Prediction fails in Handwritten Digit Recognition example

[ZhaotangLuo]

When I am trying to reproduce the example of Handwritten Digit Recognition with python and prebuild MXNet version 20160321, I encounter a preblem in prodiction, i.e. executing the following code
prob = model.predict(val_img[0:1].astype(np.float32)/255)[0]

I copied the code from http://mxnet.io/tutorials/python/mnist.html and here is my environment info and error message.

Environment info

Operating System: Windows 10

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

MXNet version: windows binary build 20160321 (20160321_win10_x64_gpu)

Python version and distribution: python 2.7

CUDA version: 8.0.44

CUDNN version: CUDNN v3

Error Message:

AssertionError                            Traceback (most recent call last)
<ipython-input-6-eec631db75f6> in <module>()
      2 plt.axis('off')
      3 plt.show()
----> 4 prob = model.predict(val_img[0:1].astype(np.float32)/255)[0]
      5 print 'Classified as %d with probability %f' % (prob.argmax(), max(prob))

E:\WinPython-64bit-2.7.10.3\python-2.7.10.amd64\lib\site-packages\mxnet-0.5.0-py2.7.egg\mxnet\model.pyc in predict(self, X, num_batch, return_data, reset)
    586             The predicted value of the output.
    587         """
--> 588         X = self._init_iter(X, None, is_train=False)
    589 
    590         if reset:

E:\WinPython-64bit-2.7.10.3\python-2.7.10.amd64\lib\site-packages\mxnet-0.5.0-py2.7.egg\mxnet\model.pyc in _init_iter(self, X, y, is_train)
    549                                       shuffle=is_train, last_batch_handle='roll_over')
    550             else:
--> 551                 return io.NDArrayIter(X, y, self.numpy_batch_size, shuffle=False)
    552         if not isinstance(X, io.DataIter):
    553             raise TypeError('X must be DataIter, NDArray or numpy.ndarray')

E:\WinPython-64bit-2.7.10.3\python-2.7.10.amd64\lib\site-packages\mxnet-0.5.0-py2.7.egg\mxnet\io.pyc in __init__(self, data, label, batch_size, shuffle, last_batch_handle)
    360         self.num_data = self.data_list[0].shape[0]
    361         assert self.num_data >= batch_size, \
--> 362             "batch_size need to be smaller than data size when not padding."
    363         self.cursor = -batch_size
    364         self.batch_size = batch_size

AssertionError: batch_size need to be smaller than data size when not padding.

What have you tried to solve it?

I tried other versions of prebuild MXNet. When using version 20160223, the problem is the same.
When I used later version 20160531 or 20160419 (with CUDNN v5.1 or v3), prediction was fine for CPU training of multilayer perceptron, but the GPU training of CNN yielded a poor accuracy (same as https://github.com/dmlc/mxnet/issues/1228)

How can I solve this problem?

[kevinthesun]

Are you running the exact same example in tutorial mnist?

[ZhaotangLuo]

Yes. I have carefully checked my code and they are exactly the same. Another tricky point I find is, when visualizing the network by mx.viz.plot_network(symbol=mlp, shape=shape), I cannot get the "data" block (but other blocks are fine).

[kevinthesun]

I can't reproduce this error. Is prediction fine for both mlp and cnn when using cpu?
For network visualizing, you need Graphviz library.

[ZhaotangLuo]

Prediction fails for CPU. If I update my prebuild MXNet to later version, prediction is fine for CPU, but the training accuracy for GPU is really poor.

[yajiedesign]

This issue is closed due to lack of activity in the last 90 days. Feel free to reopen if this is still an active issue. Thanks!