def main(args):
set_seed(args.seed)
# Load Checkpoint if exists
start_epoch = 0
if args.load:
try:
print("Loading models: {}".format(args.model_path))
checkpoint = torch.load(args.model_path)
start_epoch = checkpoint['epoch'] + 1
model = checkpoint['model']
model_optimizer = checkpoint['model_optimizer']
print("Model Loaded")
except:
print("Model couldn't be loaded. Aborting")
exit(0)
else:
model = model()
model_optimizer = torch.optim.Adam(params=filter(lambda p: p.requires_grad, model.parameters()), lr=args.lr)
print("Model Loaded")
# Deal With CUDA
if args.cuda:
device = args.cuda_device
cudnn.benchmark = True
if torch.cuda.device_count() > 1:
print("There are", torch.cuda.device_count(), "GPUs!")
model = torch.nn.DataParallel(model)
else:
device = 'cpu'
model = model.to(device)
criterion = nn.CrossEntropyLoss().to(device)
if args.do_eval:
print("Loading Validation Dataset")
val_dataset = CustomDataset(args.data_dir, 'validation')
val_loader = torch.utils.data.DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True)
print("There are {} validation data examples".format(len(val_dataset)))
print("Evaluating Model")
validation_score = validate(args)
print("Model performance is {}".format(validation_score))
if args.do_train:
validation_score, is_best = 0, False #in case we arent evaluating
print("Loading Training Dataset")
train_dataset = CustomDataset(args.data_dir, 'training')
print("There are {} training data examples".format(len(train_dataset)))
train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=True)
print("Traing model")
for epoch in range(start_epoch, args.epochs):
print("Starting epoch {}".format(epoch))
train(args)
if args.do_eval:
validation_score = validate(args)
is_best= False
if validation_score > best_validation_score:
is_best = True
best_validation_score = validation_score
save_checkpoint(args, epoch, validation_score, is_best)
def main(args):
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model = load_model(base_model_path=args.base_model_path,
model_checkpoint=args.model_checkpoint,
device=device)
dataset = CustomDataset(root=args.input_image_folder,
transform=get_transforms(),
return_paths=True)
dataloader = DataLoader(dataset=dataset,
shuffle=False,
batch_size=32,
pin_memory=True,
num_workers=8)
output_folder = './classification_output_files'
if not os.path.exists(output_folder):
os.makedirs(output_folder)
label_file = open(
os.path.join(
output_folder,
'correctly_classified_images_' + args.output_filename + '.txt'),
'w')
results_file_path = os.path.join(
output_folder, 'classified_images_' + args.output_filename + '.csv')
tot = 0
correct_ = 0
results = pd.DataFrame()
progress_bar = tqdm(dataloader,
total=len(dataloader),
desc='Classifying: {}'.format(
args.inputImageFolder.split('/')[-1]))
for x, l, p in progress_bar:
tot += x.shape[0]
x = x.to(device)
y_hat = model(x, True).max(-1)[1].cpu()
correct = torch.where(y_hat == l)[0]
correct_ += y_hat.eq(l).sum().item()
progress_bar.set_postfix(accuracy=correct_)
[label_file.write(p[idx] + '\n') for idx in correct]
for i in range(len(l)):
result = pd.DataFrame(dict(label=l[i].item(),
predicted=y_hat[i].item()),
index=[0])
results = results.append(result, ignore_index=True)
results.to_csv(results_file_path, index=False)
label_file.close()
print("Accuracy: {:.2f}%".format(correct_ / tot * 100))
def test():
print("=== Test ===")
args = get_args()
print(args)
data_dir = f"./../../asset/{args.dataset}/"
if args.train :
test_labels, test_texts = read_train_data(data_dir)
else :
test_labels, test_texts = read_test_data(data_dir)
# test_texts = list(test_texts)[:100]
# test_labels = list(test_labels)[:100]
test_texts = list(test_texts)
test_labels = list(test_labels)
model_name = args.model
tokenizer = AutoTokenizer.from_pretrained(model_name)
test_encodings = tokenizer(
test_texts, truncation=True, padding=True, max_length=512)
test_dataset = CustomDataset(test_encodings, test_labels)
checkpoint_dir = f"./models/{args.task}/{args.model}/"
best_checkpoint = find_best_checkpoint(checkpoint_dir)
model = AutoModelForSequenceClassification.from_pretrained(best_checkpoint)
test_trainer = Trainer(model)
test_loader = DataLoader(
test_dataset, batch_size=args.batch_size, shuffle=False)
raw_pred, _, _ = test_trainer.prediction_loop(
test_loader, description="prediction")
# Preprocess raw predictions
y_pred = np.argmax(raw_pred, axis=1)
metrics = compute_metrics(y_pred, test_labels)
print(metrics)
if args.train :
fpath = os.path.join(data_dir, f"train-predictions/{args.model}.pkl")
else :
fpath = os.path.join(data_dir, f"predictions/{args.model}.pkl")
parent_dir = "/".join(str(fpath).split('/')[:-1])
if not os.path.exists(parent_dir):
os.makedirs(parent_dir)
with open(fpath, 'wb') as f:
pickle.dump(y_pred, f)
def bertPreprocessing(dataset_name, data_df, MAX_LEN=128, save_all=True):
"""
Dataset preparation for Bert Model. It is splitted (0.8 train, 0.1 valid and 0.1 test) and sets are returned. Every
set is a CustomDataset class (see utils.py) that return data in Bert format.
