Cannot reproduce a model using the model parameters.

Environment info

Operating System: CentOS 6.6

Compiler: gcc-4.8

Package used (Python/R/Scala/Julia): Python

MXNet version:

Or if installed from source: from source

MXNet commit hash (git rev-parse HEAD):

If you are using python package, please provide

Python version and distribution: Python 2.7.12 |Anaconda custom (64-bit)

Error Message:

Cannot reproduce a model using the model parameters.

Minimum reproducible example

I'm trying to reproduce the feed-forward step of an autoencoder by myself using python. I first train the model and save the params in files. Then I implement the feed-forward step and load the params. Then I feed data into it. If correctly implemented, the feed-forward step should yield the same output as that yielded by the original model in mxnet. However, the result I get is different. I don't understand what each argument means, as show in the example here)

args = {'encoder_%d_weight'%istack: mx.nd.empty((num_hidden, num_input,global_kx,global_ky), self.xpu),
                'encoder_%d_bias'%istack: mx.nd.empty((num_hidden,), self.xpu),
                'decoder_%d_weight'%istack: mx.nd.empty((num_hidden, num_input,global_kx,global_ky), self.xpu),
                'decoder_%d_bias'%istack: mx.nd.empty((num_hidden,), self.xpu),}
args_grad = {'encoder_%d_weight'%istack: mx.nd.empty((num_hidden, num_input,global_kx,global_ky), self.xpu),
                     'encoder_%d_bias'%istack: mx.nd.empty((num_hidden,), self.xpu),
                     'decoder_%d_weight'%istack: mx.nd.empty((num_hidden, num_input,global_kx,global_ky), self.xpu),
                     'decoder_%d_bias'%istack: mx.nd.empty((num_hidden,), self.xpu),}
args_mult = {'encoder_%d_weight'%istack: 1.0,
                     'encoder_%d_bias'%istack: 2.0,
                     'decoder_%d_weight'%istack: 1.0,
                     'decoder_%d_bias'%istack: 2.0,}
auxs = {}
if encoder_act == 'sigmoid' and sparseness_penalty:
    auxs['sparse_encoder_%d_moving_avg' % istack] = mx.nd.ones((num_hidden), self.xpu) * 0.5
if decoder_act == 'sigmoid' and sparseness_penalty:
    auxs['sparse_decoder_%d_moving_avg' % istack] = mx.nd.ones((num_hidden), self.xpu) * 0.5
init = mx.initializer.Uniform(0.07)
for k,v in args.items():
    init(k,v)

I don't understand what each term means, i.e. args, args_grad, args_mult, auxs. I only know that the args is used to multiply the data items. I only save the args in files and load it into my own implementation. I did not load other params. I think this would be the reason why my implementation is different from mxnet. But what these params mean? Could someone tell me? Thank you all for helping me!!!

apache / mxnet