Created by: aprilvkuo

pytorch: https://pytorch.org/docs/1.1.0/nn.html?highlight=lstm#torch.nn.LSTM

paddle: https://www.paddlepaddle.org.cn/documentation/docs/zh/1.3/api_cn/layers_cn.html#dynamic-lstm

目前将pytorch的模型参数通过numpy保存， 再从paddle上加载，发现相同输入，效果不一致

pytorch代码

# 模型参数保存
  lstm_weight_ih = self.lstm.weight_ih_l0.detach().numpy()
  lstm_weight_hh = self.lstm.weight_hh_l0.detach().numpy()
  lstm_bias_ih = self.lstm.bias_ih_l0.detach().numpy()
  lstm_bias_hh = self.lstm.bias_hh_l0.detach().numpy()

#模型
 dropout = 0.0
 self.lstm = nn.LSTM(emb_size+32+32, hidden_size, layer_size,
                            bidirectional=True, batch_first=True, dropout=dropout)
 out, _ = self.lstm(out)

paddle代码

    def __lstm_reshape(self, data):
        """
        pytorch : W_ii|W_if|W_ig|W_io
        paddle : ii if io ig

        Arguments:
            data {[type]} -- [description]
        """
        dims = data.shape
        print(dims)
        assert dims[1] % 4 == 0
        each_dim = dims[1] / 4
        new_column_order = range(2 * each_dim) + range(3 * each_dim, 4 * each_dim) + range(2 * each_dim, 3 * each_dim)

        new_data = data[:, new_column_order]
        return new_data

#参数加载
     def load(self):
        self.lstm_weight_hh =  np.load(os.path.join(model_dir, "lstm_weight_hh.npy")).T
        self.lstm_weight_ih =  np.load(os.path.join(model_dir, "lstm_weight_ih.npy")).T
        self.lstm_bias_hh =  np.load(os.path.join(model_dir, "lstm_bias_hh.npy")).T.reshape((1, -1))
        self.lstm_bias_ih =  np.load(os.path.join(model_dir, "lstm_bias_ih.npy")).T.reshape((1, -1))


        self.lstm_weight_hh = self.__lstm_reshape(self.lstm_weight_hh)
        self.lstm_weight_ih = self.__lstm_reshape(self.lstm_weight_ih)
        self.lstm_bias_hh = self.__lstm_reshape(self.lstm_bias_hh)
        self.lstm_bias_ih = self.__lstm_reshape(self.lstm_bias_ih).flatten()
        # self.lstm_bias_ih = self.lstm_bias_ih.flatten()

        for item in [self.lstm_weight_hh, self.lstm_weight_ih, self.lstm_bias_hh, self.lstm_bias_ih]:
            print("data")
            print(item.shape)
        self.lstm_weight_hh =  fluid.initializer.NumpyArrayInitializer(self.lstm_weight_hh)
        self.lstm_weight_ih =  fluid.initializer.NumpyArrayInitializer(self.lstm_weight_ih)
        self.lstm_bias_hh =  fluid.initializer.NumpyArrayInitializer(self.lstm_bias_hh)
        self.lstm_bias_ih =  fluid.initializer.NumpyArrayInitializer(self.lstm_bias_ih)


# 建图
lstm_input = fluid.layers.fc(out, size=self.lstm_hidden_size*4, param_attr=self.lstm_weight_ih, bias_attr=self.lstm_bias_ih)
out_1, _  = fluid.layers.dynamic_lstm(input=lstm_input, size=self.lstm_hidden_size*4, \
                                            param_attr=self.lstm_weight_hh, bias_attr=self.lstm_bias_hh, \
                                            use_peepholes=False, is_reverse=False) # # hidden_state, cell_state
out_2, _  = fluid.layers.dynamic_lstm(input=lstm_input, size=self.lstm_hidden_size*4, \
                                            param_attr=self.lstm_weight_hh, bias_attr=self.lstm_bias_hh, \
                                            use_peepholes=False, is_reverse=True) # # hidden_state, cell_state