:param dataset_name: string of dataset name.
:param data_df: dataset in dataframe pandas format.
:param MAX_LEN: it represents total words represented in bert encoding (other words will be ignored).
:param save_all: boolean that specifies if save all data for time saving before training or network evaluating.
:return: training_set, validation_set, test_set in CustomDataset format.
"""
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
train_size = 0.8
train_dataset = data_df.sample(frac=train_size,
random_state=200).reset_index(drop=True)
tmp_dataset = data_df.drop(train_dataset.index).reset_index(drop=True)
test_dataset = tmp_dataset.sample(frac=0.5,
random_state=200).reset_index(drop=True)
val_dataset = tmp_dataset.drop(test_dataset.index).reset_index(drop=True)
print("FULL Dataset: {}".format(data_df.shape))
print("TRAIN Dataset: {}".format(train_dataset.shape))
print("TEST Dataset: {}".format(test_dataset.shape))
print("VALID Dataset: {}".format(val_dataset.shape))
training_set = CustomDataset(train_dataset, tokenizer, MAX_LEN)
validation_set = CustomDataset(val_dataset, tokenizer, MAX_LEN)
test_set = CustomDataset(test_dataset, tokenizer, MAX_LEN)
if save_all is True:
os.makedirs(os.path.dirname('datasets/' + dataset_name +
'_bert_cleaned.txt'),
exist_ok=True)
with open('datasets/' + dataset_name + '_bert_cleaned.txt', 'wb') as f:
pickle.dump([training_set, validation_set, test_set, MAX_LEN], f)
return training_set, validation_set, test_set
def bertPredict(dataset_name, n_classes, model_path, text, label):
"""
This function predicts label of instance of a bert pretrained dataset, using cpu, printing results.
:param dataset_name: string of dataset name.
:param n_classes: int of number of dataset classes.
:param model_path: path of saved pytorch model fine tuned bert network.
:param text: string to classify.
:param label: true label (int) associated to the text in input.
:return: None
"""
device = torch.device('cpu')
MAX_LEN = 128
TEST_BATCH_SIZE = 8
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
pred_data = pd.DataFrame(data={'text': [text], 'label': [label]})
predict = CustomDataset(pred_data, tokenizer, MAX_LEN)
test_params = {'batch_size': TEST_BATCH_SIZE,
'shuffle': True,
'num_workers': 0
}
testing_loader = DataLoader(predict, **test_params)
model = BertModel(n_classes=n_classes, dropout=0.3)
model.load_state_dict(torch.load(model_path, map_location=device))
model.eval()
for batch in testing_loader:
ids = batch['ids']
mask = batch['mask']
token_type_ids = batch['token_type_ids']
output = model(ids, mask, token_type_ids)
output = torch.softmax(output, dim=1).detach().numpy()
output = np.array(output)
print(output)
print(text)
print("True Label: {:}".format(label))
print("Predicted Label: {:}".format(output.argmax(axis=1) + 1))
from Multimodal import AskAttendAnswer
from Image_Features import GridFeatures
from Text_Features import TextModel
from utils import CustomDataset
import os
import torch
if __name__ == '__main__' :
data_path = "/home/alex/Desktop/4-2/Text_VQA/Data/"
ID_path = os.path.join(data_path,"train/train_ids.txt")
json_path = os.path.join(data_path,"train/cleaned.json")
dataset = CustomDataset(data_path,ID_path,json_path,(448,448),set_="train")
# Image Extractor
resnet = GridFeatures("resnet101")
# Text Extractor
bert = TextModel()
#print("Check resnet : {}".format(resnet(torch.randn((1,3,448,448))).size()))
print("Check bert : {}".format(bert(torch.tensor([1,2,4,0,0])).size()))
# embed_dim,img_ft_dims,batch_size,max_seq_len,num_outs
mixer = AskAttendAnswer(300,(2048,14,14),10,64,15000)
text_fts = torch.nn.init.normal_(torch.empty(10,64,300))
img_fts = torch.nn.init.normal_(torch.empty(10,3,448,448))
