def main(args):
if not os.path.isdir(args.checkpoint):
try:
os.makedirs(args.checkpoint)
except OSError as e:
if e.errno != errno.EEXIST:
raise
#ini log
log = open(os.path.join(args.checkpoint, 'log.txt'), 'w')
log.write('Epoch')
log.write('\t')
log.write('LR')
log.write('\t')
log.write('Loss')
log.write('\t')
log.write('Acc')
log.write('\t')
log.write('\n')
log.flush()
model = MyModel(classes=train_cfg.num_class)
# model.build(input_shape=(None,train_cfg.data_shape[0],train_cfg.data_shape[1],3))
#model._set_inputs((None,train_cfg.data_shape[0],train_cfg.data_shape[1],3))
optimizer = tf.keras.optimizers.Adam(learning_rate=5e-4)
criterion = tf.keras.losses.SparseCategoricalCrossentropy()
train_data, train_len = MyDataset(train_cfg, train_mode=0).get_dataset()
val_data, val_len = MyDataset(val_cfg, train_mode=1).get_dataset()
for epoch in range(args.start_epoch, train_cfg.epochs):
lr = adjust_learning_rate(optimizer, epoch)
print('epoch {},lr={}'.format(epoch, lr))
print('begin the {}th epoch'.format(epoch))
train_loss = train(train_data, model, criterion, optimizer)
print('avg_train_loss:{:.5f}'.format(train_loss / train_len))
acc = eval(val_data, model)
print('accuracy:{:.5f}'.format(acc))
log.write('{}'.format(epoch))
log.write('\t')
log.write('{:.5f}'.format(lr))
log.write('\t')
log.write('{:.5f}'.format(train_loss / val_len))
log.write('\t')
log.write('{:.5f}'.format(acc))
log.write('\t')
log.write('\n')
log.flush()
filename = 'epoch' + str(epoch + 1) + 'checkpoint.h5'
filepath = os.path.join(args.checkpoint, filename)
#model.save(filepath=filepath,overwrite=False,include_optimizer=True,save_format='tf')
model.save_weights(filepath=filepath,
overwrite=False,
save_format='h5')
log.close()
def main(args):
model = MyModel(classes=test_cfg.num_class)
test_data, test_len = MyDataset(test_cfg, train_mode=2).get_dataset()
predict_np = []
checkpoint_file = os.path.join(test_cfg.checkpoint_path,
args.checkpoint + '.h5')
model.load_weights(checkpoint_file)
print('successful loaded checkpoint:{} (epoch{})'.format(
checkpoint_file, args.checkpoint[5]))
print('testing.........')
for i, (inputs, targets) in tqdm(enumerate(test_data)):
out = model(inputs)
out = out.numpy()
prediction = numpy.argmax(out, axis=1)
predict_np.append(prediction)
csv_path = os.path.join(test_cfg.img_folder, '..', test_cfg.csv_name)
data = pd.read_csv(csv_path)
predict_np = numpy.concatenate(predict_np)
data['labels'] = pd.DataFrame(predict_np)
data.to_csv(csv_path, index=False)
print('successful write the predict results!')
def main(args):
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model=networks(num_class=test_cfg.num_class, pretrained = False).to(device)
test_loader = torch.utils.data.DataLoader(
MyDataset(test_cfg,train_mode=2),
batch_size=test_cfg.batch_size, shuffle=False,
num_workers=args.workers, pin_memory=True)
predict_np=[]
checkpoint_file = os.path.join(test_cfg.checkpoint_path,args.checkpoint +'.pth.tar')
checkpoint = torch.load(checkpoint_file)
model.load_state_dict(checkpoint['state_dict'])
print('successful loaded checkpoint:{} (epoch{})'.format(checkpoint_file,checkpoint['epoch']))
model.eval()
print('testing.........')
for i, inputs in tqdm(enumerate(test_loader)):
with torch.no_grad():
input_var = torch.autograd.Variable(inputs)
out = model(input_var)
_, prediction = torch.max(out.data, 1)
pre = prediction.numpy()
predict_np.append(pre)
csv_path = os.path.join(test_cfg.img_folder,'..',test_cfg.csv_name)
data = pd.read_csv(csv_path)
predict_np=numpy.concatenate(predict_np)
data['labels'] = pd.DataFrame(predict_np)
data.to_csv(csv_path,index=False)
print('successful write the predict results!')
def train():
xml_dir = args.base_path + "data/label"
img_dir = args.base_path + "data/img"
test_dir = args.base_path + "data/test"
save_path = args.base_path + "result"
dataset_class = ['Item']
colors = ((0, 0, 0), (255, 0, 0))
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
dataset = MyDataset(img_dir, xml_dir, dataset_class)
train_dataloader = torch.utils.data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=True,
collate_fn=collate_fn)
model = FasterRCNN(num_classes=2).to(device)
params = [p for p in model.parameters() if p.requires_grad]
optimizer = torch.optim.SGD(params,
lr=args.lr,
momentum=0.9,
weight_decay=0.0005)
for epoch in range(args.num_epochs):
model.train()
train_loss = 0
for i, batch in enumerate(train_dataloader):
images, targets, image_ids = batch
images = list(image.to(device) for image in images)
targets = [{k: v.to(device)
for k, v in t.items()} for t in targets]
loss_dict = model(images, targets)
losses = sum(loss for loss in loss_dict.values())
train_loss += losses.item() * len(images)
optimizer.zero_grad()
losses.backward()
optimizer.step()
print(i, train_loss)
if i >= 0:
break
print(
f"epoch {epoch+1} loss: {train_loss / len(train_dataloader.dataset)}"
)
torch.save(model, 'model.pt')
test(model, dataset_class, colors, test_dir, save_path, device)
def test(cfg):
# transform
transform_test_list = [
transforms.Resize(size=cfg.INPUT.SIZE_TEST, interpolation=3),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]
# prepare dataset
test_dataset = MyDataset(root=root,
transform=transforms.Compose(transform_test_list),
type='test')
test_loader = DataLoader(test_dataset,
batch_size=1,
shuffle=True,
num_workers=8,
pin_memory=False)
num_classes = cfg.MODEL.HEADS.NUM_CLASSES
# prepare model
def load_network(network):
save_path = cfg.LOAD_FROM
checkpoint = torch.load(save_path)
if 'model' in checkpoint:
network.load_state_dict(checkpoint['model'])
else:
network.load_state_dict(checkpoint)
return network
