Deconvolutions in Chainer vs. Tensorflow

[Rob Bergen]

I was recently trying to reproduce a Chainer model in Tensorflow and was running into problems reproducing outputs in 1D deconvolution layers. For example, in the code snippet below, I use the same weights, strides, paddings, etc and get different outputs for TF deconvs and chainer deconvs:

import numpy as np

import chainer.links as L

from tensorflow.keras.layers import Conv1DTranpose

chainer_deconv = L.DeconvolutionND(1, 100, 1024, 1, 1, 0, initialW=fsgen_weights['dc0/W'],initial_bias=fsgen_weights['dc0/b'])

tf_deconv = Conv1DTranspose(1024, 1, strides=1,data_format='channels_first',kernel_initializer=tf.constant_initializer(fsgen_weights['dc0/W']), bias_initializer=tf.constant_initializer(fsgen_weights['dc0/b']))

batch_size = 5

test_in = np.random.rand(batch_size,100,1)*2-1

#print a portion of the outputs

print(chainer_deconv(test_in)[0][0:4])  

--> prints variable([[-0.00434582] [ 0.08588306] [ 0.01726495] [ 0.0137401 ] ])

print(tf_deconv(test_in)[0][0:4])

--> prints tf.Tensor( [[-0.03280227] [ 0.08922134] [-0.03946912] [-0.10054161] ])

Can anybody tell me why these layers return different values if they are using identical weights and inputs?