mixer.combine(text_fts,img_fts)
save_weights_only=False,
mode='auto',
period=1)
#model = resnet.ResnetBuilder.build_resnet_18((18, 32, 32), nb_classes)
network = model.create_model('resnet50',
input_shape=(18, 32, 32),
num_outputs=nb_classes)
network.compile(loss='sparse_categorical_crossentropy',
optimizer=optimizers.SGD(lr=0.1, momentum=0.9, nesterov=True),
metrics=['accuracy'])
mean = unpickle("mean_channal.pkl")
trfs = T.Compose([T.Normalize(mean), T.RandomHorizontalFlip(0.5)])
training_data = CustomDataset('/mnt/img1/yangqh/Germany_cloud/training.h5',
transform=None)
validation_data = CustomDataset('/mnt/img1/yangqh/Germany_cloud/training.h5',
transform=None)
##############此处改成train set 的目录
network.fit_generator(
training_data.load_data(batch_size=batch_size),
steps_per_epoch=len(training_data) // batch_size,
validation_data=validation_data.load_data(batch_size=batch_size),
validation_steps=len(validation_data) // batch_size,
epochs=nb_epoch,
verbose=1,
max_q_size=100,
callbacks=[checkpoint, lr_reducer, early_stopper, csv_logger])
network.save('final.h5')
from torch.utils.data import DataLoader
from utils import CustomDataset
dataset = CustomDataset()
dataloader = DataLoader(
dataset,
batch_size=2,
shuffle=True,
)
def main(args):
# Setting
warnings.simplefilter("ignore", UserWarning)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# Args Parser
hj_method = args.hj_method
kr_method = args.kr_method
batch_size = args.batch_size
beam_size = args.beam_size
hidden_size = args.hidden_size
embed_size = args.embed_size
vocab_size = args.vocab_size
max_len = args.max_len
padding_index = args.pad_id
n_layers = args.n_layers
stop_ix = args.stop_ix
# Load saved model & Word2vec
save_path = 'save_{}_{}_{}_maxlen_{}'.format(vocab_size, hj_method,
kr_method, max_len)
save_list = sorted(glob.glob(f'./save/{save_path}/*.*'))
save_pt = save_list[-1]
print('Will load {} pt file...'.format(save_pt))
word2vec_hj = Word2Vec.load('./w2v/word2vec_hj_{}_{}.model'.format(
vocab_size, hj_method))
# SentencePiece model load
spm_kr = spm.SentencePieceProcessor()
spm_kr.Load("./spm/m_korean_{}.model".format(vocab_size))
# Test data load
with open('./test_dat.pkl', 'rb') as f:
test_dat = pickle.load(f)
test_dataset = CustomDataset(test_dat['test_hanja'],
test_dat['test_korean'])
test_loader = getDataLoader(test_dataset,
pad_index=padding_index,
shuffle=False,
batch_size=batch_size)
# Model load
print('Model loading...')
encoder = Encoder(vocab_size,
embed_size,
hidden_size,
word2vec_hj,
n_layers=n_layers,
padding_index=padding_index)
decoder = Decoder(embed_size,
hidden_size,
vocab_size,
n_layers=n_layers,
padding_index=padding_index)
seq2seq = Seq2Seq(encoder, decoder, beam_size).cuda()
#optimizer = optim.Adam(seq2seq.parameters(), lr=lr, weight_decay=w_decay)
#scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=scheduler_step_size, gamma=lr_decay)
print(seq2seq)
print('Testing...')
start_time = time.time()
results = test(seq2seq,
test_loader,
vocab_size,
load_pt=save_pt,
stop_ix=stop_ix)
print(time.time() - start_time)
print('Done!')
print("Decoding...")
pred_list = list()
for result_text in tqdm(results):
text = torch.Tensor(result_text).squeeze().tolist()
text = [int(x) for x in text]
prediction_sentence = spm_kr.decode_ids(
text).strip() # Decode with strip
pred_list.append(prediction_sentence)
ref_list = list()
for ref_text in tqdm(test_dat['test_korean'][:stop_ix]):
ref_list.append(spm_kr.decode_ids(ref_text).strip())
print('Done!')
with open(f'./save/{save_path}/test_result.pkl', 'wb') as f:
pickle.dump({
'pred': pred_list,
'reference': ref_list,
}, f)
print('Save file; /test_dat.pkl')
# Calculate BLEU Score
print('Calculate BLEU4, METEOR, Rogue-L...')