model = build_model(cfg, num_classes)
model = load_network(model)
model = model.cuda()
# for data in tqdm(test_loader):
for data in test_loader:
model.train(False)
inputs, labels = data
# print(inputs.shape)
inputs = Variable(inputs.cuda().detach())
with torch.no_grad():
out = model(inputs)
# print(logits)
score, preds = torch.max(out['pred_class_logits'], 1)
# print(score, preds)
if preds.int() == 0:
cat = "No"
elif preds.int() == 1:
cat = "Yes"
print(preds.cpu().numpy().item(), labels.cpu().numpy().item())
return
# training config
epochs = 500
batch_size = 16 # mini-batch size
sample_size = 128 # number of samples to train d&g over one epoch
learning_rate = {'pretrain': 1e-5, 'g': 1e-3, 'd': 1e-3}
k_step = 1 # step of training discriminator over one epoch
sample_rate = 50 # save result every sample_reate epochs
dropout = 0
pretrain_epochs = 30 # 0 is not to do pretrain
teacher_forcing_rate = 0.7
save_path = '/home/wu/projects/emo-gan/chkpt/rnngan'
# load data
train_data_path = "/home/wu/mounts/Emo-gesture/train_set.pkl"
with open(train_data_path, 'rb') as f:
data = pickle.load(f)
dataset = MyDataset(data, max_len=max_len, num_joints=3, dim=dim)
gan = GAN(latent_code_size,
hidden_size,
generator_output_size,
discriminator_output_size,
num_layers,
bidirectional,
relu_slope,
dropout,
max_len=300)
gan.fit(dataset, epochs, batch_size, sample_size, learning_rate, k_step,
sample_rate, pretrain_epochs, teacher_forcing_rate, save_path)
batch_input = batch_input.cuda()
target = target.cuda()
output, _ = clf(batch_input)
pred = torch.max(output, 1)[1]
correct += (pred == target).sum()
acc = correct.item() / len(valid_loader.dataset)
print("Acc {:.3f}".format(acc))
return acc
if __name__ == '__main__':
# Dataloader
train_data = MyDataset('./dataset/metadata/train.csv',
'train',
transform=transform)
valid_data = MyDataset('./dataset/metadata/dev.csv',
'dev',
transform=test_transform)
train_loader = DataLoader(train_data,
batch_size=32,
shuffle=True,
num_workers=8)
valid_loader = DataLoader(valid_data,
batch_size=32,
shuffle=False,
num_workers=8)
# Model
clf = classifier().cuda()
from dataloader import MyDataset
checkpoint = '/raid/user-data/lscheucher/tmp/my_instance_segmentation/checkpoints/350my_instance_segmentation_SGD.pth'
parent_dir = checkpoint.rsplit('/', maxsplit=1)[0]
eval_folder = 'eval_' + checkpoint.rsplit('/', maxsplit=1)[1].split('.')[0]
eval_dir = os.path.join(parent_dir, eval_folder)
if not os.path.isdir(eval_dir):
os.mkdir(eval_dir)
image_dir = '/raid/group-data/uc150429/AID_DATA_201905/batch-123/original_image/'
segvecs_dir = '/raid/group-data/uc150429/AID_DATA_201905/batch-123/center_vectors/'
classes_dir = '/raid/group-data/uc150429/AID_DATA_201905/batch-123/pixelwise_annotation_xml'
dataset_eval = MyDataset(phase='val',
image_dir=image_dir,
classes_dir=classes_dir,
segvecs_dir=segvecs_dir)
sampler = None
eval_loader = torch.utils.data.DataLoader(dataset_eval,
batch_size=3,
shuffle=(sampler is None),
num_workers=2,
pin_memory=True,
sampler=sampler)
model = smp.Unet('resnet34', classes=4)
model.load_state_dict(torch.load(checkpoint))
model.eval()
# Loop over evaluation images
data_split = (train, dev, test)
np.random.seed(42)
img_list = np.random.permutation(int(sum(data_split)))
# img_list = range(0, int(sum(data_split)+1), 1)
# print(img_list)
# sys.exit()
model = Mobile_UNet()
if args.training:
train_dataset = MyDataset('train',
root,
img_list,
data_split,
transform=img_transform)
train_loader_args = dict(batch_size=batch_size,
shuffle=True,
num_workers=8)
train_loader = data.DataLoader(train_dataset, **train_loader_args)
# print(train_dataset.__len__())
dev_dataset = MyDataset('dev',
root,
img_list,
data_split,
transform=test_transform)
dev_loader_args = dict(batch_size=batch_size,
shuffle=True,
def main_worker(gpu, n_gpus_per_node, args):
global best_meteor
global scores_record
# distribute init
print(f"Use GPU: {gpu} for training"
) if args.verbose and gpu is not None else False
distributed = args.world_size > 1
args.gpu = gpu
if distributed:
args.rank = args.rank * n_gpus_per_node + gpu
dist.init_process_group('nccl',
init_method=args.dist_url,
world_size=args.world_size,
rank=args.rank)
# dataset
print('loading training dataset') if args.verbose else False
train_dataset = MyDataset(args.train_dataset,
args.text_feature,
args.audio_feature,
args.video_feature,
None,
args.min_freq,
args.modality,
max_len=args.seq_len,
context_len=args.context_len,
context=args.context,
on_memory=args.on_memory)
print('loading validation dataset') if args.verbose else False
test_dataset = MyDataset(args.test_dataset,
args.text_feature,
args.audio_feature,
args.video_feature,
train_dataset,
args.min_freq,
args.modality,
max_len=args.seq_len,
context_len=args.context_len,
context=args.context,
on_memory=args.on_memory)
# model
print('loading model') if args.verbose else False
if args.model == 'base':
target_model = SimpleEncoderDecoderCat
elif args.model == 'uni_tricoder':
target_model = UnifiedTriEncoderTransformer
else:
raise ValueError(f'Unknown model : {args.model}')
dim_feature = args.dim_audio if args.modality == 'a' else args.dim_video
model = target_model(len(train_dataset.caption_vocab),
dim_feature,
args.dim_model,
args.dim_ff,
args.head,
args.n_layer,
args.dropout,
args.modality,
n_src_vocab=len(train_dataset.text_vocab),
args=args)
torch.cuda.set_device(gpu)
model.cuda(gpu)
total_params = sum(p.numel() for p in model.parameters())
trainable_params = sum(p.numel() for p in model.parameters()
if p.requires_grad)
print(f"total parameters : {total_params}") if args.verbose else False
