mx.viz.plot_network KeyError: 'ctc_loss0_output'

[MenglingD]

Description

I builded my network with mx.sym.ctc_loss for the final loss on speech recognition task, and with mx.viz.plot_network to make sure the right struct of network. But I got an exception as I invoke mx.viz.plot_network, even if the symbol of network is as simple as possible.

Environment info (Required)

operating system: ubuntu-16.04 LTS
python version: python 3.5
mxnet version: 1.4.0-cpu

Error Message:

Traceback (most recent call last):
  File "ctc_plot.py", line 21, in <module>
    mx.viz.plot_network(sym, shape=shape).view()
  File "/home/windog/.local/lib/python3.5/site-packages/mxnet/visualization.py", line 382, in plot_network
    shape = shape_dict[key][1:]
KeyError: 'ctc_loss0_output'

Minimum reproducible example

(If you are using your own code, please provide a short script that reproduces the error. Otherwise, please provide link to the existing example.)

ctc_plot.py:

import mxnet as mx

def ctc_sym():
    data = mx.sym.Variable('data')
    label = mx.sym.Variable('label')
    data_len = mx.sym.Variable('data_len')
    label_len = mx.sym.Variable('label_len')

    ctc_loss = mx.sym.ctc_loss(data, label, data_len, label_len, True, True)
    ctc_loss = mx.sym.MakeLoss(ctc_loss)

    return ctc_loss


if __name__ == '__main__':
    shape = {'data': (400, 32, 100),    # [max_seq_len, batch_size, num_cls]
             'label': (32, 50),         # [batch_size, max_label_len]
             'data_len': (32,),         # [batch_size,]
             'label_len': (32,)}        # [batch_size,]
    sym = ctc_sym()
    mx.viz.plot_network(sym, shape=shape).view()

Steps to reproduce

(Paste the commands you ran that produced the error.)

1. python3 ctc_plot.py

What have you tried to solve it?

1. there is a related issue: #3221
2. I have debugged into incubator-mxnet/python/mxnet/visualization.py in which throw a exception, and got a wrong key for shape_dict in following position in visualization.py:

And then, I found out that the input name of symbol of MakeLoss(ctc_loss), "ctc_loss0_output0", has a bit difference from key = input_name + "_output" which is "ctc_loss0_output". And shape variable shape_dict is given by following code in visualization.py:

The definition of list_outputs which is located at line 734 of mxnet/model/symbol.py:

I could not deep into as its' low-level implement, so is there anyone to help me out. What does raise that exception? Incorrect invoking of ctc_loss or plot_network, or some inner mistakes in mxnet?

[MenglingD]

@szha

[vdantu]

@mxnet-label-bot add [question]

[szha]

@windoman sorry for the late reply. The index comes from that ctc_loss op is set to return more than one output: https://github.com/apache/incubator-mxnet/blob/master/src/operator/nn/ctc_loss.cc#L103

[MenglingD]

@windoman sorry for the late reply. The index comes from that ctc_loss op is set to return more than one output: https://github.com/apache/incubator-mxnet/blob/master/src/operator/nn/ctc_loss.cc#L103

@szha Thank you very much for your reply. The ctc_loss have two outputs, a loss symbol for per sample in batch and gradient output, which are accounted https://github.com/apache/incubator-mxnet/blob/master/src/operator/nn/ctc_loss.cc#L105, but only one output is visible https://github.com/apache/incubator-mxnet/blob/master/src/operator/nn/ctc_loss.cc#L109.