chencherry = SmoothingFunction()
bleu4 = corpus_bleu(test_dat['reference'],
test_dat['pred'],
smoothing_function=chencherry.method4)
print('BLEU Score is {}'.format(bleu4))
# Calculate METEOR Score
meteor = meteor_score(test_dat['reference'], test_dat['pred'])
print('METEOR Score is {}'.format(meteor))
# Calculate Rouge-L Score
r = Rouge()
total_test_length = len(test_dat['reference'])
precision_all = 0
recall_all = 0
f_score_all = 0
for i in range(total_test_length):
[precision, recall, f_score] = r.rouge_l([test_dat['reference'][i]],
[test_dat['pred'][i]])
precision_all += precision
recall_all += recall
f_score_all += f_score
print('Precision : {}'.foramt(round(precision_all / total_test_length, 4)))
print('Recall : {}'.foramt(round(recall_all / total_test_length, 4)))
print('F Score : {}'.foramt(round(f_score_all / total_test_length, 4)))
def predict():
print("=== Predict ===")
args = get_args()
print(args)
if args.bias_type != "":
data_dir = f"./../../data/{args.mutation_tool}/{args.bias_type}/{args.mutant}/"
else:
data_dir = f"./../../data/{args.mutation_tool}/{args.mutant}/"
if args.type == "mutant":
test_labels, test_texts = read_test_data(data_dir)
elif args.type == "original":
generate_original_data(data_dir, mutation_tool=args.mutation_tool)
test_labels, test_texts = read_original_data(data_dir)
else:
raise ValueError("Unknown type that needs to be tested")
# test_texts = list(test_texts)[:100]
# test_labels = list(test_labels)[:100]
test_texts = list(test_texts)
test_labels = list(test_labels)
model_name = args.model
tokenizer = AutoTokenizer.from_pretrained(model_name)
if args.task == "imdb" and args.type == "mutant" and (
args.bias_type == "occupation" or args.bias_type == "country"):
test_encodings = batch_tokenizer(tokenizer,
test_texts,
batch_size=10000)
else:
test_encodings = tokenizer(test_texts,
truncation=True,
padding=True,
max_length=512)
test_dataset = CustomDataset(test_encodings, test_labels)
checkpoint_dir = f"./models/{args.task}/{args.model}/"
best_checkpoint = find_best_checkpoint(checkpoint_dir)
model = AutoModelForSequenceClassification.from_pretrained(best_checkpoint)
test_trainer = Trainer(model)
test_loader = DataLoader(test_dataset,
batch_size=args.batch_size,
shuffle=False)
raw_pred, _, _ = test_trainer.prediction_loop(test_loader,
description="prediction")
# Preprocess raw predictions
y_pred = np.argmax(raw_pred, axis=1)
fpath = os.path.join(data_dir, f"{args.type}-predictions/{args.model}.pkl")
parent_dir = "/".join(str(fpath).split('/')[:-1])
if not os.path.exists(parent_dir):
os.makedirs(parent_dir)
with open(fpath, 'wb') as f:
pickle.dump(y_pred, f)
def fine_tune():
print("=== Fine-tune ===")
args = get_args()
print(args)
if args.task == "imdb":
data_dir = "./../../asset/imdb/"
train_labels, train_texts = read_imdb_train(data_dir)
elif args.task == "twitter_semeval":
data_dir = "./../../asset/twitter_semeval/"
train_labels, train_texts = read_twitter_train(data_dir)
elif args.task == "twitter_s140":
data_dir = "./../../asset/twitter_s140/"
train_labels, train_texts = read_twitter_train(data_dir)
# check_data()
train_texts, val_texts, train_labels, val_labels = train_test_split(
train_texts, train_labels, test_size=args.test_size)
## IF HAVE MUCH TIME, try to increase test size because the fine-tuning run fast
train_texts = list(train_texts)
val_texts = list(val_texts)
train_labels = list(train_labels)
val_labels = list(val_labels)
model_name = args.model
# model_name = "bert-base-cased"
# model_name = "roberta-base"
# model_name = "microsoft/deberta-large-mnli"
# model_name = "bert-base-uncased"
tokenizer = AutoTokenizer.from_pretrained(model_name)
# check_data()
train_encodings = tokenizer(train_texts,
truncation=True,
padding=True,
max_length=512)
val_encodings = tokenizer(val_texts,
truncation=True,
padding=True,
max_length=512)
train_dataset = CustomDataset(train_encodings, train_labels)
val_dataset = CustomDataset(val_encodings, val_labels)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
training_args = TrainingArguments(
# output directory
output_dir=f'./models/{args.task}/{model_name}/',
num_train_epochs=args.epochs, # total number of training epochs
per_device_train_batch_size=args.
train_bs, # batch size per device during training
per_device_eval_batch_size=64, # batch size for evaluation
warmup_steps=args.
