def run_test(model, cfg):
print("----run test---", cfg['model'], cfg['step'])
model.eval()
with torch.no_grad():
test_loader = DataLoader(cfg['test'], batch_size=cfg['batch'],
shuffle=False, num_workers=cfg['nworker'],
collate_fn=cfg['collate'])
np_label = []
np_pd = []
np_score = []
np_sgene = []
for i_batch, sample_batched in enumerate(test_loader):
sgene, img, label = sample_batched
inputs = torch.from_numpy(img).type(torch.cuda.FloatTensor)
predict = model(inputs)
test_pd = util.threshold_tensor_batch(predict)
np_pd.append(test_pd.data.cpu().numpy())
np_sgene.extend(sgene)
np_label.append(label)
np_score.append(predict.data.cpu().numpy())
np_label = np.concatenate(np_label)
np_pd = np.concatenate(np_pd)
np_score = np.concatenate(np_score)
util.torch_metrics(np_label, np_pd, cfg['writer'],
cfg['step'], mode='test', score=np_score)
np_target = [' '.join([str(x) for x in np.where(item)[0]])
for item in np_pd]
df = pd.DataFrame({'Gene': np_sgene, 'Predicted': np_target})
result = os.path.join(cfg['model_dir'], "%s_%d.csv"
% (cfg['model'], cfg['step']))
df.to_csv(result, header=True, sep=',', index=False)
def run_val(model, cfg):
print("----run val---", cfg['model'], cfg['step'])
model.eval()
with torch.no_grad():
val_loader = DataLoader(cfg['val'],
batch_size=cfg['batch'],
shuffle=False,
num_workers=cfg['nworker'],
collate_fn=cfg['collate'])
criterion = cfg['criterion']
np_label = []
np_pd = []
np_score = []
tot_loss = 0.0
for i_batch, sample_batched in enumerate(val_loader):
sgene, predict, gt = cfg['instance'](model, sample_batched)
loss = criterion(predict, gt)
tot_loss += loss
val_pd = util.threshold_tensor_batch(predict)
np_pd.append(val_pd.data.cpu().numpy())
np_score.append(predict.data.cpu().numpy())
np_label.append(gt.data.cpu().numpy())
np_label = np.concatenate(np_label)
np_pd = np.concatenate(np_pd)
np_score = np.concatenate(np_score)
tot_loss = tot_loss / len(val_loader)
cfg['writer'].add_scalar("val loss", tot_loss.item(), cfg['step'])
lab_f1_macro = util.torch_metrics(np_label,
np_pd,
cfg['writer'],
cfg['step'],
score=np_score)
return tot_loss.item(), lab_f1_macro
def run_train(model, cfg):
train_loader = DataLoader(cfg['train'],
batch_size=cfg['batch'],
shuffle=True,
num_workers=cfg['nworker'],
collate_fn=cfg['collate'])
model_pth = os.path.join(cfg['model_dir'], "model.pth")
writer = tensorboardX.SummaryWriter(cfg['model_dir'])
cfg['writer'] = writer
criterion = cfg['criterion']
optimizer = torch.optim.Adam(model.parameters(),
lr=cfg['lr'],
weight_decay=cfg['decay'])
if cfg['scheduler']:
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer, 'max', factor=cfg['factor'], patience=cfg['patience'])
step = cfg['step'] * len(train_loader)
min_loss = 1e8
max_f1 = 0.0
for e in range(cfg['step'], cfg['epochs']):
print("----run train---", cfg['model'], e)
model.train()
st = time.time()
cfg['step'] = e
np_label = []
np_pd = []
np_score = []
for i_batch, sample_batched in tqdm(enumerate(train_loader),
total=len(train_loader)):
sgene, predict, gt = cfg['instance'](model, sample_batched)
loss = criterion(predict, gt)
loss.backward()
optimizer.step()
writer.add_scalar("loss", loss, step)
step += 1
val_pd = util.threshold_tensor_batch(predict)
np_pd.append(val_pd.data.cpu().numpy())
np_score.append(predict.data.cpu().numpy())
np_label.append(gt.data.cpu().numpy())
np_label = np.concatenate(np_label)
np_pd = np.concatenate(np_pd)
np_score = np.concatenate(np_score)
et = time.time()
writer.add_scalar("train time", et - st, e)
util.torch_metrics(np_label,
np_pd,
cfg['writer'],
cfg['step'],
mode='train',
score=np_score)
val_loss, lab_f1_macro = run_val(model, cfg)
print("val loss:", val_loss, "\tf1:", lab_f1_macro)
for g in optimizer.param_groups:
writer.add_scalar("lr", g['lr'], e)
