def Train_GACNN(params_dict, chrom_i, args, cfg, log, epoches=20):
# args = parser()
# cfg = Config.fromfile(args.config)
# log = Logger(cfg.PARA.utils_paths.log_path + 'GA_MyCNN' + '_log.txt', level='info')
log.logger.info('==> Preparing dataset <==')
cifar10 = Cifar10(batch_size = cfg.PARA.train.batch_size)
# subtrain_loader, subvalid_loader = cifar10.Download_SubTrain_SubValid()
# print(len(subtrain_loader), len(subvalid_loader))
train_loader, valid_loader = cifar10.Download_Train_Valid()
test_loader = cifar10.Download_Test()
log.logger.info('==> Loading model <==')
# net = vgg16()
net = MyModel(params_dict, chrom_i)
if torch.cuda.device_count() > 1:
net = nn.DataParallel(net)
net.to(device)
print(net)
criterion = nn.CrossEntropyLoss().to(device)
optimizer = optim.SGD(net.parameters(), lr=cfg.PARA.train.lr) # , momentum=cfg.PARA.train.momentum
log.logger.info('==> Waiting Train <==')
loss = train_valid(net=net, dict=params_dict, criterion=criterion, optimizer=optimizer, train_loader=train_loader,
valid_loader=valid_loader, args=args, log=log, cfg=cfg, epoches=epoches)
# log.logger.info("==> Waiting Test <==")
# # with open(cfg.PARA.utils_paths.checkpoint_path + 'GACNN/' + 'best_net_params.pkl', 'rb') as f:
# # best_net_params = pkl.load(f)
# checkpoint = torch.load(cfg.PARA.utils_paths.checkpoint_path + 'GACNN/' + 'best_ckpt.pth')
# # 进行测试时，net 和保存的 net 是不一样的，所以需要重新设置 net 可是之前就可以，说明没问题啊。。。奇怪
# net2 = MyCNN(best_net_params)
# net2.to(device)
# net2.load_state_dict(checkpoint['net'])
# test_acc = test(net=net2, test_loader=test_loader)
# net.load_state_dict(checkpoint['net'])
# test_acc = test(net=net, test_loader=test_loader)
# log.logger.info('Test ACC = {:.5f}'.format(test_acc))
# log.logger.info('==> One Train & Valid & Test End <==')
return loss
def train(dataset_path, lr, epoch, batch_size, scaler_flag, state_name):
train_loss_curve = []
best = -1
# load model
device = torch.device('cuda:1' if torch.cuda.is_available() else 'cpu')
model = MyModel()
model = model.to(device)
model.train()
# dataset and dataloader
full_dataset = Visitor_Dataset(dataset_path, scaler_flag)
train_dataloader = DataLoader(dataset=full_dataset,
batch_size=batch_size,
shuffle=True)
# loss function and optimizer
criterion = RMSLELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
# start training
for e in range(epoch):
train_loss, train_rmsle = 0.0, 0.0
print(f'\nEpoch: {e+1}/{epoch}')
print('-' * len(f'Epoch: {e+1}/{epoch}'))
# tqdm to disply progress bar
for inputs, labels in tqdm(train_dataloader):
# data from data_loader
inputs = inputs.float().to(device)
labels = labels.float().to(device)
outputs = model(inputs)
# RMSLE Loss
loss = criterion(outputs, labels)
# weights update
optimizer.zero_grad()
loss.backward()
optimizer.step()
# loss and rmsle calculate
train_loss += loss.item()
# save the best model weights as .pth file
train_loss_epoch = sqrt(train_loss / len(full_dataset))
if best == -1 or train_loss_epoch < best:
best_loss = train_loss_epoch
best_epoch = e
torch.save(model.state_dict(), f'mymodel_{state_name}.pth')
print(f'Training Loss: {train_loss_epoch:.6f}')
# save loss and RMSLE every epoch
train_loss_curve.append(train_loss_epoch)
# print the best RMSLE
print(f"Final Training RMSLE Loss = {best_loss:.6f}")
visualize(value=train_loss_curve,
title='Train Loss Curve',
filename=f'rmsle_{state_name}.png')
return full_dataset
def train(lr=0.001, epoch=600, batch_size=32):
train_loss_curve = []
train_wrmse_curve = []
valid_loss_curve = []
valid_wrmse_curve = []
# load model
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = MyModel()
