def test(pretrain):
print('TESTING MSDN...')
full_transform=transforms.Compose([ToTensor(),
Normalize(cfg.BGR_MEAN,cfg.BGR_STD)])
params = {'batch_size': 1,
'shuffle':False,
'num_workers':cfg.TRAIN.NUM_WORKERS}
print(params)
print('score thress: {}'.format(cfg.TEST.THRESS ))
my_dataset = MyDataset(imgs_csv=cfg.TEST.IMGS_CSV,rois_csv=cfg.TEST.ROIS_CSV,
root_dir=cfg.TEST.ROOT_DIR, ther_path=cfg.TEST.THERMAL_PATH,transform = full_transform,train=False)
print(my_dataset.__len__())
dataloader = DataLoader(my_dataset, **params)
MSDN_net = MyMSDN()
MSDN_net.to(cfg.DEVICE)
print('pretrain: ' + pretrain)
MSDN_net.load_state_dict(torch.load(pretrain))
running_loss = 0.0
st = time.time()
test_file = 'mymodel/MSDN/test1/test70_{}.txt'.format(pretrain.split('/')[-1])
f = open(test_file,'a')
def train(epoch, pos_weight = None, batch_size = 100):
t = time.time()
model.train()
train_feat_files, train_edge_files = generate_filenames('train')
train_dataset = MyDataset('train', train_feat_files, train_edge_files)
train_iter = Data.DataLoader(dataset = train_dataset, batch_size = 1)
img_num = 1000
train_output = None
train_labels = None
for img_id, (x, y) in enumerate(train_iter):
# print(img_id)
x = x[0].numpy()
y = y[0].numpy()
adj, features, labels, idx_train = load_data(x, y)
if args.cuda:
features = features.cuda()
adj = adj.cuda()
labels = labels.cuda()
idx_train = idx_train.cuda()
features, adj, labels = Variable(features), Variable(adj), Variable(labels)
output = model(features, adj)
idx_train = sample_training(labels, idx_train)
if train_output is None:
train_output = output[idx_train]
train_labels = labels[idx_train]
else:
train_output = torch.cat((train_output, output[idx_train]), 0)
train_labels = torch.cat((train_labels, labels[idx_train]), 0)
if (img_id+1) % batch_size == 0 or (img_id+1) == img_num:
if pos_weight is None:
loss_train = F.nll_loss(train_output, train_labels)
else:
loss_train = F.nll_loss(train_output, train_labels, weight=pos_weight)
loss_data = loss_train.data.item()
acc_train, recall_bg, recall_nobg, precision_bg, precision_nobg \
= evaluation_train(train_output, train_labels)
optimizer.zero_grad()
loss_train.backward()
optimizer.step()
train_output = None
train_labels = None
print('Epoch: {:04d}'.format(epoch+1),
'loss_train: {:.6f}'.format(loss_data),
'acc_train: {:.6f}'.format(acc_train.data.item()),
'recall_bg: {:.6f}'.format(recall_bg.data.item()),
'recall_nobg: {:.6f}'.format(recall_nobg.data.item()),
'precision_bg: {:.6f}'.format(precision_bg.data.item()),
'precision_nobg: {:.6f}'.format(precision_nobg.data.item()),
'time: {:.4f}s'.format(time.time() - t))
return loss_data
def __init__(self, args):
# parameters
self.epoch = args.epoch
self.sample_num = 100
self.batch_size = args.batch_size
self.save_dir = args.save_dir
self.result_dir = args.result_dir
self.dataset = args.dataset
self.log_dir = args.log_dir
self.gpu_mode = args.gpu_mode
self.model_name = args.gan_type
self.input_size = args.input_size
self.pre_train_model = args.model
self.output_dim = 8
self.z_dim = 100
self.max_length = 64
self.lambda_ = 10
self.n_critic = 5 # the number of iterations of the critic per generator iteration
# load dataset
self.dataset = MyDataset()
self.data_loader = DataLoader(self.dataset,
batch_size=self.batch_size,
shuffle=True)
self.iter = self.data_loader.__iter__()
# networks init
self.G = G(self.gpu_mode)
self.D = D(self.gpu_mode)
self.G_optimizer = optim.Adam(self.G.parameters(),
lr=args.lrG,
betas=(args.beta1, args.beta2))
self.D_optimizer = optim.Adam(self.D.parameters(),
lr=args.lrD,
betas=(args.beta1, args.beta2))
if self.pre_train_model:
dic_D = torch.load(self.pre_train_model)
dic_G = torch.load(
self.pre_train_model[:self.pre_train_model.rfind('D')] +
'G.pth')
self.G.load_state_dict(dic_G['state_dict'])
self.G_optimizer.load_state_dict(dic_G['optimizer'])
self.D.load_state_dict(dic_D['state_dict'])
self.D_optimizer.load_state_dict(dic_D['optimizer'])
if self.gpu_mode:
self.G.cuda()
self.D.cuda()
print('---------- Networks architecture -------------')
utils.print_network(self.G)
utils.print_network(self.D)
print('-----------------------------------------------')
# fixed noise
self.sample_z_ = torch.rand((self.batch_size, self.z_dim))
if self.gpu_mode:
self.sample_z_ = self.sample_z_.cuda()
def main():
x_test = LoadData(sys.argv[1])
val_transform = transforms.Compose([
transforms.ToPILImage(),
transforms.CenterCrop(44),
transforms.ToTensor(),
transforms.Normalize(mean=[0],
std=[255])
])
batch_size = 128
test_set = MyDataset(x_test, None, val_transform)
test_loader = DataLoader(test_set, batch_size=batch_size, shuffle=False, num_workers=8)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)
model = Classifier().to(device)
model = loading_compression(model)
ans = np.array([])
for i, data in enumerate(test_loader):
pred = model(data.to(device))
ans = np.append(ans,np.argmax(pred.data.cpu().numpy(), axis=1))
ans = ans.astype(int)
df = [[cnt,i] for cnt, i in enumerate(ans)]
df = pd.DataFrame(df, columns=['id', 'label'])
df.to_csv(sys.argv[2], index=None)
def result_write(transformer, n):
data_w = open("result.txt", "w")
pair = pairs_devtest
inputs, outputs = map(list, zip(*pair))
test_pairs = [tensorsFromPair(inputs[i].split(" "), outputs[i].split(" "))
for i in range(n)]
dataset = MyDataset(test_pairs)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=50, shuffle=False, num_workers=0, collate_fn=my_collate_fn)
# idx : [inputs, targets], [ip_len, op_len]
# 1iter : [batch_size, length]
# batch=96, text_size=1784
for idx in tqdm(dataloader, ascii = True):
batch = len(idx[0][0])
test_pair = idx[0]
lengths = idx[1]
op_len = lengths[1]
input_tensor = torch.tensor(test_pair[0], dtype=torch.long, device=device)