warmup_steps, # number of warmup steps for learning rate scheduler
weight_decay=args.weight_decay, # strength of weight decay
# directory for storing logs
logging_dir=f'./logs/{args.task}/{model_name}/',
logging_steps=args.logging_steps,
learning_rate=args.learning_rate,
seed=0,
evaluation_strategy="steps",
eval_steps=args.eval_steps,
save_total_limit=5,
save_steps=args.save_steps,
load_best_model_at_end=True)
# trainer = Trainer(
# # the instantiated 🤗 Transformers model to be trained
# model=model,
# args=training_args, # training arguments, defined above
# train_dataset=train_dataset, # training dataset
# eval_dataset=val_dataset, # evaluation dataset
# compute_metrics=compute_metrics,
# )
trainer = Trainer(
# the instantiated 🤗 Transformers model to be trained
model=model,
args=training_args, # training arguments, defined above
train_dataset=train_dataset, # training dataset
eval_dataset=val_dataset, # evaluation dataset
compute_metrics=compute_metrics,
callbacks=[EarlyStoppingCallback(early_stopping_patience=7)],
)
trainer.train()
args = parse_args()
param = {
'max_depth': 4,
'eta': 1,
'silent': 1,
'objective': 'binary:logistic'
}
param['nthread'] = 12
param['eval_metric'] = 'auc'
name = os.path.join(args.output_root, '{}'.format(args.model_type))
exclude = args.exclude_feature.split(',')
train_dataset = CustomDataset(args.dataset_root, 'train_data.pkl', exclude)
valid_dataset = CustomDataset(args.dataset_root, 'valid_data.pkl', exclude)
test_dataset = CustomDataset(args.dataset_root, 'test_data.pkl', exclude)
dtrain = xgb.DMatrix(np.array(train_dataset.static),
label=train_dataset.label)
dvalid = xgb.DMatrix(np.array(valid_dataset.static),
label=valid_dataset.label)
dtest = xgb.DMatrix(np.array(test_dataset.static),
label=test_dataset.label)
evallist = [(dvalid, 'eval'), (dtrain, 'train')]
bst = xgb.train(param, dtrain, 10, evallist)
pred = bst.predict(dtest)
# Split dataset
X_train, X_test, y_train, y_test = train_test_split(images,
labels,
random_state=42,
shuffle=True,
train_size=0.75,
test_size=0.25)
# Definte transforms
transform = torchvision.transforms.Compose(
[torchvision.transforms.ToTensor()])
# Init DataLoader
loaders = {}
train_dataset = CustomDataset(X_train, y_train, transform)
train_sampler = torch.utils.data.RandomSampler(train_dataset)
train_loader = DataLoader(train_dataset,
batch_size=32,
shuffle=True,
collate_fn=custom_collate_fn,
num_workers=6,
drop_last=True)
eval_dataset = CustomDataset(X_test, y_test, transform)
eval_sampler = torch.utils.data.RandomSampler(eval_dataset)
eval_loader = DataLoader(eval_dataset,
batch_size=32,
shuffle=False,
collate_fn=custom_collate_fn,
num_workers=4,
def main(root_dir=cfg.root_dir,batch_size=cfg.batch_size,lr=cfg.lr,model_name=cfg.model_name,\
weight_decay=cfg.weight_decay,n_epochs=cfg.n_epochs,log_dir=cfg.log_dir,k_fold=cfg.k_fold,\
patience=cfg.patience,steplr_step_size=cfg.steplr_step_size,steplr_gamma=cfg.steplr_gamma,\
save_dir=cfg.save_dir):
#Seed
torch.manual_seed(42)
np.random.seed(42)
# Available CUDA
use_cuda= True if torch.cuda.is_available() else False
device=torch.device('cuda:0' if use_cuda else 'cpu') # CPU or GPU
#Transforms
train_transforms=A.Compose([
A.CenterCrop(230,230),
A.RandomCrop(224,224),
A.ElasticTransform(),
A.IAAPerspective(),
A.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
ToTensorV2(),
])
valid_transforms=A.Compose([
A.CenterCrop(224,224),
A.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
ToTensorV2()
])
#Dataset
dataset=CustomDataset(root_dir,transforms=train_transforms)
# Stratified K-fold Cross Validation
kf=StratifiedKFold(n_splits=k_fold,shuffle=True,random_state=42)
# Tensorboard Writer
writer=SummaryWriter(log_dir)
for n_fold, (train_indices,test_indices) in enumerate(kf.split(dataset.x_data,dataset.y_data),start=1):
print(f'=====Stratified {k_fold}-fold : {n_fold}=====')
#Dataloader
train_sampler=torch.utils.data.SubsetRandomSampler(train_indices)
valid_sampler=torch.utils.data.SubsetRandomSampler(test_indices)
train_loader=torch.utils.data.DataLoader(dataset,batch_size=batch_size,sampler=train_sampler)
valid_loader=torch.utils.data.DataLoader(dataset,batch_size=batch_size,sampler=valid_sampler)
#Model,Criterion,Optimizer,scheduler,regularization
model=Net(model_name).to(device)
criterion=nn.CrossEntropyLoss().to(device)
optimizer=AdamP(model.parameters(),lr=lr)
scheduler_steplr=optim.lr_scheduler.StepLR(optimizer,step_size=steplr_step_size,gamma=steplr_gamma)
regularization=EarlyStopping(patience=patience)
#Train
model.train()
for epoch in range(n_epochs):
print(f'Learning rate : {optimizer.param_groups[0]["lr"]}')
train_metric_monitor=MetricMonitor()
train_stream=tqdm(train_loader)
for batch_idx,sample in enumerate(train_stream,start=1):