model = model.to(device)
model.train()
# dataset and dataloader
# can use torch random_split to create the validation dataset
dataset = MLDataset()
train_size = int(0.9 * len(dataset))
valid_size = len(dataset) - train_size
train_dataset, valid_dataset = random_split(dataset,
[train_size, valid_size])
train_dataloader = DataLoader(dataset=train_dataset,
batch_size=batch_size,
shuffle=True)
valid_dataloader = DataLoader(dataset=valid_dataset,
batch_size=batch_size,
shuffle=True)
# loss function and optimizer
# can change loss function and optimizer you want
criterion = nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
best = 100
# start training
for e in range(epoch):
train_loss = 0.0
train_wrmse = 0.0
valid_loss = 0.0
valid_wrmse = 0.0
print(f'\nEpoch: {e+1}/{epoch}')
print('-' * len(f'Epoch: {e+1}/{epoch}'))
# tqdm to disply progress bar
for inputs, labels in tqdm(train_dataloader):
# data from data_loader
inputs = inputs.float().to(device)
labels = labels.float().to(device)
outputs = model(inputs)
# MSE loss and WRMSE
loss = criterion(outputs, labels)
wrmse = WRMSE(outputs, labels, device)
# weights update
optimizer.zero_grad()
loss.backward()
optimizer.step()
# loss calculate
train_loss += loss.item()
train_wrmse += wrmse
# =================================================================== #
# If you have created the validation dataset,
# you can refer to the for loop above and calculate the validation loss
for inputs, labels in tqdm(valid_dataloader):
# data from data_loader
inputs = inputs.float().to(device)
labels = labels.float().to(device)
outputs = model(inputs)
# MSE loss and WRMSE
loss = criterion(outputs, labels)
wrmse = WRMSE(outputs, labels, device)
# loss calculate
valid_loss += loss.item()
valid_wrmse += wrmse
# =================================================================== #
# save the best model weights as .pth file
loss_epoch = train_loss / len(train_dataset)
wrmse_epoch = math.sqrt(train_wrmse / len(train_dataset))
valid_loss_epoch = valid_loss / len(valid_dataset)
valid_wrmse_epoch = math.sqrt(valid_wrmse / len(valid_dataset))
if valid_wrmse_epoch < best:
best = valid_wrmse_epoch
torch.save(model.state_dict(), 'mymodel.pth')
print(f'Training loss: {loss_epoch:.4f}')
print(f'Training WRMSE: {wrmse_epoch:.4f}')
print(f'Valid loss: {valid_loss_epoch:.4f}')
print(f'Valid WRMSE: {valid_wrmse_epoch:.4f}')
# save loss and wrmse every epoch
train_loss_curve.append(loss_epoch)
train_wrmse_curve.append(wrmse_epoch)
valid_loss_curve.append(valid_loss_epoch)
valid_wrmse_curve.append(valid_wrmse_epoch)
# generate training curve
visualize(train_loss_curve, valid_loss_curve, 'Train Loss')
visualize(train_wrmse_curve, valid_wrmse_curve, 'Train WRMSE')
print("\nBest Validation WRMSE:", best)
def train(lr=0.001, epoch=200, batch_size=64):
train_loss_curve = []
train_wrmse_curve = []
valid_loss_curve = []
valid_wrmse_curve = []
best = 100
# load model
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = MyModel()
model = model.to(device)
model.train()
# dataset and dataloader
# load data
full_dataset = pd.read_csv('train.csv', encoding='utf-8')
# can use torch random_split to create the validation dataset
lengths = [
int(round(len(full_dataset) * 0.8)),
int(round(len(full_dataset) * 0.2))
]
train_set, valid_set = random_split(full_dataset, lengths)
train_dataset = MLDataset(full_dataset.iloc[train_set.indices])
valid_dataset = MLDataset(full_dataset.iloc[valid_set.indices])
train_dataloader = DataLoader(dataset=train_dataset,
batch_size=batch_size,
shuffle=True)