if cfg['scheduler'] and g['lr'] > 1e-6:
scheduler.step(lab_f1_macro)
break
# if val_loss > 2 * min_loss:
# print("early stopping at %d" % e)
# break
# run_test(model, cfg)
if min_loss > val_loss or lab_f1_macro > max_f1:
if min_loss > val_loss:
min_loss = val_loss
print("----save best epoch:%d, loss:%f---" % (e, val_loss))
if lab_f1_macro > max_f1:
max_f1 = lab_f1_macro
print("----save best epoch:%d, f1:%f---" % (e, max_f1))
# torch.save(model.state_dict(), model_pth)
torch.save({'epoch': e, 'model': model.state_dict()}, model_pth)
run_test(model, cfg)
def run_train(model, cfg):
model_pth = os.path.join(cfg['model_dir'], "model.pth")
writer = tensorboardX.SummaryWriter(cfg['model_dir'])
cfg['writer'] = writer
criterion = cfg['criterion']
optimizer = torch.optim.Adam(model.parameters(),
lr=cfg['lr'],
weight_decay=cfg['decay'])
if cfg['scheduler']:
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer, 'max', factor=cfg['factor'], patience=cfg['patience'])
nsample = len(cfg['train'])
sample_weights = np.ones(nsample, dtype='float')
# origin bce loss
sample_loss = np.zeros(nsample, dtype='float')
step = cfg['step'] * len(cfg['train'])
min_loss = 1e8
max_f1 = 0.0
for e in range(cfg['step'], cfg['epochs']):
print("----run train---", cfg['model'], e)
model.train()
st = time.time()
cfg['step'] = e
np_label = []
np_pd = []
np_score = []
shuffle_idx = np.array(range(len(cfg['train'])))
np.random.shuffle(shuffle_idx)
train_dataset = Subset(cfg['train'], shuffle_idx)
train_loader = DataLoader(train_dataset,
batch_size=cfg['batch'],
shuffle=False,
num_workers=cfg['nworker'],
collate_fn=cfg['collate'])
for i_batch, sample_batched in tqdm(enumerate(train_loader),
total=len(train_loader)):
sgene, predict, gt = cfg['instance'](model, sample_batched)
st_idx = i_batch * cfg['batch']
end_idx = st_idx + len(sgene)
batch_idx = shuffle_idx[st_idx:end_idx]
# orig_loss = F.binary_cross_entropy(predict, gt, reduction='none')
orig_loss = criterion(predict, gt)
orig_loss = torch.sum(orig_loss, dim=1) / predict.shape[-1]
weight_loss = torch.from_numpy(sample_weights[batch_idx]).type(
torch.cuda.FloatTensor) * orig_loss
loss = torch.mean(weight_loss)
writer.add_scalar("loss_weight", loss, step)
writer.add_scalar("loss_orig", torch.mean(orig_loss), step)
# loss = criterion(predict, gt)
loss.backward()
optimizer.step()
# writer.add_scalar("loss", loss, step)
step += 1
val_pd = util.threshold_tensor_batch(predict)
np_pd.append(val_pd.data.cpu().numpy())
np_score.append(predict.data.cpu().numpy())
np_label.append(gt.data.cpu().numpy())
sample_loss[batch_idx] = orig_loss.clone().cpu().data.numpy()
exp_loss = np.exp(sample_loss)
exp_tot = np.sum(exp_loss)
sample_weights = nsample * (exp_loss / exp_tot)
sample_loss = np.zeros(nsample)
np_label = np.concatenate(np_label)
np_pd = np.concatenate(np_pd)
np_score = np.concatenate(np_score)
et = time.time()
writer.add_scalar("train time", et - st, e)
util.torch_metrics(np_label,
np_pd,
cfg['writer'],
cfg['step'],
mode='train',
score=np_score)
val_loss, lab_f1_macro = run_val(model, cfg)
print("val loss:", val_loss, "\tf1:", lab_f1_macro)
for g in optimizer.param_groups:
writer.add_scalar("lr", g['lr'], e)
if cfg['scheduler'] and g['lr'] > 1e-5:
scheduler.step(lab_f1_macro)
break
# if val_loss > 2 * min_loss:
# print("early stopping at %d" % e)
# break
# run_test(model, cfg)
if min_loss > val_loss or lab_f1_macro > max_f1:
if min_loss > val_loss:
min_loss = val_loss
print("----save best epoch:%d, loss:%f---" % (e, val_loss))
if lab_f1_macro > max_f1:
max_f1 = lab_f1_macro
print("----save best epoch:%d, f1:%f---" % (e, max_f1))
# torch.save(model.state_dict(), model_pth)
torch.save({'epoch': e, 'model': model.state_dict()}, model_pth)
run_test(model, cfg)