output_words = evaluate(transformer, input_tensor, batch)
for s in output_words:
if len(s) != 0:
data_w.write(s+"\n")
else:
data_w.write("\n")
data_w.close()
def get_loaders(dataroot, val_batch_size, train_batch_size, input_size,
workers):
val_data = MyDataset(txt_path=os.path.join(dataroot, 'valid_label1.txt'),
transform=get_transform(False, input_size))
val_loader = torch.utils.data.DataLoader(val_data,
batch_size=val_batch_size,
shuffle=False,
num_workers=workers,
pin_memory=True)
train_data = MyDataset(txt_path=os.path.join(dataroot, 'train_label1.txt'),
transform=get_transform(input_size=input_size))
train_loader = torch.utils.data.DataLoader(train_data,
batch_size=train_batch_size,
shuffle=True,
num_workers=workers,
pin_memory=True)
return train_loader, val_loader
def Validation(transformer, n, epoch):
all_BLEUscore = 0
new_n = int(n/max_epoch)
start = 0
stop = n
with torch.no_grad():
pair = pairs_dev
inputs, outputs = map(list, zip(*pair))
dev_pairs = [tensorsFromPair(inputs[i].split(" "), outputs[i].split(" "))
for i in range(n)]
dataset = MyDataset(dev_pairs)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=50, shuffle=False, num_workers=0, collate_fn=my_collate_fn)
cnt = 0
res = []
hy = []
chencherry = bleu_score.SmoothingFunction()
for idx in tqdm(dataloader, ascii = True):
batch = len(idx[0][0])
dev_pair = idx[0]
input_tensor = torch.tensor(dev_pair[0], dtype=torch.long, device=device)
re_tensor = torch.tensor(dev_pair[1], dtype=torch.long, device=device)
output_words = evaluate(transformer, input_tensor, batch, 0)
#output_words = evaluate(transformer, input_tensor, batch, -1) #debag
for i in range(len(output_words)):
re = outputs[cnt].split(" ")
res.append([ re ])
if len(output_words[i]) != 0: # 空判定
hy.append(output_words[i].split(" "))
else:
hy.append([""])
cnt += 1
#all_BLEUscore += bleu_score.corpus_bleu(res, hy, smoothing_function=chencherry.method4) # smoothing_function=chencherry.method4
all_BLEUscore += bleu_score.corpus_bleu(res, hy)
all_BLEUscore *= 100
return all_BLEUscore
def compute_test(epoch, pos_weight = None):
model.eval()
test_feat_files, test_edge_files = generate_filenames('test')
test_dataset = MyDataset('test', test_feat_files, test_edge_files)
test_iter = Data.DataLoader(dataset = test_dataset, batch_size = 1)
test_output = None
test_labels = None
for _, (x, y) in enumerate(test_iter):
x = x[0].numpy()
y = y[0].numpy()
adj, features, labels, idx_test = load_data(x, y)
if args.cuda:
features = features.cuda()
adj = adj.cuda()
labels = labels.cuda()
idx_test = idx_test.cuda()
output = model(features, adj)
preds = output[idx_test].max(1)[1].type_as(labels)
if test_output is None:
test_output = preds
test_labels = labels[idx_test]
else:
test_output = torch.cat((test_output, preds), 0)
test_labels = torch.cat((test_labels, labels[idx_test]), 0)
acc_test, recall_bg, recall_nobg, precision_bg, precision_nobg \
= evaluation_test(test_output, test_labels)
print('Epoch: {:04d}'.format(epoch+1),
'acc_test: {:.6f}'.format(acc_test.data.item()),
'recall_bg: {:.6f}'.format(recall_bg.data.item()),
'recall_nobg: {:.6f}'.format(recall_nobg.data.item()),
'precision_bg: {:.6f}'.format(precision_bg.data.item()),
'precision_nobg: {:.6f}'.format(precision_nobg.data.item()))
if word not in word_dict:
word_dict[len(word_dict)] = word
vec_str = data_pair[1]
tmp = vec_str.replace('\t',
'').strip('\n').strip('[').strip(']').split(', ')
vec = [round(float(i), 5) for i in tmp]
# for i in tmp:
# print(float(i))
word_vec_dict[word] = vec
return word_dict, word_vec_dict
word_dict, word_vec_dict = read_data_vec(filename)
#vecs = list(word_vec_dict.values())
vecs = [torch.FloatTensor(i) for i in word_vec_dict.values()]
dataset = MyDataset(vecs, vecs)
train_loader = DataLoader(dataset, batch_size=BATCH_SIZE, shuffle=True)
test_data = vecs[:10]
class AutoEncoder(nn.Module):
def __init__(self, embdding_size):
super(AutoEncoder, self).__init__()
self.encoder = nn.Sequential(
nn.Linear(embdding_size, 58),
nn.Tanh(),
nn.Linear(58, 29),
nn.Tanh(),
nn.Linear(29, 15),
#nn.Tanh(),
def main(opt):
if opt.manual_seed is None:
opt.manual_seed = random.randint(1, 10000)
print('Random Seed: ', opt.manual_seed)
random.seed(opt.manual_seed)
torch.manual_seed(opt.manual_seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(opt.manual_seed)
model = Net(num_classes=opt.num_classes,
num_layers=opt.num_layers,
feat_dim=opt.feat_dim,
embed_dim=opt.embed_dim,
jk_layer=opt.jk_layer,
process_step=opt.process_step,
dropout=opt.dropout)
print(model)
model = model.to(device)
criterion = torch.nn.CrossEntropyLoss()
df = pd.read_csv(opt.lpath)
nfold = len(df['group'].unique())
dataset = MyDataset(opt.path)
for k in range(nfold):
trainings = df[df['group'] != k + 1]['instance'].tolist()
validations = df[df['group'] == k + 1]['instance'].tolist()
total = [f[:-3] for f in os.listdir(os.path.join(opt.path, 'raw'))]
train_ids = [total.index(x) for x in trainings]
val_ids = [total.index(x) for x in validations]
train_ids = torch.tensor(train_ids, dtype=torch.long)
val_ids = torch.tensor(val_ids, dtype=torch.long)
train_dataset = dataset[train_ids]
val_dataset = dataset[val_ids]
train_loader = DataLoader(train_dataset,
batch_size=opt.batch_size,
shuffle=True,
drop_last=True)
val_loader = DataLoader(val_dataset,
batch_size=opt.batch_size,
drop_last=True)
tr_losses = np.zeros((opt.num_epochs, ))
tr_accs = np.zeros((opt.num_epochs, ))
val_losses = np.zeros((opt.num_epochs, ))
val_accs = np.zeros((opt.num_epochs, ))
model.reset_parameters()
optimizer = torch.optim.Adam(model.parameters(),
lr=opt.lr,
weight_decay=opt.weight_decay)
# optimizer = torch.optim.SGD(model.parameters(), lr=opt.lr, momentum=0.9, weight_decay=opt.weight_decay, nesterov=True)
best_val_loss = 1e6