img,label=sample['img'].to(device),sample['label'].to(device)
output=model(img)
optimizer.zero_grad()
loss=criterion(output,label)
_,preds=torch.max(output,dim=1)
correct=torch.sum(preds==label.data)
train_metric_monitor.update('Loss',loss.item())
train_metric_monitor.update('Accuracy',100.*correct/len(img))
loss.backward()
optimizer.step()
train_stream.set_description(
f'Train epoch : {epoch} | {train_metric_monitor}'
)
# Valid
valid_metric_monitor=MetricMonitor()
valid_stream=tqdm(valid_loader)
model.eval()
with torch.no_grad():
for batch_idx,sample in enumerate(valid_stream):
img,label=sample['img'].to(device),sample['label'].to(device)
output=model(img)
loss=criterion(output,label)
_,preds=torch.max(output,dim=1)
correct=torch.sum(preds==label.data)
valid_metric_monitor.update('Loss',loss.item())
valid_metric_monitor.update('Accuracy',100.*correct/len(img))
valid_stream.set_description(
f'Test epoch : {epoch} | {valid_metric_moniotr}'
)
# Tensorboard
train_loss=train_metric_monitor.metrics['Loss']['avg']
train_accuracy=train_metric_monitor.metrics['Accuracy']['avg']
valid_loss=valid_metric_monitor.metrics['Loss']['avg']
valid_accuracy=valid_metric_monitor.metrics['Accuracy']['avg']
writer.add_scalars(f'{n_fold}-fold Loss',{'train':train_loss,'valid':valid_loss},epoch)
writer.add_scalars(f'{n_fold}-fold Accuracy',{'train':train_accuracy,'valid':valid_accuracy},epoch)
#Save Model
if regularization.early_stopping:
break
regularization.path=os.path.join(save_dir,f'{n_fold}_fold_{epoch}_epoch.pt')
regularization(val_loss=valid_loss,model=model)
scheduler_steplr.step()
writer.close()
def train(config):
if not os.path.exists(config.out):
os.makedirs(config.out)
comp_transform = transforms.Compose([
transforms.CenterCrop(config.crop),
transforms.Resize(config.resize),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
domain_a_train = CustomDataset(os.path.join(config.root, 'trainA.txt'), transform=comp_transform)
domain_b_train = CustomDataset(os.path.join(config.root, 'trainB.txt'), transform=comp_transform)
a_label = torch.full((config.bs,), 1)
b_label = torch.full((config.bs,), 0)
b_separate = torch.full((config.bs,
config.sep,
config.resize // (2 ** (config.n_blocks + 1)),
config.resize // (2 ** (config.n_blocks + 1))), 0)
# build networks
e1 = E1(sep=config.sep, size=config.resize)
e2 = E2(n_feats=config.n_tot_feats, sep=config.sep)
decoder = Decoder(n_feats=config.n_tot_feats)
disc = Disc(size=config.resize, sep=config.sep)
rho_clipper = RhoClipper(0., 1.)
mse = nn.MSELoss()
bce = nn.BCELoss()
if torch.cuda.is_available():
e1 = e1.cuda()
e2 = e2.cuda()
decoder = decoder.cuda()
disc = disc.cuda()
a_label = a_label.cuda()
b_label = b_label.cuda()
b_separate = b_separate.cuda()
mse = mse.cuda()
bce = bce.cuda()
ae_params = list(e1.parameters()) + list(e2.parameters()) + list(decoder.parameters())
ae_optimizer = optim.Adam(ae_params, lr=config.lr,
betas=(config.beta1, config.beta2), eps=config.eps)
disc_params = disc.parameters()
disc_optimizer = optim.Adam(disc_params, lr=config.d_lr,
betas=(config.beta1, config.beta2), eps=config.eps)
_iter: int = 0
if config.load != '':
save_file = os.path.join(config.load, 'checkpoint')
_iter = load_model(save_file, e1, e2, decoder, ae_optimizer, disc, disc_optimizer)
e1 = e1.train()
e2 = e2.train()
decoder = decoder.train()
disc = disc.train()
print('[*] Started training...')
while True:
domain_a_loader = torch.utils.data.DataLoader(domain_a_train, batch_size=config.bs,
shuffle=True, num_workers=config.n_threads)
domain_b_loader = torch.utils.data.DataLoader(domain_b_train, batch_size=config.bs,
shuffle=True, num_workers=config.n_threads)
if _iter >= config.iters:
break
for domain_a_img, domain_b_img in zip(domain_a_loader, domain_b_loader):
if domain_a_img.size(0) != config.bs or domain_b_img.size(0) != config.bs:
break
domain_a_img = Variable(domain_a_img)
domain_b_img = Variable(domain_b_img)
if torch.cuda.is_available():
domain_a_img = domain_a_img.cuda()
domain_b_img = domain_b_img.cuda()
domain_a_img = domain_a_img.view((-1, 3, config.resize, config.resize))
domain_b_img = domain_b_img.view((-1, 3, config.resize, config.resize))
ae_optimizer.zero_grad()
a_common = e1(domain_a_img)
a_separate = e2(domain_a_img)
a_encoding = torch.cat([a_common, a_separate], dim=1)
b_common = e1(domain_b_img)
b_encoding = torch.cat([b_common, b_separate], dim=1)
a_decoding = decoder(a_encoding)
b_decoding = decoder(b_encoding)
g_loss = mse(a_decoding, domain_a_img) + mse(b_decoding, domain_b_img)
preds_a = disc(a_common)
preds_b = disc(b_common)
g_loss += config.adv_weight * (bce(preds_a, b_label) + bce(preds_b, b_label))
g_loss.backward()
torch.nn.utils.clip_grad_norm_(ae_params, 5.)
