def fine_tune_stage_2(model):
label_dict = get_TNEWS_label_dict()
train = Mydataset("data_split/TNEWS_dev.csv", label_dict)
eval = Mydataset("data_split/TNEWS_train_small.csv", label_dict)
training_args = TrainingArguments(
output_dir='exp/TNEWS/model', # output directory
num_train_epochs=5, # total # of training epochs
per_device_train_batch_size=16, # batch size per device during training
per_device_eval_batch_size=16, # batch size for evaluation
warmup_steps=500, # number of warmup steps for learning rate scheduler
weight_decay=0.01, # strength of weight decay
save_total_limit=2,
eval_steps=200,
learning_rate=1e-5,
logging_dir='exp/TNEWS/logs', # directory for storing logs
evaluation_strategy='steps',
load_best_model_at_end=True,
metric_for_best_model="marco_f1_score",
)
logging.info(model)
trainer = Trainer(
model=model, # the instantiated 🤗 Transformers model to be trained
args=training_args, # training arguments, defined above
train_dataset=train, # training dataset
eval_dataset=eval, # evaluation dataset
data_collator= data_collator,
compute_metrics=compute_metrics,
callbacks=[EarlyStoppingCallback(early_stopping_patience=3)],
)
trainer.train()
return trainer.model
def __init__(self, root, img_size):
dataset = Mydataset(root, img_size)
self.dataloader = DataLoader(dataset, batch_size=1024, shuffle=True)
torch.cuda.empty_cache()
if img_size == 12:
self.net = PNet()
self.net.load_state_dict(torch.load("../param/test_param/net.pt"))
elif img_size == 24:
self.net = RNet()
# self.net.load_state_dict(torch.load("../param/0_pnet.pt"))
elif img_size == 48:
self.net = ONet()
# self.net.load_state_dict(torch.load("../param/0_pnet.pt"))
else:
print("img_size error!", img_size)
exit()
self.img_size = img_size
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.net.to(self.device)
self.opt = optim.Adam(self.net.parameters())
self.off_loss_fn = torch.nn.MSELoss()
self.conf_loss_fn = torch.nn.BCEWithLogitsLoss()
self.summary = SummaryWriter("./logs")
def inference(model):
import json
index = 0
task_name = "OCEMOTION"
task_label_dict = get_OCEMOTION_label_dict()
file = open("%s_predict.json" % task_name.lower(), "w+")
test_dataset = Mydataset("/tcdata/%s_test_B.csv" % task_name.lower(),
task_label_dict,
is_test=True)
dataloader = torch.utils.data.DataLoader(test_dataset,
batch_size=16,
collate_fn=data_collator)
device = torch.device("cuda")
model.to(device)
model.eval()
inverse_label_dict = {v: k for k, v in task_label_dict.items()}
for batch in dataloader:
logits = model(batch["input_ids"].to(device),
batch["attention_mask"].to(device))[0]
predict_label = logits.argmax(-1)
labels = batch["labels"].flatten().cpu().numpy()
predict_label = predict_label.flatten().cpu().numpy()
for i in range(labels.shape[0]):
json.dump(
{
"id": int(index),
"label": inverse_label_dict[int(predict_label[i])]
}, file)
file.write("\n")
index += 1
print("task %s complete " % task_name)
file.close()
model_dict = {}
file_name = os.walk("exp/TNEWS/model/").__next__()[1][-1]
model_dict["TNEWS"] = "exp/TNEWS/model/%s"%file_name
file_name = os.walk("exp/OCNLI/model/").__next__()[1][-1]
model_dict["OCNLI"] = "exp/OCNLI/model/%s"%file_name
file_name = os.walk("exp/OCEMOTION/model/").__next__()[1][-1]
model_dict["OCEMOTION"] = "exp/OCEMOTION/model/%s"%file_name
for task_name in task_name_list:
index = 0
file = open("%s_predict.json"%task_name.lower(), "w+")
if task_name == "OCNLI":
test_dataset = OCNLI_dataset("/tcdata/ocnli_test_B.csv", task_label_dict["OCNLI"],is_test=True)
else:
test_dataset = Mydataset("/tcdata/%s_test_B.csv" % task_name.lower(), task_label_dict[task_name], is_test=True)
dataloader = torch.utils.data.DataLoader(test_dataset, batch_size=16, collate_fn=data_collator)
print("logging model from %s"%model_dict[task_name])
model = BertForSequenceClassification.from_pretrained(model_dict[task_name], num_labels=len(task_label_dict[task_name]))
model.to(device)
model.eval()
inverse_label_dict = {v:k for k,v in task_label_dict[task_name].items()}
for batch in dataloader:
logits = model(batch["input_ids"].to(device), batch["attention_mask"].to(device)).logits
predict_label = logits.argmax(-1)
labels = batch["labels"].flatten().cpu().numpy()
predict_label = predict_label.flatten().cpu().numpy()
for i in range(labels.shape[0]):
json.dump({"id":int(index), "label": inverse_label_dict[int(predict_label[i])]}, file)
file.write("\n")
index+=1
if epoch % 5 == 0:
torch.save(net.state_dict(), "./ckpt/{}.pkl".format(epoch))
if __name__ == "__main__":
net = TinySSD()
net.cuda()
loss_fn = MultiBoxLoss(3.)