print('===================Fold {} starts==================='.format(k +
1))
for epoch in range(opt.num_epochs):
s = time.time()
model.train()
losses = 0
acc = 0
for i, data in enumerate(train_loader):
data = data.to(device)
optimizer.zero_grad()
output = model(data)
loss = criterion(output, data.y.squeeze())
loss.backward()
optimizer.step()
y_true = data.y.squeeze().cpu().numpy()
y_pred = output.data.cpu().numpy().argmax(axis=1)
acc += accuracy_score(y_true, y_pred) * 100
losses += loss.data.cpu().numpy()
tr_losses[epoch] = losses / (i + 1)
tr_accs[epoch] = acc / (i + 1)
model.eval()
v_losses = 0
v_acc = 0
y_preds = []
y_trues = []
for j, data in enumerate(val_loader):
data = data.to(device)
with torch.no_grad():
output = model(data)
loss = criterion(output, data.y.squeeze())
y_pred = output.data.cpu().numpy().argmax(axis=1)
y_true = data.y.squeeze().cpu().numpy()
y_trues += y_true.tolist()
y_preds += y_pred.tolist()
v_acc += accuracy_score(y_true, y_pred) * 100
v_losses += loss.data.cpu().numpy()
cnf = confusion_matrix(y_trues, y_preds)
val_losses[epoch] = v_losses / (j + 1)
val_accs[epoch] = v_acc / (j + 1)
current_val_loss = v_losses / (j + 1)
if current_val_loss < best_val_loss:
best_val_loss = current_val_loss
torch.save(
model.state_dict(),
os.path.join(output_path,
'best_model_fold{}.ckpt'.format(k + 1)))
print(
'Epoch: {:03d} | time: {:.4f} seconds\n'
'Train Loss: {:.4f} | Train accuracy {:.4f}\n'
'Validation Loss: {:.4f} | Validation accuracy {:.4f} | Best {:.4f}'
.format(epoch + 1,
time.time() - s, losses / (i + 1), acc / (i + 1),
v_losses / (j + 1), v_acc / (j + 1), best_val_loss))
print('Validation confusion matrix:')
print(cnf)
print('===================Fold {} ends==================='.format(k +
1))
np.save(os.path.join(log_path, 'train_loss_{}.npy'.format(k + 1)),
tr_losses)
np.save(os.path.join(log_path, 'train_acc_{}.npy'.format(k + 1)),
tr_accs)
np.save(os.path.join(log_path, 'val_loss_{}.npy'.format(k + 1)),
val_losses)
np.save(os.path.join(log_path, 'val_acc_{}.npy'.format(k + 1)),
val_accs)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--max-epoch', type=int, default=200)
parser.add_argument('--save-epoch', type=int, default=50)
parser.add_argument('--shot', type=int, default=5)
parser.add_argument('--query', type=int, default=1)
parser.add_argument('--way', type=int, default=21)
parser.add_argument('--save-path', default='./save/proto-me-200')
parser.add_argument('--gpu', default='0')
args = parser.parse_args()
pprint(vars(args))
set_gpu(args.gpu)
ensure_path(args.save_path)
trainset = MyDataset('train', './data20_me/')
train_sampler = TrainSampler(trainset.label, 25, args.way,
args.shot + args.query)
train_loader = DataLoader(dataset=trainset,
batch_sampler=train_sampler,
pin_memory=True)
valset = MyDataset('val', './data20_me/')
val_sampler = TrainSampler(valset.label, 50, args.way,
args.shot + args.query)
val_loader = DataLoader(dataset=valset,
batch_sampler=val_sampler,
pin_memory=True)
model = Protonet().cuda()
def train():
THERMAL_PATH = '/storageStudents/K2015/duyld/dungnm/dataset/KAIST/train/images_train_tm/'
ROOT_DIR = '/storageStudents/K2015/duyld/dungnm/dataset/KAIST/train/images_train'
IMGS_CSV = 'mydata/imgs_train.csv'
ROIS_CSV = 'mydata/rois_trainKaist_thr70_MSDN.csv'
full_transform = transforms.Compose(
[RandomHorizontalFlip(),
ToTensor(), my_normalize()])
# Normalize(rgb_mean,rgb_std)])
device = torch.device("cuda:0")
params = {'batch_size': 2, 'shuffle': True, 'num_workers': 24}
print(params)
max_epoch = 5
print('max_epoch', max_epoch)
LR = 1e-3 #learning rate
print('learning_rate', LR)
MT = 0.9 #momentum
my_dataset = MyDataset(imgs_csv=IMGS_CSV,
rois_csv=ROIS_CSV,
root_dir=ROOT_DIR,
ther_path=THERMAL_PATH,
transform=full_transform)
print(my_dataset.__len__())
dataloader = DataLoader(my_dataset, **params)
# print(list(aa.front_subnetB.parameters())[2])
device = torch.device("cuda:0")
MSDN_net = MyMSDN()
MSDN_net.to(device)
criterion = nn.MSELoss()
optimizer = optim.SGD(filter(lambda p: p.requires_grad,
MSDN_net.parameters()),
lr=LR,
momentum=MT)
for epoch in range(max_epoch): # Lặp qua bộ dữ liệu huấn luyện nhiều lần
running_loss = 0.0
st = time.time()
for i, sample in enumerate(dataloader):
label = sample['label']
bbb = sample['bb']
gt_rois = sample['gt_roi']
# label,bbb,gt_rois = label.to(device),bbb.to(device),gt_rois.to(device)
bbox_label, bbox_targets, bbox_inside_weights, bbox_outside_weights = createTarget(
label, bbb, gt_rois)
bbox_label, bbox_targets, bbox_inside_weights, bbox_outside_weights = bbox_label.to(
device), bbox_targets.to(device), bbox_inside_weights.to(
device), bbox_outside_weights.to(device)
sam = sample['image']
# bbb = sample['bb']
num = bbb.size(1)
bbb = bbb.view(-1, 5)
ind = torch.arange(params['batch_size'],
requires_grad=False).view(-1, 1)
ind = ind.repeat(1, num).view(-1, 1)
bbb[:, 0] = ind[:, 0]
# print(bbb.shape)
# print(tm.shape)
sam = sam.to(device)
bbb = bbb.to(device)
cls_score, bbox_pred = MSDN_net(sam, bbb)
RCNN_loss_cls = F.cross_entropy(cls_score, bbox_label)
# print(RCNN_loss_cls.mean())
RCNN_loss_bbox = smooth_l1_loss(bbox_pred, bbox_targets,
bbox_inside_weights,
bbox_outside_weights)
# print(RCNN_loss_bbox.mean())
loss = RCNN_loss_cls.mean() + RCNN_loss_bbox.mean()
# print('loss at {}: '.format(i),loss.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
running_loss += loss.item()
if i % 10 == 9: # In mỗi 2000 mini-batches.
text = '[{}, {}] loss: {:.3f} time: {:.3f}'.format(
epoch + 1, i + 1, running_loss / 10,
time.time() - st)
print(text)
with open('models/MSDN/model1/log1.txt', 'a') as f:
f.write(text + '\n')
running_loss = 0.0
st = time.time()
torch.save(
MSDN_net.state_dict(),
'models/MSDN/model1/model1_lr_1e-3_bz_6_NBS_128_norm_epoch_{}.pth'.
