]
dsets["source"] = TensorDataset(pristine_x_train, pristine_y_train)
dsets["target"] = TensorDataset(noisy_x_train, noisy_y_train)
dset_loaders["source"] = DataLoader(dsets["source"],
batch_size=128,
shuffle=True,
num_workers=1)
dset_loaders["target"] = DataLoader(dsets["target"],
batch_size=128,
shuffle=True,
num_workers=1)
config["num_iterations"] = len(
dset_loaders["source"]) * config["epochs"] + 1
config["test_interval"] = len(dset_loaders["source"])
config["snapshot_interval"] = len(
dset_loaders["source"]) * config["epochs"] * .25
config["log_iter"] = len(dset_loaders["source"])
## set base network
net_config = config["network"]
base_network = net_config["name"](**net_config["params"])
if config["ckpt_path"] is not None:
ckpt = torch.load(config['ckpt_path'] + '/best_model.pth.tar',
map_location=torch.device('cpu'))
base_network.load_state_dict(ckpt['base_network'])
use_gpu = torch.cuda.is_available()
if use_gpu:
base_network = base_network.cuda()
## add additional network for some methods
ad_net = network.AdversarialNetwork(base_network.output_num())
gradient_reverse_layer = network.AdversarialLayer(
high_value=config["high"])
if use_gpu:
ad_net = ad_net.cuda()
## collect parameters
if "DeepMerge" in config["net"]:
parameter_list = [{
"params": base_network.parameters(),
"lr_mult": 1,
'decay_mult': 2
}]
parameter_list.append({
"params": ad_net.parameters(),
"lr_mult": .1,
'decay_mult': 2
})
parameter_list.append({
"params": center_criterion.parameters(),
"lr_mult": 10,
'decay_mult': 1
})
elif "ResNet18" in config["net"]:
parameter_list = [{
"params": base_network.parameters(),
"lr_mult": 1,
'decay_mult': 2
}]
parameter_list.append({
"params": ad_net.parameters(),
"lr_mult": .1,
'decay_mult': 2
})
parameter_list.append({
"params": center_criterion.parameters(),
"lr_mult": 10,
'decay_mult': 1
})
if net_config["params"]["new_cls"]:
if net_config["params"]["use_bottleneck"]:
parameter_list = [{"params":base_network.feature_layers.parameters(), "lr_mult":1, 'decay_mult':2}, \
{"params":base_network.bottleneck.parameters(), "lr_mult":10, 'decay_mult':2}, \
{"params":base_network.fc.parameters(), "lr_mult":10, 'decay_mult':2}]
parameter_list.append({
"params": ad_net.parameters(),
"lr_mult": config["ad_net_mult_lr"],
'decay_mult': 2
})
parameter_list.append({
"params": center_criterion.parameters(),
"lr_mult": 10,
'decay_mult': 1
})
else:
parameter_list = [{"params":base_network.feature_layers.parameters(), "lr_mult":1, 'decay_mult':2}, \
{"params":base_network.fc.parameters(), "lr_mult":10, 'decay_mult':2}]
parameter_list.append({
"params": ad_net.parameters(),
"lr_mult": config["ad_net_mult_lr"],
'decay_mult': 2
})
parameter_list.append({
"params": center_criterion.parameters(),
"lr_mult": 10,
'decay_mult': 1
})
else:
parameter_list = [{
"params": base_network.parameters(),
"lr_mult": 1,
'decay_mult': 2
}]
parameter_list.append({
"params": ad_net.parameters(),
"lr_mult": config["ad_net_mult_lr"],
'decay_mult': 2
})
parameter_list.append({
"params": center_criterion.parameters(),
"lr_mult": 10,
'decay_mult': 1
})
## set optimizer
optimizer_config = config["optimizer"]
optimizer = optim_dict[optimizer_config["type"]](parameter_list, \
**(optimizer_config["optim_params"]))
param_lr = []
for param_group in optimizer.param_groups:
param_lr.append(param_group["lr"])
schedule_param = optimizer_config["lr_param"]
lr_scheduler = lr_schedule.schedule_dict[optimizer_config["lr_type"]]
## train
len_train_source = len(dset_loaders["source"])
len_train_target = len(dset_loaders["target"])
transfer_loss_value = classifier_loss_value = total_loss_value = 0.0
best_acc = 0.0
for i in range(config["num_iterations"]):
## train one iter
base_network.train(True)
if i % config["log_iter"] == 0:
optimizer = lr_scheduler(param_lr, optimizer, i,
config["log_iter"], config["frozen lr"],
config["cycle_length"], **schedule_param)
if config["optimizer"]["lr_type"] == "one-cycle":
optimizer = lr_scheduler(param_lr, optimizer, i,
config["log_iter"], config["frozen lr"],
config["cycle_length"], **schedule_param)
elif config["optimizer"]["lr_type"] == "linear":
optimizer = lr_scheduler(param_lr, optimizer, i,
config["log_iter"], config["frozen lr"],
config["cycle_length"], **schedule_param)
optim = optimizer.state_dict()
optimizer.zero_grad()
try:
inputs_source, labels_source = iter(dset_loaders["source"]).next()
inputs_target, labels_target = iter(dset_loaders["target"]).next()
except StopIteration:
iter(dset_loaders["source"])
iter(dset_loaders["target"])
if use_gpu:
inputs_source, inputs_target, labels_source = \
Variable(inputs_source).cuda(), Variable(inputs_target).cuda(), \
Variable(labels_source).cuda()
else:
inputs_source, inputs_target, labels_source = Variable(inputs_source), \
Variable(inputs_target), Variable(labels_source)
inputs = torch.cat((inputs_source, inputs_target), dim=0)
source_batch_size = inputs_source.size(0)
if config['loss']['ly_type'] == 'cosine':
features, logits = base_network(inputs)
source_logits = logits.narrow(0, 0, source_batch_size)
elif config['loss']['ly_type'] == 'euclidean':
features, _ = base_network(inputs)
logits = -1.0 * loss.distance_to_centroids(
features, center_criterion.centers.detach())
source_logits = logits.narrow(0, 0, source_batch_size)
ad_net.train(True)
weight_ad = torch.ones(inputs.size(0))
transfer_loss = transfer_criterion(features, ad_net, gradient_reverse_layer, \
weight_ad, use_gpu)
ad_out, _ = ad_net(features.detach())
ad_acc, source_acc_ad, target_acc_ad = domain_cls_accuracy(ad_out)
# source domain classification task loss
classifier_loss = class_criterion(source_logits, labels_source.long())
# fisher loss on labeled source domain
if config["fisher_or_no"] == 'no':
total_loss = loss_params["trade_off"] * transfer_loss \
+ classifier_loss
total_loss = classifier_loss + classifier_loss * (
0.5 - source_acc_ad)**2 + classifier_loss * (0.5 -
target_acc_ad)**2
total_loss.backward()
optimizer.step()
else:
fisher_loss, fisher_intra_loss, fisher_inter_loss, center_grad = center_criterion(
features.narrow(0, 0, int(inputs.size(0) / 2)),
labels_source,
inter_class=config["loss"]["inter_type"],
intra_loss_weight=loss_params["intra_loss_coef"],
inter_loss_weight=loss_params["inter_loss_coef"])
# entropy minimization loss
em_loss = loss.EntropyLoss(nn.Softmax(dim=1)(logits))
# final loss
total_loss = loss_params["trade_off"] * transfer_loss \
+ fisher_loss \
+ loss_params["em_loss_coef"] * em_loss \
+ classifier_loss
total_loss = classifier_loss + fisher_loss + loss_params[
"em_loss_coef"] * em_loss + classifier_loss * (
0.5 - source_acc_ad)**2 + classifier_loss * (
0.5 - target_acc_ad)**2
total_loss.backward()
if center_grad is not None:
# clear mmc_loss
center_criterion.centers.grad.zero_()
# Manually assign centers gradients other than using autograd
center_criterion.centers.backward(center_grad)
optimizer.step()
return (-1 * total_loss.cpu().float().item())
def train(config):
## set up summary writer
writer = SummaryWriter(config['output_path'])
# set up early stop
early_stop_engine = EarlyStopping(config["early_stop_patience"])
## set pre-process
prep_dict = {}
prep_config = config["prep"]
prep_dict["source"] = prep.image_train( \
resize_size=prep_config["resize_size"], \
crop_size=prep_config["crop_size"])
prep_dict["target"] = prep.image_train( \
resize_size=prep_config["resize_size"], \
crop_size=prep_config["crop_size"])
if prep_config["test_10crop"]:
prep_dict["test"] = prep.image_test_10crop( \
resize_size=prep_config["resize_size"], \
crop_size=prep_config["crop_size"])
else:
prep_dict["test"] = prep.image_test( \
resize_size=prep_config["resize_size"], \
crop_size=prep_config["crop_size"])
## set loss
class_num = config["network"]["params"]["class_num"]
loss_params = config["loss"]
class_criterion = nn.CrossEntropyLoss()
transfer_criterion = loss.PADA
center_criterion = loss_params["loss_type"](num_classes=class_num,
feat_dim=config["network"]["params"]["bottleneck_dim"])
## prepare data
dsets = {}
dset_loaders = {}
data_config = config["data"]
dsets["source"] = ImageList(stratify_sampling(open(data_config["source"]["list_path"]).readlines(), prep_config["source_size"]), \
transform=prep_dict["source"])
dset_loaders["source"] = util_data.DataLoader(dsets["source"], \
batch_size=data_config["source"]["batch_size"], \
shuffle=True, num_workers=1)
dsets["target"] = ImageList(stratify_sampling(open(data_config["target"]["list_path"]).readlines(), prep_config["target_size"]), \
transform=prep_dict["target"])
dset_loaders["target"] = util_data.DataLoader(dsets["target"], \
batch_size=data_config["target"]["batch_size"], \
shuffle=True, num_workers=1)
if prep_config["test_10crop"]:
for i in range(10):
dsets["test"+str(i)] = ImageList(stratify_sampling(open(data_config["test"]["list_path"]).readlines(), ratio=prep_config['target_size']), \
transform=prep_dict["test"]["val"+str(i)])