transform = transforms.Compose([
transforms.Resize((512, 512)),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
# transform = Compose([
# ConvertFromInts(),
# PhotometricDistort(),
# Expand([123, 117, 104]),
# RandomSampleCrop(),
# RandomMirror(),
# ToPercentCoords(),
# Resize(300),
# SubtractMeans([123, 117, 104]),
# ToTensor(),
# ])
optm = optim.Adam(net.parameters(), lr=1e-3)
dtset = Mydataset(img_path="./dataset", transform=transform)
dataloader = DataLoader(dtset, batch_size=8, shuffle=True)
train(dataloader, net, loss_fn, optm)
def main():
global args, best_prec1
print('parsing args...')
args = parser.parse_args()
# create model
print("=> creating model '{}'".format(args.arch))
if args.arch.lower().startswith('wideresnet'):
# a customized resnet model with last feature map size as 14x14 for better class activation mapping
model = wideresnet.resnet50(num_classes=args.num_classes)
else:
model = models.__dict__[args.arch](num_classes=args.num_classes)
if args.arch.lower().startswith('alexnet') or args.arch.lower().startswith(
'vgg'):
model.features = torch.nn.DataParallel(model.features)
model.cuda()
else:
model = torch.nn.DataParallel(model).cuda()
model = model.cuda()
print(model)
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
cudnn.benchmark = True
# Data loading code
#train_loader = Provider(phase = 'train', batch_size=args.batch_size, workers=args.workers)
#val_loader = Provider(phase = 'test', batch_size=args.batch_size)
train_loader = torch.utils.data.DataLoader(Mydataset(phase='train'),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(Mydataset(phase='test'),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
# define loss function (criterion) and pptimizer
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
if args.evaluate:
validate(val_loader, model, criterion)
return
for epoch in range(args.start_epoch, args.epochs):
adjust_learning_rate(optimizer, epoch)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch)
# evaluate on validation set
prec1 = validate(val_loader, model, criterion)
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
save_checkpoint(
{
'epoch': epoch,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
}, is_best, './snapshot/' + args.arch.lower() + '_' + str(epoch))
loss_obj_center = center_loss_fn(output_obj[:, 0:2], target_obj[:, 0:2])
loss_obj_wh = wh_loss_fn(output_obj[:, 2:4], target_obj[:, 2:4])
loss_obj_cls = cls_loss_fn(output_obj[:, 5:], target_obj[:, 5].long())
loss_obj = loss_obj_conf + loss_obj_center + loss_obj_wh + loss_obj_cls
# 负样本的时候只需要计算置信度损失
mask_noobj = target[..., 4] == 0
output_noobj, target_noobj = output[mask_noobj], target[mask_noobj]
loss_noobj = conf_loss_fn(output_noobj[:, 4], target_noobj[:, 4])
loss = alpha * loss_obj + (1 - alpha) * loss_noobj
return loss
if __name__ == '__main__':
datas = Mydataset()
imageDataloader = data.DataLoader(dataset=datas,
batch_size=1,
shuffle=True)
net = MainNet(14).to(device)
if os.path.exists(savepath):
net.load_state_dict(torch.load(savepath))
optim = torch.optim.Adam(net.parameters(), weight_decay=4e-4)
base_opt = torch.optim.Adam(net.parameters(), lr=1e-3,
betas=(0.9, 0.999)) # Any optimizer
lookahead = Lookahead(base_opt, k=5, alpha=0.5) # Initialize Lookahead
losses = []
def main():
global args, best_prec1
args = parser.parse_args()
final = pd.read_csv(
'/home/cytuser/code/panorama/pretrained-models.pytorch-master/classification_data/final.csv'
)
final['label'] = final['label'].apply(lambda x: classes(x))
y = final['label'].values
skf = StratifiedKFold(n_splits=5, shuffle=True)
kfold = []
for train_index, test_index in skf.split(y, y):
#print(len(train_index))
#print(len(test_index))
kfold.append([train_index, test_index])
# create model
print("=> creating model '{}'".format(args.arch))
#if args.pretrained.lower() not in ['false', 'none', 'not', 'no', '0']:
# print("=> using pre-trained parameters '{}'".format(args.pretrained))
# model = pretrainedmodels.__dict__[args.arch](num_classes=1000,
# pretrained=args.pretrained)
#else:
# model = pretrainedmodels.__dict__[args.arch]()
model = pretrainedmodels.__dict__['se_resnet50'](num_classes=1000,
pretrained='imagenet')
n_inputs = model.last_linear.in_features
classifier = nn.Sequential(
OrderedDict([('classifier', nn.Linear(n_inputs, 596))]))
#model.conv0.conv = nn.Conv2d(1, 96, kernel_size=(3, 3), stride=(2, 2), bias=False)
model.layer0.conv1 = nn.Conv2d(1,
64,
kernel_size=(7, 7),
stride=(2, 2),
padding=(3, 3),
bias=False)
model.last_linear = classifier
#model = EfficientNetB0()
#model.conv1 = nn.Conv2d(1, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
cudnn.benchmark = True
# Data loading code
traindir = os.path.join(args.data, 'classification_data')
#print(traindir)
#valdir = os.path.join(args.data, 'val')
# transform = transforms.Compose([
# transforms.Resize((224,224)),
# transforms.ToTensor(),
# transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))])
transform = transforms.Compose([
#transforms.Resize((224,224)),
transforms.ToTensor(),
transforms.Normalize([0.5], [0.5])
])
# if 'scale' in pretrainedmodels.pretrained_settings[args.arch][args.pretrained]:
# scale = pretrainedmodels.pretrained_settings[args.arch][args.pretrained]['scale']
# else:
# scale = 0.875
scale = 0.875
#print('Images transformed from size {} to {}'.format(
# int(round(max(model.input_size) / scale)),
# model.input_size))
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
model = torch.nn.DataParallel(model).cuda()
for epoch in range(args.start_epoch, args.epochs):
adjust_learning_rate(optimizer, epoch)
i = np.random.choice(5)
train_idx, val_idx = kfold[i][0], kfold[i][1]
train_dataset = Mydataset(traindir, train_idx, transform)
#print('batch size is :',args.batch_size)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
val_dataset = Mydataset(traindir, val_idx, transform)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch)
# evaluate on validation set
prec1 = validate(val_loader, model, criterion)
if epoch % 5 == 0:
validate(val_loader, model, criterion)
# remember best prec@1 and save checkpoint
is_best = prec1[0] > best_prec1
best_prec1 = max(prec1[0], best_prec1)
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
}, is_best)
logging.basicConfig(
level=logging.INFO,
format='%(asctime)s %(filename)s[line:%(lineno)d] %(levelname)s %(message)s'
)
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "None"
task_name = ["OCEMOTION", "OCNLI", "TNEWS"]
label_len = [3, 7, 17]
task_label_dict = {}
task_label_dict["OCEMOTION"] = {
0: "100",
}
train = Mydataset(path="OCEMOTION_train1128.csv")
label_dict = {0: "O", 1: "，", 2: "。", 3: "：", 4: "、"}
model = BertForSequenceClassification.from_pretrained("bert-base-chinese")
training_args = TrainingArguments(
output_dir='exp/bert_base/model', # output directory
num_train_epochs=25, # total # of training epochs
per_device_train_batch_size=2, # batch size per device during training
per_device_eval_batch_size=2, # batch size for evaluation
warmup_steps=500, # number of warmup steps for learning rate scheduler
weight_decay=0.01, # strength of weight decay
save_steps=20000,
eval_steps=5,
logging_dir='exp/bert_base/logs', # directory for storing logs
evaluation_strategy='steps',
#prediction_loss_only=True,
level=logging.INFO,
format='%(asctime)s %(filename)s[line:%(lineno)d] %(levelname)s %(message)s'
)
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
label_dict = {
'sadness': 0,
'happiness': 1,
'disgust': 2,
'anger': 3,
'like': 4,
'surprise': 5,
'fear': 6
}
train = Mydataset("data/OCEMOTION_train1128.csv", label_dict)
eval = Mydataset("data/OCEMOTION_train1128.csv", label_dict)
logging.info(torch.cuda.is_available())