format(epoch))
print('Huấn luyện xong')
import os
import torchvision.transforms as transforms
import torchvision.datasets as datasets
import torch
from mydataset import MyDataset
import torch.nn as nn
import numpy as np
# 定义时的接口与之前的ImageFolder一样
traindir = '/data2/yujin/cadene_cnn/data/train_val/train'
mydataset = MyDataset(traindir, mode='train')
myloader = torch.utils.data.DataLoader(mydataset, batch_size=4, shuffle=True)
# related to mix_up
def mixup_data(x, y, alpha=1.0, use_cuda=True):
'''Returns mixed inputs, pairs of targets, and lambda'''
if alpha > 0:
lam = np.random.beta(alpha, alpha)
else:
lam = 1
batch_size = x.size()[0]
if use_cuda:
index = torch.randperm(batch_size).cuda()
else:
index = torch.randperm(batch_size)
mixed_x = lam * x + (1 - lam) * x[index, :]
y_a, y_b = y, y[index]
args = parser.parse_args()
args.cuda = not args.no_cuda and torch.cuda.is_available()
# 出力先のフォルダーを作成
try:
os.makedirs('results/%s' % args.run_name)
except OSError:
pass
torch.manual_seed(args.seed)
cudnn.benchmark = True
device = torch.device("cuda" if args.cuda else "cpu")
train_dataset = MyDataset(args.csv_path, args.num_primary_color, mode='train')
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
worker_init_fn=lambda x: np.random.seed(),
drop_last=True,
pin_memory=True)
val_dataset = MyDataset(args.csv_path, args.num_primary_color, mode='val')
val_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=1,
shuffle=False,
num_workers=0,
import torch.nn.parallel
from torch.autograd import Variable
import torch
import model
from mydataset import MyDataset, ToTensor
from torchvision.transforms import transforms
from torch.utils.data import DataLoader
batchsize = 1
epochsize = 500
learningRate = 0.001
print_step = 50
trainPath = r'data/trainingDigits'
train_dataset = MyDataset(trainPath,
transform=transforms.Compose([ToTensor()]))
train_loader = DataLoader(dataset=train_dataset,
batch_size=batchsize,
shuffle=True)
SEED = 0
torch.manual_seed(SEED)
torch.cuda.manual_seed(SEED)
net = model.LeNet().cuda()
criteron = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(net.parameters(),
lr=learningRate,
weight_decay=0.0001)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=100,
def main():
batch_size = 2
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("using {} device.".format(device))
train_db = MyDataset('E:\\ai_learning_resource\\hwdb\\HWDB1\\train', 224, num_clazz=num_clazz, mode='train')
val_db = MyDataset('E:\\ai_learning_resource\\hwdb\\HWDB1\\train', 224, num_clazz=num_clazz, mode='val')
train_loader = DataLoader(train_db, batch_size=batch_size, shuffle=True, num_workers=nw)
val_loader = DataLoader(val_db, batch_size=batch_size, num_workers=nw // 2)
print("using {} images for training, {} images for validation.".format(len(train_db), len(val_db)))
# net = GoogLeNet(num_classes=num_clazz, aux_logits=True, init_weights=True)
net = models.densenet121(pretrained=True, bn_size=batch_size, drop_rate=0.6, num_classes=num_clazz)
net.to(device)
optimizer = optim.Adam(net.parameters(), lr=lr)
criteon = nn.CrossEntropyLoss().to(device)
start = time.time()
print('start time: ', start)
train_steps = len(train_loader)
# best_acc, best_epoch = 0, 0
for epoch in range(epochs):
net.train()
print('start train')
x, y = next(iter(train_loader))
print(x[0].numpy(), y[0])
running_loss = 0.0
# train_bar = tqdm(train_loader)
for step, data in enumerate(train_loader):
print('training')
x, y = data
x, y = x.to(device), y.to(device)
optimizer.zero_grad()
logits, aux_logits2, aux_logits1 = net(x)
loss0 = criteon(logits, y)
loss1 = criteon(aux_logits1, y)
loss2 = criteon(aux_logits2, y)
loss = loss0 + loss1 * 0.3 + loss2 * 0.3
loss.backward()
optimizer.step()
# print statistics
running_loss += loss.item()
print('Step {}/{}'.format(step, len(train_loader)))
# train_bar.desc = "train epoch[{}/{}] loss:{:.3f}".format(epoch + 1, epochs, loss)
# validate
net.eval()
acc = 0.0 # accumulate accurate number / epoch
with torch.no_grad():
# val_bar = tqdm(val_loader)
for val_x, val_y in val_loader:
val_x, val_y = val_x.to(device), val_y.to(device)
outputs = net(val_x) # eval model only have last output layer
predict_y = torch.max(outputs, dim=1)[1]
acc += torch.eq(predict_y, val_y).sum().item()
val_accurate = acc / len(val_db)
print('[epoch %d] train_loss: %.3f val_accuracy: %.3f' %
(epoch + 1, running_loss / train_steps, val_accurate))
print('Finished Training')
print('\n{} epoch cost time {:f}s'.format(epochs, time.time() - start))
if word in word_to_ix:
vec[word_to_ix[word]] = 1
else:
vec[len(word_to_ix)] = 1
return vec
data_words = []
data_labels = []
for i in range(len(sentence_set) -2):
word = sentence_set[i]
label = sentence_set[i+1]
data_words.append(make_word_to_ix(word,word_to_ix))
data_labels.append(make_word_to_ix(label,word_to_ix))
dataset = MyDataset(data_words, data_labels)
train_loader = DataLoader(dataset, batch_size=BATCH_SIZE,shuffle=True)
'''
for x in enumerate(train_loader):
print("word_batch------------>\n")
print(batch[0])
print("label batch----------->\n")
print(batch[1])
'''
class RNNModel(nn.Module):
def __init__(self, embdding_size, hidden_size):
super(RNNModel, self).__init__()
self.rnn = nn.RNN(embdding_size, hidden_size,num_layers=1,nonlinearity='relu')
loss_func_r = nn.L1Loss(reduce=False)
loss_func_c = nn.CrossEntropyLoss(reduce=False)
optimizer = optim.Adam(model.parameters(), lr=lr)
#optimizer = optim.SGD(model.parameters(), lr=lr, momentum=0.8)
