使用MindSpore Cell的construct报错AttributeError: For 'Cell', the method 'construct' is not defined.

经验分享 Tech Blogs 功能调试-Function Debugging
1

1 系统环境

硬件环境(Ascend/GPU/CPU): Ascend/GPU/CPU
MindSpore版本: mindspore=2.0.0
执行模式（PyNative/ Graph）:不限
Python版本: Python=3.7
操作系统平台: 不限

2 报错信息

2.1 问题描述

在pytorch迁移MindSpore过程中，错误信息如下

File "main_attention.py", line 88, in <module>  
    autofe.fit_attention(args)  
  File "/root/FETCH-main/autofe.py", line 251, in fit_attention  
    actions, log_probs, m1_output, m2_output, m3_output, action_softmax = self.ppo.choose_action_c(  
  File "/root/FETCH-main/feature_engineer/attention_searching/ppo.py", line 48, in choose_action_c  
    action_softmax, m1_output, m2_output, m3_output = self.actor_c(input_c, step)  
  File "/root/miniconda3/lib/python3.8/site-packages/mindspore/nn/cell.py", line 659, in __call__  
    raise AttributeError("For 'Cell', the method 'construct' is not defined.")  
AttributeError: For 'Cell', the method 'construct' is not defined.

2.2 脚本信息

两段相关的脚本如下：

actions, log_probs, m1_output, m2_output, m3_output, action_softmax = self.ppo.choose_action_c(workers_c[i].states[-1], step, epoch, c_ops, sample_rule)  
actions, log_probs, m1_output, m2_output, m3_output, action_softmax = ppo.choose_action_c(state_c, step, epoch, c_ops, sample_rule)

脚本定义如下：

self.actor_c = Actor(args, data_nums, operations_c, d_model, d_k, d_v, d_ff, n_heads, dropout=dropout, enc_load_pth=args.enc_c_pth)  
class Actor(nn.Cell):  
    def __init__(self, args, data_nums, operations, d_model, d_k, d_v, d_ff, n_heads, dropout=None, enc_load_pth=None):  
        super(Actor, self).__init__()  
        self.args = args  
        self.reduction_dimension = ReductionDimension(data_nums, d_model)  
        self.encoder = EncoderLayer(d_model, d_k, d_v, d_ff, n_heads, dropout)  
        logging.info(f"Randomly initial encoder")  
        if os.path.exists(enc_load_pth):  
            self.encoder.load_state_dict(mindspore.load(enc_load_pth))  
            logging.info(f"Successfully load encoder, enc_load_pth:{enc_load_pth}")  
        self.select_operation = SelectOperations(d_model, operations)  
        self.c_nums = len(args.c_columns)  
        self.layernorm = nn.LayerNorm(normalized_shape=(data_nums,))  
    
    def forward(self, input, step):  
        input_norm = self.layernorm(input)  
        data_reduction_dimension = self.reduction_dimension(input_norm)  
        data_reduction_dimension = mindspore.ops.where(mindspore.ops.isnan(data_reduction_dimension),  
                                                mindspore.ops.full_like(data_reduction_dimension, 0), data_reduction_dimension)  
        encoder_output = self.encoder(data_reduction_dimension)  
        encoder_output = mindspore.ops.where(mindspore.ops.isnan(encoder_output), mindspore.ops.full_like(encoder_output, 0), encoder_output)  
        output = self.select_operation(encoder_output)  
        output = mindspore.ops.where(mindspore.ops.isnan(output), mindspore.ops.full_like(output, 0), output)  
        operation_softmax = mindspore.ops.softmax(output, dim=-1)  
        return operation_softmax, data_reduction_dimension.squeeze(), \  
                encoder_output.squeeze(), output

3 根因分析

根据文档：https://www.mindspore.cn/docs/zh-CN/r2.0/migration_guide/typical_api_comparision.html?highlight=device#nn.Module
使用 PyTorch 构建网络结构时，我们会用到nn.Module 类，通常将网络中的元素定义在__init__ 函数中并对其初始化，将网络的图结构表达定义在forward 函数中，通过调用这些类的对象完成整个模型的构建和训练。nn.Module 不仅为我们提供了构建图接口，它还为我们提供了一些常用的 API ，来帮助我们执行更复杂逻辑。
MindSpore 中的 nn.Cell 类发挥着和 PyTorch 中 nn.Module 相同的作用，都是用来构建图结构的模块，MindSpore 也同样提供了丰富的 API 供开发者使用，虽然名字不能一一对应，但 nn.Module 中常用的功能都可以在nn.Cell 中找到映射

4 解决方案

import mindspore as ms  
from mindspore import ops, nn  
class Actor(nn.Cell):  
    def __init__(self, args, data_nums, operations, d_model, d_k, d_v, d_ff, n_heads, dropout=None, enc_load_pth=None):  
        super(Actor, self).__init__()  
        pass # 原代码  
    def construct(self, input, step):  
        pass # 原代码