译（四十九）-Pytorch计算模型参数量

stackoverflow热门问题目录

如有翻译问题欢迎评论指出，谢谢。

计算Pytorch模型参数量

• Fábio Perez asked:

  • 怎么计算 Pytorch 模型的参数量？类似 Keras 的 model.count_params() 那样的函数。

• Answers:

  • Fábio Perez - vote: 198

  • Pytorch 没有类似 Keras 计算参数量的函数，但可以通过每个参数组的求和得出参数量：

  • pytorch_total_params = sum(p.numel() for p in model.parameters())

  • 如果希望只计算可训练的参数：

  • pytorch_total_params = sum(p.numel() for p in model.parameters() if p.requires_grad)

  • 参考 Pytorch 论坛回答。

  • 注：这个回答是自问自答，如果你们有更好的解决方法，请分享一下。

  • Fábio Perez - vote: 46

  • 为了像 Keras 一样计算每层的参数量，Pytorch 的 model.named_paramters() 能返回一个迭代器，包含参数名与参数本身。

  • 举例说明：

  • from prettytable import PrettyTable
    #
    def count_parameters(model):
      table = PrettyTable(["Modules", "Parameters"])
      total_params = 0
      for name, parameter in model.named_parameters():
          if not parameter.requires_grad: continue
          params = parameter.numel()
          table.add_row([name, params])
          total_params+=params
      print(table)
      print(f"Total Trainable Params: {total_params}")
      return total_params
    #
    count_parameters(net)

  • 输出如下：

  • +-------------------+------------+
    |      Modules      | Parameters |
    +-------------------+------------+
    | embeddings.weight |   922866   |
    |    conv1.weight   |  1048576   |
    |     conv1.bias    |    1024    |
    |     bn1.weight    |    1024    |
    |      bn1.bias     |    1024    |
    |    conv2.weight   |  2097152   |
    |     conv2.bias    |    1024    |
    |     bn2.weight    |    1024    |
    |      bn2.bias     |    1024    |
    |    conv3.weight   |  2097152   |
    |     conv3.bias    |    1024    |
    |     bn3.weight    |    1024    |
    |      bn3.bias     |    1024    |
    |    lin1.weight    |  50331648  |
    |     lin1.bias     |    512     |
    |    lin2.weight    |   265728   |
    |     lin2.bias     |    519     |
    +-------------------+------------+
    Total Trainable Params: 56773369

  • Thong Nguyen - vote: 12

  • 如果希望避免重复计算共享的参数，可以用 torch.Tensor.data_ptr，即：

  • sum(dict((p.data_ptr(), p.numel()) for p in model.parameters()).values())

  • 下面是更繁琐的实现，添加了一个选项，用于跳过不被训练的参数：

  • def numel(m: torch.nn.Module, only_trainable: bool = False):
      """
      returns the total number of parameters used by m (only counting
      shared parameters once); if only_trainable is True, then only
      includes parameters with requires_grad = True
      """
      parameters = list(m.parameters())
      if only_trainable:
          parameters = [p for p in parameters if p.requires_grad]
      unique = {p.data_ptr(): p for p in parameters}.values()
      return sum(p.numel() for p in unique)

Check the total number of parameters in a PyTorch model

• Fábio Perez asked:

  • How to count the total number of parameters in a PyTorch model? Something similar to model.count_params() in Keras.
    怎么计算 Pytorch 模型的参数量？类似 Keras 的 model.count_params() 那样的函数。

• Answers:

  • Fábio Perez - vote: 198

  • PyTorch doesn\'t have a function to calculate the total number of parameters as Keras does, but it\'s possible to sum the number of elements for every parameter group:
    Pytorch 没有类似 Keras 计算参数量的函数，但可以通过每个参数组的求和得出参数量：

  • pytorch_total_params = sum(p.numel() for p in model.parameters())

  • If you want to calculate only the trainable parameters:
    如果希望只计算可训练的参数：

  • pytorch_total_params = sum(p.numel() for p in model.parameters() if p.requires_grad)

  • Answer inspired by this answer on PyTorch Forums.
    参考 Pytorch 论坛回答。

  • Note: I\'m answering my own question. If anyone has a better solution, please share with us.
    注：这个回答是自问自答，如果你们有更好的解决方法，请分享一下。

  • Fábio Perez - vote: 46

  • To get the parameter count of each layer like Keras, PyTorch has model.named_paramters() that returns an iterator of both the parameter name and the parameter itself.
    为了像 Keras 一样计算每层的参数量，Pytorch 的 model.named_paramters() 能返回一个迭代器，包含参数名与参数本身。

  • Here is an example:
    举例说明：

  • from prettytable import PrettyTable
    #
    def count_parameters(model):
      table = PrettyTable(["Modules", "Parameters"])
      total_params = 0
      for name, parameter in model.named_parameters():
          if not parameter.requires_grad: continue
          params = parameter.numel()
          table.add_row([name, params])
          total_params+=params
      print(table)
      print(f"Total Trainable Params: {total_params}")
      return total_params
    #
    count_parameters(net)

  • The output would look something like this:
    输出如下：

  • +-------------------+------------+
    |      Modules      | Parameters |
    +-------------------+------------+
    | embeddings.weight |   922866   |
    |    conv1.weight   |  1048576   |
    |     conv1.bias    |    1024    |
    |     bn1.weight    |    1024    |
    |      bn1.bias     |    1024    |
    |    conv2.weight   |  2097152   |
    |     conv2.bias    |    1024    |
    |     bn2.weight    |    1024    |
    |      bn2.bias     |    1024    |
    |    conv3.weight   |  2097152   |
    |     conv3.bias    |    1024    |
    |     bn3.weight    |    1024    |
    |      bn3.bias     |    1024    |
    |    lin1.weight    |  50331648  |
    |     lin1.bias     |    512     |
    |    lin2.weight    |   265728   |
    |     lin2.bias     |    519     |
    +-------------------+------------+
    Total Trainable Params: 56773369

  • Thong Nguyen - vote: 12

  • If you want to avoid double counting shared parameters, you can use torch.Tensor.data_ptr. E.g.:
    如果希望避免重复计算共享的参数，可以用 torch.Tensor.data_ptr，即：

  • sum(dict((p.data_ptr(), p.numel()) for p in model.parameters()).values())

  • Here\'s a more verbose implementation that includes an option to filter out non-trainable parameters:
    下面是更繁琐的实现，添加了一个选项，用于跳过不被训练的参数：

  • def numel(m: torch.nn.Module, only_trainable: bool = False):
      """
      returns the total number of parameters used by m (only counting
      shared parameters once); if only_trainable is True, then only
      includes parameters with requires_grad = True
      """
      parameters = list(m.parameters())
      if only_trainable:
          parameters = [p for p in parameters if p.requires_grad]
      unique = {p.data_ptr(): p for p in parameters}.values()
      return sum(p.numel() for p in unique)