#train_dataset = MyDataset(train=True, transform=transforms.ToTensor())
#train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
#when inference
batch_size = 1
epochs = 1
th = 1
val_dataset = MyDataset(train=False, transform=transforms.ToTensor())
val_loader = DataLoader(val_dataset, batch_size=batch_size, shuffle=False)
#when inference
dloader = [
val_loader,
]
is_train = 0
#when train
#dloader = [train_loader, ]
#is_train = 1
tag = ['val', 'train']
for epoch in range(epochs):
#recode running_loss[1] is train, 0 is val
torch.cuda.manual_seed(seed) #gpu
np.random.seed(seed) #numpy
random.seed(seed)
torch.backends.cudnn.deterministic = True # cudnn
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--load', default='./save/proto-5-1000/max-acc.pth')
parser.add_argument('--way', type=int, default=20)
parser.add_argument('--shot', type=int, default=5)
parser.add_argument('--query', type=int, default=1)
args = parser.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
#### preprocess
dataset = MyDataset('test', './data20_v2/')
sampler = TestSampler(dataset.label, 1, args.way, args.shot + args.query)
loader = DataLoader(dataset,
batch_sampler=sampler,
num_workers=8,
pin_memory=True)
model = Protonet().cuda()
model.load_state_dict(torch.load(args.load))
model.eval()
data, l = [_.cuda() for _ in next(iter(loader))]
k = args.way * args.shot
data_shot = data[:k]
x = model(data_shot)
loss.backward()
optimizer.step()
epoch_loss += loss.item()
print("%d/%d,train_loss:%0.3f" %
(step,
(dt_size - 1) // dataload.batch_size + 1, loss.item()))
break
print("epoch %d loss:%0.3f" % (epoch, epoch_loss / step))
return
torch.save(model.state_dict(), 'weights_%d.pth' % epoch)
return model
#训练模型
model = Unet(3, 1).to(device)
batch_size = 1
criterion = nn.BCEWithLogitsLoss()
optimizer = optim.Adam(model.parameters())
dataset = MyDataset(image_path,
label_path,
train_data,
transform=x_transforms,
target_transform=y_transforms)
data_loader = data.DataLoader(dataset,
batch_size=1,
shuffle=True,
num_workers=0)
train_model(model, criterion, optimizer, data_loader)
def trainIters(transformer, n_iters, print_every, learning_rate=0):
start = time.time()
print_loss_total = 0 # Reset every print_every
#transformer_optimizer = optim.SGD(transformer.parameters(), lr=learning_rate)
transformer_optimizer = optim.Adam(transformer.parameters(), lr=learning_rate, weight_decay=1e-06, betas=(0.9, 0.98), eps=1e-09)
#weight_decay=1e-06 , 0.002
criterion = LabelSmoothing(PAD_token, 0.1)
#criterion = nn.NLLLoss(ignore_index=PAD_token, reduction='sum')
global max_epoch
BLEUs = {}
all_step_cnt = 1
inputs, outputs = map(list, zip(*pairs))
training_pairs = [tensorsFromPair(inputs[i].split(" "), outputs[i].split(" "))
for i in range(text_size)]
dataset = MyDataset(training_pairs)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=batch_size, shuffle=True, num_workers=0, collate_fn=my_collate_fn)
#num_workers=3
for epoch in range(1, max_epoch+1):
start_epo = time.time()
print_loss_avg_total = 0
iter = 1
transformer.train()
# idx : [inputs, targets], [ip_len, op_len]
# 1iter : [batch_size, length]
for idx in dataloader:
batch = len(idx[0][0])
training_pair = idx[0]
sq_lengths = idx[1]
input_tensor = torch.tensor(training_pair[0], dtype=torch.long, device=device)
target_tensor = torch.tensor(training_pair[1], dtype=torch.long, device=device)
loss = train(input_tensor, target_tensor, sq_lengths,
transformer, transformer_optimizer, criterion, all_step_cnt, batch)
print_loss_total += loss
if iter % print_every == 0:
print_loss_avg = print_loss_total / print_every
print_loss_avg_total += print_loss_avg
print_loss_total = 0
print('%s (%d %d%%) %.4f' % (timeSince(start, start_epo, iter / n_iters),
iter, iter / n_iters * 100, print_loss_avg))
all_step_cnt += 1
iter += 1
print("Epoch %d Loss: %.4f" % (epoch, print_loss_avg_total/10))
model_name = "./models/seq2seq_model_v{}.pt".format(epoch)
torch.save({
"transformer_model": transformer.state_dict(),
}, model_name)
all_BLEUscore = Validation(transformer, len(pairs_dev), epoch-1)
print("Epoch %d BLEU : %.4f" % (epoch, all_BLEUscore))
# Early Stopping
if epoch >= 6:
BLEU_list = list(BLEUs.values())
if (sum(BLEU_list) / len(BLEU_list)) >= all_BLEUscore:
max_BLEU_epoch = max(BLEUs, key=BLEUs.get)
print("Max Epoch:", max_BLEU_epoch)
max_epoch = max_BLEU_epoch
model_name = "./models/seq2seq_model.pt"
torch.save({
"transformer_model": transformer.state_dict(),
}, model_name)
break
else:
min_BLEU_epoch = min(BLEUs, key=BLEUs.get)
del BLEUs[min_BLEU_epoch]
BLEUs[epoch] = all_BLEUscore
np.random.uniform(low=0.0, high=1.0, size=x.shape[0]))
epsilon = epsilon.type(torch.FloatTensor)
return 2 * x + 8 + epsilon.unsqueeze(1)
X, Y = gen_data(N_DATA + N_Val)
XTrain, YTrain = X[:N_DATA], Y[:N_DATA]
XTest, YTest = X[N_DATA:], Y[N_DATA:]
XTrain = Variable(torch.FloatTensor(XTrain))
YTrain = Variable(torch.FloatTensor(YTrain))
XTest = Variable(torch.FloatTensor(XTest))
YTest = Variable(torch.FloatTensor(YTest))
train_set = MyDataset(x=XTrain, y=YTrain)
train_loader = DataLoader(train_set, batch_size=BATCH_SIZE, shuffle=True)
model = MyNet(n_in=1, n_out=1).to(device)
model.eval()
YPRED = model.forward(XTest)
criterion = nn.MSELoss()
optimizer = optim.Adam(model.parameters(), lr=INITIAL_LEARNING_RATE)
scheduler = StepLR(optimizer, 20)
for epoch in range(N_EPOCHS):
model.train()
epoch_loss = 0
for i, (x, y) in enumerate(train_loader):
def train(pretrain):
print('TRAINING MSDN...')