dset_loaders["test"+str(i)] = util_data.DataLoader(dsets["test"+str(i)], \
batch_size=data_config["test"]["batch_size"], \
shuffle=False, num_workers=1)
dsets["target"+str(i)] = ImageList(stratify_sampling(open(data_config["target"]["list_path"]).readlines(), ratio=prep_config['target_size']), \
transform=prep_dict["test"]["val"+str(i)])
dset_loaders["target"+str(i)] = util_data.DataLoader(dsets["target"+str(i)], \
batch_size=data_config["test"]["batch_size"], \
shuffle=False, num_workers=1)
else:
dsets["test"] = ImageList(stratify_sampling(open(data_config["test"]["list_path"]).readlines(), ratio=prep_config['target_size']), \
transform=prep_dict["test"])
dset_loaders["test"] = util_data.DataLoader(dsets["test"], \
batch_size=data_config["test"]["batch_size"], \
shuffle=False, num_workers=1)
dsets["target_test"] = ImageList(stratify_sampling(open(data_config["target"]["list_path"]).readlines(), ratio=prep_config['target_size']), \
transform=prep_dict["test"])
dset_loaders["target_test"] = MyDataLoader(dsets["target_test"], \
batch_size=data_config["test"]["batch_size"], \
shuffle=False, num_workers=1)
config['out_file'].write("dataset sizes: source={}, target={}\n".format(
len(dsets["source"]), len(dsets["target"])))
## set base network
net_config = config["network"]
base_network = net_config["name"](**net_config["params"])
use_gpu = torch.cuda.is_available()
if use_gpu:
base_network = base_network.cuda()
## collect parameters
if net_config["params"]["new_cls"]:
if net_config["params"]["use_bottleneck"]:
parameter_list = [{"params":base_network.feature_layers.parameters(), "lr_mult":1, 'decay_mult':2}, \
{"params":base_network.bottleneck.parameters(), "lr_mult":10, 'decay_mult':2}, \
{"params":base_network.fc.parameters(), "lr_mult":10, 'decay_mult':2}]
else:
parameter_list = [{"params":base_network.feature_layers.parameters(), "lr_mult":1, 'decay_mult':2}, \
{"params":base_network.fc.parameters(), "lr_mult":10, 'decay_mult':2}]
else:
parameter_list = [{"params":base_network.parameters(), "lr_mult":1, 'decay_mult':2}]
## add additional network for some methods
ad_net = network.AdversarialNetwork(base_network.output_num())
gradient_reverse_layer = network.AdversarialLayer(high_value=config["high"]) #,
#max_iter_value=config["num_iterations"])
if use_gpu:
ad_net = ad_net.cuda()
parameter_list.append({"params":ad_net.parameters(), "lr_mult":10, 'decay_mult':2})
parameter_list.append({"params":center_criterion.parameters(), "lr_mult": 10, 'decay_mult':1})
## set optimizer
optimizer_config = config["optimizer"]
optimizer = optim_dict[optimizer_config["type"]](parameter_list, \
**(optimizer_config["optim_params"]))
param_lr = []
for param_group in optimizer.param_groups:
param_lr.append(param_group["lr"])
schedule_param = optimizer_config["lr_param"]
lr_scheduler = lr_schedule.schedule_dict[optimizer_config["lr_type"]]
## train
len_train_source = len(dset_loaders["source"]) - 1
len_train_target = len(dset_loaders["target"]) - 1
transfer_loss_value = classifier_loss_value = total_loss_value = 0.0
best_acc = 0.0
for i in range(config["num_iterations"]):
if i % config["test_interval"] == 0:
base_network.train(False)
if config['loss']['ly_type'] == "cosine":
temp_acc = image_classification_test(dset_loaders, \
base_network, test_10crop=prep_config["test_10crop"], \
gpu=use_gpu)
elif config['loss']['ly_type'] == "euclidean":
temp_acc, _ = distance_classification_test(dset_loaders, \
base_network, center_criterion.centers.detach(), test_10crop=prep_config["test_10crop"], \
gpu=use_gpu)
else:
raise ValueError("no test method for cls loss: {}".format(config['loss']['ly_type']))
snapshot_obj = {'step': i,
"base_network": base_network.state_dict(),
'precision': temp_acc,
}
if config["loss"]["loss_name"] != "laplacian" and config["loss"]["ly_type"] == "euclidean":
snapshot_obj['center_criterion'] = center_criterion.state_dict()
if temp_acc > best_acc:
best_acc = temp_acc
# save best model
torch.save(snapshot_obj,
osp.join(config["output_path"], "best_model.pth.tar"))
log_str = "iter: {:05d}, {} precision: {:.5f}\n".format(i, config['loss']['ly_type'], temp_acc)
config["out_file"].write(log_str)
config["out_file"].flush()
writer.add_scalar("precision", temp_acc, i)
if early_stop_engine.is_stop_training(temp_acc):
config["out_file"].write("no improvement after {}, stop training at step {}\n".format(
config["early_stop_patience"], i))
# config["out_file"].write("finish training! \n")
break
if (i+1) % config["snapshot_interval"] == 0:
torch.save(snapshot_obj,
osp.join(config["output_path"], "iter_{:05d}_model.pth.tar".format(i)))
## train one iter
base_network.train(True)
optimizer = lr_scheduler(param_lr, optimizer, i, **schedule_param)
optimizer.zero_grad()
if i % len_train_source == 0:
iter_source = iter(dset_loaders["source"])
if i % len_train_target == 0:
iter_target = iter(dset_loaders["target"])
inputs_source, labels_source = iter_source.next()
inputs_target, labels_target = iter_target.next()
if use_gpu:
inputs_source, inputs_target, labels_source = \
Variable(inputs_source).cuda(), Variable(inputs_target).cuda(), \
Variable(labels_source).cuda()
else:
inputs_source, inputs_target, labels_source = Variable(inputs_source), \
Variable(inputs_target), Variable(labels_source)
inputs = torch.cat((inputs_source, inputs_target), dim=0)
source_batch_size = inputs_source.size(0)
if config['loss']['ly_type'] == 'cosine':
features, logits = base_network(inputs)
source_logits = logits.narrow(0, 0, source_batch_size)
elif config['loss']['ly_type'] == 'euclidean':
features, _ = base_network(inputs)
logits = -1.0 * loss.distance_to_centroids(features, center_criterion.centers.detach())
source_logits = logits.narrow(0, 0, source_batch_size)
ad_net.train(True)
weight_ad = torch.ones(inputs.size(0))
transfer_loss = transfer_criterion(features, ad_net, gradient_reverse_layer, \
weight_ad, use_gpu)
ad_out, _ = ad_net(features.detach())
ad_acc, source_acc_ad, target_acc_ad = domain_cls_accuracy(ad_out)
# source domain classification task loss
classifier_loss = class_criterion(source_logits, labels_source)
# fisher loss on labeled source domain
fisher_loss, fisher_intra_loss, fisher_inter_loss, center_grad = center_criterion(features.narrow(0, 0, int(inputs.size(0)/2)), labels_source, inter_class=config["loss"]["inter_type"],
intra_loss_weight=loss_params["intra_loss_coef"], inter_loss_weight=loss_params["inter_loss_coef"])
# entropy minimization loss
em_loss = loss.EntropyLoss(nn.Softmax(dim=1)(logits))
# final loss
total_loss = loss_params["trade_off"] * transfer_loss \
+ fisher_loss \
+ loss_params["em_loss_coef"] * em_loss \
+ classifier_loss
total_loss.backward()
if center_grad is not None:
# clear mmc_loss
center_criterion.centers.grad.zero_()
# Manually assign centers gradients other than using autograd
center_criterion.centers.backward(center_grad)
optimizer.step()
if i % config["log_iter"] == 0:
config['out_file'].write('iter {}: total loss={:0.4f}, transfer loss={:0.4f}, cls loss={:0.4f}, '
'em loss={:0.4f}, '
'mmc loss={:0.4f}, intra loss={:0.4f}, inter loss={:0.4f}, '
'ad acc={:0.4f}, source_acc={:0.4f}, target_acc={:0.4f}\n'.format(
i, total_loss.data.cpu().float().item(), transfer_loss.data.cpu().float().item(), classifier_loss.data.cpu().float().item(),
em_loss.data.cpu().float().item(),
fisher_loss.cpu().float().item(), fisher_intra_loss.cpu().float().item(), fisher_inter_loss.cpu().float().item(),
ad_acc, source_acc_ad, target_acc_ad,
))
config['out_file'].flush()
writer.add_scalar("total_loss", total_loss.data.cpu().float().item(), i)
writer.add_scalar("cls_loss", classifier_loss.data.cpu().float().item(), i)
writer.add_scalar("transfer_loss", transfer_loss.data.cpu().float().item(), i)
writer.add_scalar("ad_acc", ad_acc, i)
writer.add_scalar("d_loss/total", fisher_loss.data.cpu().float().item(), i)
writer.add_scalar("d_loss/intra", fisher_intra_loss.data.cpu().float().item(), i)
writer.add_scalar("d_loss/inter", fisher_inter_loss.data.cpu().float().item(), i)
return best_acc
def train(config):
# Fix random seed and unable deterministic calcualtions
torch.manual_seed(config["seed"])
torch.cuda.manual_seed(config["seed"])
np.random.seed(config["seed"])
torch.backends.cudnn.enabled = False
torch.backends.cudnn.deterministic = True
## In case of highly imbalanced classes one can add weights inside Cross-entropy loss
## made by: [1 - (x / sum(nSamples)) for x in nSamples]
#weights = [0.936,0.064]
#class_weights = torch.FloatTensor(weights).cuda()
# Set up summary writer
writer = SummaryWriter(config['output_path'])
class_num = config["network"]["params"]["class_num"]
class_criterion = nn.CrossEntropyLoss(
) # optionally add "weight=class_weights" in case of higly imbalanced classes
transfer_criterion = config["loss"]["name"]
center_criterion = config["loss"]["discriminant_loss"](
num_classes=class_num,
feat_dim=config["network"]["params"]["bottleneck_dim"])
loss_params = config["loss"]
