v0.1.0.post2

Author: Tongjilibo

.gitignore

@@ -9,4 +9,5 @@ backup
 test
 .DS_Store
 *.pt
-*.log
+*.log
+ckpt

README.md

@@ -65,7 +65,8 @@ pip install git+https://github.com/Tongjilibo/torch4keras.git
 - 简单示例: [turorials_mnist](https://github.com/Tongjilibo/torch4keras/blob/master/examples/turorials_mnist.py)
 
 ## 4. 版本说明
-- **v0.1.0**: 允许调整进度条的显示参数, 进度条和日志同步（如果进度条平滑了则日志也平滑）, 自动把tensor转到model.device上, 允许打印第一个batch来检查样本
+- **v0.1.0.post2**: 20230725 修复v0.1.0的bug，主要是进度条和log的标签平滑的问题
+- **v0.1.0**: 20230724 允许调整进度条的显示参数, 进度条和日志同步（如果进度条平滑了则日志也平滑）, 自动把tensor转到model.device上, 允许打印第一个batch来检查样本
 - **v0.0.9**：20230716 增加auto_set_cuda_devices自动选择显卡，增加log_info，log_warn, log_error等小函数
 - **v0.0.8**：20230625 增加EmailCallback和WandbCallback, 增加AccelerateTrainer和DeepSpeedTrainer, grad_accumulation_steps内算一个batch，修改Trainer中部分成员函数
 - **v0.0.7.post3**: 20230517 修复保存scheduler
@@ -80,7 +81,8 @@ pip install git+https://github.com/Tongjilibo/torch4keras.git
 - **v0.0.1**：20221019 初始版本
 
 ## 5. 更新：
-- **20230721**: 允许调整进度条的显示参数, 进度条和日志同步（如果进度条平滑了则日志也平滑）, 自动把tensor转到model.device上, 允许打印第一个batch来检查样本
+- **20230725**: 修复v0.1.0的bug，主要是进度条和log的标签平滑的问题
+- **20230724**: 允许调整进度条的显示参数, 进度条和日志同步（如果进度条平滑了则日志也平滑）, 自动把tensor转到model.device上, 允许打印第一个batch来检查样本
 - **20230716**：增加auto_set_cuda_devices自动选择显卡，增加log_info，log_warn, log_error等小函数
 - **20230625**：增加EmailCallback和WandbCallback, 增加AccelerateTrainer和DeepSpeedTrainer, grad_accumulation_steps内算一个batch，修改Trainer中部分成员函数
 - **20230517**：Checkpoint Calback增加保存scheduler, save_weights可自行创建目录，Logger, Tensorboard模块加入lr, 修改predict和add_trainer

examples/tutorials_mnist.py

@@ -4,7 +4,7 @@
 import torchvision
 from torch4keras.model import BaseModel, Trainer
 from torch4keras.snippets import seed_everything
-from torch4keras.callbacks import Checkpoint, Evaluator, EarlyStopping, Summary, Logger, EmailCallback, WandbCallback
+from torch4keras.callbacks import Checkpoint, Evaluator, EarlyStopping, Summary, Logger, EmailCallback, WandbCallback, Tensorboard
 from transformers.optimization import get_linear_schedule_with_warmup
 from torch.utils.data import TensorDataset, DataLoader
 from tqdm import tqdm
@@ -54,7 +54,7 @@
 model = Trainer(net.to(device))
 optimizer = optim.Adam(net.parameters())
 scheduler = get_linear_schedule_with_warmup(optimizer, steps_per_epoch, steps_per_epoch*epochs)
-model.compile(optimizer=optimizer, scheduler=scheduler, loss=nn.CrossEntropyLoss(), metrics=['acc'], bar='tqdm')
+model.compile(optimizer=optimizer, scheduler=scheduler, loss=nn.CrossEntropyLoss(), metrics=['acc'])
 