full_transform=transforms.Compose([RandomHorizontalFlip(),
ToTensor(),
Normalize(cfg.BGR_MEAN,cfg.BGR_STD)])
params = {'batch_size': cfg.TRAIN.BATCH_SIZE,
'shuffle': cfg.TRAIN.SHUFFLE,
'num_workers': cfg.TRAIN.NUM_WORKERS}
print(params)
max_epoch = cfg.TRAIN.MAX_EPOCH
print('max_epoch',max_epoch)
LR = cfg.TRAIN.LEARNING_RATE #learning rate
print('learning_rate',LR)
MT = cfg.TRAIN.MOMENTUM #momentum
W_DECAY = cfg.TRAIN.WEIGHT_DECAY
print('weight decay',W_DECAY)
my_dataset = MyDataset(imgs_csv=cfg.TRAIN.IMGS_CSV,rois_csv=cfg.TRAIN.ROIS_CSV,
root_dir=cfg.TRAIN.ROOT_DIR, ther_path=cfg.TRAIN.THERMAL_PATH,transform = full_transform)
print(my_dataset.__len__())
dataloader = DataLoader(my_dataset, **params)
MSDN_net = MyMSDN()
MSDN_net.to(cfg.DEVICE)
#load pretrain model
if pretrain is not None:
print('pretrain: ' + pretrain)
MSDN_net.load_state_dict(torch.load(pretrain))
criterion = nn.MSELoss()
optimizer = optim.SGD(filter(lambda p: p.requires_grad,MSDN_net.parameters()), lr=LR, momentum=MT,weight_decay=W_DECAY)
f = open('models/MSDN/model12/log12.txt','a')
for epoch in range(max_epoch): # Lặp qua bộ dữ liệu huấn luyện nhiều lần
running_loss = 0.0
st = time.time()
for i, sample in enumerate(dataloader):
label = sample['label']
bbb = sample['bb']
gt_rois = sample['gt_roi']
# label,bbb,gt_rois = label.to(cfg.DEVICE),bbb.to(cfg.DEVICE),gt_rois.to(cfg.DEVICE)
bbox_label,bbox_targets,bbox_inside_weights,bbox_outside_weights = createTarget(label,bbb,gt_rois)
bbox_label,bbox_targets,bbox_inside_weights,bbox_outside_weights = bbox_label.to(cfg.DEVICE),bbox_targets.to(cfg.DEVICE),bbox_inside_weights.to(cfg.DEVICE),bbox_outside_weights.to(cfg.DEVICE)
sam = sample['image']
num=bbb.size(1)
bbb=bbb.view(-1, 5)
ind = torch.arange(params['batch_size'],requires_grad=False).view(-1,1)
ind = ind.repeat(1,num).view(-1,1)
bbb[:,0] = ind[:,0]
sam = sam.to(cfg.DEVICE)
bbb = bbb.to(cfg.DEVICE)
cls_score, bbox_pred = MSDN_net(sam,bbb)
RCNN_loss_cls = F.cross_entropy(cls_score, bbox_label)
RCNN_loss_bbox = smooth_l1_loss(bbox_pred, bbox_targets, bbox_inside_weights, bbox_outside_weights)
# print(RCNN_loss_bbox.mean())
loss = RCNN_loss_cls.mean() + RCNN_loss_bbox.mean()
# print('loss at {}: '.format(i),loss.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
running_loss += loss.item()
if i % 10 == 9: # In mỗi 10 mini-batches.
text = '[{}, {}] loss: {:.3f} time: {:.3f}'.format(epoch + 1, i + 1, running_loss / 10,time.time()-st)
print(text)
#write txt file
f.write(text + '\n')
running_loss = 0.0
st = time.time()
name_model = 'models/MSDN/model12/model12_lr_1e-4_bz_2_decay_epoch_{}.pth'.format(epoch)
torch.save(MSDN_net.state_dict(), name_model)
f.close()
print('model saved: ' + name_model)
print('Huấn luyện xong')
def train():
THERMAL_PATH = '/storageStudents/K2015/duyld/dungnm/dataset/KAIST/train/images_train_tm/'
ROOT_DIR = '/storageStudents/K2015/duyld/dungnm/dataset/KAIST/train/images_train'
IMGS_CSV = 'mydata/imgs_train.csv'
ROIS_CSV = 'mydata/rois_trainKaist_thr70_1.csv'
params = {'batch_size': 6, 'shuffle': True, 'num_workers': 12}
print(params)
max_epoch = 10
print('max_epoch', max_epoch)
LR = 1e-6 #learning rate
print('learning_rate', LR)
MT = 0.9 #momentum
device = torch.device("cuda:0")
# cudnn.benchmark = True
# transform = ToTensor()
full_transform = transforms.Compose(
[RandomHorizontalFlip(),
ToTensor(), my_normalize()])
# Normalize(rgb_mean,rgb_std)])
my_dataset = MyDataset(imgs_csv=IMGS_CSV,
rois_csv=ROIS_CSV,
root_dir=ROOT_DIR,
ther_path=THERMAL_PATH,
transform=full_transform)
dataloader = DataLoader(my_dataset, **params)
RRN_net = MyRRN()
RRN_net.to(device)
RRN_net.load_state_dict(
torch.load(
'models/model24/model24_lr_1e-6_bz_6_NBS_128_norm_epoch_9.pth'))
criterion = nn.MSELoss()
optimizer = optim.SGD(filter(lambda p: p.requires_grad,
RRN_net.parameters()),
lr=LR,
momentum=MT)
for epoch in range(max_epoch): # Lặp qua bộ dữ liệu huấn luyện nhiều lần
running_loss = 0.0
st = time.time()
for i, data in enumerate(dataloader):
# Lấy dữ liệu
sample = data
sam = sample['image']
bbb = sample['bb']
num = bbb.size(1)
bbb = bbb.view(-1, 5)
#reset id
# bbb[:, 0] = bbb[:, 0] - bbb[0, 0]
ind = torch.arange(params['batch_size'],
requires_grad=False).view(-1, 1)
ind = ind.repeat(1, num).view(-1, 1)
bbb[:, 0] = ind[:, 0]
tm = sample['tm']
# print(bbb.shape)
# print(tm.shape)
sam, bbb, tm = sam.to(device), bbb.to(device), tm.to(device)
# roi_pool = ROIPool((50, 50), 1/1)
# labels_output = roi_pool(tm,bbb)
labels_output = resizeThermal(tm, bbb)
labels_output = labels_output.to(device)
# Xoá giá trị đạo hàm
optimizer.zero_grad()
# Tính giá trị tiên đoán, đạo hàm, và dùng bộ tối ưu hoá để cập nhật trọng số.
out_RRN = RRN_net(sam, bbb)
loss = criterion(out_RRN, labels_output)
# print('loss at {}: '.format(i),loss.item())
loss.backward()
optimizer.step()
# In ra số liệu trong quá trình huấn luyện
running_loss += loss.item()
if i % 10 == 9: # In mỗi 2000 mini-batches.
text = '[{}, {}] loss: {:.3f} time: {:.3f}'.format(
epoch + 1, i + 1, running_loss / 10,
time.time() - st)
print(text)
with open('models/model26/log26.txt', 'a') as f:
f.write(text + '\n')
running_loss = 0.0
st = time.time()
torch.save(
RRN_net.state_dict(),
'models/model26/model26_lr_1e-6_bz_6_NBS_128_norm_epoch_{}.pth'.