print(f"trainable parameters : {trainable_params}"
) if args.verbose else False
if distributed:
args.batch_size = args.batch_size // n_gpus_per_node
args.n_worker = (args.n_worker + n_gpus_per_node -
1) // n_gpus_per_node
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu], find_unused_parameters=True)
torch.backends.cudnn.benchmark = True
# dataloader
if distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset)
else:
train_sampler = None
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
num_workers=args.n_worker,
sampler=train_sampler,
shuffle=train_sampler is None)
test_loader = DataLoader(
test_dataset,
batch_size=args.batch_size,
num_workers=args.n_worker,
)
# scheduler
print('loading scheduler') if args.verbose else False
scheduler = Scheduler(model, train_dataset.pad_idx, args)
# epoch runner
print('loading epoch runner') if args.verbose else False
trainer = EpochRunner(model, train_loader, test_loader, scheduler, args)
min_loss = float('inf')
# run epoch
for i in range(args.epoch):
if train_sampler:
train_sampler.set_epoch(i)
loss = trainer.train(i)
scores_record['epoch'].append(i)
scores_record['train_loss'].append(loss)
if i < args.warm_up:
scores_record['eval_loss'].append(0)
scores_record['bleu4'].append(0)
scores_record['meteor'].append(0)
scores_record['rouge_l'].append(0)
scores_record['cider'].append(0)
continue
scores = trainer.eval(i, min_loss)
min_loss = max(min_loss, scores['eval_loss'])
if scores:
best_meteor = max(best_meteor, scores['meteor'])
is_best = best_meteor == scores['meteor']
if args.save_model and (i % args.log_freq == 0 or is_best):
save_checkpoint(
{
'epoch': i,
'state_dict': model.state_dict(),
'scores': scores,
'optimizer': scheduler.optimizer.state_dict(),
}, is_best, i, args.log_path)
print(
'**************************************************************'
)
print(f'epoch({i}): scores {scores}') if args.verbose else False
print(
'**************************************************************'
)
for each in scores:
scores_record[each].append(scores[each])
if scores['bleu4'] != 0:
record_path = os.path.join(args.log_path,
'score_record' + str(i) + '.csv')
pd.DataFrame(scores_record).to_csv(record_path)
print(f'best_meteor : {best_meteor}')
transforms.Resize(100),
transforms.CenterCrop(100),
transforms.ToTensor(),
normalize,
])
# Create train, validation, and test data set
#TRAIN DATA
s=np.arange(39209) # total number of training data is 39209
s_train, s_val = train_test_split(s, test_size=0.35, random_state=1)
#s_train, s_test = train_test_split(s_train, test_size=0.25, random_state=1)
ds_train = MyDataset(data_path,'Train',train_transform, s_train)
ds_val = MyDataset(data_path,'Train',train_transform, s_val)
# TEST DATA
s=np.arange(12630) #total number of test data is 12630
#np.random.seed(43)
#np.random.shuffle(s)
ds_test = MyDataset(data_path,'Test',train_transform, s)
print(ds_train.names.shape)
print(ds_val.names.shape)
print(ds_test.names.shape)
train_loader = torch.utils.data.DataLoader(dataset=ds_train,
label_ls_test = text_data.get_label_ls('test')
word2idx = vocab.get_word2idx()
word2idx = int_dict_increment_by(vocab.get_word2idx(),2)
word2idx["<pad>"] = 0
word2idx["<unk>"] = 1
idx2word = dict_kv2vk(word2idx)
print("dataloaded...time taken: " + str(time.time() - start))
train_x = text_2_int_list(text_ls_train, word2idx, args["max_doc_len"])
valid_x = text_2_int_list(text_ls_val, word2idx, args["max_doc_len"])
test_x = text_2_int_list(text_ls_test, word2idx, args["max_doc_len"])
# datasets
train = MyDataset(train_x, label_ls_train)
valid = MyDataset(valid_x, label_ls_val)
test = MyDataset(test_x, label_ls_test)
# hyper-parameters
task = config["preprocess"]["dataset"]
embed_dim = args["embed_dim"]
nhead = args["nhead"]
nhid = args["nhid"]
nlayers = args["nlayers"]
vocab_size = len(word2idx)
bs = args["batch_size"]
nclass = len(label2idx)
lr = args["lr"]
grad_clip = args["grad_clip"]
print_iter = args["print_iter"]
from dataloader import MyDataset
def count_acc(logits, label):
pred = torch.argmax(logits, dim=1)
return (pred == label).type(torch.cuda.FloatTensor).mean().item()
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--max_epoch', type=int, default=50)
parser.add_argument('--batch_size', type=int, default=256)
parser.add_argument('--lr', type=float, default=0.01)
parser.add_argument('--step_size', type=int, default=5)
parser.add_argument('--gamma', type=float, default=0.5)
args = parser.parse_args()
trainset = MyDataset('train')
train_loader = DataLoader(dataset=trainset, num_workers=4, batch_size=args.batch_size, shuffle=True, drop_last=True, pin_memory=True)
valset = MyDataset('val')
val_loader = DataLoader(dataset=valset, num_workers=4, batch_size=args.batch_size, pin_memory=True)
testset = MyDataset('test')
test_loader = DataLoader(dataset=testset, num_workers=4, batch_size=args.batch_size, pin_memory=True)
model = ConvNet()
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=args.step_size, gamma=args.gamma)
model = model.cuda()
best_acc = 0.0
for epoch in range(args.max_epoch):
def run_train():
learning_rate = 1e-3
folder_path = "D:/Dataset/VOC_Dataset/"
train_dataset = MyDataset(folder_path=folder_path + 'VOC2012_trainval',
train_category='train')
val_dataset = MyDataset(folder_path=folder_path + 'VOC2012_trainval',
train_category='val')
batch_size = 8
train_data_loader = DataLoader(dataset=train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4)
val_data_loader = DataLoader(dataset=val_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4)
my_model = model.YOLO_v2().cuda()
my_model.train()
optimizer = optim.Adam(my_model.parameters(), lr=1e-4, weight_decay=1e-5)
scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.95)
training_epoch = 135
grid_size = 13
criterion = losses.DetectionLoss().cuda()
train_loss = []
val_loss = []
sum_correct, sum = 0, 0
for epoch in range(training_epoch):
x_line = [i + 1 for i in range(epoch + 1)]
epoch_loss = []
print('------training------')
for it, data in enumerate(train_data_loader):
x = data[0].cuda()
y = data[1]
path = data[2][0]
y_pred = my_model(x).cuda()
gt = list(y)
loss, obj_loss, no_obj_loss, conf_loss = criterion(y_pred, y)
epoch_loss.append(loss.item())
if it % 100 == 0: #and it > 0:
print(
'step [{0:} / {1}] \t loss : {2:3.4f} \t obj_loss : {3:3.4f} \t no_obj_loss : {4:3.4f} \t conf_loss : {5:3.4f}'
.format(it, len(train_data_loader), loss, obj_loss,
no_obj_loss, conf_loss))
optimizer.zero_grad()
loss.backward()
scheduler.optimizer.step()
train_loss.append(np.mean(epoch_loss))
epoch_loss = []
print('------validation------')
for it, data in enumerate(val_data_loader):
x = data[0].cuda()
img = x.clone()
y = data[1]
path = data[2]
y_pred = my_model(x)
loss, obj_loss, no_obj_loss, conf_loss = criterion(y_pred, y)
loss.cuda()
epoch_loss.append(loss.item())
if it % 300 == 0 and it > 0:
print(
'step [{0:} / {1}] \t loss : {2:3.4f} \t obj_loss : {3:3.4f} \t no_obj_loss : {4:3.4f} \t conf_loss : {5:3.4f}'
.format(it, len(val_data_loader), loss, obj_loss,
no_obj_loss, conf_loss))
# pred, pred_img = NonMaxSupression(y_pred, path, grid_size)
# print(pred.shape, pred_img.shape)
# cv.imshow("pred", pred_img)
# cv.waitKey(0)
val_loss.append(np.mean(epoch_loss))
print(
'step [{0:} / {1}] \t loss : {2:3.4f} \t obj_loss : {3:3.4f} \t no_obj_loss : {4:3.4f} \t conf_loss : {5:3.4f}'
.format(epoch, training_epoch, loss, obj_loss, no_obj_loss,
conf_loss))
plt.plot(x_line, train_loss, 'r-', label='train')
plt.plot(x_line, val_loss, 'b-', label='val')
plt.xlabel('Epoch')
plt.ylabel('Loss')
plt.title('YOLO_v2')
# plt.show()
plt.savefig('YOLO_v2_loss.png', dpi=300)
torch.save(my_model.state_dict(),
'Weights/YOLO_v2_{}.pt'.format(epoch + 1))
def main(args=None):
train_transform = transforms.Compose([
transforms.Resize(96),
transforms.RandomHorizontalFlip(p=0.5),
#transforms.RandomRotation(15),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])
test_transform = transforms.Compose([
transforms.Resize(96),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])
dataset_train = MyDataset(label_path=args.train_txt,
transform=train_transform)
dataset_val = MyDataset(label_path=args.val_txt, transform=test_transform)
dataloader_train = DataLoader(dataset=dataset_train,
batch_size=args.batch_size,
shuffle=True)
dataloader_val = DataLoader(dataset=dataset_val,
batch_size=args.batch_size,
shuffle=False)
batch_num = len(dataset_train) // args.batch_size
model = mobilenetv3()
print('network:')
print(model)
save_path = './checkpoint'
if not os.path.exists(save_path):
os.mkdir(save_path)
use_cuda = torch.cuda.is_available() and len(args.gpus) > 0
if use_cuda:
torch.cuda.set_device(args.gpus[0])
torch.cuda.manual_seed(args.seed)
#model = torch.nn.DataParallel(model, device_ids=args.gpus)
model.to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=1e-4, weight_decay=0.0005)
milestones = [50, 80, 120, 150]
scheduler = optim.lr_scheduler.MultiStepLR(optimizer,
milestones,
gamma=0.1,
last_epoch=-1)
writer = SummaryWriter(log_dir='./summary')
for epoch in range(args.start_epoch, args.epochs + 1):
train_loss = train(dataloader_train, model, criterion, optimizer,
epoch, scheduler, batch_num)
test_loss = test(dataloader_val, model, criterion)
scheduler.step()
model_name = 'mask_detection'
save_name = '{}/{}_{}.pth.tar'.format(save_path, model_name, epoch)
save_checkpoint(
{
'epoch': epoch,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict()
},
filename=save_name)
writer.add_scalars('scalar/loss', {
'train_loss': train_loss,
'test_loss': test_loss
}, epoch)
writer.export_scalars_to_json('./summary/' + 'pretrain' +
'all_scalars.json')
writer.close()
train_instances = prepare_instance(dicts, args.data_path, args,
args.MAX_LENGTH)
print("train_instances {}".format(len(train_instances)))
if args.version != 'mimic2':
dev_instances = prepare_instance(
dicts, args.data_path.replace('train', 'dev'), args,
args.MAX_LENGTH)
print("dev_instances {}".format(len(dev_instances)))
else:
dev_instances = None
test_instances = prepare_instance(dicts,
args.data_path.replace('train', 'test'),
args, args.MAX_LENGTH)
print("test_instances {}".format(len(test_instances)))
train_loader = DataLoader(MyDataset(train_instances),
args.batch_size,
shuffle=True,
collate_fn=my_collate)
if args.version != 'mimic2':
dev_loader = DataLoader(MyDataset(dev_instances),
1,
shuffle=False,
collate_fn=my_collate)
else:
dev_loader = None
test_loader = DataLoader(MyDataset(test_instances),
1,
shuffle=False,
collate_fn=my_collate)
def main():
# training set
data_set = MyDataset(200, dir_path=train_data_dir)
data_set_val = MyDatasetEval(200, dir_path=train_data_dir)
# dataloader = torch.utils.data.DataLoader(data_set, batch_size=16, shuffle=True)
label_list = []
for i in range(len(data_set)):
label_list.append(data_set[i][1])
total_loss = 0
total_size = 0
#cross_validationごとのclassification_report
report1 = ""
report2 = ""
report3 = ""
conf1 = ""
conf2 = ""
conf3 = ""
pre_score = 0
re_score = 0
f_score = 0
#cross_validation
kf = StratifiedKFold(n_splits=3, shuffle=True, random_state=0)
#loss function
criterion = nn.CrossEntropyLoss()
# criterion = FocalLoss(alpha=0.25, gamma=2)
for fold_idx, idx in enumerate(kf.split(data_set, label_list)):
# モデルを構築
# net = models.compare(pretrained=True)
net = models.resnet34(pretrained=True)
