New Issue in FCN-XS example!

[ghost]

Hi all,

I originally ran the fcn-xs example well, but I met the errors when I updated mxnet this morning. The issue is following:

INFO:root:Namespace(epoch=27, init_type='fcnxs', model='fcn8s', prefix='FCN16s_VGG16', retrain=False)
INFO:root:Start training with gpu(0)
Traceback (most recent call last):
File "fcn_xs.py", line 73, in
main()
File "fcn_xs.py", line 57, in main
epoch_end_callback = mx.callback.do_checkpoint(fcnxs_model_prefix))
File "/home/x/xx/solver.py", line 71, in fit
aux_states=self.aux_params)
File "/home/xxx/mxnet/python/mxnet/symbol.py", line 718, in bind
args_handle, args = self._get_ndarray_inputs('args', args, listed_arguments, False)
File "/home/xxx/mxnet/python/mxnet/symbol.py", line 585, in _get_ndarray_inputs
raise ValueError('Must specify all the arguments in %s' % arg_key)
ValueError: Must specify all the arguments in args

Could anyone please tell me what's wrong with it? Thanks a lot in advance!

[piiswrong]

@pluskid Possibly due to the recent executor manager refactor?

[pluskid]

I don't see a connection, it seems the fcn-xs example is not using the FeedForward model API. This might be due to incompatibility between the saved model in the checkpoint file and the model trying to use. Could you double check that args.model is the same as what you saved for checkpoints? If it doesn't resolve the issue, maybe before the line

raise ValueError('Must specify all the arguments in %s' % arg_key)

in symbol.py, add a line that print(name) could help diagnosing.

[ghost]

@pluskid I add print(name) before ValueError line and it shows 'bigscore_bias'. What dose that mean? Thanks.

[pluskid]

@Rafel-Mou I'm not sure. It says that you are trying to initialize an executor but the variable bigscore_bias is not provided. Are you running from scratch or loading some existing models?

[tornadomeet]

@Rafel-Mou , do you use the original code in fcn-xs example? because big_score only lie in deconvolution operator, and bias is not included by default, so we no name bigsocre_bias there, i think you may use your own code, pleanse refere this line: https://github.com/dmlc/mxnet/blob/master/example/fcn-xs/symbol_fcnxs.py#L140

[ghost]

@tornadomeet Hi, Mr. OriginalAuthor. :D I use the original code in the example directly and do not make any changes. Formerly I can run the code well, but this error happens after I updated mxnet the day before yesterday...

[ghost]

@pluskid I'm running from scratch.

[tornadomeet]

@oh, i see it here: 27a3382
@piiswrong has changed the default no_bias from true to false, but in deconvolution, we usually not use bias like in caffe. so can you @Rafel-Mou give a pr for these?that is set no_bias = True.

[ghost]

@tornadomeet Could you please tell me which part of code that I should change? Thsnks!

[ghost]

@tornadomeet I get it done! I set DMLC_DECLARE_FIELD(no_bias).set_default(true) instead of DMLC_DECLARE_FIELD(no_bias).set_default(false) in convolution-inl.h and deconvolution-inl.h according to your suggestion. Thanks.

[tornadomeet]

@Rafel-Mou only need change in deconvolution-inl.h, not convolution.
because someone may changes the code in the newest mxnet, you may use my branch: https://github.com/tornadomeet/mxnet/tree/seg

[ghost]

@tornadomeet Okay, thank you. :)