valid_dataloader = DataLoader(dataset=valid_dataset,
batch_size=batch_size,
shuffle=False)
# loss function and optimizer
# can change loss function and optimizer you want
criterion = nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
# start training
for e in tqdm(range(epoch)):
train_loss, valid_loss = 0.0, 0.0
train_wrmse, valid_wrmse = 0.0, 0.0
print(f'\nEpoch: {e+1}/{epoch}')
print('-' * len(f'Epoch: {e+1}/{epoch}'))
# tqdm to disply progress bar
for inputs, labels in train_dataloader:
# data from data_loader
inputs = inputs.float().to(device)
labels = labels.float().to(device)
outputs = model(inputs)
# MSE loss and WRMSE
loss = criterion(outputs, labels)
wrmse = WRMSE(outputs, labels, device)
# weights update
optimizer.zero_grad()
loss.backward()
optimizer.step()
# loss calculate
train_loss += loss.item()
train_wrmse += wrmse
# =================================================================== #
# If you have created the validation dataset,
# you can refer to the for loop above and calculate the validation loss
# tqdm to disply progress bar
for inputs, labels in valid_dataloader:
# data from data_loader
inputs = inputs.float().to(device)
labels = labels.float().to(device)
outputs = model(inputs)
# MSE loss and WRMSE
loss = criterion(outputs, labels)
wrmse = WRMSE(outputs, labels, device)
# loss calculate
valid_loss += loss.item()
valid_wrmse += wrmse
# =================================================================== #
# save the best model weights as .pth file
train_loss_epoch = train_loss / len(train_dataset)
train_wrmse_epoch = math.sqrt(train_wrmse / len(train_dataset))
valid_loss_epoch = valid_loss / len(valid_dataset)
valid_wrmse_epoch = math.sqrt(valid_wrmse / len(valid_dataset))
if train_wrmse_epoch < best:
best_wrmse = train_wrmse_epoch
best_loss = train_loss_epoch
best_epoch = e
torch.save(model.state_dict(), 'mymodel.pth')
print(f'Training loss: {train_loss_epoch:.6f}')
print(f'Training WRMSE: {train_wrmse_epoch:.6f}')
print(f'Valid loss: {valid_loss_epoch:.6f}')
print(f'Valid WRMSE: {valid_wrmse_epoch:.6f}')
# save loss and wrmse every epoch
train_loss_curve.append(train_loss_epoch)
train_wrmse_curve.append(train_wrmse_epoch)
valid_loss_curve.append(valid_loss_epoch)
valid_wrmse_curve.append(valid_wrmse_epoch)
# print the best wrmse
print(f"\nBest Epoch = {best_epoch}")
print(f"Best Loss = {best_loss:.4f}")
print(f"Best WRMSE = {best_wrmse:.4f}\n")
# generate training curve
visualize(train=train_loss_curve,
valid=valid_loss_curve,
title='Loss Curve',
filename='loss.png',
best=(e, best_loss))
visualize(train_wrmse_curve,
valid_wrmse_curve,
title='WRMSE Curve',
filename='wrmse.png',
best=(e, best_wrmse),
wrmse=True)
def main():
_i, _j, _k = 2,3,3
dataset = MyDataset(_i,_j,_k)
dtype = torch.float
device = torch.device("cpu")
# device = torch.device("cuda:0")
#batch, input, hidden, output
N, D_in, H, D_out = 10, _i+_j+_k, 16, _i*_j*_k
msg_len = 10
x, y = dataset.get_frame()
x = torch.tensor(x, dtype=dtype, device=device)
#x = torch.cat((x,x,x,x,x),0)
y = torch.tensor(y, dtype=torch.long, device=device).squeeze()
#y = torch.cat((y,y,y,y,y),0)
print(x.size(), y.size())
#x = torch.zeros(N, D_in, device=device, dtype=dtype)
#y = torch.zeros(N, device=device, dtype=dtype)
model = MyModel(D_in, H, D_out)
#model = torch.nn.Linear(D_in, D_out)
loss_fn = torch.nn.CrossEntropyLoss(reduce=None)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-2)
for t in range(10001):
if True: #reinforce
y_pred = model(x)
probs = F.softmax(y_pred, dim=1)
m = Categorical(probs)
action = m.sample()