# print(net)
#vgg16の全結合
# num_ftrs = net.classifier[6].in_features
# net.classifier[6] = nn.Linear(num_ftrs, 2)
#resnetの全結合
num_ftrs = net.fc.in_features
net.fc = nn.Linear(num_ftrs, 2)
# print(net)
net = net.to(device)
print("ネットワーク設定完了：学習をtrainモードで開始します")
# モデルの重みを読み込み
# net.load_state_dict(torch.load(weight_pash), strict=False)
# 最適化
optimizer = optim.SGD(net.parameters(), lr=0.001)
train_idx, valid_idx = idx
train_loader = torch.utils.data.DataLoader(Subset(data_set, train_idx),
batch_size=16,
shuffle=True)
valid_loader = torch.utils.data.DataLoader(Subset(
data_set_val, valid_idx),
batch_size=1,
shuffle=False)
# trainモードで開始
net.train()
# modelの全体像を表示
summary(net, input_size=(3, 200, 200))
train_loss_value = [] # trainingのlossを保持するlist
train_acc_value = [] # trainingのaccuracyを保持するlist
# 同じデータを回学習します
for epoch in range(10):
print("epoch =", epoch + 1)
# 今回の学習効果を保存するための変数
running_loss = 0.0
for batch_idx, data in enumerate(
train_loader): # dataがラベルと画像情報の2つの情報を持つ
# データ整理
inputs, labels = data
inputs = inputs.to(device)
labels = labels.to(device)
# 前回の勾配情報をリセット
optimizer.zero_grad()
# 予測
outputs = net(inputs)
# 予測結果と教師ラベルを比べて損失を計算
loss = criterion(outputs, labels)
running_loss += loss.item()
total_loss += loss.item()
total_size += inputs.size(0)
# 損失に基づいてネットワークのパラメーターを更新
loss.backward()
optimizer.step()
# if batch_idx % 1 == 0:
# now = datetime.datetime.now()
# print('[{}] Train Epoch: {} [{}/{} ({:.0f}%)]\tAverage loss: {:.6f}'.format(
# now, epoch, batch_idx * len(inputs), len(dataloader.dataset), 100. * batch_idx / len(dataloader),
# total_loss / total_size))
# train_loss_value.append(running_loss * 50 / len(dataloader.dataset)) # traindataのlossをグラフ描画のためにlistに保持
# print("running_loss=", running_loss * 50 / len(dataloader.dataset))
if batch_idx % 1 == 0:
now = datetime.datetime.now()
print(
'[{}] Train Epoch: {} [{}/{} ({:.0f}%)]\tAverage loss: {:.6f}'
.format(now, epoch + 1, batch_idx * len(inputs),
len(train_loader.dataset),
100. * batch_idx / len(train_loader),
total_loss / total_size))
train_loss_value.append(
running_loss /
len(train_loader.dataset)) # traindataのlossをグラフ描画のためにlistに保持
print("train_loss=", running_loss / len(train_loader))
#評価モード
net.eval()
print("ネットワークをevalに変更します")
#test
with torch.no_grad():
pred = []
Y = []
target_name = ["correct 0", "broken 1"]
for batch_idx, data in enumerate(valid_loader):
input_val, label_val, image_name = data
input_val = input_val.to(device)
label_val = label_val.to(device)
image_name = image_name[0].replace("'", "").replace(",", "")
output_val = net(input_val)
pred += [int(l.argmax()) for l in output_val]
Y += [int(l) for l in label_val]
if Y[-1] != pred[-1]:
if Y[-1] == 1:
mis_image = cv2.imread(
os.path.join(
train_data_dir, "broken/{filename}".format(
filename=image_name)))
# print(os.path.join(train2, "{filename}".format(filename=image_name)))
cv2.imwrite(
os.path.join(
img_save_dir, "broken/{filename}".format(
filename=image_name)), mis_image)
else:
mis_image = cv2.imread(
os.path.join(
train_data_dir, "correct/{filename}".format(
filename=image_name)))
# print(os.path.join(train2, "{filename}".format(filename=image_name)))
cv2.imwrite(
os.path.join(
img_save_dir, "correct/{filename}".format(
filename=image_name)), mis_image)
# save_imageはネットの出力を保存する
# save_image(output_val, os.path.join(img_save_dir, "{filename}".format(filename=image_name)), nrow=1)
pre_score += precision_score(Y, pred)
re_score += recall_score(Y, pred)
f_score += f1_score(Y, pred)
print(classification_report(Y, pred, target_names=target_name))
print(confusion_matrix(Y, pred))
if fold_idx == 0:
report1 = classification_report(Y,
pred,
target_names=target_name)
conf1 = confusion_matrix(Y, pred)
# モデルを保存
model_path = 'model_cross1.pth'
torch.save(net.state_dict(), model_path)
elif fold_idx == 1:
report2 = classification_report(Y,
pred,
target_names=target_name)
conf2 = confusion_matrix(Y, pred)
# モデルを保存
model_path = 'model_cross2.pth'
torch.save(net.state_dict(), model_path)
else:
report3 = classification_report(Y,
pred,
target_names=target_name)
conf3 = confusion_matrix(Y, pred)
# モデルを保存
model_path = 'model_cross3.pth'
torch.save(net.state_dict(), model_path)
conf_exp = "TN FP\nFN TP"
print("====cross1の結果====")
print(report1)
print(conf1, "\n", conf_exp)
print("====cross2の結果====")
print(report2)
print(conf2, "\n", conf_exp)
print("====cross3の結果====")
print(report3)
print(conf3, "\n", conf_exp)
print("broken_precision=", pre_score / 3)
print("broken_recall=", re_score / 3)
print("broken_f1=", f_score / 3)
model = smp.Unet('resnet34', classes=params.numclasses+2, encoder_weights='imagenet')
model = model.to(device)
#output = model.forward(inputs)
#print("output-shape:", output.shape)
preprocessing_fn = smp.encoders.get_preprocessing_fn(encoder_name = 'resnet34', pretrained='imagenet')
"""dataloader"""
image_dir = '/raid/group-data/uc150429/AID_DATA_201905/batch-123/original_image/'
segvecs_dir = '/raid/group-data/uc150429/AID_DATA_201905/batch-123/center_vectors/'
classes_dir = '/raid/group-data/uc150429/AID_DATA_201905/batch-123/pixelwise_annotation_xml'
dataset_train = MyDataset(phase='train', image_dir = image_dir, classes_dir = classes_dir, segvecs_dir = segvecs_dir)
dataset_val = MyDataset(phase='val', image_dir = image_dir, classes_dir = classes_dir, segvecs_dir = segvecs_dir)
#print(x.shape)
train_sampler = None
train_loader = torch.utils.data.DataLoader(
dataset_train,
batch_size=params.batch_size, shuffle=(train_sampler is None),
num_workers=20, pin_memory=True, sampler=train_sampler)
val_loader = torch.utils.data.DataLoader(
dataset_val,