 class MyEvaluator(Evaluator):
     # 重构评价函数
@@ -78,7 +78,9 @@ def evaluate(self):
                       scheduler_path='./ckpt/{epoch}/scheduler_{epoch}_{test_acc:.5f}.pt',
                       steps_params_path='./ckpt/{epoch}/steps_params_{epoch}_{test_acc:.5f}.pt')
     early_stop = EarlyStopping(monitor='test_acc', verbose=1)
-    logger = Logger('./ckpt/log.log', interval=100)
-    email = EmailCallback(receivers='[email protected]')
-    wandb = WandbCallback(save_code=True)
-    hist = model.fit(train_dataloader, steps_per_epoch=steps_per_epoch, epochs=epochs, callbacks=[Summary(), evaluator, logger, ckpt, early_stop])
+    logger = Logger('./ckpt/log.log', interval=100)  # log文件
+    ts_board = Tensorboard('./ckpt/tensorboard', method='step', interval=100)  # tensorboard
+    email = EmailCallback(receivers='[email protected]')  # 发送邮件
+    wandb = WandbCallback(save_code=True)  # wandb
+    hist = model.fit(train_dataloader, steps_per_epoch=steps_per_epoch, epochs=epochs, 
+                     callbacks=[Summary(), evaluator, logger, ts_board, ckpt, early_stop])

setup.py

@@ -7,7 +7,7 @@
 
 setup(
     name='torch4keras',
-    version='v0.1.0',
+    version='v0.1.0.post2',
     description='Use torch like keras',
     long_description=long_description,
     long_description_content_type="text/markdown",

torch4keras/callbacks.py

@@ -62,7 +62,7 @@ def update(self, current, values=None):
             if k not in self.stateful_metrics:
                 if k not in self._values:
                     self._values[k] = [v * (current - self._seen_so_far), current - self._seen_so_far]
-                elif (self.smooth_interval is not None) and (current % self.smooth_interval == 0):
+                elif (self.smooth_interval is not None) and (current % self.smooth_interval == 1):
                     # 如果定义了累积smooth_interval，则需要重新累计
                     self._values[k] = [v, 1]
                 else:
@@ -435,7 +435,7 @@ def smooth_values(self, current, values=None):
             if k not in self.stateful_metrics:
                 if k not in self._values:
                     self._values[k] = [v * (current - self._seen_so_far), current - self._seen_so_far]
-                elif (self.smooth_interval is not None) and (current % self.smooth_interval == 0):
+                elif (self.smooth_interval is not None) and (current % self.smooth_interval == 1):
                     # 如果定义了累积smooth_interval，则需要重新累计
                     self._values[k] = [v, 1]
                 else:
@@ -887,15 +887,13 @@ class Tensorboard(Callback):
     :param method: str, 控制是按照epoch还是step来计算，默认为'epoch', 可选{'step', 'epoch'}
     :param interval: int, 保存tensorboard的间隔
     :param prefix: str, tensorboard分栏的前缀，默认为'train'
-    :param on_epoch_end_scalar_epoch: bool, epoch结束后是横轴是按照epoch还是global_step来记录
     '''
-    def __init__(self, log_dir, method='epoch', interval=10, prefix='train', on_epoch_end_scalar_epoch=True, **kwargs):
+    def __init__(self, log_dir, method='epoch', interval=10, prefix='train', **kwargs):
         super(Tensorboard, self).__init__(**kwargs)
         assert method in {'step', 'epoch'}, 'Args `method` only support `step` or `epoch`'
         self.method = method
         self.interval = interval
         self.prefix = prefix+'/' if len(prefix.strip()) > 0 else ''  # 控制默认的前缀，用于区分栏目
-        self.on_epoch_end_scalar_epoch = on_epoch_end_scalar_epoch  # 控制on_epoch_end记录的是epoch还是global_step
 
         from tensorboardX import SummaryWriter
         os.makedirs(log_dir, exist_ok=True)
@@ -904,8 +902,7 @@ def __init__(self, log_dir, method='epoch', interval=10, prefix='train', on_epoc
     def on_epoch_end(self, global_step, epoch, logs=None):
         if self.method == 'epoch':
             # 默认记录的是epoch
-            log_step = epoch+1 if self.on_epoch_end_scalar_epoch else global_step+1
-            self.process(log_step, logs)
+            self.process(epoch+1, logs)
 