ae_optimizer.step()
disc_optimizer.zero_grad()
a_common = e1(domain_a_img)
b_common = e1(domain_b_img)
disc_a = disc(a_common)
disc_b = disc(b_common)
d_loss = bce(disc_a, a_label) + bce(disc_b, b_label)
d_loss.backward()
torch.nn.utils.clip_grad_norm_(disc_params, 5.)
disc_optimizer.step()
decoder.apply(rho_clipper)
if _iter % config.progress_iter == 0:
print('[*] [%07d/%07d] d_loss : %.4f, g_loss : %.4f' %
(_iter, config.iters, d_loss, g_loss))
if _iter % config.display_iter == 0:
e1 = e1.eval()
e2 = e2.eval()
decoder = decoder.eval()
save_images(config, e1, e2, decoder, _iter)
e1 = e1.train()
e2 = e2.train()
decoder = decoder.train()
if _iter % config.save_iter == 0:
save_file = os.path.join(config.out, 'checkpoint')
save_model(save_file, e1, e2, decoder, ae_optimizer, disc, disc_optimizer, _iter)
_iter += 1
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.RandomRotation(20),
transforms.ColorJitter(brightness=0.1, contrast=0.1, hue=0.1),
transforms.ToTensor(),
transforms.Normalize(normMean, normStd)
])
# define the transformation of the val images
val_transform = transforms.Compose([
transforms.Resize((input_size, input_size)),
transforms.ToTensor(),
transforms.Normalize(normMean, normStd)
])
# Define training set using train_df and our defined transitions (train_transform)
training_set = CustomDataset(df_train, transform=train_transform)
train_loader = DataLoader(training_set,
batch_size=32,
shuffle=True,
num_workers=4)
# Same for validation set:
validation_set = CustomDataset(df_val, transform=train_transform)
val_loader = DataLoader(validation_set,
batch_size=32,
shuffle=False,
num_workers=4)
model.to(device)
optimizer = optim.Adam(model.parameters(), lr=1e-3)
criterion = nn.CrossEntropyLoss().to(device)
def train(args):
if not os.path.exists(args.out):
os.makedirs(args.out)
_iter = 0
comp_transformA = transforms.Compose([
transforms.CenterCrop(args.cropA),
transforms.Resize(args.resize),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
comp_transformB = transforms.Compose([
transforms.CenterCrop(args.cropB),
transforms.Resize(args.resize),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
domA_train = CustomDataset(os.path.join(args.root, 'trainA.txt'), transform=comp_transformA)
domB_train = CustomDataset(os.path.join(args.root, 'trainB.txt'), transform=comp_transformB)
A_label = torch.full((args.bs,), 1)
B_label = torch.full((args.bs,), 0)
B_separate = torch.full((args.bs, args.sep * (args.resize // 64) * (args.resize // 64)), 0)
e1 = E1(args.sep, args.resize // 64)
e2 = E2(args.sep, args.resize // 64)
decoder = Decoder(args.resize // 64)
disc = Disc(args.sep, args.resize // 64)
mse = nn.MSELoss()
bce = nn.BCELoss()
if torch.cuda.is_available():
e1 = e1.cuda()
e2 = e2.cuda()
decoder = decoder.cuda()
disc = disc.cuda()
A_label = A_label.cuda()
B_label = B_label.cuda()
B_separate = B_separate.cuda()
mse = mse.cuda()
bce = bce.cuda()
ae_params = list(e1.parameters()) + list(e2.parameters()) + list(decoder.parameters())
ae_optimizer = optim.Adam(ae_params, lr=args.lr, betas=(0.5, 0.999))
disc_params = disc.parameters()
disc_optimizer = optim.Adam(disc_params, lr=args.disclr, betas=(0.5, 0.999))
if args.load != '':
save_file = os.path.join(args.load, 'checkpoint')
_iter = load_model(save_file, e1, e2, decoder, ae_optimizer, disc, disc_optimizer)
e1 = e1.train()
e2 = e2.train()
decoder = decoder.train()
disc = disc.train()
print('Started training...')