# Prepare image data. Image shuffling is fixed with the random seed choice.
# Train:validation:test = 70:10:20
dsets = {}
dset_loaders = {}
pristine_indices = torch.randperm(len(pristine_x))
# Train sample
pristine_x_train = pristine_x[
pristine_indices[:int(np.floor(.7 * len(pristine_x)))]]
pristine_y_train = pristine_y[
pristine_indices[:int(np.floor(.7 * len(pristine_x)))]]
# Validation sample --- gets passed into test functions in train file
pristine_x_valid = pristine_x[pristine_indices[
int(np.floor(.7 * len(pristine_x))):int(np.floor(.8 *
len(pristine_x)))]]
pristine_y_valid = pristine_y[pristine_indices[
int(np.floor(.7 * len(pristine_x))):int(np.floor(.8 *
len(pristine_x)))]]
# Test sample for evaluation file
pristine_x_test = pristine_x[
pristine_indices[int(np.floor(.8 * len(pristine_x))):]]
pristine_y_test = pristine_y[
pristine_indices[int(np.floor(.8 * len(pristine_x))):]]
noisy_indices = torch.randperm(len(noisy_x))
# Train sample
noisy_x_train = noisy_x[noisy_indices[:int(np.floor(.7 * len(noisy_x)))]]
noisy_y_train = noisy_y[noisy_indices[:int(np.floor(.7 * len(noisy_x)))]]
# Validation sample --- gets passed into test functions in train file
noisy_x_valid = noisy_x[noisy_indices[int(np.floor(.7 * len(noisy_x))
):int(np.floor(.8 *
len(noisy_x)))]]
noisy_y_valid = noisy_y[noisy_indices[int(np.floor(.7 * len(noisy_x))
):int(np.floor(.8 *
len(noisy_x)))]]
# Test sample for evaluation file
noisy_x_test = noisy_x[noisy_indices[int(np.floor(.8 * len(noisy_x))):]]
noisy_y_test = noisy_y[noisy_indices[int(np.floor(.8 * len(noisy_x))):]]
dsets["source"] = TensorDataset(pristine_x_train, pristine_y_train)
dsets["target"] = TensorDataset(noisy_x_train, noisy_y_train)
dsets["source_valid"] = TensorDataset(pristine_x_valid, pristine_y_valid)
dsets["target_valid"] = TensorDataset(noisy_x_valid, noisy_y_valid)
dsets["source_test"] = TensorDataset(pristine_x_test, pristine_y_test)
dsets["target_test"] = TensorDataset(noisy_x_test, noisy_y_test)
dset_loaders["source"] = DataLoader(dsets["source"],
batch_size=128,
shuffle=True,
num_workers=1)
dset_loaders["target"] = DataLoader(dsets["target"],
batch_size=128,
shuffle=True,
num_workers=1)
dset_loaders["source_valid"] = DataLoader(dsets["source_valid"],
batch_size=64,
shuffle=True,
num_workers=1)
dset_loaders["target_valid"] = DataLoader(dsets["target_valid"],
batch_size=64,
shuffle=True,
num_workers=1)
dset_loaders["source_test"] = DataLoader(dsets["source_test"],
batch_size=64,
shuffle=True,
num_workers=1)
dset_loaders["target_test"] = DataLoader(dsets["target_test"],
batch_size=64,
shuffle=True,
num_workers=1)
config['out_file'].write("dataset sizes: source={}, target={}\n".format(
len(dsets["source"]), len(dsets["target"])))
# Set number of epochs, and logging intervals
config["num_iterations"] = len(
dset_loaders["source"]) * config["epochs"] + 1
config["test_interval"] = len(dset_loaders["source"])
config["snapshot_interval"] = len(
dset_loaders["source"]) * config["epochs"] * .25
config["log_iter"] = len(dset_loaders["source"])
# Print the configuration you are using
config["out_file"].write("config: {}\n".format(config))
config["out_file"].flush()
# Set up early stop
early_stop_engine = EarlyStopping(config["early_stop_patience"])
# Set base network
net_config = config["network"]
base_network = net_config["name"](**net_config["params"])
#Loading trained model if we want:
if config["ckpt_path"] is not None:
print('load model from {}'.format(config['ckpt_path']))
ckpt = torch.load(config['ckpt_path'] + '/best_model.pth.tar')
base_network.load_state_dict(ckpt['base_network'])
use_gpu = torch.cuda.is_available()
if use_gpu:
base_network = base_network.cuda()
# Collect parameters for the chosen network to be trained
if "DeepMerge" in args.net:
parameter_list = [{
"params": base_network.parameters(),
"lr_mult": 1,
'decay_mult': 2
}]
elif net_config["params"]["new_cls"]:
if net_config["params"]["use_bottleneck"]:
parameter_list = [{"params":base_network.feature_layers.parameters(), "lr_mult":1, 'decay_mult':2}, \
{"params":base_network.bottleneck.parameters(), "lr_mult":10, 'decay_mult':2}, \
{"params":base_network.fc.parameters(), "lr_mult":10, 'decay_mult':2}]
else:
parameter_list = [{"params":base_network.feature_layers.parameters(), "lr_mult":1, 'decay_mult':2}, \
{"params":base_network.fc.parameters(), "lr_mult":10, 'decay_mult':2}]
else:
parameter_list = [{
"params": base_network.parameters(),
"lr_mult": 10,
'decay_mult': 2
}]
# Class weights in case we need them, hewe we have balanced sample so weights are 1.0
class_weight = torch.from_numpy(np.array([1.0] * class_num))
if use_gpu:
class_weight = class_weight.cuda()
parameter_list.append({
"params": center_criterion.parameters(),
"lr_mult": 10,
'decay_mult': 1
})
# Set optimizer
optimizer_config = config["optimizer"]
optimizer = optim_dict[optimizer_config["type"]](parameter_list, \
**(optimizer_config["optim_params"]))
# Set learning rate scheduler
param_lr = []
for param_group in optimizer.param_groups:
param_lr.append(param_group["lr"])
schedule_param = optimizer_config["lr_param"]
lr_scheduler = lr_schedule.schedule_dict[optimizer_config["lr_type"]]
scan_lr = []
scan_loss = []
###################
###### TRAIN ######
###################
len_train_source = len(dset_loaders["source"])
len_train_target = len(dset_loaders["target"])
len_valid_source = len(dset_loaders["source_valid"])
len_valid_target = len(dset_loaders["target_valid"])
transfer_loss_value = classifier_loss_value = total_loss_value = 0.0
best_acc = 0.0
for i in range(config["num_iterations"]):
if i % config["test_interval"] == 0:
base_network.train(False)
if config['loss']['ly_type'] == "cosine":
temp_acc, _ = image_classification_test(dset_loaders, 'source_valid', \
base_network, gpu=use_gpu, verbose = False, save_where = None)
train_acc, _ = image_classification_test(dset_loaders, 'source', \
base_network, gpu=use_gpu, verbose = False, save_where = None)
elif config['loss']['ly_type'] == "euclidean":
temp_acc, _ = distance_classification_test(dset_loaders, 'source_valid', \
base_network, center_criterion.centers.detach(), gpu=use_gpu, verbose = False, save_where = None)
train_acc, _ = distance_classification_test(dset_loaders, 'source', \
base_network, center_criterion.centers.detach(), gpu=use_gpu, verbose = False, save_where = None)
else:
raise ValueError("no test method for cls loss: {}".format(
config['loss']['ly_type']))
snapshot_obj = {
'epoch': i / len(dset_loaders["source"]),
"base_network": base_network.state_dict(),
'valid accuracy': temp_acc,
'train accuracy': train_acc,
}
snapshot_obj['center_criterion'] = center_criterion.state_dict()
if (i + 1) % config["snapshot_interval"] == 0:
torch.save(
snapshot_obj,
osp.join(
config["output_path"], "epoch_{}_model.pth.tar".format(
i / len(dset_loaders["source"]))))
if temp_acc > best_acc:
best_acc = temp_acc
# Save best model
torch.save(
snapshot_obj,
osp.join(config["output_path"], "best_model.pth.tar"))
# Write to log file
log_str = "epoch: {}, {} validation accuracy: {:.5f}, {} training accuracy: {:.5f}\n".format(
i / len(dset_loaders["source"]), config['loss']['ly_type'],
temp_acc, config['loss']['ly_type'], train_acc)
config["out_file"].write(log_str)
config["out_file"].flush()
writer.add_scalar("validation accuracy", temp_acc,
i / len(dset_loaders["source"]))
writer.add_scalar("training accuracy", train_acc,
i / len(dset_loaders["source"]))
## Train one iteration
base_network.train(True)
if i % config["log_iter"] == 0:
optimizer = lr_scheduler(param_lr, optimizer, i,
config["log_iter"], config["frozen lr"],
config["cycle_length"], **schedule_param)
if config["optimizer"]["lr_type"] == "one-cycle":
optimizer = lr_scheduler(param_lr, optimizer, i,
config["log_iter"], config["frozen lr"],
config["cycle_length"], **schedule_param)
if config["optimizer"]["lr_type"] == "linear":
optimizer = lr_scheduler(param_lr, optimizer, i,
config["log_iter"], config["frozen lr"],
config["cycle_length"], **schedule_param)
optim = optimizer.state_dict()
scan_lr.append(optim['param_groups'][0]['lr'])
optimizer.zero_grad()
try:
inputs_source, labels_source = iter(dset_loaders["source"]).next()
inputs_target, labels_target = iter(dset_loaders["target"]).next()
except StopIteration:
iter(dset_loaders["source"])
iter(dset_loaders["target"])
if use_gpu:
inputs_source, inputs_target, labels_source = \
Variable(inputs_source).cuda(), Variable(inputs_target).cuda(), \
Variable(labels_source).cuda()
else:
inputs_source, inputs_target, labels_source = Variable(inputs_source), \
Variable(inputs_target), Variable(labels_source)
inputs = torch.cat((inputs_source, inputs_target), dim=0)
source_batch_size = inputs_source.size(0)
# Distance type. We use cosine.
if config['loss']['ly_type'] == 'cosine':
features, logits = base_network(inputs)
source_logits = logits.narrow(0, 0, source_batch_size)
elif config['loss']['ly_type'] == 'euclidean':
features, _ = base_network(inputs)
logits = -1.0 * loss.distance_to_centroids(
features, center_criterion.centers.detach())
source_logits = logits.narrow(0, 0, source_batch_size)
# Transfer loss - MMD
transfer_loss = transfer_criterion(features[:source_batch_size],
features[source_batch_size:])
# Source domain classification task loss
classifier_loss = class_criterion(source_logits, labels_source.long())
# Final loss in case we do not want to add Fisher loss and Entropy minimization
if config["fisher_or_no"] == 'no':
total_loss = loss_params["trade_off"] * transfer_loss \
+ classifier_loss
scan_loss.append(total_loss.cpu().float().item())
total_loss.backward()
#################
# Plot embeddings periodically. tSNE plots
if args.blobs is not None and i / len(
dset_loaders["source"]) % 20 == 0:
visualizePerformance(base_network,
dset_loaders["source"],
dset_loaders["target"],
batch_size=128,
num_of_samples=2000,
imgName='embedding_' +
str(i / len(dset_loaders["source"])),
save_dir=osp.join(config["output_path"],
"blobs"))
#################
optimizer.step()
if i % config["log_iter"] == 0:
# In case we want to do a learning rate scane to find best lr_cycle lengh:
if config['lr_scan'] != 'no':
if not osp.exists(
osp.join(config["output_path"],
"learning_rate_scan")):
os.makedirs(
osp.join(config["output_path"],
"learning_rate_scan"))
plot_learning_rate_scan(
scan_lr, scan_loss, i / len(dset_loaders["source"]),
osp.join(config["output_path"], "learning_rate_scan"))
# In case we want to visualize gradients:
if config['grad_vis'] != 'no':
if not osp.exists(
osp.join(config["output_path"], "gradients")):
os.makedirs(
osp.join(config["output_path"], "gradients"))
plot_grad_flow(
osp.join(config["output_path"], "gradients"),
i / len(dset_loaders["source"]),
base_network.named_parameters())
# Logging:
config['out_file'].write(
'epoch {}: train total loss={:0.4f}, train transfer loss={:0.4f}, train classifier loss={:0.4f}\n'
.format(
i / len(dset_loaders["source"]),
total_loss.data.cpu().float().item(),
transfer_loss.data.cpu().float().item(),
classifier_loss.data.cpu().float().item(),
))
config['out_file'].flush()
# Logging for tensorboard
writer.add_scalar("training total loss",
total_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar("training classifier loss",
classifier_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar("training transfer loss",
transfer_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
#################
# Validation step
#################
for j in range(0, len(dset_loaders["source_valid"])):
base_network.train(False)
with torch.no_grad():
try:
inputs_valid_source, labels_valid_source = iter(
dset_loaders["source_valid"]).next()
inputs_valid_target, labels_valid_target = iter(
dset_loaders["target_valid"]).next()
except StopIteration:
iter(dset_loaders["source_valid"])
iter(dset_loaders["target_valid"])
if use_gpu:
inputs_valid_source, inputs_valid_target, labels_valid_source = \
Variable(inputs_valid_source).cuda(), Variable(inputs_valid_target).cuda(), \
Variable(labels_valid_source).cuda()
else:
inputs_valid_source, inputs_valid_target, labels_valid_source = Variable(inputs_valid_source), \
Variable(inputs_valid_target), Variable(labels_valid_source)
valid_inputs = torch.cat(
(inputs_valid_source, inputs_valid_target), dim=0)
valid_source_batch_size = inputs_valid_source.size(0)