model = BertForSequenceClassification.from_pretrained(
"hfl/chinese-roberta-wwm-ext-large", num_labels=len(label_dict))
training_args = TrainingArguments(
output_dir='exp/OCEMOTION/model', # output directory
num_train_epochs=6, # total # of training epochs
per_device_train_batch_size=8, # batch size per device during training
per_device_eval_batch_size=8, # batch size for evaluation
warmup_steps=500, # number of warmup steps for learning rate scheduler
weight_decay=0.01, # strength of weight decay
save_steps=1000,
save_total_limit=1,
pred = torch.argmax(pred, dim=1)
return pred
data_dir = 'data/data1.txt'
model_dir = 'model\scnn_0_512_32fr_1.pkl'
if __name__ == "__main__":
scnn = SignalCNN(16, 32, 2)
scnn.cuda()
optimizer = optim.Adam(scnn.parameters(), lr=LR, weight_decay=LR_DECAY)
# scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size = 100, gamma = 0.1, last_epoch=-1)
loss_func = nn.CrossEntropyLoss()
dataSet = Mydataset(data_dir)
# testSet = Mydataset('data/sptestdata.txt')
# test_total = len(testSet)
total = len(dataSet)
print(total)
train_size = int(0.8*total)
valid_size = total - train_size
train_dataset, valid_dataset = torch.utils.data.random_split(dataSet, [train_size, valid_size])
train_loader = DataLoader(dataset=train_dataset, batch_size=BATCH_SIZE, shuffle=True)
valid_loader = DataLoader(dataset=valid_dataset, batch_size=BATCH_SIZE, shuffle=True)
# test_loader = DataLoader(dataset=testSet, batch_size=60, shuffle=True)
for epoch in range(N_EPOCH):
scnn.zero_grad()
# scheduler.step()
for batch, (x, y, c, index) in enumerate(train_loader):
def fine_tune_stage():
label_dict = {}
for i in range(17):
if i < 10:
label_dict["10%d" % i] = i
else:
label_dict["1%d" % i] = i
train = Mydataset("data_split/TNEWS_train.csv", label_dict)
eval = Mydataset("data_split/TNEWS_dev.csv", label_dict)
model = BertClassification("hfl/chinese-roberta-wwm-ext-large",
num_labels=len(label_dict))
for name, param in model.named_parameters():
if param.requires_grad:
logging.info(name)
#model = BertForSequenceClassification.from_pretrained("exp/TNEWS_pretrain/model/checkpoint-6000", num_labels=len(label_dict))
#model = BertForSequenceClassification.from_pretrained("hfl/chinese-roberta-wwm-ext-large", num_labels=len(label_dict))
training_args = TrainingArguments(
output_dir='exp/TNEWS/model', # output directory
num_train_epochs=1, # total # of training epochs
per_device_train_batch_size=16, # batch size per device during training
per_device_eval_batch_size=16, # batch size for evaluation
warmup_steps=500, # number of warmup steps for learning rate scheduler
weight_decay=0.01, # strength of weight decay
save_total_limit=1,
save_steps=1000,
eval_steps=1000,
learning_rate=1e-5,
logging_dir='exp/TNEWS/logs', # directory for storing logs
evaluation_strategy='steps',
#load_best_model_at_end=True,
#metric_for_best_model="marco_f1_score",
#prediction_loss_only=True,
do_eval=False,
)
#logging.info(model)
from sklearn.metrics import precision_recall_fscore_support, f1_score, confusion_matrix, classification_report
def compute_metrics(pred):
labels = pred.label_ids
preds = pred.predictions.argmax(-1)
labels = labels.flatten()
preds = preds.flatten()
marco_f1_score = f1_score(labels, preds, average='macro')
logging.info(marco_f1_score)
#logging.info(f"{'confusion_matrix':*^80}")
#logging.info(confusion_matrix(labels, preds, ))
#logging.info(f"{'classification_report':*^80}")
#logging.info(classification_report(labels, preds, ))
res = {"marco_f1_score": marco_f1_score}
return res
trainer = Trainer(
model=model, # the instantiated 🤗 Transformers model to be trained
args=training_args, # training arguments, defined above
train_dataset=train, # training dataset
eval_dataset=eval, # evaluation dataset
data_collator=data_collator,
compute_metrics=compute_metrics,
)
trainer.train()
