def main(config, resume):
# parameters
batch_size = config.get('batch_size', 32)
start_epoch = config['epoch']['start']
max_epoch = config['epoch']['max']
lr = config.get('lr', 0.0005)
use_conf = config.get('use_conf', False)
## path
save_path = config['save_path']
timestamp = datetime.now().strftime(r"%Y-%m-%d_%H-%M-%S")
save_path = os.path.join(save_path, timestamp)
result_path = os.path.join(save_path, 'result')
if not os.path.exists(result_path):
os.makedirs(result_path)
model_path = os.path.join(save_path, 'model')
if not os.path.exists(model_path):
os.makedirs(model_path)
dest = shutil.copy('train.py', save_path)
print("save to: ", dest)
## cuda or cpu
if config['n_gpu'] == 0 or not torch.cuda.is_available():
device = torch.device("cpu")
print("using CPU")
else:
device = torch.device("cuda:0")
## dataloader
dataset = Dataset(phase='train', do_augmentations=False)
data_loader = DataLoader(
dataset,
batch_size=int(batch_size),
num_workers=1,
shuffle=True,
drop_last=True,
pin_memory=True,
# **loader_kwargs,
)
val_dataset = Dataset(phase='val', do_augmentations=False)
val_data_loader = DataLoader(
val_dataset,
batch_size=int(batch_size),
num_workers=1,
shuffle=True,
drop_last=True,
pin_memory=True,
# **loader_kwargs,
)
## few shot
do_few_shot = True
if do_few_shot:
fs_dataset = Dataset(
phase='train',
do_augmentations=False,
metafile_path='metadata/detection_train_images.json')
fs_data_loader = DataLoader(
fs_dataset,
batch_size=int(128),
num_workers=1,
shuffle=True,
pin_memory=True,
# **loader_kwargs,
)
## CNN model
output_dim = 3
model = MyNet(output_dim)
model = model.to(device)
model.train()
print(model)
## loss
criterion = nn.CrossEntropyLoss(reduction='none')
## optimizer
params = list(filter(lambda p: p.requires_grad, model.parameters()))
optim_params = {
'lr': lr,
'weight_decay': 0,
'amsgrad': False,
}
optimizer = torch.optim.Adam(params, **optim_params)
lr_params = {
'milestones': [10],
'gamma': 0.1,
}
lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, **lr_params)
loss_avg = AverageMeter()
acc_avg = AverageMeter()
fs_loss_avg = AverageMeter()
fs_acc_avg = AverageMeter()
logger = SimpleLogger(['train_loss', 'train_acc', 'val_loss', 'val_acc'])
## loop
for epoch in range(start_epoch, max_epoch):
loss_avg.reset()
for batch_idx, batch in tqdm(
enumerate(data_loader),
total=len(data_loader),
ncols=80,
desc=f'training epoch {epoch}',
):
data = batch[0].to(device)
gt_lbls = batch[1].to(device)
gt_gt_lbls = batch[2].to(device)
## set zerograd
optimizer.zero_grad()
## run forward pass
out = model(data) ## logits: [B, NC]; conf: [B, 1]
preds = torch.max(out, dim=-1)[1]
# print("out shape: ", out.shape)
weights = model.compute_entropy_weight(out)
# print("weights shape: ", weights.shape)
## compute loss
class_loss = criterion(out, gt_lbls) ## [B, 1]
# print("class_loss shape: ", class_loss.shape)
if use_conf:
loss = (class_loss * (weights**2) + (1 - weights)**2).mean()
else:
loss = class_loss.mean()
## record
loss_avg.update(loss.item(), batch_size)
positive = ((gt_lbls == preds) + (gt_gt_lbls > 2)).sum()
batch_acc = positive.to(torch.float) / batch_size
acc_avg.update(batch_acc.item(), batch_size)
## run backward pass
loss.backward()
optimizer.step() ## update
## each epoch
logger.update(loss_avg.avg, 'train_loss')
logger.update(acc_avg.avg, 'train_acc')
print("train loss: ", loss_avg.avg)
print("train acc: ", acc_avg.avg)
if do_few_shot and fs_data_loader is not None:
for batch_idx, batch in tqdm(
enumerate(fs_data_loader),
total=len(fs_data_loader),
ncols=80,
desc=f'training epoch {epoch}',
):
data = batch[0].to(device)
gt_lbls = batch[1].to(device)
gt_gt_lbls = batch[2].to(device)
## set zerograd
optimizer.zero_grad()
## run forward pass
out = model(data) ## logits: [B, NC]; conf: [B, 1]
preds = torch.max(out, dim=-1)[1]
# print("out shape: ", out.shape)
weights = model.compute_entropy_weight(out)
# print("weights shape: ", weights.shape)
## compute loss
class_loss = criterion(out, gt_lbls) ## [B, 1]
# print("class_loss shape: ", class_loss.shape)
if use_conf:
loss = (class_loss * (weights**2) +
(1 - weights)**2).mean()
else:
loss = class_loss.mean()
## record
positive = ((gt_lbls == preds) + (gt_gt_lbls > 2)).sum()
batch_acc = positive.to(torch.float) / data.shape[0]
fs_loss_avg.update(loss.item(), data.shape[0])
fs_acc_avg.update(batch_acc.item(), data.shape[0])
## run backward pass
loss = loss * 1.0
loss.backward()
optimizer.step() ## update
# print(f"\nfew-shot: {preds}, {gt_gt_lbls}")
## each epoch
print("fs train loss: ", fs_loss_avg.avg)
print("fs train acc: ", fs_acc_avg.avg)
if val_data_loader is not None:
log = evaluate(model.eval(),
val_data_loader,
device,
use_conf=use_conf)
model.train()
logger.update(log['loss'], 'val_loss')
logger.update(log['acc'], 'val_acc')
print("val loss: ", log['loss'])
print("val acc: ", log['acc'])
best_idx = logger.get_best('val_acc', best='max')
if best_idx == epoch:
print('save ckpt')
## save ckpt
_save_checkpoint(model_path, epoch, model)
lr_scheduler.step()
print()
## save final model
_save_checkpoint(model_path, epoch, model)
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):
x.to(device)
y.to(device)
# set_trace()
optimizer.zero_grad()
y_pred = model(x)
loss = criterion(y_pred, y)
epoch_loss += loss
# plt.figure(1)
# plt.plot(i, loss.detach().numpy(), '*')
# plt.pause(1)
loss.backward()
optimizer.step()
epoch_loss = epoch_loss / len(train_loader)