How can I feed .csv training data to a convolutional neural network in mxnet?

Question

I recently installed mxnet (python package) with GPU support on Windows 10 and Python 3.5. I run through a couple of examples and they seem to work fine.

I am used to scikit-learn style machine learning packages and very new to Python deep learning packages such as Mxnet although I have already used Mxnet in R. I'm having a hard time understanding how to feed .csv training data to the model.

I would like to feed to a simple CNN some images. The images are 28x28 pixel and stored as flattened arrays in a .csv. I have two .csv files, one for training and the other for testing. Each .csv file has the following structure:

label, pixel1, pixel2, ..., pixel784
0,...
1,...

There are 10 labels in total and around 1000/300 images in the training set/test set.

I am using the following code to load the data and train the model:

import mxnet as mx
import pandas as pd
import numpy as np
import os

path = "C://users//me//data"
os.chdir(path)

df_train = pd.read_csv("train_28.csv")
df_test = pd.read_csv("test_28.csv")

keys = ['pixel.'+str(i) for i in range(1,785)]

X_train = df_train[keys].get_values().T
X_train = X_train.reshape((1200,28,28,1))
y_train = df_train['label'].get_values().reshape((1200,1))
#y_train = y_train.reshape((28,28,1,1200))



data = mx.symbol.Variable('data')

# First conv layer
conv1 = mx.symbol.Convolution(data=data, kernel=(5,5), num_filter=20)
tanh1 = mx.symbol.Activation(data=conv1, act_type="tanh")
pool1 = mx.symbol.Pooling(data=tanh1, pool_type="max",
                              kernel=(2,2), stride=(2,2))

# Second conv layer
conv2 = mx.symbol.Convolution(data=pool1, kernel=(5,5), num_filter=50)
tanh2 = mx.symbol.Activation(data=conv2, act_type="tanh")
pool2 = mx.symbol.Pooling(data=tanh2, pool_type="max",
                              kernel=(2,2), stride=(2,2))

# First fully connected
flatten = mx.symbol.Flatten(data=pool2)
fc1 = mx.symbol.FullyConnected(data=flatten, num_hidden=500)
tanh3 = mx.symbol.Activation(data=fc1, act_type="tanh")
# second fullc
fc2 = mx.symbol.FullyConnected(data=tanh3, num_hidden=10)
# loss
lenet = mx.symbol.SoftmaxOutput(data=fc2, name='softmax')                              

device = mx.gpu()                              

model = mx.model.FeedForward.create(lenet,
                                    X = X_train,
                                    y = y_train,
                                    ctx = device,
                                    num_epoch = 30)

I am using this approach which is similar to the one I was using with mxnet in R, (btw on R it works perfectly, however I cannot use the GPU on R so I need to use Python for better performances...) however I am getting the following error:

[16:54:11] D:\chhong\mxnet\dmlc-core\include\dmlc/logging.h:235: [16:54:11] d:\chhong\mxnet\src\operator\./convolution-inl.h:347: Check failed: ksize_x <= dshape[3] && ksize_y <= dshape[2] kernel size exceed input
Traceback (most recent call last):
  File "C:\Users\Me\Desktop\esempio_lenet.py", line 57, in <module>
    num_epoch = 30)
  File "C:\Users\Me\Anaconda3\lib\site-packages\mxnet-0.7.0-py3.5.egg\mxnet\model.py", line 901, in create
    eval_batch_end_callback=eval_batch_end_callback)
  File "C:\Users\Me\Anaconda3\lib\site-packages\mxnet-0.7.0-py3.5.egg\mxnet\model.py", line 745, in fit
    self._init_params(dict(data.provide_data+data.provide_label))
  File "C:\Users\Me\Anaconda3\lib\site-packages\mxnet-0.7.0-py3.5.egg\mxnet\model.py", line 485, in _init_params
    arg_shapes, _, aux_shapes = self.symbol.infer_shape(**input_shapes)
  File "C:\Users\Me\Anaconda3\lib\site-packages\mxnet-0.7.0-py3.5.egg\mxnet\symbol.py", line 453, in infer_shape
    return self._infer_shape_impl(False, *args, **kwargs)
  File "C:\Users\Me\Anaconda3\lib\site-packages\mxnet-0.7.0-py3.5.egg\mxnet\symbol.py", line 513, in _infer_shape_impl
    ctypes.byref(complete)))
  File "C:\Users\Me\Anaconda3\lib\site-packages\mxnet-0.7.0-py3.5.egg\mxnet\base.py", line 77, in check_call
    raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: InferShape Error in convolution0: [16:54:11] d:\chhong\mxnet\src\operator\./convolution-inl.h:347: Check failed: ksize_x <= dshape[3] && ksize_y <= dshape[2] kernel size exceed input

And I cannot figure out what I am doing wrong. Could please someone tell me what is this error about and provide me with a clear set of instructions on how to load .csv files with the same structure as above and train a mxnet model? I took a look at the documentation but could not figure out on my own how to load .csv files properly...

The reason why I am asking for a procedure for loading such .csv files is that I mostly deal with data in that format and it would be very valuable to me to be able to run a script against a folder with these .csv files and have them ready to be used for training a deep convolutional neural network.

A set of train and test .csv files are available here should you need them to reproduce the example code I wrote.

Thank you

Answer 1

The problems is in this line:

X_train = X_train.reshape((1200,28,28,1))

In mxnet the second dimension is feature maps, while the third and fourth dimensions are width and height, so it should be:

X_train = X_train.reshape((1200,1,28,28))