reward = torch.eq(action, y).to(torch.float)
reward = (reward - reward.mean())
loss = -m.log_prob(action) * reward
model.zero_grad()
loss.sum().backward()
#loss.backward(loss)
optimizer.step()
elif True:
y_pred = model(x)
else: # supervised
y_pred = model(x)
loss = loss_fn(y_pred, y)
model.zero_grad()
loss.backward()
optimizer.step()
if t % 100 == 0:
with torch.no_grad():
y_pred = model(x)
eq = torch.eq(torch.argmax(y_pred, dim=1), y)
print("t: {}, acc: {}/{} = {}".format(t, torch.sum(eq).item(), eq.numel(), torch.sum(eq).item() / eq.numel()))
torch.save({'epoch': t,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'loss': loss
}, "checkpoints.tar")
def train(
batch_size=16,
pretrain_model_path='',
name='',
model_type='mlp',
after_bert_choice='last_cls',
dim=1024,
lr=1e-5,
epoch=12,
smoothing=0.05,
sample=False,
#open_ad='',
dialog_name='xxx'):
if not pretrain_model_path or not name:
assert 1 == -1
print('\n********** model type:', model_type, '**********')
print('batch_size:', batch_size)
# load dataset
train_file = '/kaggle/input/dataset/my_train.csv'
dev_file = '/kaggle/input/dataset/my_dev.csv'
train_num = len(pd.read_csv(train_file).values.tolist())
val_num = len(pd.read_csv(dev_file).values.tolist())
print('train_num: %d, dev_num: %d' % (train_num, val_num))
# 选择模型
if model_type in ['siam', 'esim', 'sbert']:
assert 1 == -1
else:
train_iter = MyDataset(file=train_file,
is_train=True,
sample=sample,
pretrain_model_path=pretrain_model_path)
train_iter = get_dataloader(train_iter,
batch_size,
shuffle=True,
drop_last=True)
dev_iter = MyDataset(file=dev_file,
is_train=True,
sample=sample,
pretrain_model_path=pretrain_model_path)
dev_iter = get_dataloader(dev_iter,
batch_size,
shuffle=False,
drop_last=False)
if model_type == 'mlp':
model = MyModel(dim=dim,
pretrain_model_path=pretrain_model_path,
smoothing=smoothing,
after_bert_choice='last_cls')
elif model_type == 'cnn':
model = MyTextCNNModel(dim=dim,
pretrain_model_path=pretrain_model_path,
smoothing=smoothing)
elif model_type == 'rcnn':
model = MyRCNNModel(dim=dim,
pretrain_model_path=pretrain_model_path,
smoothing=smoothing)
#模型加载到gpu
model.to(device)
model_param_num = 0
##### 3.24 muppti-gpu-training
if n_gpu > 1:
model = torch.nn.DataParallel(model)
for p in model.parameters():
if p.requires_grad:
model_param_num += p.nelement()
print('param_num:%d\n' % model_param_num)
# 加入对抗训练，提升泛化能力；但是训练速度明显变慢 (插件式调用)
# 3.12 change to FGM 更快！
"""
if open_ad == 'fgm':
fgm = FGM(model)
elif open_ad == 'pgd':
pgd = PGD(model)
K = 3
"""
# model-store-path
#model_path = '/kaggle/output/' + name + '.pkl' # 输出模型默认存放在当前路径下
output_dir = 'output'
state = {}
time0 = time.time()
best_loss = 999
early_stop = 0
for e in range(epoch):
print("*" * 100)
print("Epoch:", e)
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_grouped_parameters = [{
'params': [
p for n, p in param_optimizer
if not any(nd in n for nd in no_decay)
],
'weight_decay':
0.01
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.0
}]
optimizer = BertAdam(optimizer_grouped_parameters,
lr=lr,
warmup=0.05,
t_total=len(train_iter)) # 设置优化器
train_loss = 0
train_c = 0
train_right_num = 0
model.train() # 将模型设置成训练模式（Sets the module in training mode）
print('training..., %s, e:%d, lr:%7f' % (name, e, lr))
for batch in tqdm(train_iter): # 每一次返回 batch_size 条数据
optimizer.zero_grad() # 清空梯度
batch = [b.to(device) for b in batch] # cpu -> GPU
# 正常训练
labels = batch[-1].view(-1).cpu().numpy()
loss, bert_enc = model(batch, task='train',
epoch=epoch) # 进行前向传播，真正开始训练；计算 loss
right_num = count_right_num(bert_enc, labels)