batch_size=params.batch_size, shuffle=False,
from model import classifier, classifier_sep
import argparse
if __name__ == '__main__':
# Command parser
parser = argparse.ArgumentParser()
parser.add_argument("-p",
'--checkpoint',
type=str,
required=True,
help="model checkpoint")
args = parser.parse_args()
# Dataloader
test_data = MyDataset('./dataset/metadata/test.csv',
'test',
transform=test_transform)
test_loader = DataLoader(test_data,
batch_size=32,
shuffle=False,
num_workers=8)
clf = classifier()
clf.load_state_dict(torch.load(args.checkpoint))
clf = clf.cuda()
clf.eval()
f = open('test.csv', 'w')
preds = []
for i, (batch_input, target) in enumerate(test_loader):
def main(args):
if not os.path.isdir(args.checkpoint):
try:
os.makedirs(args.checkpoint)
except OSError as e:
if e.errno != errno.EEXIST:
raise
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = networks(num_class=train_cfg.num_class, pretrained=True).to(device)
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(),
lr=5e-4,
weight_decay=1e-5)
train_loader = torch.utils.data.DataLoader(MyDataset(train_cfg,
train_mode=0),
batch_size=train_cfg.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
eval_loader = torch.utils.data.DataLoader(MyDataset(val_cfg, train_mode=1),
batch_size=val_cfg.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
#ini log
log = open(os.path.join(args.checkpoint, 'log.txt'), 'w')
log.write('Epoch')
log.write('\t')
log.write('LR')
log.write('\t')
log.write('Avg Train Loss')
log.write('\t')
log.write('Acc')
log.write('\t')
log.write('\n')
log.flush()
print('Total params: %.2fMB' %
(sum(p.numel() for p in model.parameters()) / (1024 * 1024) * 4))
for epoch in range(args.start_epoch, train_cfg.epochs):
lr = adjust_learning_rate(optimizer, epoch)
print('epoch {},lr={}'.format(epoch, lr))
print('begin the {}th epoch'.format(epoch))
train_loss = train(train_loader, model, criterion, optimizer)
print('avg_train_loss:{:.5f}'.format(train_loss /
train_loader.__len__()))
acc = eval(eval_loader, model)
print('accuracy:{:.5f}'.format(acc))
log.write('{}'.format(epoch))
log.write('\t')
log.write('{:.5f}'.format(lr))
log.write('\t')
log.write('{:.5f}'.format(train_loss / train_loader.__len__()))
log.write('\t')
log.write('{:.5f}'.format(acc))
log.write('\t')
log.write('\n')
log.flush()
filename = 'epoch' + str(epoch + 1) + 'checkpoint.pth.tar'
filepath = os.path.join(args.checkpoint, filename)
torch.save(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
}, filepath)
log.close()
def main(world_size,
epochs,
rank,
batch_size=200,
backend='nccl',
data_path='/dataset',
lr=1e-3,
momentum=0.01,
weight_decay=1e-5,
no_cuda=False,
seed=35,
aggregation_method='naive',
load_model=False,
load_path='/data'):
'''Main Function'''
use_cuda = not no_cuda and torch.cuda.is_available(
) # 使用的使用多个gpu进行训练，需要改进一下
torch.manual_seed(seed)
timeline = time.strftime('%m%d%Y_%H:%M', time.localtime())
device = torch.device("cuda" if use_cuda else "cpu")
ratio = 0.8551957853612336
weight = torch.FloatTensor([ratio, 1 - ratio]).to(device)
Loss = NN.BCELoss(weight=weight)
start_epoch = 1
result_dir = 'logs' + '/' + timeline + '/' + 'results' + '/' + '{:s}' + '/'
model_dir = 'logs' + '/' + timeline + '/' + 'models' + '/' + '{:s}' + '/'
param_dir = 'logs' + '/' + timeline + '/'
csvname = '{:s}_log.csv'
modelname = 'model_{:d}.pth'
paramname = 'param.csv'
param = {
'world_size': world_size,
'batch_size': batch_size,
'bachend': backend,
'lr': lr,
'momentum': momentum,
'weight_decay': weight_decay,
'seed': seed,
'aggregation': aggregation_method
}
if rank == 0:
name = 'master'
if not os.path.exists(param_dir):
os.makedirs(param_dir)
with open(param_dir + paramname, 'a', newline='') as p:
fieldnames = param.keys()
writer = csv.DictWriter(p, fieldnames=fieldnames)
writer.writeheader()
writer.writerow(param)
p.close()
if aggregation_method == 'distillation':
DistillationData = MyDataset(root=data_path,
train=True,
data_root='distillation.csv')
distillation_dataloader = DataLoader(dataset=DistillationData,
batch_size=batch_size,
shuffle=True,
drop_last=True)
model_set = []
for worker_id in range(world_size):
model_set.append(CNN_Model_withDropout().to(device))
opt_set = []
for worker_id in range(world_size):
opt_set.append(
optim.SGD(model_set[worker_id].parameters(),
lr=lr,
momentum=momentum,
weight_decay=weight_decay))
if load_model:
if aggregation_method == 'distillation':
raise ValueError('Unexpected model')
checkpoint = torch.load(load_path)
for worker_id in range(world_size):
model_set[worker_id].load_state_dict(
checkpoint['model_state_dict'])
opt_set[worker_id].load_state_dict(
checkpoint['opt_state_dict'])
start_epoch = checkpoint['epoch']
model = SF.Master(model=model_set[0],
backend=backend,
rank=rank,
world_size=world_size,
learning_rate=lr,
device=device,
aggregation_method=aggregation_method)
for epoch in range(start_epoch, epochs + 1):
model.train()
model.step(model_buffer=model_set)
model.update(model_set[1:])
for worder_id in range(world_size):
adjust_learning_rate(opt_set[worker_id], epoch, lr)
if aggregation_method == 'distillation':
distillation(NN_set=model_set[1:],
opt_set=opt_set[1:],
dataset=distillation_dataloader,
world_size=world_size,
epoch=epoch,
device=device)
# best_idx = choose_best(NN_set=model_set[1:], name=name, dataset=dataloader,world_size=world_size,
# epoch=epoch, Loss=Loss, time=timeline)
# best_state_dict = model_set[best_idx+1].state_dict()
# model_set[0].load_state_dict(best_state_dict)
# 这里要回传所有的模型
else:
name = 'worker' + str(rank)
DataSet_train = MyDataset(root=data_path,
train=True,
data_root='{}.csv'.format(name))
