Convert a net from TF to Torch

1

Hi I’m new to deep learning. Currently have an issue converting TensorFlow Net to PyTorch one… Since I don’t find any “template” that explains how “tf to pytorch”.
The TensorFlow Net is like follows:

from tensorflow.keras.layers import BatchNormalization
from tensorflow.keras.layers import Dense, Dropout, Input
from tensorflow.keras.layers import Bidirectional, Multiply
from tensorflow.keras.layers import Concatenate, LSTM, GRU

def dnn_model():
    
    x_input = Input(shape=(train.shape[-2:]))
    
    x1 = Bidirectional(LSTM(units=768, return_sequences=True))(x_input)
    x2 = Bidirectional(LSTM(units=512, return_sequences=True))(x1)
    x3 = Bidirectional(LSTM(units=384, return_sequences=True))(x2)
    x4 = Bidirectional(LSTM(units=256, return_sequences=True))(x3)
    x5 = Bidirectional(LSTM(units=128, return_sequences=True))(x4)
    
    z2 = Bidirectional(GRU(units=384, return_sequences=True))(x2)
    
    z31 = Multiply()([x3, z2])
    z31 = BatchNormalization()(z31)
    z3 = Bidirectional(GRU(units=256, return_sequences=True))(z31)
    
    z41 = Multiply()([x4, z3])
    z41 = BatchNormalization()(z41)
    z4 = Bidirectional(GRU(units=128, return_sequences=True))(z41)
    
    z51 = Multiply()([x5, z4])
    z51 = BatchNormalization()(z51)
    z5 = Bidirectional(GRU(units=64, return_sequences=True))(z51)
    
    x = Concatenate(axis=2)([x5, z2, z3, z4, z5])
    
    x = Dense(units=128, activation='selu')(x)
    
    x_output = Dense(units=1)(x)

    model = Model(inputs=x_input, outputs=x_output, 
                  name='DNN_Model')
    return model

've referred to both tf docs and pytorch ones. The parameters are pretty different.

For example tensorflow.keras.layers.LSTM has parameter units as the dimensionality of the output space, what Torch.nn.LSTM parameter does this correspond to?

Can I please have a guide?

2

hi,

Can you check GitHub - ysh329/deep-learning-model-convertor: The convertor/conversion of deep learning models for different deep learning frameworks/softwares.

Hope this works

4

This doesn’t solve the issue that transferring a manual Keras network to Pytorch structure, which is my purpose.
Thanks for sharing the repo still.

5

I manually transferred it, by referring to both Keras and Pytorch docs step by step.

import torch
from torch import nn
from torch.nn import LSTM
from torch.nn import GRU
from torch.nn import BatchNorm2d

class dnn_moduel(nn.Module):
    def __init__(self,input_size):
        super(vpp_moduel, self).__init__()
        self.x1 = LSTM(input_size = input_size, hidden_size = 768, bidirectional=True)
        self.x2 = LSTM(input_size = 768*2, hidden_size = 512, bidirectional=True)
        self.x3 = LSTM(input_size = 512*2, hidden_size = 384, bidirectional=True)
        self.x4 = LSTM(input_size = 384*2, hidden_size = 256, bidirectional=True)
        self.x5 = LSTM(input_size = 256*2, hidden_size = 128, bidirectional=True)

        self.z2 = GRU(input_size = 512, hidden_size = 384, bidirectional=True)

        self.z31b = BatchNorm2d(num_features=384)
        self.z3 = GRU(input_size = 384, hidden_size = 256, bidirectional=True)

        self.z41b = BatchNorm2d(num_features=256)
        self.z4 = GRU(input_size = 256, hidden_size = 128, bidirectional=True)

        self.z51b = BatchNorm2d(num_features=128)
        self.z5 = GRU(input_size = 128, hidden_size = 64, bidirectional=True)

        self.dense = torch.nn.Linear(1920,128)
        self.dense2 = torch.nn.Linear(128,1)
            
            
    def forward(self, x):
        
        x1 = self.x1(x)
        x2 = self.x2(x1)
        x3 = self.x3(x2)
        x4 = self.x4(x3)
        x5 = self.x5(x4)
        
        z2 = self.z2(x2)
        
        z31 = torch.multiply(x3,z2)
        z31b = self.z31b(z31)
        z3 = self.z3(z31b)
        
        z41 = torch.multiply(x4,z3)
        z41b = self.z41b(z41)
        z4 = self.z4(z41b)
        
        z51 = torch.multiply(x5,z4)
        z51b = self.z51b(z51)
        z5 = self.z5(z51b)
        
        c = torch.cat((x5, z2, z3, z4, z5)) #check dimention
        d = nn.SELU(self.dense(c))
        
        output = self.dense(d)
        
        return output