# Distance type. We use cosine.
if config['loss']['ly_type'] == 'cosine':
features, logits = base_network(valid_inputs)
source_logits = logits.narrow(
0, 0, valid_source_batch_size)
elif config['loss']['ly_type'] == 'euclidean':
features, _ = base_network(valid_inputs)
logits = -1.0 * loss.distance_to_centroids(
features, center_criterion.centers.detach())
source_logits = logits.narrow(
0, 0, valid_source_batch_size)
# Transfer loss - MMD
transfer_loss = transfer_criterion(
features[:valid_source_batch_size],
features[valid_source_batch_size:])
# Source domain classification task loss
classifier_loss = class_criterion(
source_logits, labels_valid_source.long())
# Final loss in case we do not want to add Fisher loss and Entropy minimization
total_loss = loss_params["trade_off"] * transfer_loss \
+ classifier_loss
# Logging:
if j % len(dset_loaders["source_valid"]) == 0:
config['out_file'].write(
'epoch {}: valid total loss={:0.4f}, valid transfer loss={:0.4f}, valid classifier loss={:0.4f}\n'
.format(
i / len(dset_loaders["source"]),
total_loss.data.cpu().float().item(),
transfer_loss.data.cpu().float().item(),
classifier_loss.data.cpu().float().item(),
))
config['out_file'].flush()
# Logging for tensorboard:
writer.add_scalar("validation total loss",
total_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar(
"validation classifier loss",
classifier_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar(
"validation transfer loss",
transfer_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
# Early stop in case we see overfitting
if early_stop_engine.is_stop_training(
classifier_loss.cpu().float().item()):
config["out_file"].write(
"no improvement after {}, stop training at step {}\n"
.format(config["early_stop_patience"],
i / len(dset_loaders["source"])))
sys.exit()
# In case we want to add Fisher loss and Entropy minimizaiton
else:
fisher_loss, fisher_intra_loss, fisher_inter_loss, center_grad = center_criterion(
features.narrow(0, 0, int(inputs.size(0) / 2)),
labels_source,
inter_class=loss_params["inter_type"],
intra_loss_weight=loss_params["intra_loss_coef"],
inter_loss_weight=loss_params["inter_loss_coef"])
# Entropy minimization loss
em_loss = loss.EntropyLoss(nn.Softmax(dim=1)(logits))
# Final loss is the sum of all losses
total_loss = loss_params["trade_off"] * transfer_loss \
+ fisher_loss \
+ loss_params["em_loss_coef"] * em_loss \
+ classifier_loss
scan_loss.append(total_loss.cpu().float().item())
total_loss.backward()
#################
# Plot embeddings periodically.
if args.blobs is not None and i / len(
dset_loaders["source"]) % 20 == 0:
visualizePerformance(base_network,
dset_loaders["source"],
dset_loaders["target"],
batch_size=128,
num_of_samples=2000,
imgName='embedding_' +
str(i / len(dset_loaders["source"])),
save_dir=osp.join(config["output_path"],
"blobs"))
#################
if center_grad is not None:
# Clear mmc_loss
center_criterion.centers.grad.zero_()
# Manually assign centers gradients other than using autograd
center_criterion.centers.backward(center_grad)
optimizer.step()
if i % config["log_iter"] == 0:
# In case we want to do a learning rate scane to find best lr_cycle lengh:
if config['lr_scan'] != 'no':
if not osp.exists(
osp.join(config["output_path"],
"learning_rate_scan")):
os.makedirs(
osp.join(config["output_path"],
"learning_rate_scan"))
plot_learning_rate_scan(
scan_lr, scan_loss, i / len(dset_loaders["source"]),
osp.join(config["output_path"], "learning_rate_scan"))
# In case we want to visualize gradients:
if config['grad_vis'] != 'no':
if not osp.exists(
osp.join(config["output_path"], "gradients")):
os.makedirs(
osp.join(config["output_path"], "gradients"))
plot_grad_flow(
osp.join(config["output_path"], "gradients"),
i / len(dset_loaders["source"]),
base_network.named_parameters())
# Logging
config['out_file'].write(
'epoch {}: train total loss={:0.4f}, train transfer loss={:0.4f}, train classifier loss={:0.4f}, '
'train entropy min loss={:0.4f}, '
'train fisher loss={:0.4f}, train intra-group fisher loss={:0.4f}, train inter-group fisher loss={:0.4f}\n'
.format(
i / len(dset_loaders["source"]),
total_loss.data.cpu(),
transfer_loss.data.cpu().float().item(),
classifier_loss.data.cpu().float().item(),
em_loss.data.cpu().float().item(),
fisher_loss.cpu().float().item(),
fisher_intra_loss.cpu().float().item(),
fisher_inter_loss.cpu().float().item(),
))
config['out_file'].flush()
# Logging for tensorboard
writer.add_scalar("training total loss",
total_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar("training transfer loss",
transfer_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar("training classifier loss",
classifier_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar("training entropy minimization loss",
em_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar("training total fisher loss",
fisher_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar("training intra-group fisher",
fisher_intra_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar("training inter-group fisher",
fisher_inter_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
#################
# Validation step
#################
for j in range(0, len(dset_loaders["source_valid"])):
base_network.train(False)
with torch.no_grad():
try:
inputs_valid_source, labels_valid_source = iter(
dset_loaders["source_valid"]).next()
inputs_valid_target, labels_valid_target = iter(
dset_loaders["target_valid"]).next()
except StopIteration:
iter(dset_loaders["source_valid"])
iter(dset_loaders["target_valid"])
if use_gpu:
inputs_valid_source, inputs_valid_target, labels_valid_source = \
Variable(inputs_valid_source).cuda(), Variable(inputs_valid_target).cuda(), \
Variable(labels_valid_source).cuda()
else:
inputs_valid_source, inputs_valid_target, labels_valid_source = Variable(inputs_valid_source), \
Variable(inputs_valid_target), Variable(labels_valid_source)
valid_inputs = torch.cat(
(inputs_valid_source, inputs_valid_target), dim=0)
valid_source_batch_size = inputs_valid_source.size(0)
# Distance type. We use cosine.
if config['loss']['ly_type'] == 'cosine':
features, logits = base_network(valid_inputs)
source_logits = logits.narrow(
0, 0, valid_source_batch_size)
elif config['loss']['ly_type'] == 'euclidean':
features, _ = base_network(valid_inputs)
logits = -1.0 * loss.distance_to_centroids(
features, center_criterion.centers.detach())
source_logits = logits.narrow(
0, 0, valid_source_batch_size)
# Transfer loss - MMD
transfer_loss = transfer_criterion(
features[:valid_source_batch_size],
features[valid_source_batch_size:])
# Source domain classification task loss
classifier_loss = class_criterion(
source_logits, labels_valid_source.long())
# Fisher loss on labeled source domain
fisher_loss, fisher_intra_loss, fisher_inter_loss, center_grad = center_criterion(
features.narrow(0, 0,
int(valid_inputs.size(0) / 2)),
labels_valid_source,
inter_class=loss_params["inter_type"],
intra_loss_weight=loss_params["intra_loss_coef"],
inter_loss_weight=loss_params["inter_loss_coef"])
# Entropy minimization loss
em_loss = loss.EntropyLoss(nn.Softmax(dim=1)(logits))
# Final loss
total_loss = loss_params["trade_off"] * transfer_loss \
+ fisher_loss \
+ loss_params["em_loss_coef"] * em_loss \
+ classifier_loss
# Logging
if j % len(dset_loaders["source_valid"]) == 0:
config['out_file'].write(
'epoch {}, valid transfer loss={:0.4f}, valid classifier loss={:0.4f}, '
'valid entropy min loss={:0.4f}, '
'valid fisher loss={:0.4f}, valid intra-group fisher loss={:0.4f}, valid inter-group fisher loss={:0.4f}\n'
.format(
i / len(dset_loaders["source"]),
transfer_loss.data.cpu().float().item(),
classifier_loss.data.cpu().float().item(),
em_loss.data.cpu().float().item(),
fisher_loss.cpu().float().item(),
fisher_intra_loss.cpu().float().item(),
fisher_inter_loss.cpu().float().item(),
))
config['out_file'].flush()
# Logging for tensorboard
writer.add_scalar("validation total loss",
total_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar(
"validation transfer loss",
transfer_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar(
"validation classifier loss",
classifier_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar(
"validation entropy minimization loss",
em_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar(
"validation total fisher loss",
fisher_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar(
"validation intra-group fisher",
fisher_intra_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar(
"validation inter-group fisher",
fisher_inter_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
# Early stop in case we see overfitting
if early_stop_engine.is_stop_training(
classifier_loss.cpu().float().item()):
config["out_file"].write(
"no improvement after {}, stop training at epoch {}\n"
.format(config["early_stop_patience"],
i / len(dset_loaders["source"])))
sys.exit()
return best_acc
def train(config):
## set pre-process
prep_dict = {}
prep_config = config["prep"]
prep_dict["source"] = prep.image_train( \
resize_size=prep_config["resize_size"], \
crop_size=prep_config["crop_size"])
prep_dict["target"] = prep.image_train( \
resize_size=prep_config["resize_size"], \
crop_size=prep_config["crop_size"])
if prep_config["test_10crop"]:
prep_dict["test"] = prep.image_test_10crop( \
resize_size=prep_config["resize_size"], \
crop_size=prep_config["crop_size"])
else:
prep_dict["test"] = prep.image_test( \
resize_size=prep_config["resize_size"], \
crop_size=prep_config["crop_size"])
## set loss
class_criterion = nn.CrossEntropyLoss()
transfer_criterion = loss.SAN
loss_params = config["loss"]
## prepare data
dsets = {}
dset_loaders = {}
data_config = config["data"]
dsets["source"] = ImageList(open(data_config["source"]["list_path"]).readlines(), \
transform=prep_dict["source"])
dset_loaders["source"] = util_data.DataLoader(dsets["source"], \
batch_size=data_config["source"]["batch_size"], \
shuffle=True, num_workers=4)
dsets["target"] = ImageList(open(data_config["target"]["list_path"]).readlines(), \
transform=prep_dict["target"])
dset_loaders["target"] = util_data.DataLoader(dsets["target"], \
batch_size=data_config["target"]["batch_size"], \
shuffle=True, num_workers=4)
if prep_config["test_10crop"]:
for i in range(10):
dsets["test"+str(i)] = ImageList(open(data_config["test"]["list_path"]).readlines(), \
transform=prep_dict["test"]["val"+str(i)])
dset_loaders["test"+str(i)] = util_data.DataLoader(dsets["test"+str(i)], \
batch_size=data_config["test"]["batch_size"], \
shuffle=False, num_workers=4)
else:
dsets["test"] = ImageList(open(data_config["test"]["list_path"]).readlines(), \
transform=prep_dict["test"])
dset_loaders["test"] = util_data.DataLoader(dsets["test"], \
batch_size=data_config["test"]["batch_size"], \
shuffle=False, num_workers=4)
class_num = config["class_num"]
## set base network
net_config = config["network"]
base_network = net_config["name"](**net_config["params"])
use_gpu = torch.cuda.is_available()
if use_gpu:
base_network = base_network.cuda()
## collect parameters
if net_config["params"]["new_cls"]:
if net_config["params"]["use_bottleneck"]:
parameter_list = [{"params":base_network.feature_layers.parameters(), "lr":1}, \
{"params":base_network.bottleneck.parameters(), "lr":10}, \
{"params":base_network.fc.parameters(), "lr":10}]
else:
parameter_list = [{"params":base_network.feature_layers.parameters(), "lr":1}, \
{"params":base_network.fc.parameters(), "lr":10}]
else:
parameter_list = [{"params": base_network.parameters(), "lr": 1}]
## add additional network for some methods
ad_net_list = []
gradient_reverse_layer_list = []
class_weight = torch.from_numpy(np.array([1.0] * class_num))
if use_gpu:
class_weight = class_weight.cuda()
for i in range(class_num):
if config["dataset"] == "caltech":
ad_net = network.SmallAdversarialNetwork(base_network.output_num())
elif config["dataset"] == "imagenet":
ad_net = network.LittleAdversarialNetwork(
base_network.output_num())
else:
ad_net = network.AdversarialNetwork(base_network.output_num())
gradient_reverse_layer = network.AdversarialLayer()
gradient_reverse_layer_list.append(gradient_reverse_layer)
if use_gpu:
ad_net = ad_net.cuda()
ad_net_list.append(ad_net)
parameter_list.append({"params": ad_net.parameters(), "lr": 10})
## set optimizer
optimizer_config = config["optimizer"]
optimizer = optim_dict[optimizer_config["type"]](parameter_list, \
**(optimizer_config["optim_params"]))
param_lr = []
for param_group in optimizer.param_groups:
param_lr.append(param_group["lr"])
schedule_param = optimizer_config["lr_param"]
lr_scheduler = lr_schedule.schedule_dict[optimizer_config["lr_type"]]
## train
len_train_source = len(dset_loaders["source"]) - 1
len_train_target = len(dset_loaders["target"]) - 1
transfer_loss_value = classifier_loss_value = total_loss_value = 0.0
best_acc = 0.0
for i in range(config["num_iterations"]):
if i % config["test_interval"] == 0:
base_network.train(False)
temp_acc = image_classification_test(dset_loaders, \
base_network, test_10crop=prep_config["test_10crop"], \
gpu=use_gpu)
temp_model = nn.Sequential(base_network)
if temp_acc > best_acc:
best_acc = temp_acc
best_model = temp_model
log_str = "iter: {:05d}, precision: {:.5f}".format(i, temp_acc)
config["out_file"].write(log_str)
config["out_file"].flush()
print(log_str)
if i % config["snapshot_interval"] == 0:
torch.save(nn.Sequential(base_network), osp.join(config["output_path"], \
"iter_{:05d}_model.pth.tar".format(i)))
## train one iter
base_network.train(True)
optimizer = lr_scheduler(param_lr, optimizer, i, **schedule_param)
optimizer.zero_grad()
if i % len_train_source == 0:
iter_source = iter(dset_loaders["source"])
if i % len_train_target == 0:
iter_target = iter(dset_loaders["target"])
inputs_source, labels_source = iter_source.next()
inputs_target, labels_target = iter_target.next()
if use_gpu:
inputs_source, inputs_target, labels_source = \
Variable(inputs_source).cuda(), Variable(inputs_target).cuda(), \
Variable(labels_source).cuda()
else:
inputs_source, inputs_target, labels_source = Variable(inputs_source), \
Variable(inputs_target), Variable(labels_source)
inputs = torch.cat((inputs_source, inputs_target), dim=0)
features, outputs = base_network(inputs)
## switch between different transfer loss
ad_net.train(True)
softmax_out = nn.Softmax(dim=1)(outputs).detach()
class_weight = torch.mean(softmax_out.data, 0)
class_weight = (class_weight / torch.max(class_weight)).cuda().view(-1)
for i in range(class_num):
gradient_reverse_layer_list[i].high = class_weight[i]
transfer_loss = transfer_criterion([features, softmax_out], ad_net_list, \
gradient_reverse_layer_list, class_weight, use_gpu) \
+ loss_params["entropy_trade_off"] * \
loss.EntropyLoss(nn.Softmax(dim=1)(outputs))
classifier_loss = class_criterion(
outputs.narrow(0, 0,
inputs.size(0) / 2), labels_source)
total_loss = loss_params["trade_off"] * transfer_loss + classifier_loss
total_loss.backward()
optimizer.step()
torch.save(best_model, osp.join(config["output_path"],
"best_model.pth.tar"))
return best_acc
def train(config):
## set up summary writer
writer = SummaryWriter(config['output_path'])
# set up early stop
early_stop_engine = EarlyStopping(config["early_stop_patience"])
class_num = config["network"]["params"]["class_num"]
class_criterion = nn.CrossEntropyLoss()
# transfer_criterion = config["loss"]["name"]
#center_criterion = config["loss"]["discriminant_loss"](num_classes=class_num, feat_dim=config["network"]["params"]["bottleneck_dim"])
loss_params = config["loss"]
## prepare data
dsets = {}
dset_loaders = {}
#sampling WOR, i guess we leave the 10 in the middle to validate?
pristine_indices = torch.randperm(len(pristine_x))
#train
pristine_x_train = pristine_x[pristine_indices[:int(np.floor(.7*len(pristine_x)))]]
pristine_y_train = pristine_y[pristine_indices[:int(np.floor(.7*len(pristine_x)))]]
#validate --- gets passed into test functions in train file
pristine_x_valid = pristine_x[pristine_indices[int(np.floor(.7*len(pristine_x))) : int(np.floor(.8*len(pristine_x)))]]
pristine_y_valid = pristine_y[pristine_indices[int(np.floor(.7*len(pristine_x))) : int(np.floor(.8*len(pristine_x)))]]
#test for evaluation file
pristine_x_test = pristine_x[pristine_indices[int(np.floor(.8*len(pristine_x))):]]
pristine_y_test = pristine_y[pristine_indices[int(np.floor(.8*len(pristine_x))):]]
# noisy_indices = torch.randperm(len(noisy_x))
# #train
# noisy_x_train = noisy_x[noisy_indices[:int(np.floor(.7*len(noisy_x)))]]
# noisy_y_train = noisy_y[noisy_indices[:int(np.floor(.7*len(noisy_x)))]]
# #validate --- gets passed into test functions in train file
# noisy_x_valid = noisy_x[noisy_indices[int(np.floor(.7*len(noisy_x))) : int(np.floor(.8*len(noisy_x)))]]
# noisy_y_valid = noisy_y[noisy_indices[int(np.floor(.7*len(noisy_x))) : int(np.floor(.8*len(noisy_x)))]]
# #test for evaluation file
# noisy_x_test = noisy_x[noisy_indices[int(np.floor(.8*len(noisy_x))):]]
# noisy_y_test = noisy_y[noisy_indices[int(np.floor(.8*len(noisy_x))):]]
dsets["source"] = TensorDataset(pristine_x_train, pristine_y_train)
# dsets["target"] = TensorDataset(noisy_x_train, noisy_y_train)
dsets["source_valid"] = TensorDataset(pristine_x_valid, pristine_y_valid)
# dsets["target_valid"] = TensorDataset(noisy_x_valid, noisy_y_valid)
dsets["source_test"] = TensorDataset(pristine_x_test, pristine_y_test)
# dsets["target_test"] = TensorDataset(noisy_x_test, noisy_y_test)
# data_config = config["data"]
# dsets["source"] = ImageList(stratify_sampling(open(data_config["source"]["list_path"]).readlines(), ratio=prep_config["source_size"]), \
# transform=prep_dict["source"])
#put your dataloaders here
#i stole batch size numbers from below
dset_loaders["source"] = DataLoader(dsets["source"], batch_size = 36, shuffle = True, num_workers = 1)
# dset_loaders["target"] = DataLoader(dsets["target"], batch_size = 36, shuffle = True, num_workers = 1)
#guessing batch size based on what was done for testing in the original file
dset_loaders["source_valid"] = DataLoader(dsets["source_valid"], batch_size = 4, shuffle = True, num_workers = 1)
# dset_loaders["target_valid"] = DataLoader(dsets["target_valid"], batch_size = 4, shuffle = True, num_workers = 1)
dset_loaders["source_test"] = DataLoader(dsets["source_test"], batch_size = 4, shuffle = True, num_workers = 1)
# dset_loaders["target_test"] = DataLoader(dsets["target_test"], batch_size = 4, shuffle = True, num_workers = 1)
config['out_file'].write("dataset sizes: source={}\n".format(
len(dsets["source"]))) #TODO: change this too
## set base network
net_config = config["network"]
base_network = net_config["name"](**net_config["params"])
use_gpu = torch.cuda.is_available()
if use_gpu:
base_network = base_network.cuda()
## collect parameters
if net_config["params"]["new_cls"]:
if net_config["params"]["use_bottleneck"]:
parameter_list = [{"params":base_network.feature_layers.parameters(), "lr_mult":1, 'decay_mult':2}, \
{"params":base_network.bottleneck.parameters(), "lr_mult":10, 'decay_mult':2}, \
{"params":base_network.fc.parameters(), "lr_mult":10, 'decay_mult':2}]
else:
parameter_list = [{"params":base_network.feature_layers.parameters(), "lr_mult":1, 'decay_mult':2}, \
{"params":base_network.fc.parameters(), "lr_mult":10, 'decay_mult':2}]
else:
parameter_list = [{"params":base_network.parameters(), "lr_mult":1, 'decay_mult':2}]
## add additional network for some methods
class_weight = torch.from_numpy(np.array([1.0] * class_num))
if use_gpu:
class_weight = class_weight.cuda()
#parameter_list.append({"params":center_criterion.parameters(), "lr_mult": 10, 'decay_mult':1})
## set optimizer
optimizer_config = config["optimizer"]
optimizer = optim_dict[optimizer_config["type"]](parameter_list, \
**(optimizer_config["optim_params"]))
param_lr = []
for param_group in optimizer.param_groups:
param_lr.append(param_group["lr"])
schedule_param = optimizer_config["lr_param"]
lr_scheduler = lr_schedule.schedule_dict[optimizer_config["lr_type"]]
## train
len_train_source = len(dset_loaders["source"]) - 1
# len_train_target = len(dset_loaders["target"]) - 1
len_valid_source = len(dset_loaders["source_valid"]) - 1
# len_valid_target = len(dset_loaders["target_valid"]) - 1
classifier_loss_value = 0.0
# transfer_loss_value = classifier_loss_value = total_loss_value = 0.0
best_acc = 0.0
for i in range(config["num_iterations"]):
if i % config["test_interval"] == 0:
base_network.train(False)
if config['loss']['ly_type'] == "cosine":
temp_acc, _ = image_classification_test(dset_loaders, 'source_valid', \
base_network, \
gpu=use_gpu)
train_acc, _ = image_classification_test(dset_loaders, 'source', \
base_network, \
gpu=use_gpu)
# elif config['loss']['ly_type'] == "euclidean":
# temp_acc, _ = distance_classification_test(dset_loaders, 'source_valid', \
# base_network, center_criterion.centers.detach(), \
# gpu=use_gpu)
# train_acc, _ = distance_classification_test(dset_loaders, 'source', \
# base_network, \
# gpu=use_gpu)
else:
raise ValueError("no test method for cls loss: {}".format(config['loss']['ly_type']))
snapshot_obj = {'step': i,
"base_network": base_network.state_dict(),
'valid accuracy': temp_acc,
'train accuracy' : train_acc,
}
# snapshot_obj['center_criterion'] = center_criterion.state_dict()
if temp_acc > best_acc:
best_acc = temp_acc
# save best model
torch.save(snapshot_obj,
osp.join(config["output_path"], "best_model.pth.tar"))
log_str = "iter: {:05d}, {} validation accuracy: {:.5f}, {} training accuracy: {:.5f}\n".format(i, config['loss']['ly_type'], temp_acc, config['loss']['ly_type'], train_acc)
config["out_file"].write(log_str)
config["out_file"].flush()
writer.add_scalar("validation accuracy", temp_acc, i)
writer.add_scalar("training accuracy", train_acc, i)
if early_stop_engine.is_stop_training(temp_acc):
config["out_file"].write("no improvement after {}, stop training at step {}\n".format(
config["early_stop_patience"], i))
# config["out_file"].write("finish training! \n")
break
if (i+1) % config["snapshot_interval"] == 0:
torch.save(snapshot_obj,
osp.join(config["output_path"], "iter_{:05d}_model.pth.tar".format(i)))
## train one iter
base_network.train(True)
optimizer = lr_scheduler(param_lr, optimizer, i, **schedule_param)
optimizer.zero_grad()
if i % len_train_source == 0:
iter_source = iter(dset_loaders["source"])
# if i % len_train_target == 0:
# iter_target = iter(dset_loaders["target"])
try:
inputs_source, labels_source = iter_source.next()
# inputs_target, labels_target = iter_target.next()
except StopIteration:
iter_source = iter(dset_loaders["source"])
# iter_target = iter(dset_loaders["target"])
if use_gpu:
inputs_source, labels_source = Variable(inputs_source).cuda(), Variable(labels_source).cuda()