# multi-gpu training!
if n_gpu > 1:
loss = loss.mean()
loss.backward() # 反向传播计算参数的梯度
#"""
if open_ad == 'fgm':
# 对抗训练
fgm.attack() # 在embedding上添加对抗扰动
if model_type == 'multi-task':
loss_adv, _, _ = model(batch, task='train')
else:
loss_adv, _ = model(batch, task='train')
if n_gpu > 1:
loss_adv = loss_adv.mean()
loss_adv.backward() # 反向传播，并在正常的grad基础上，累加对抗训练的梯度
fgm.restore() # 恢复embedding参数
elif open_ad == 'pgd':
pgd.backup_grad()
# 对抗训练
for t in range(K):
pgd.attack(is_first_attack=(
t == 0
)) # 在embedding上添加对抗扰动, first attack时备份param.data
if t != K - 1:
optimizer.zero_grad()
else:
pgd.restore_grad()
if model_type == 'multi-task':
loss_adv, _, _ = model(batch, task='train')
else:
loss_adv, _ = model(batch, task='train')
if n_gpu > 1:
loss_adv = loss_adv.mean()
loss_adv.backward() # 反向传播，并在正常的grad基础上，累加对抗训练的梯度
pgd.restore() # 恢复embedding参数
#"""
optimizer.step() # 更新参数
train_loss += loss.item() # loss 求和
train_c += 1
train_right_num += right_num
val_loss = 0
val_c = 0
val_right_num = 0
model.eval()
print('eval...')
with torch.no_grad(): # 不进行梯度的反向传播
for batch in tqdm(dev_iter): # 每一次返回 batch_size 条数据
batch = [b.to(device) for b in batch]
labels = batch[-1].view(-1).cpu().numpy()
loss, bert_enc = model(batch, task='train',
epoch=epoch) # 进行前向传播，真正开始训练；计算 loss
right_num = count_right_num(bert_enc, labels)
if n_gpu > 1:
loss = loss.mean()
val_c += 1
val_loss += loss.item()
val_right_num += right_num
train_acc = train_right_num / train_num
val_acc = val_right_num / val_num
print('train_acc: %.4f, val_acc: %.4f' % (train_acc, val_acc))
print('train_loss: %.4f, val_loss: %.4f, time: %d' %
(train_loss / train_c, val_loss / val_c, time.time() - time0))
if val_loss / val_c < best_loss:
early_stop = 0
best_loss = val_loss / val_c
best_acc = val_acc
# 3.24 update 多卡训练时模型保存避坑:
if not os.path.exists(output_dir):
os.makedirs(output_dir)
model_to_save = model.module if hasattr(model, 'module') else model
state['model_state'] = model_to_save.state_dict()
state['loss'] = val_loss / val_c
state['acc'] = val_acc
state['e'] = e
state['time'] = time.time() - time0
state['lr'] = lr
output_model_file = os.path.join(output_dir, name + '.pkl')
torch.save(state, output_model_file)
#torch.save(state, model_path)
best_epoch = e
cost_time = time.time() - time0
tmp_train_acc = train_acc
best_model = model
else:
early_stop += 1
if early_stop == 2:
break
model = best_model
lr = lr * 0.5
print("best_loss:", best_loss)
# 3.12 add 打印显示最终的最优结果
print('-' * 30)
print('best_epoch:', best_epoch, 'best_loss:', best_loss, 'best_acc:',
best_acc, 'reach time:', cost_time, '\n')
# model-clean
del model
gc.collect()
# 实验结果写入日志
"""