dataloader_train = DataLoader(dataset=DataSet_train,
batch_size=batch_size,
shuffle=True,
drop_last=True)
DataSet_test = MyDataset(root=data_path,
train=True,
data_root='{}.csv'.format('test'))
dataloader_test = DataLoader(dataset=DataSet_test,
batch_size=batch_size,
shuffle=True,
drop_last=True)
model_set = []
for worker_id in range(world_size):
model_set.append(CNN_Model_withDropout().to(device))
train_model = model_set[0]
optimizer = optim.SGD(train_model.parameters(),
lr=lr,
momentum=momentum,
weight_decay=weight_decay)
if load_model:
checkpoint = torch.load(load_path)
for worker_id in range(world_size):
model_set[worker_id].load_state_dict(
checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['opt_state_dict'])
start_epoch = checkpoint['epoch']
train_model = model_set[0]
backup_model = CNN_Model_withDropout().to(device)
model = SF.Servent(model=train_model,
backend=backend,
rank=rank,
world_size=world_size,
device=device,
aggregation_method=aggregation_method)
for epoch in range(start_epoch, epochs + 1):
model.train()
model.step(model_buffer=model_set, rank=rank)
best_state_dict = train_model.state_dict()
backup_model.load_state_dict(best_state_dict)
adjust_learning_rate(optimizer, epoch, lr)
train(dataloader=dataloader_train,
model=train_model,
optimizer=optimizer,
Loss=Loss,
epoch=epoch,
time=timeline,
result_dir=result_dir.format(name),
model_dir=model_dir.format(name),
device=device,
csvname=csvname.format(name),
modelname=modelname)
model.update(backup_model)
test(dataloader=dataloader_test,
model=train_model,
epoch=epoch,
Loss=Loss,
time=timeline,
result_dir=result_dir.format(name),
model_dir=model_dir.format(name),
csvname=csvname.format(name),
modelname=modelname,
device=device)
with open('./configs/train_params.yaml') as config_file:
config = yaml.load(config_file, Loader=yaml.FullLoader)
config = Dict(config)
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
print('Device: ', device)
model = MyModel()
model.to(device)
train_df = pd.read_csv('../data/train_dataset.csv')
test_df = pd.read_csv('../data/test_dataset.csv')
train_dataset = MyDataset(train_df,
image_folder_path='../data/images',
augmentations=get_training_augmentation())
test_dataset = MyDataset(test_df,
image_folder_path='../data/images',
augmentations=get_validation_augmentation())
train_dataloader = DataLoader(train_dataset, batch_size=config.batch_size)
test_dataloader = DataLoader(test_dataset, batch_size=1)
optimizer = Adam(model.parameters(), lr=config.learning_rate)
loss_function = CrossEntropyLoss()
for epoch in range(config.epochs):
package_name = 'net_lstm'
model_name = 'LSTMClassifaierAndDenoise'
model_arg = 39
# batch_size = 10000
batch_size = 1
optimazier = 'Adam'
lr = 0.001
root_path = '/media/ophir/DATA1/Ophir/DeepLearning/project/data_for_lstm'
# mfcc_path = os.path.join(root_path, 'TRAIN', 'MFCC')
# stft_path = os.path.join(root_path, 'TRAIN', 'STFT')
# map_path = os.path.join(root_path, 'TRAIN', 'MAP')
main_path_train = os.path.join(root_path, 'TRAIN')
# main_path_train = '/media/ophir/DATA1/Ophir/DeepLearning/project/data_in_batchs'
train_data = MyDataset(main_path_train)
trainloader = DataLoader(train_data,
batch_size=batch_size,
shuffle=False,
num_workers=0)
# valid_mfcc_path = os.path.join(root_path, 'TEST', 'MFCC')
# valid_stft_path = os.path.join(root_path, 'TEST', 'STFT')
# valid_map_path = os.path.join(root_path, 'TEST', 'MAP')
main_path_val = os.path.join(root_path, 'TEST')
valid_data = MyDataset(main_path_val)
validloader = DataLoader(train_data,
batch_size=batch_size,
shuffle=False,
num_workers=0)
def train(args):
print('Start')
if torch.cuda.is_available():
device = 'cuda'
torch.set_default_tensor_type('torch.cuda.FloatTensor')
else:
device = 'cpu'
train_epoch = args.train_epoch
lr = args.lr
beta1 = args.beta1
beta2 = args.beta2
batch_size = args.batch_size
noise_var = args.noise_var
h_dim = args.h_dim
images_path = glob.glob(args.data_dir+'/face_images/*/*.png')
random.shuffle(images_path)
split_num = int(len(images_path)*0.8)
train_path = images_path[:split_num]
test_path = images_path[split_num:]
result_path = images_path[-15:]
train_dataset = MyDataset(train_path)
train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
test_dataset = MyDataset(test_path)
test_dataloader = torch.utils.data.DataLoader(test_dataset, batch_size=batch_size, shuffle=True)
result_dataset = MyDataset(result_path)
result_dataloader = torch.utils.data.DataLoader(result_dataset, batch_size=result_dataset.__len__(), shuffle=False)
result_images = next(iter(result_dataloader))
model = AutoEncoder(h_dim=h_dim).to(device)
criterion = nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr, (beta1, beta2))
out_path = args.model_dir
train_loss_list = []
test_loss_list = []
for epoch in range(train_epoch):
model.to(device)
loss_train = 0
for x in train_dataloader:
noised_x = add_noise(x, noise_var)
recon_x = model(noised_x)
loss = criterion(recon_x, x)
optimizer.zero_grad()
loss.backward()
optimizer.step()
loss_train += loss.item()
loss_train /= train_dataloader.__len__()
train_loss_list.append(loss_train)
if epoch % 1 == 0:
with torch.no_grad():
model.eval()
loss_test = 0
for x_test in test_dataloader:
recon_x_test = model(x_test)
loss_test += criterion(recon_x_test, x_test).item()
loss_test /= test_dataloader.__len__()
test_loss_list.append(loss_test)
np.save(os.path.join(out_path, 'train_loss.npy'), np.array(train_loss_list))
np.save(os.path.join(out_path, 'test_loss.npy'), np.array(test_loss_list))
model.train()
def train(cfg):
# logger
logger = logging.getLogger(name="merlin.baseline.train")
logger.info("training...")