else:
inputs_source, labels_source = Variable(inputs_source), Variable(labels_source)
inputs = inputs_source
source_batch_size = inputs_source.size(0)
# if config['loss']['ly_type'] == 'cosine':
features, logits = base_network(inputs)
source_logits = logits.narrow(0, 0, source_batch_size)
# elif config['loss']['ly_type'] == 'euclidean':
# features, _ = base_network(inputs)
# logits = -1.0 * loss.distance_to_centroids(features, center_criterion.centers.detach())
# source_logits = logits.narrow(0, 0, source_batch_size)
# transfer_loss = transfer_criterion(features[:source_batch_size], features[source_batch_size:])
# source domain classification task loss
classifier_loss = class_criterion(source_logits, labels_source.long())
# fisher loss on labeled source domain
# fisher_loss, fisher_intra_loss, fisher_inter_loss, center_grad = center_criterion(features.narrow(0, 0, int(inputs.size(0)/2)), labels_source, inter_class=loss_params["inter_type"],
# intra_loss_weight=loss_params["intra_loss_coef"], inter_loss_weight=loss_params["inter_loss_coef"])
# entropy minimization loss
em_loss = loss.EntropyLoss(nn.Softmax(dim=1)(logits))
total_loss = loss_params["em_loss_coef"] * em_loss + classifier_loss
total_loss.backward() #we need to fix this
# if center_grad is not None:
# # clear mmc_loss
# center_criterion.centers.grad.zero_()
# # Manually assign centers gradients other than using autograd
# center_criterion.centers.backward(center_grad)
optimizer.step()
if i % config["log_iter"] == 0:
config['out_file'].write('iter {}: train total loss={:0.4f}, train classifier loss={:0.4f}, '
'train entropy min loss={:0.4f}\n'.format(
i, total_loss.data.cpu(), classifier_loss.data.cpu().float().item(), em_loss.data.cpu().float().item(),
))
config['out_file'].flush()
writer.add_scalar("training total loss", total_loss.data.cpu().float().item(), i)
writer.add_scalar("training classifier loss", classifier_loss.data.cpu().float().item(), i)
writer.add_scalar("training entropy minimization loss", em_loss.data.cpu().float().item(), i)
# writer.add_scalar("training transfer loss", transfer_loss.data.cpu().float().item(), i)
# writer.add_scalar("training total fisher loss", fisher_loss.data.cpu().float().item(), i)
# writer.add_scalar("training intra-group fisher", fisher_intra_loss.data.cpu().float().item(), i)
# writer.add_scalar("training inter-group fisher", fisher_inter_loss.data.cpu().float().item(), i)
#attempted validation step
#if i < len_valid_source:
base_network.eval()
with torch.no_grad():
if i % len_valid_source == 0:
iter_source = iter(dset_loaders["source_valid"])
# if i % len_valid_target == 0:
# iter_target = iter(dset_loaders["target_valid"])
try:
inputs_source, labels_source = iter_source.next()
# inputs_target, labels_target = iter_target.next()
except StopIteration:
iter_source = iter(dset_loaders["source_valid"])
# iter_target = iter(dset_loaders["target_valid"])
if use_gpu:
inputs_source, labels_source = Variable(inputs_source).cuda(), Variable(labels_source).cuda()
else:
inputs_source,labels_source = Variable(inputs_source), Variable(labels_source)
inputs = inputs_source
source_batch_size = inputs_source.size(0)
# if config['loss']['ly_type'] == 'cosine':
features, logits = base_network(inputs)
source_logits = logits.narrow(0, 0, source_batch_size)
# elif config['loss']['ly_type'] == 'euclidean':
# features, _ = base_network(inputs)
# logits = -1.0 * loss.distance_to_centroids(features, center_criterion.centers.detach())
# source_logits = logits.narrow(0, 0, source_batch_size)
# transfer_loss = transfer_criterion(features[:source_batch_size], features[source_batch_size:])
# source domain classification task loss
classifier_loss = class_criterion(source_logits, labels_source.long())
# fisher loss on labeled source domain
# fisher_loss, fisher_intra_loss, fisher_inter_loss, center_grad = center_criterion(features.narrow(0, 0, int(inputs.size(0)/2)), labels_source, inter_class=loss_params["inter_type"],
# intra_loss_weight=loss_params["intra_loss_coef"], inter_loss_weight=loss_params["inter_loss_coef"])
# entropy minimization loss
em_loss = loss.EntropyLoss(nn.Softmax(dim=1)(logits))
# final loss
total_loss = loss_params["em_loss_coef"] * em_loss + classifier_loss
#total_loss.backward() no backprop on the eval mode
if i % config["log_iter"] == 0:
config['out_file'].write('iter {}: validation total loss={:0.4f}, validation classifier loss={:0.4f}, '
'validation entropy min loss={:0.4f}\n'.format(
i, total_loss.data.cpu(), classifier_loss.data.cpu().float().item(), em_loss.data.cpu().float().item(),
))
config['out_file'].flush()
writer.add_scalar("validation total loss", total_loss.data.cpu().float().item(), i)
writer.add_scalar("validation classifier loss", classifier_loss.data.cpu().float().item(), i)
writer.add_scalar("validation entropy minimization loss", em_loss.data.cpu().float().item(), i)
# writer.add_scalar("validation transfer loss", transfer_loss.data.cpu().float().item(), i)
# writer.add_scalar("validation total fisher loss", fisher_loss.data.cpu().float().item(), i)
# writer.add_scalar("validation intra-group fisher", fisher_intra_loss.data.cpu().float().item(), i)
# writer.add_scalar("validation inter-group fisher", fisher_inter_loss.data.cpu().float().item(), i)
return best_acc
def train(config):
## set up summary writer
writer = SummaryWriter(config['output_path'])
class_num = config["network"]["params"]["class_num"]
loss_params = config["loss"]
class_criterion = nn.CrossEntropyLoss()
transfer_criterion = loss.PADA
center_criterion = loss_params["loss_type"](
num_classes=class_num,
feat_dim=config["network"]["params"]["bottleneck_dim"])
## prepare data
dsets = {}
dset_loaders = {}
#sampling WOR, i guess we leave the 10 in the middle to validate?
pristine_indices = torch.randperm(len(pristine_x))
#train
pristine_x_train = pristine_x[
pristine_indices[:int(np.floor(.7 * len(pristine_x)))]]
pristine_y_train = pristine_y[
pristine_indices[:int(np.floor(.7 * len(pristine_x)))]]
#validate --- gets passed into test functions in train file
pristine_x_valid = pristine_x[pristine_indices[
int(np.floor(.7 * len(pristine_x))):int(np.floor(.8 *
len(pristine_x)))]]
pristine_y_valid = pristine_y[pristine_indices[
int(np.floor(.7 * len(pristine_x))):int(np.floor(.8 *
len(pristine_x)))]]
#test for evaluation file
pristine_x_test = pristine_x[
pristine_indices[int(np.floor(.8 * len(pristine_x))):]]
pristine_y_test = pristine_y[
pristine_indices[int(np.floor(.8 * len(pristine_x))):]]
noisy_indices = torch.randperm(len(noisy_x))
#train
noisy_x_train = noisy_x[noisy_indices[:int(np.floor(.7 * len(noisy_x)))]]
noisy_y_train = noisy_y[noisy_indices[:int(np.floor(.7 * len(noisy_x)))]]
#validate --- gets passed into test functions in train file
noisy_x_valid = noisy_x[noisy_indices[int(np.floor(.7 * len(noisy_x))
):int(np.floor(.8 *
len(noisy_x)))]]
noisy_y_valid = noisy_y[noisy_indices[int(np.floor(.7 * len(noisy_x))
):int(np.floor(.8 *
len(noisy_x)))]]
#test for evaluation file
noisy_x_test = noisy_x[noisy_indices[int(np.floor(.8 * len(noisy_x))):]]
noisy_y_test = noisy_y[noisy_indices[int(np.floor(.8 * len(noisy_x))):]]
dsets["source"] = TensorDataset(pristine_x_train, pristine_y_train)
dsets["target"] = TensorDataset(noisy_x_train, noisy_y_train)
dsets["source_valid"] = TensorDataset(pristine_x_valid, pristine_y_valid)
dsets["target_valid"] = TensorDataset(noisy_x_valid, noisy_y_valid)
dsets["source_test"] = TensorDataset(pristine_x_test, pristine_y_test)
dsets["target_test"] = TensorDataset(noisy_x_test, noisy_y_test)
#put your dataloaders here
#i stole batch size numbers from below
dset_loaders["source"] = DataLoader(dsets["source"],
batch_size=128,
shuffle=True,
num_workers=1)
dset_loaders["target"] = DataLoader(dsets["target"],
batch_size=128,
shuffle=True,
num_workers=1)
#guessing batch size based on what was done for testing in the original file
dset_loaders["source_valid"] = DataLoader(dsets["source_valid"],
batch_size=64,
shuffle=True,
num_workers=1)
dset_loaders["target_valid"] = DataLoader(dsets["target_valid"],
batch_size=64,
shuffle=True,
num_workers=1)
dset_loaders["source_test"] = DataLoader(dsets["source_test"],
batch_size=64,
shuffle=True,
num_workers=1)
dset_loaders["target_test"] = DataLoader(dsets["target_test"],
batch_size=64,
shuffle=True,
num_workers=1)
config['out_file'].write("dataset sizes: source={}, target={}\n".format(
len(dsets["source"]), len(dsets["target"])))
config["num_iterations"] = len(
dset_loaders["source"]) * config["epochs"] + 1
config["test_interval"] = len(dset_loaders["source"])
config["snapshot_interval"] = len(
dset_loaders["source"]) * config["epochs"] * .25
config["log_iter"] = len(dset_loaders["source"])
#print the configuration you are using
config["out_file"].write("config: {}\n".format(config))
config["out_file"].flush()
# set up early stop
early_stop_engine = EarlyStopping(config["early_stop_patience"])
## set base network
net_config = config["network"]
base_network = net_config["name"](**net_config["params"])
use_gpu = torch.cuda.is_available()
if use_gpu:
base_network = base_network.cuda()
## add additional network for some methods
ad_net = network.AdversarialNetwork(base_network.output_num())
gradient_reverse_layer = network.AdversarialLayer(
high_value=config["high"]) #,
#max_iter_value=config["num_iterations"])
if use_gpu:
ad_net = ad_net.cuda()
## collect parameters
if "DeepMerge" in args.net:
parameter_list = [{
"params": base_network.parameters(),
"lr_mult": 1,
'decay_mult': 2
}]
parameter_list.append({
"params": ad_net.parameters(),
"lr_mult": .1,
'decay_mult': 2
})
parameter_list.append({
"params": center_criterion.parameters(),
"lr_mult": 10,
'decay_mult': 1
})
elif "ResNet18" in args.net:
parameter_list = [{
"params": base_network.parameters(),
"lr_mult": 1,
'decay_mult': 2
}]
parameter_list.append({
"params": ad_net.parameters(),
"lr_mult": .1,
'decay_mult': 2
})
parameter_list.append({
"params": center_criterion.parameters(),
"lr_mult": 10,
'decay_mult': 1
})
if net_config["params"]["new_cls"]:
if net_config["params"]["use_bottleneck"]:
parameter_list = [{"params":base_network.feature_layers.parameters(), "lr_mult":1, 'decay_mult':2}, \
{"params":base_network.bottleneck.parameters(), "lr_mult":10, 'decay_mult':2}, \
{"params":base_network.fc.parameters(), "lr_mult":10, 'decay_mult':2}]
parameter_list.append({
"params": ad_net.parameters(),
"lr_mult": config["ad_net_mult_lr"],
'decay_mult': 2
})
parameter_list.append({
"params": center_criterion.parameters(),
"lr_mult": 10,
'decay_mult': 1
})
else:
parameter_list = [{"params":base_network.feature_layers.parameters(), "lr_mult":1, 'decay_mult':2}, \
{"params":base_network.fc.parameters(), "lr_mult":10, 'decay_mult':2}]
parameter_list.append({
"params": ad_net.parameters(),
"lr_mult": config["ad_net_mult_lr"],
'decay_mult': 2
})
parameter_list.append({
"params": center_criterion.parameters(),
"lr_mult": 10,
'decay_mult': 1
})
else:
parameter_list = [{
"params": base_network.parameters(),
"lr_mult": 1,
'decay_mult': 2
}]
parameter_list.append({
"params": ad_net.parameters(),
"lr_mult": config["ad_net_mult_lr"],
'decay_mult': 2
})
parameter_list.append({
"params": center_criterion.parameters(),
"lr_mult": 10,
'decay_mult': 1
})
#Should I put lr_mult here as 1 for DeepMerge too? Probably!