format(epoch))
print('Huấn luyện xong')
# manual_color_2,
# manual_color_3,
# manual_color_4,
# manual_color_5,
# ]
# )
# / 255
# )
manual_colors = model.cluster_centers_ / 255
try:
os.makedirs("results/%s/%s" % (run_name, img_name))
except OSError:
pass
test_dataset = MyDataset(csv_path, num_primary_color, mode="test")
test_loader = torch.utils.data.DataLoader(
test_dataset,
batch_size=1,
shuffle=False,
num_workers=0,
)
device = "cuda"
# define model
mask_generator = MaskGenerator(num_primary_color).to(device)
residue_predictor = ResiduePredictor(num_primary_color).to(device)
# load params
path_mask_generator = "results/" + run_name + "/mask_generator.pth"
def main():
x_train, val_data, x_label, val_label = LoadData(sys.argv[1])
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)
data_transform = transforms.Compose([
transforms.ToPILImage(),
transforms.RandomResizedCrop(44),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0], std=[255])
])
val_transform = transforms.Compose([
transforms.ToPILImage(),
transforms.CenterCrop(44),
transforms.ToTensor(),
transforms.Normalize(mean=[0], std=[255])
])
train_set = MyDataset(x_train, x_label, data_transform)
val_set = MyDataset(val_data, val_label, val_transform)
batch_size = 128
train_loader = DataLoader(train_set,
batch_size=batch_size,
shuffle=True,
num_workers=8)
val_loader = DataLoader(val_set,
batch_size=batch_size,
shuffle=False,
num_workers=8)
model = Classifier().to(device)
model.initialize_weights()
print(model.eval())
loss = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
scheduler = ReduceLROnPlateau(optimizer, 'min')
best_acc = 0.0
num_epoch = 1000
for epoch in range(num_epoch):
adjust_learning_rate(optimizer, epoch)
epoch_start_time = time.time()
train_acc = 0.0
train_loss = 0.0
val_acc = 0.0
val_loss = 0.0
model.train()
for i, data in enumerate(train_loader):
optimizer.zero_grad()
train_pred = model(data[0].to(device))
batch_loss = loss(train_pred, data[1].to(device))
batch_loss.backward()
optimizer.step()
train_acc += np.sum(
np.argmax(train_pred.cpu().data.numpy(), axis=1) ==
data[1].numpy())
train_loss += batch_loss.item()
progress = ('#' * int(float(i) / len(train_loader) * 40)).ljust(40)
print ('[%03d/%03d] %2.2f sec(s) | %s |' % (epoch+1, num_epoch, \
(time.time() - epoch_start_time), progress), end='\r', flush=True)
model.eval()
for i, data in enumerate(val_loader):
val_pred = model(data[0].cuda())
batch_loss = loss(val_pred, data[1].to(device))
val_acc += np.sum(
np.argmax(val_pred.cpu().data.numpy(), axis=1) ==
data[1].numpy())
val_loss += batch_loss.item()
progress = ('#' * int(float(i) / len(val_loader) * 40)).ljust(40)
print ('[%03d/%03d] %2.2f sec(s) | %s |' % (epoch+1, num_epoch, \
(time.time() - epoch_start_time), progress), end='\r', flush=True)
val_acc = val_acc / val_set.__len__()
print('[%03d/%03d] %2.2f sec(s) Train Acc: %3.6f Loss: %3.6f | Val Acc: %3.6f loss: %3.6f' % \
(epoch + 1, num_epoch, time.time()-epoch_start_time, \
train_acc/train_set.__len__(), train_loss, val_acc, val_loss))
# scheduler.step(val_loss)
if (val_acc > best_acc):
best_acc = val_acc
if not os.path.exists("static_dict/"):
os.system('mkdir static_dict/')
saving_compression(model)
os.system("du static_dict/ --apparent-size --bytes --max-depth=0")
with open('acc.txt', 'w') as f:
f.write(str(epoch) + '\t' + str(val_acc) + '\t')
# torch.save(model.state_dict(), 'save/model.pth')
print('Model Saved!')
def test():
TEST_THERMAL_PATH = '/storageStudents/K2015/duyld/dungnm/dataset/KAIST/test/images_test_tm/'
TEST_ROOT_DIR = '/storageStudents/K2015/duyld/dungnm/dataset/KAIST/test/images_test'
IMGS_CSV = 'mydata/imgs_test.csv'
ROIS_CSV = 'mydata/rois_testKaist_thr70_0.csv'
params = {'batch_size': 2, 'shuffle': True, 'num_workers': 24}
device = torch.device("cuda:0")
full_transform = transforms.Compose([
RandomHorizontalFlip(),
ToTensor(),
])
# Normalize(rgb_mean,rgb_std)])
my_dataset = MyDataset(imgs_csv=IMGS_CSV,
rois_csv=ROIS_CSV,
root_dir=TEST_ROOT_DIR,
ther_path=TEST_THERMAL_PATH,
transform=full_transform)
print(my_dataset.__len__())
dataloader = DataLoader(my_dataset, **params)
RRN_net = MyRRN()
RRN_net.to(device)
RRN_net.load_state_dict(
torch.load(
'models/model21/model21_lr_1e-7_bz_6_NBS_128_data_True_epoch_7.ptx'
))
st = time.time()
running_loss = 0.0
total_loss = []
criterion = nn.MSELoss()
for i, sample in enumerate(dataloader):
# Lấy dữ liệu
sam = sample['image']
bbb = sample['bb']
num = bbb.size(1)
bbb = bbb.view(-1, 5)
#reset id
# bbb[:, 0] = bbb[:, 0] - bbb[0, 0]
ind = torch.arange(params['batch_size'],
requires_grad=False).view(-1, 1)
ind = ind.repeat(1, num).view(-1, 1)
bbb[:, 0] = ind[:, 0]
tm = sample['tm']
# print(bbb.shape)
# print(tm.shape)
sam, bbb, tm = sam.to(device), bbb.to(device), tm.to(device)
# roi_pool = ROIPool((50, 50), 1/1)
# labels_output = roi_pool(tm,bbb)
labels_output = resizeThermal(tm, bbb)
labels_output = labels_output.to(device)
# print('label shape',labels_output.shape)
# Tính giá trị tiên đoán, đạo hàm, và dùng bộ tối ưu hoá để cập nhật trọng số.
out_RRN = RRN_net(sam, bbb)
loss = criterion(out_RRN, labels_output)
# In ra số liệu trong quá trình huấn luyện
running_loss += loss.item()
if i % 10 == 9: # In mỗi 9 mini-batches.