# transform
transform_train_list = [
# transforms.RandomResizedCrop(size=128, scale=(0.75,1.0), ratio=(0.75,1.3333), interpolation=3), #Image.BICUBIC)
transforms.Resize(size=cfg.INPUT.SIZE_TRAIN, interpolation=1),
transforms.Pad(32),
transforms.RandomCrop(cfg.INPUT.SIZE_TRAIN),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]
transform_val_list = [
transforms.Resize(size=cfg.INPUT.SIZE_TEST, interpolation=3),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]
# prepare dataset
train_dataset = MyDataset(root=cfg.DATA.ROOT, transform=transforms.Compose(transform_train_list), type='train')
val_dataset = MyDataset(root=cfg.DATA.ROOT, transform=transforms.Compose(transform_val_list), type='val')
train_loader = DataLoader(train_dataset,
batch_size=cfg.SOLVER.BATCH_SIZE,
shuffle=True,
num_workers=8,
pin_memory=False)
val_loader = DataLoader(val_dataset,
batch_size=cfg.SOLVER.BATCH_SIZE,
shuffle=True,
num_workers=8,
pin_memory=False)
num_classes = cfg.MODEL.HEADS.NUM_CLASSES
# prepare model
model = build_model(cfg, num_classes)
model = model.cuda()
model = nn.DataParallel(model)
# prepare solver
optimizer = make_optimizer(cfg, model)
scheduler = WarmupMultiStepLR(optimizer, cfg.SOLVER.STEPS, cfg.SOLVER.GAMMA, cfg.SOLVER.WARMUP_FACTOR,
cfg.SOLVER.WARMUP_ITERS, cfg.SOLVER.WARMUP_METHOD)
start_epoch = 0
# Train and val
since = time.time()
for epoch in range(start_epoch, cfg.SOLVER.MAX_EPOCHS):
model.train(True)
logger.info("Epoch {}/{}".format(epoch, cfg.SOLVER.MAX_EPOCHS - 1))
logger.info('-' * 10)
running_loss = 0.0
# Iterate over data
it = 0
running_acc = 0
for data in train_loader:
it += 1
# get the inputs
inputs, labels = data
now_batch_size, c, h, w = inputs.shape
if now_batch_size < cfg.SOLVER.BATCH_SIZE: # skip the last batch
continue
# wrap them in Variable
inputs = Variable(inputs.cuda().detach())
labels = Variable(labels.cuda().detach())
# zero the parameter gradients
optimizer.zero_grad()
# forward
out = model(inputs)
loss_dict = get_loss(cfg, outs=out, label=labels)
loss = sum(loss_dict.values())
loss.backward()
optimizer.step()
scheduler.step()
# statistics
with torch.no_grad():
_, preds = torch.max(out['pred_class_logits'], 1)
running_loss += loss
running_acc += torch.sum(preds == labels.data).float().item() / cfg.SOLVER.BATCH_SIZE
if it % 50 == 0:
logger.info(
'epoch {}, iter {}, loss: {:.3f}, acc: {:.3f}, lr: {:.5f}'.format(
epoch, it, running_loss / it, running_acc / it,
optimizer.param_groups[0]['lr']))
epoch_loss = running_loss / it
epoch_acc = running_acc / it
logger.info('epoch {} loss: {:.4f} Acc: {:.4f}'.format(epoch, epoch_loss, epoch_acc))
# save checkpoint
if epoch % cfg.SOLVER.CHECKPOINT_PERIOD == 0:
checkpoint = {'epoch': epoch + 1,
'model': model.module.state_dict() if (len(cfg.MODEL.DEVICE_ID) - 2) > 1 else model.state_dict(),
'optimizer': optimizer.state_dict()
}
save_checkpoint(checkpoint, epoch, cfg)
# evaluate
if epoch % cfg.SOLVER.EVAL_PERIOD == 0:
logger.info('evaluate...')
model.train(False)
total = 0.0
correct = 0.0
for data in val_loader:
inputs, labels = data
inputs = Variable(inputs.cuda().detach())
labels = Variable(labels.cuda().detach())
with torch.no_grad():
out = model(inputs)
_, preds = torch.max(out['pred_class_logits'], 1)
c = (preds == labels).squeeze()
total += c.size(0)
correct += c.float().sum().item()
acc = correct / total
logger.info('eval acc:{:.4f}'.format(acc))
time_elapsed = time.time() - since
logger.info('Training complete in {:.0f}m {:.0f}s\n'.format(
time_elapsed // 60, time_elapsed % 60))
return model
import torch
import glob
from torch.utils.data import DataLoader
from torchvision import transforms
from dataloader import MyDataset
from fmeasure import calculateF1Measure
composed = transforms.Compose([transforms.Grayscale(1), transforms.ToTensor()])
test_dataset = MyDataset('/path/*', '/path/*', transform=composed)
test_dataloader = DataLoader(test_dataset,
batch_size=1,
shuffle=False,
num_workers=8)
filenames = sorted(glob.glob('/path/*.png'))
filenames = [x.split('/')[-1] for x in filenames]
with torch.no_grad():
sum_val_loss_g1 = 0
sum_val_false_ratio_g1 = 0
sum_val_detect_ratio_g1 = 0
sum_val_F1_g1 = 0
g1_time = 0
sum_val_loss_g2 = 0
sum_val_false_ratio_g2 = 0
sum_val_detect_ratio_g2 = 0
sum_val_F1_g2 = 0