## set optimizer
optimizer_config = config["optimizer"]
optimizer = optim_dict[optimizer_config["type"]](parameter_list, \
**(optimizer_config["optim_params"]))
param_lr = []
for param_group in optimizer.param_groups:
param_lr.append(param_group["lr"])
schedule_param = optimizer_config["lr_param"]
lr_scheduler = lr_schedule.schedule_dict[optimizer_config["lr_type"]]
scan_lr = []
scan_loss = []
## train
len_train_source = len(dset_loaders["source"])
len_train_target = len(dset_loaders["target"])
len_valid_source = len(dset_loaders["source_valid"])
len_valid_target = len(dset_loaders["target_valid"])
transfer_loss_value = classifier_loss_value = total_loss_value = 0.0
best_acc = 0.0
for i in range(config["num_iterations"]):
if i % config["test_interval"] == 0:
base_network.train(False)
if config['loss']['ly_type'] == "cosine":
temp_acc, _ = image_classification_test(dset_loaders, 'source_valid', \
base_network, gpu=use_gpu, verbose = False, save_where = None)
train_acc, _ = image_classification_test(dset_loaders, 'source', \
base_network, gpu=use_gpu, verbose = False, save_where = None)
elif config['loss']['ly_type'] == "euclidean":
temp_acc, _ = distance_classification_test(dset_loaders, 'source_valid', \
base_network, center_criterion.centers.detach(), gpu=use_gpu, verbose = False, save_where = None)
train_acc, _ = distance_classification_test(
dset_loaders,
'source',
base_network,
center_criterion.centers.detach(),
gpu=use_gpu,
verbose=False,
save_where=None)
else:
raise ValueError("no test method for cls loss: {}".format(
config['loss']['ly_type']))
snapshot_obj = {
'epoch': i / len(dset_loaders["source"]),
"base_network": base_network.state_dict(),
'valid accuracy': temp_acc,
'train accuracy': train_acc,
}
if (i + 1) % config["snapshot_interval"] == 0:
torch.save(
snapshot_obj,
osp.join(
config["output_path"], "epoch_{}_model.pth.tar".format(
i / len(dset_loaders["source"]))))
if config["loss"]["loss_name"] != "laplacian" and config["loss"][
"ly_type"] == "euclidean":
snapshot_obj['center_criterion'] = center_criterion.state_dict(
)
if temp_acc > best_acc:
best_acc = temp_acc
# save best model
torch.save(
snapshot_obj,
osp.join(config["output_path"], "best_model.pth.tar"))
log_str = "epoch: {}, {} validation accuracy: {:.5f}, {} training accuracy: {:.5f}\n".format(
i / len(dset_loaders["source"]), config['loss']['ly_type'],
temp_acc, config['loss']['ly_type'], train_acc)
config["out_file"].write(log_str)
config["out_file"].flush()
writer.add_scalar("validation accuracy", temp_acc,
i / len(dset_loaders["source"]))
writer.add_scalar("training accuracy", train_acc,
i / len(dset_loaders["source"]))
## train one iter
base_network.train(True)
if i % config["log_iter"] == 0:
optimizer = lr_scheduler(param_lr, optimizer, i,
config["log_iter"], config["frozen lr"],
config["cycle_length"], **schedule_param)
if config["optimizer"]["lr_type"] == "one-cycle":
optimizer = lr_scheduler(param_lr, optimizer, i,
config["log_iter"], config["frozen lr"],
config["cycle_length"], **schedule_param)
if config["optimizer"]["lr_type"] == "linear":
optimizer = lr_scheduler(param_lr, optimizer, i,
config["log_iter"], config["frozen lr"],
config["cycle_length"], **schedule_param)
optim = optimizer.state_dict()
scan_lr.append(optim['param_groups'][0]['lr'])
optimizer.zero_grad()
try:
inputs_source, labels_source = iter(dset_loaders["source"]).next()
inputs_target, labels_target = iter(dset_loaders["target"]).next()
except StopIteration:
iter(dset_loaders["source"])
iter(dset_loaders["target"])
if use_gpu:
inputs_source, inputs_target, labels_source = \
Variable(inputs_source).cuda(), Variable(inputs_target).cuda(), \
Variable(labels_source).cuda()
else:
inputs_source, inputs_target, labels_source = Variable(inputs_source), \
Variable(inputs_target), Variable(labels_source)
inputs = torch.cat((inputs_source, inputs_target), dim=0)
source_batch_size = inputs_source.size(0)
if config['loss']['ly_type'] == 'cosine':
features, logits = base_network(inputs)
source_logits = logits.narrow(0, 0, source_batch_size)
elif config['loss']['ly_type'] == 'euclidean':
features, _ = base_network(inputs)
logits = -1.0 * loss.distance_to_centroids(
features, center_criterion.centers.detach())
source_logits = logits.narrow(0, 0, source_batch_size)
ad_net.train(True)
weight_ad = torch.ones(inputs.size(0))
transfer_loss = transfer_criterion(features, ad_net, gradient_reverse_layer, \
weight_ad, use_gpu)
ad_out, _ = ad_net(features.detach())
ad_acc, source_acc_ad, target_acc_ad = domain_cls_accuracy(ad_out)
# source domain classification task loss
classifier_loss = class_criterion(source_logits, labels_source.long())
# fisher loss on labeled source domain
if config["fisher_or_no"] == 'no':
total_loss = loss_params["trade_off"] * transfer_loss \
+ classifier_loss
scan_loss.append(total_loss.cpu().float().item())
total_loss.backward()
