Lua version Deep Learning from scratch Part 5.5 [Making pkl files available in Lua Torch]

Past articles

Lua version Deep Learning from scratch Part 1 [Implementation of Perceptron] Lua version Deep Learning from scratch 2 [Activation function] Lua version Deep Learning from scratch Part 3 [Implementation of 3-layer neural network] [Lua version Deep Learning from scratch 4 [Implementation of softmax function]] (http://qiita.com/Kazuki-Nakamae/items/20e53a02a8b759583d31) Lua version Deep Learning from scratch Part 5 [Display MNIST image]

Introduction

This time it has nothing to do with the main story, but I will show you how to make the data stored in the pkl file available in Lua. 　 STEP1: Convert pkl file to npz file

You can do it with the following script.

pkl2npz.py

#!/usr/local/bin/python3
# coding: utf-8

"""
Output the contents of the pkl file to the npz file.
"""
__author__ = "Kazuki Nakamae <[email protected]>"
__version__ = "0.00"
__date__ = "22 Jun 2017"

import sys
import numpy as np
import pickle

def pkl2npz(infn, outfn):
    """
    @function   pkl2npz();
Output the contents of the pkl file to the npz file.
    @param  {string} infn :Input file name
    @param  {string} outfn :Output file name
    """

    with open(infn, 'rb') as f:
            ndarr = pickle.load(f)
            np.savez(outfn, W1=ndarr['W1'],W2=ndarr['W2'],W3=ndarr['W3'],b1=ndarr['b1'],b2=ndarr['b2'],b3=ndarr['b3'])

if __name__ == '__main__':
    argvs = sys.argv
    argc = len(argvs)

    if (argc != 3):   # Checking input
        print("USAGE : python3 pkl2npz.py <INPUT_PKLFILE> <OUTPUT_NPZFILE>")
        quit()

    pkl2npz(str(argvs[1]),str(argvs[2]))
quit()

pkl2npz.Run py

$ python3 pkl2npz.py sample_weight.pkl sample_weight.npz

One caveat is that you need to know the elements in advance. In this case, it is a pkl file that stores the weights (W1, W2, W3) and biases (b1, b2, b3) of the three-layer NN. 　 STEP2: Read the npz file

Now let's load the created sample_weight.npz on Lua. No hassle. The following people have created a package (npy4th) for that purpose. htwaijry/npy4th

npy4th installation

$ git clone https://github.com/htwaijry/npy4th.git
$ cd npy4th
$ luarocks make

It's easy to use and you can load it just by inserting loadnpz ([filename]).

loadnpz()How to use

npy4th = require 'npy4th'

-- read a .npz file into a table
tbl = npy4th.loadnpz('sample_weight.npz')

print(tbl["W1"])

Output result

Columns 1 to 6
-7.4125e-03 -7.9044e-03 -1.3075e-02  1.8526e-02 -1.5346e-03 -8.7649e-03
-1.0297e-02 -1.6167e-02 -1.2284e-02 -1.7926e-02  3.3988e-03 -7.0708e-02
-1.3092e-02 -2.4475e-03 -1.7722e-02 -2.4240e-02 -2.2041e-02 -5.0149e-03
-1.0008e-02  1.9586e-02 -5.6170e-03  3.8307e-02 -5.2507e-02 -2.3568e-02
(Omitted)
 1.1210e-02  1.0272e-02
-1.2299e-02  2.4070e-02
 7.4309e-03 -4.0211e-02
[torch.FloatTensor of size 784x50]

It has been properly converted to Tensor type. I didn't introduce it this time, but it is possible to read npy files in the same way. However, although it does load, please note that there is no guarantee that it will be in the form of a matrix that can be used as is. What it looks like depends on what it was on the numpy side. If necessary, you will need to resize () etc. after this. 　 in conclusion

Torch is not popular in Japan, but I think it would be easier to use if numpy resources could be used in this way. that's all. Thank you very much.