text = '[{}, {}] loss: {:.3f} time: {:.3f}'.format(
0 + 1, i + 1, running_loss / 10,
time.time() - st)
print(text)
with open('test2_model21_epoch7.txt', 'a') as f:
f.write(text + '\n')
total_loss.append(running_loss)
running_loss = 0.0
st = time.time()
print("TOTAL: ", sum(total_loss) / len(total_loss))
from torch import optim, nn
from torch.utils.data import DataLoader
from mydataset import MyDataset
from model import GoogLeNet
batchsz = 20
lr = 1e-3
epochs = 10
num_clazz = 200
device = torch.device('cuda')
root = "E:\\ai_learning_resource\\hwdb\\HWDB1\\train"
root_ = "E:\\ai_learning_resource\\hwdb\\HWDB1\\test"
train_db = MyDataset(root, 224, num_clazz, mode='train')
val_db = MyDataset(root, 224, num_clazz, mode='val')
test_db = MyDataset(root_, 224, num_clazz, mode='test')
train_loader = DataLoader(train_db,
batch_size=batchsz,
shuffle=True,
num_workers=8)
val_loader = DataLoader(val_db, batch_size=batchsz, num_workers=8)
test_loader = DataLoader(test_db, batch_size=batchsz, num_workers=8)
def evalute(model, loader):
correct = 0
total = len(loader.dataset)
for x, y in loader:
def main():
parser = argparse.ArgumentParser(
description="PyTorch Object Detection Training")
parser.add_argument(
"--config-file",
default="",
metavar="FILE",
help="path to config file",
type=str,
)
parser.add_argument(
"--save_dir",
default="./data",
help="path to save dataset",
type=str,
)
parser.add_argument(
"-d",
"--data_name",
default=['icpr'],
help="path to save dataset",
type=str,
nargs='+',
)
parser.add_argument("--is_training",
'-t',
default=False,
help="training or evaluation",
action='store_true')
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument(
"--skip-test",
dest="skip_test",
help="Do not test the final model",
action="store_true",
)
parser.add_argument(
"--ckpt",
default=None,
help="path to save dataset",
type=str,
)
parser.add_argument(
"opts",
help="Modify config options using the command-line",
default=None,
nargs=argparse.REMAINDER,
)
args = parser.parse_args()
args.num_layers = 2
args.featrue_layers = 512
args.hidden_dim = 512
args.vocab_size = 36
args.out_seq_len = 30
args.hidden_dim_de = 512
args.embedding_size = 512
args.batch_size = 32
args.multi_bias = True
######
args.lr = 0.02
args.checkpoint_dir = './model/' # adam_lowlr/
args.optim = 'sgd'
os.environ["CUDA_VISIBLE_DEVICES"] = str(0)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
args.device = 'cuda'
if not os.path.exists(args.checkpoint_dir):
os.makedirs(args.checkpoint_dir)
# settings for multi-gpu training not implement yet
num_gpus = int(
os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
args.distributed = num_gpus > 1
args.distributed = False # not implement
if args.distributed:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend="nccl",
init_method="env://")
else:
pass
# trainsform for train and test
# TODO: data enhancement maybe in class dataset or in transform
if args.is_training:
transform = transforms.Compose([
transforms.Resize((48, 160)), # (h, w) 48, 160 6 40
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465),
(0.2023, 0.1994, 0.2010)),
])
else:
transform = transforms.Compose([
transforms.Resize((48, 160)), # 32 280
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465),
(0.2023, 0.1994, 0.2010)),
])
# prepare for data
# trainset = prepare_data(args.save_dir, args.data_name, args.is_training)
if args.is_training:
# trainset = MyDataset('./data/icpr/crop/',
# './data/icpr/char2num.txt', transform)
# ./2423/6/96_Flowerpots_29746.jpg flowerpots
trainset = MyDataset([
'90kDICT32px_train',
], transform)
sys.stdout = Logger(args.checkpoint_dir + '/log.txt')
train(args, args.local_rank, args.distributed, trainset)
else:
# testset= MyDataset('/data4/ydb/dataset/recognition/imgs2_east_regions', transform=transform)
# testset = MyDataset('./data/icpr/crop/',
# './data/icpr/char2num.txt', transform)
testset = MyDataset('90kDICT32px_val', transform)
test(args, args.local_rank, args.distributed, testset)
def train():
params = {
'batch_size': cfg.TRAIN.BATCH_SIZE,
'shuffle': cfg.TRAIN.SHUFFLE,
'num_workers': cfg.TRAIN.NUM_WORKERS
}
print(params)
max_epoch = cfg.TRAIN.MAX_EPOCH
print('max_epoch', max_epoch)
LR = cfg.TRAIN.LEARNING_RATE #learning rate
print('learning_rate', LR)
MT = cfg.TRAIN.MOMENTUM
# cudnn.benchmark = True
# transform = ToTensor()
full_transform = transforms.Compose([
RandomHorizontalFlip(),
ToTensor(),
Normalize(cfg.BGR_MEAN, cfg.BGR_STD)
])
my_dataset = MyDataset(imgs_csv=cfg.TRAIN.IMGS_CSV,
rois_csv=cfg.TRAIN.ROIS_CSV,
root_dir=cfg.TRAIN.ROOT_DIR,
ther_path=cfg.TRAIN.THERMAL_PATH,
transform=full_transform)
dataloader = DataLoader(my_dataset, **params)
RRN_net = MyRRN()
RRN_net.to(cfg.DEVICE)
RRN_net.load_state_dict(
torch.load(
'models/RRN/model24/model24_lr_1e-6_bz_6_NBS_128_norm_epoch_9.pth')
)
criterion = nn.MSELoss()
optimizer = optim.SGD(filter(lambda p: p.requires_grad,
RRN_net.parameters()),
lr=LR,
momentum=MT)
f = open('models/RRN/model27/log27.txt', 'a')
for epoch in range(max_epoch): # Lặp qua bộ dữ liệu huấn luyện nhiều lần
running_loss = 0.0
st = time.time()
for i, data in enumerate(dataloader):
# Lấy dữ liệu
sample = data
sam = sample['image']
bbb = sample['bb']
num = bbb.size(1)
bbb = bbb.view(-1, 5)
#reset id
# bbb[:, 0] = bbb[:, 0] - bbb[0, 0]
ind = torch.arange(params['batch_size'],
requires_grad=False).view(-1, 1)
ind = ind.repeat(1, num).view(-1, 1)
bbb[:, 0] = ind[:, 0]
tm = sample['tm']
sam, bbb, tm = sam.to(cfg.DEVICE), bbb.to(cfg.DEVICE), tm.to(
cfg.DEVICE)
# labels_output = roi_pool(tm,bbb)
labels_output = resizeThermal(tm, bbb)
labels_output = labels_output.to(cfg.DEVICE)
# Xoá giá trị đạo hàm
optimizer.zero_grad()
# Tính giá trị tiên đoán, đạo hàm, và dùng bộ tối ưu hoá để cập nhật trọng số.
out_RRN = RRN_net(sam, bbb)
loss = criterion(out_RRN, labels_output)
# print('loss at {}: '.format(i),loss.item())
loss.backward()
optimizer.step()
# In ra số liệu trong quá trình huấn luyện
running_loss += loss.item()
if i % 10 == 9: # In mỗi 2000 mini-batches.
text = '[{}, {}] loss: {:.3f} time: {:.3f}'.format(
epoch + 1, i + 1, running_loss / 10,
time.time() - st)
print(text)
f.write(text + '\n')
running_loss = 0.0
st = time.time()
torch.save(
RRN_net.state_dict(),
'models/RRN/model27/model27_lr_1e-6_bz_6_NBS_128_norm_epoch_{}.pth'
.format(epoch))
f.close()
print('Huấn luyện xong')