while True:
domA_loader = torch.utils.data.DataLoader(domA_train, batch_size=args.bs,
shuffle=True, num_workers=6)
domB_loader = torch.utils.data.DataLoader(domB_train, batch_size=args.bs,
shuffle=True, num_workers=6)
if _iter >= args.iters:
break
for domA_img, domB_img in zip(domA_loader, domB_loader):
if domA_img.size(0) != args.bs or domB_img.size(0) != args.bs:
break
domA_img = Variable(domA_img)
domB_img = Variable(domB_img)
if torch.cuda.is_available():
domA_img = domA_img.cuda()
domB_img = domB_img.cuda()
domA_img = domA_img.view((-1, 3, args.resize, args.resize))
domB_img = domB_img.view((-1, 3, args.resize, args.resize))
ae_optimizer.zero_grad()
A_common = e1(domA_img)
A_separate = e2(domA_img)
A_encoding = torch.cat([A_common, A_separate], dim=1)
B_common = e1(domB_img)
B_encoding = torch.cat([B_common, B_separate], dim=1)
A_decoding = decoder(A_encoding)
B_decoding = decoder(B_encoding)
loss = mse(A_decoding, domA_img) + mse(B_decoding, domB_img)
if args.discweight > 0:
preds_A = disc(A_common)
preds_B = disc(B_common)
loss += args.discweight * (bce(preds_A, B_label) + bce(preds_B, B_label))
loss.backward()
torch.nn.utils.clip_grad_norm_(ae_params, 5)
ae_optimizer.step()
if args.discweight > 0:
disc_optimizer.zero_grad()
A_common = e1(domA_img)
B_common = e1(domB_img)
disc_A = disc(A_common)
disc_B = disc(B_common)
loss2 = bce(disc_A, A_label) + bce(disc_B, B_label)
loss2.backward()
torch.nn.utils.clip_grad_norm_(disc_params, 5)
disc_optimizer.step()
if _iter % args.progress_iter == 0:
print('Outfile: %s | Iteration %d | loss %.6f | loss1: %.6f | loss2: %.6f' % (args.out, _iter, loss+loss2, loss, loss2))
if _iter % args.display_iter == 0:
e1 = e1.eval()
e2 = e2.eval()
decoder = decoder.eval()
save_imgs(args, e1, e2, decoder, _iter)
e1 = e1.train()
e2 = e2.train()
decoder = decoder.train()
if _iter % args.save_iter == 0:
save_file = os.path.join(args.out, 'checkpoint_%d' % _iter)
save_model(save_file, e1, e2, decoder, ae_optimizer, disc, disc_optimizer, _iter)
_iter += 1
def load_dataset(dataset, data_path):
print('==> Preparing data..')
if dataset == 'CIFAR-10':
transform_train = transforms.Compose([
transforms.ToTensor(),
# transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
transform_test = transforms.Compose([
transforms.ToTensor(),
# transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
trainset = torchvision.datasets.CIFAR10(root=data_path,
train=True,
download=True,
transform=transform_train)
testset = torchvision.datasets.CIFAR10(root=data_path,
train=False,
download=True,
transform=transform_test)
total_set = ConcatDataset((trainset, testset))
elif dataset == 'CIFAR-100':
transform_train = transforms.Compose([
transforms.ToTensor(),
# transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
transform_test = transforms.Compose([
transforms.ToTensor(),
# transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
trainset = torchvision.datasets.CIFAR100(root=data_path,
train=True,
download=True,
transform=transform_train)
testset = torchvision.datasets.CIFAR100(root=data_path,
train=False,
download=True,
transform=transform_test)
total_set = ConcatDataset((trainset, testset))
elif dataset == 'MNIST':
transform_train = transforms.Compose([
transforms.Resize(32),
transforms.ToTensor(),
# transforms.Normalize(mean=[0.5], std=[0.5]),
])
transform_test = transforms.Compose([
transforms.Resize(32),
transforms.ToTensor(),
# transforms.Normalize(mean=[0.5], std=[0.5]),
])
trainset = torchvision.datasets.MNIST(root=data_path,
train=True,
download=True,
transform=transform_train)
testset = torchvision.datasets.MNIST(root=data_path,
train=False,
download=True,
transform=transform_test)
total_set = ConcatDataset((trainset, testset))
elif dataset == 'Fashion-MNIST':
transform_train = transforms.Compose([
transforms.Resize(32),
transforms.ToTensor(),
# transforms.Normalize(mean=[0.5], std=[0.5]),
])
transform_test = transforms.Compose([
transforms.Resize(32),
transforms.ToTensor(),
# transforms.Normalize(mean=[0.5], std=[0.5]),
])
trainset = torchvision.datasets.FashionMNIST(root=data_path,
train=True,
download=True,
transform=transform_train)
testset = torchvision.datasets.FashionMNIST(root=data_path,
train=False,
download=True,
transform=transform_test)
total_set = ConcatDataset((trainset, testset))
elif dataset == 'SVHN':
transform_train = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
])
transform_test = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
])
trainset = torchvision.datasets.SVHN(root=data_path,
split='train',
download=True,
transform=transform_train)
testset = torchvision.datasets.SVHN(root=data_path,
split='test',
download=True,
transform=transform_test)
total_set = ConcatDataset((trainset, testset))
elif dataset == 'adult':
# todo : import pre processing code
dataset = np.load(os.path.join(data_path, 'preprocessed.npy'),
allow_pickle=True).item()
total_set = CustomDataset(torch.FloatTensor(dataset['data']),
dataset['label'])
elif dataset == 'location':
dataset = np.load(os.path.join(data_path, 'data_complete.npz'))
# print(np.unique(dataset['y']-1))
total_set = CustomDataset(torch.FloatTensor(dataset['x']),
dataset['y'] - 1)
return total_set