     def on_batch_end(self, global_step, local_step, logs=None):
         # 默认记录的是global_step
@@ -998,22 +995,22 @@ def __init__(self, receivers, subject='', method='epoch', interval=10, mail_host
 
     def on_epoch_end(self, global_step, epoch, logs=None):
         if self.method == 'epoch':
-            msg = json.dumps({k:f'{v:.5f}' for k,v in logs.items() if k!='size'}, indent=2, ensure_ascii=False)
+            msg = json.dumps({k:f'{v:.5f}' for k,v in logs.items() if k not in SKIP_METRICS}, indent=2, ensure_ascii=False)
             subject = f'[INFO] Epoch {epoch+1} performance'
             if self.subject != '':
                 subject = self.subject + ' | ' + subject
             self._email(subject, msg)
 
     def on_batch_end(self, global_step, local_step, logs=None):
         if (self.method == 'step') and ((global_step+1) % self.interval == 0):
-            msg = json.dumps({k:f'{v:.5f}' for k,v in logs.items() if k!='size'}, indent=2, ensure_ascii=False)
+            msg = json.dumps({k:f'{v:.5f}' for k,v in logs.items() if k not in SKIP_METRICS}, indent=2, ensure_ascii=False)
             subject = f'[INFO] Step {global_step} performance'
             if self.subject != '':
                 subject = self.subject + ' | ' + subject
             self._email(subject, msg)
 
     def on_train_end(self, logs=None):
-        msg = json.dumps({k:f'{v:.5f}' for k,v in logs.items() if k!='size'}, indent=2, ensure_ascii=False)
+        msg = json.dumps({k:f'{v:.5f}' for k,v in logs.items() if k not in SKIP_METRICS}, indent=2, ensure_ascii=False)
         subject = f'[INFO] Finish training'
         if self.subject != '':
             subject = self.subject + ' | ' + subject

torch4keras/model.py

@@ -45,7 +45,7 @@ def compile(self, loss, optimizer, scheduler=None, clip_grad_norm=None, mixed_pr
         :param metrics: str/List[str]/dict, 训练过程中需要打印的指标, loss相关指标默认会打印, 目前支持accuracy, 也支持自定义metric，形式为{key: func}
         :param grad_accumulation_steps: int, 梯度累积步数，默认为1
         :param bar: str, 使用进度条的种类，从kwargs中解析，默认为keras, 可选keras, tqdm, progressbar2
-        :param progbar_config: 进度条的配置，如果使用指标平滑会更新到后续其他callbacks中（比如Logger），实现进度条显示和日志会保持一致
+        :param progbar_config: 进度条的配置，默认是对整个epoch计算均值指标
             bar: str, 默认为keras
             stateful_metrics: List[str], 表示不使用指标平滑仅进行状态记录的metric，指标抖动会更加明显，默认为None表示使用指标平滑
             smooth_interval: int, 表示指标平滑时候的累计步数，默认为None表示对整个epoch进行平滑

torch4keras/snippets.py

@@ -246,7 +246,7 @@ def send_email(receivers, subject, msg="", mail_host=None, mail_user=None, mail_
         smtpObj.login(mail_user, mail_pwd)  # 登录到服务器
         smtpObj.sendmail(sender, receivers, message.as_string())  # 发送
         smtpObj.quit()  # 退出
-        print('[INFO] Send email success')
+        log_info('Send email success')
     except smtplib.SMTPException as e:
         log_error('Send email error : '+str(e))
         return str(e)
@@ -357,7 +357,7 @@ def print_trainable_parameters(module):
         all_param += num_params
         if param.requires_grad:
             trainable_params += num_params
-    print(f"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param}")
+    log_info(f"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param}")
 
 
 def get_parameter_device(parameter):