######################################
# Plot embeddings periodically.
if args.blobs is not None and i / len(
dset_loaders["source"]) % 50 == 0:
visualizePerformance(base_network,
dset_loaders["source"],
dset_loaders["target"],
batch_size=128,
domain_classifier=ad_net,
num_of_samples=100,
imgName='embedding_' +
str(i / len(dset_loaders["source"])),
save_dir=osp.join(config["output_path"],
"blobs"))
##########################################
# if center_grad is not None:
# # clear mmc_loss
# center_criterion.centers.grad.zero_()
# # Manually assign centers gradients other than using autograd
# center_criterion.centers.backward(center_grad)
optimizer.step()
if i % config["log_iter"] == 0:
if config['lr_scan'] != 'no':
if not osp.exists(
osp.join(config["output_path"],
"learning_rate_scan")):
os.makedirs(
osp.join(config["output_path"],
"learning_rate_scan"))
plot_learning_rate_scan(
scan_lr, scan_loss, i / len(dset_loaders["source"]),
osp.join(config["output_path"], "learning_rate_scan"))
if config['grad_vis'] != 'no':
if not osp.exists(
osp.join(config["output_path"], "gradients")):
os.makedirs(
osp.join(config["output_path"], "gradients"))
plot_grad_flow(
osp.join(config["output_path"], "gradients"),
i / len(dset_loaders["source"]),
base_network.named_parameters())
config['out_file'].write(
'epoch {}: train total loss={:0.4f}, train transfer loss={:0.4f}, train classifier loss={:0.4f},'
'train source+target domain accuracy={:0.4f}, train source domain accuracy={:0.4f}, train target domain accuracy={:0.4f}\n'
.format(
i / len(dset_loaders["source"]),
total_loss.data.cpu().float().item(),
transfer_loss.data.cpu().float().item(),
classifier_loss.data.cpu().float().item(),
ad_acc,
source_acc_ad,
target_acc_ad,
))
config['out_file'].flush()
writer.add_scalar("training total loss",
total_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar("training classifier loss",
classifier_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar("training transfer loss",
transfer_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar("training source+target domain accuracy",
ad_acc, i / len(dset_loaders["source"]))
writer.add_scalar("training source domain accuracy",
source_acc_ad,
i / len(dset_loaders["source"]))
writer.add_scalar("training target domain accuracy",
target_acc_ad,
i / len(dset_loaders["source"]))
#attempted validation step
for j in range(0, len(dset_loaders["source_valid"])):
base_network.train(False)
with torch.no_grad():
try:
inputs_valid_source, labels_valid_source = iter(
dset_loaders["source_valid"]).next()
inputs_valid_target, labels_valid_target = iter(
dset_loaders["target_valid"]).next()
except StopIteration:
iter(dset_loaders["source_valid"])
iter(dset_loaders["target_valid"])
if use_gpu:
inputs_valid_source, inputs_valid_target, labels_valid_source = \
Variable(inputs_valid_source).cuda(), Variable(inputs_valid_target).cuda(), \
Variable(labels_valid_source).cuda()
else:
inputs_valid_source, inputs_valid_target, labels_valid_source = Variable(inputs_valid_source), \
Variable(inputs_valid_target), Variable(labels_valid_source)
valid_inputs = torch.cat(
(inputs_valid_source, inputs_valid_target), dim=0)
valid_source_batch_size = inputs_valid_source.size(0)
if config['loss']['ly_type'] == 'cosine':
features, logits = base_network(valid_inputs)
source_logits = logits.narrow(
0, 0, valid_source_batch_size)
elif config['loss']['ly_type'] == 'euclidean':
features, _ = base_network(valid_inputs)
logits = -1.0 * loss.distance_to_centroids(
features, center_criterion.centers.detach())
source_logits = logits.narrow(
0, 0, valid_source_batch_size)
ad_net.train(False)
weight_ad = torch.ones(valid_inputs.size(0))
transfer_loss = transfer_criterion(features, ad_net, gradient_reverse_layer, \
weight_ad, use_gpu)
ad_out, _ = ad_net(features.detach())
ad_acc, source_acc_ad, target_acc_ad = domain_cls_accuracy(
ad_out)
# source domain classification task loss
classifier_loss = class_criterion(
source_logits, labels_valid_source.long())
#if config["fisher_or_no"] == 'no':
total_loss = loss_params["trade_off"] * transfer_loss \
+ classifier_loss
if j % len(dset_loaders["source_valid"]) == 0:
config['out_file'].write(
'epoch {}: valid total loss={:0.4f}, valid transfer loss={:0.4f}, valid classifier loss={:0.4f},'
'valid source+target domain accuracy={:0.4f}, valid source domain accuracy={:0.4f}, valid target domain accuracy={:0.4f}\n'
.format(
i / len(dset_loaders["source"]),
total_loss.data.cpu().float().item(),
transfer_loss.data.cpu().float().item(),
classifier_loss.data.cpu().float().item(),
ad_acc,
source_acc_ad,
target_acc_ad,
))
config['out_file'].flush()
writer.add_scalar("validation total loss",
total_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar(
"validation classifier loss",
classifier_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar(
"validation transfer loss",
transfer_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar(
"validation source+target domain accuracy", ad_acc,
i / len(dset_loaders["source"]))
writer.add_scalar("validation source domain accuracy",
source_acc_ad,
i / len(dset_loaders["source"]))
writer.add_scalar("validation target domain accuracy",
target_acc_ad,
i / len(dset_loaders["source"]))
if early_stop_engine.is_stop_training(
classifier_loss.cpu().float().item()):
config["out_file"].write(
"no improvement after {}, stop training at step {}\n"
.format(config["early_stop_patience"],
i / len(dset_loaders["source"])))
sys.exit()
else:
fisher_loss, fisher_intra_loss, fisher_inter_loss, center_grad = center_criterion(
features.narrow(0, 0, int(inputs.size(0) / 2)),
labels_source,
inter_class=config["loss"]["inter_type"],
intra_loss_weight=loss_params["intra_loss_coef"],
inter_loss_weight=loss_params["inter_loss_coef"])
# entropy minimization loss
em_loss = loss.EntropyLoss(nn.Softmax(dim=1)(logits))
# final loss
total_loss = loss_params["trade_off"] * transfer_loss \
+ fisher_loss \
+ loss_params["em_loss_coef"] * em_loss \
+ classifier_loss
scan_loss.append(total_loss.cpu().float().item())
total_loss.backward()
######################################
# Plot embeddings periodically.
if args.blobs is not None and i / len(
dset_loaders["source"]) % 50 == 0:
visualizePerformance(base_network,
dset_loaders["source"],
dset_loaders["target"],
batch_size=128,
num_of_samples=50,
imgName='embedding_' +
str(i / len(dset_loaders["source"])),
save_dir=osp.join(config["output_path"],
"blobs"))
##########################################
if center_grad is not None:
# clear mmc_loss
center_criterion.centers.grad.zero_()
# Manually assign centers gradients other than using autograd
center_criterion.centers.backward(center_grad)
optimizer.step()
if i % config["log_iter"] == 0:
if config['lr_scan'] != 'no':
if not osp.exists(
osp.join(config["output_path"],
"learning_rate_scan")):
os.makedirs(
osp.join(config["output_path"],
"learning_rate_scan"))
plot_learning_rate_scan(
scan_lr, scan_loss, i / len(dset_loaders["source"]),
osp.join(config["output_path"], "learning_rate_scan"))
if config['grad_vis'] != 'no':
if not osp.exists(
osp.join(config["output_path"], "gradients")):
os.makedirs(
osp.join(config["output_path"], "gradients"))
plot_grad_flow(
osp.join(config["output_path"], "gradients"),
i / len(dset_loaders["source"]),
base_network.named_parameters())
config['out_file'].write(
'epoch {}: train total loss={:0.4f}, train transfer loss={:0.4f}, train classifier loss={:0.4f}, '
'train entropy min loss={:0.4f}, '
'train fisher loss={:0.4f}, train intra-group fisher loss={:0.4f}, train inter-group fisher loss={:0.4f}, '
'train source+target domain accuracy={:0.4f}, train source domain accuracy={:0.4f}, train target domain accuracy={:0.4f}\n'
.format(
i / len(dset_loaders["source"]),
total_loss.data.cpu().float().item(),
transfer_loss.data.cpu().float().item(),
classifier_loss.data.cpu().float().item(),
em_loss.data.cpu().float().item(),
fisher_loss.cpu().float().item(),
fisher_intra_loss.cpu().float().item(),
fisher_inter_loss.cpu().float().item(),
ad_acc,
source_acc_ad,
target_acc_ad,
))
config['out_file'].flush()
writer.add_scalar("training total loss",
total_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar("training classifier loss",
classifier_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar("training transfer loss",
transfer_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar("training entropy minimization loss",
em_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar("training total fisher loss",
fisher_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar("training intra-group fisher",
fisher_intra_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar("training inter-group fisher",
fisher_inter_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar("training source+target domain accuracy",
ad_acc, i / len(dset_loaders["source"]))
writer.add_scalar("training source domain accuracy",
source_acc_ad,
i / len(dset_loaders["source"]))
writer.add_scalar("training target domain accuracy",
target_acc_ad,
i / len(dset_loaders["source"]))
#attempted validation step
for j in range(0, len(dset_loaders["source_valid"])):
base_network.train(False)
with torch.no_grad():
try:
inputs_valid_source, labels_valid_source = iter(
dset_loaders["source_valid"]).next()
inputs_valid_target, labels_valid_target = iter(
dset_loaders["target_valid"]).next()
except StopIteration:
iter(dset_loaders["source_valid"])
iter(dset_loaders["target_valid"])
if use_gpu:
inputs_valid_source, inputs_valid_target, labels_valid_source = \
Variable(inputs_valid_source).cuda(), Variable(inputs_valid_target).cuda(), \
Variable(labels_valid_source).cuda()
else:
inputs_valid_source, inputs_valid_target, labels_valid_source = Variable(inputs_valid_source), \
Variable(inputs_valid_target), Variable(labels_valid_source)
valid_inputs = torch.cat(
(inputs_valid_source, inputs_valid_target), dim=0)
valid_source_batch_size = inputs_valid_source.size(0)
if config['loss']['ly_type'] == 'cosine':
features, logits = base_network(valid_inputs)
source_logits = logits.narrow(
0, 0, valid_source_batch_size)
elif config['loss']['ly_type'] == 'euclidean':
features, _ = base_network(valid_inputs)
logits = -1.0 * loss.distance_to_centroids(
features, center_criterion.centers.detach())
source_logits = logits.narrow(
0, 0, valid_source_batch_size)
ad_net.train(False)
weight_ad = torch.ones(valid_inputs.size(0))
transfer_loss = transfer_criterion(features, ad_net, gradient_reverse_layer, \
weight_ad, use_gpu)
ad_out, _ = ad_net(features.detach())
ad_acc, source_acc_ad, target_acc_ad = domain_cls_accuracy(
ad_out)
# source domain classification task loss
classifier_loss = class_criterion(
source_logits, labels_valid_source.long())
# fisher loss on labeled source domain
fisher_loss, fisher_intra_loss, fisher_inter_loss, center_grad = center_criterion(
features.narrow(0, 0,
int(valid_inputs.size(0) / 2)),
labels_valid_source,
inter_class=loss_params["inter_type"],
intra_loss_weight=loss_params["intra_loss_coef"],
inter_loss_weight=loss_params["inter_loss_coef"])
# entropy minimization loss
em_loss = loss.EntropyLoss(nn.Softmax(dim=1)(logits))
# final loss
total_loss = loss_params["trade_off"] * transfer_loss \
+ fisher_loss \
+ loss_params["em_loss_coef"] * em_loss \
+ classifier_loss
if j % len(dset_loaders["source_valid"]) == 0:
config['out_file'].write(
'epoch {}: valid total loss={:0.4f}, valid transfer loss={:0.4f}, valid classifier loss={:0.4f}, '
'valid entropy min loss={:0.4f}, '
'valid fisher loss={:0.4f}, valid intra-group fisher loss={:0.4f}, valid inter-group fisher loss={:0.4f}, '
'valid source+target domain accuracy={:0.4f}, valid source domain accuracy={:0.4f}, valid target domain accuracy={:0.4f}\n'
.format(
i / len(dset_loaders["source"]),
total_loss.data.cpu().float().item(),
transfer_loss.data.cpu().float().item(),
classifier_loss.data.cpu().float().item(),
em_loss.data.cpu().float().item(),
fisher_loss.cpu().float().item(),
fisher_intra_loss.cpu().float().item(),
fisher_inter_loss.cpu().float().item(),
ad_acc,
source_acc_ad,
target_acc_ad,
))
config['out_file'].flush()
writer.add_scalar("validation total loss",
total_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar(
"validation classifier loss",
classifier_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar(
"validation entropy minimization loss",
em_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar(
"validation transfer loss",
transfer_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar(
"validation entropy minimization loss",
em_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar(
"validation total fisher loss",
fisher_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar(
"validation intra-group fisher",
fisher_intra_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar(
"validation inter-group fisher",
fisher_inter_loss.data.cpu().float().item(),
i / len(dset_loaders["source"]))
writer.add_scalar(
"validation source+target domain accuracy", ad_acc,
i / len(dset_loaders["source"]))
writer.add_scalar("validation source domain accuracy",
source_acc_ad,
i / len(dset_loaders["source"]))
writer.add_scalar("validation target domain accuracy",
target_acc_ad,
i / len(dset_loaders["source"]))
if early_stop_engine.is_stop_training(
classifier_loss.cpu().float().item()):
config["out_file"].write(
"no improvement after {}, stop training at step {}\n"
.format(config["early_stop_patience"],
i / len(dset_loaders["source"])))
sys.exit()
return best_acc
models = {
'describer': Describer(categorical_dim),
'translator': Translator(categorical_dim),
}
for model in models.values():
model.to(device)
model.apply(weights_init_normal)
optimizers = {
'all': create_optimizer([
models['describer'],
models['translator'],
], lr)
}
losses = {
'entropy': loss.EntropyLoss(categorical_dim),
'classification': loss.ClassificationLoss(),
}
trainer = Trainer(dataloader, dataloader_subset, testloader, models,
optimizers, losses)
for _ in range(n_epochs):
trainer.train()
save_list('data/losses.txt', trainer.losses)
save_list('data/accuracies.txt', trainer.accuracies)
save_list('data/misclassifications.txt', trainer.misclassifications)