First Deep Learning ~ Struggle ~

Hello everyone. @best_not_best. This article is a continuation of First Deep Learning ~ Preparation ~. If you haven't read it, please read it first.

Caution

This article is a product of personal desires and is not the official view of the organization to which it belongs.

Practice

3. Collect learning images (favorite entertainers)

I wanted to use the Google Image Search API, but according to the following article, usage restrictions seem to be strict. Collecting images with Web image search-Throat candy

Use the Bing Search API. I also referred to the above article for the source code.

get_images.py


#!/usr/bin/env python
# -*- coding: UTF-8 -*-

import sys
import os
import requests
import urllib
import urllib2
import json

BING_URL = 'https://api.datamarket.azure.com/Bing/Search/Image?'
MS_ACCTKEY = 'hogehoge'
QUERY = 'The name of your favorite celebrity'
OUTPUT_DIR_PATH = './talent/'

opener = urllib2.build_opener()
urllib2.install_opener(opener)

def download_urllist(urllist):
    for url in urllist:
        try:
            print 'url: ' + url
            with open(OUTPUT_DIR_PATH + '/' + os.path.basename(url), 'wb') as f:
                img = urllib2.urlopen(url, timeout = 5).read()
                f.write(img)
        except urllib2.URLError:
            print('URLError')
        except IOError:
            print('IOError')
        except UnicodeEncodeError:
            print('EncodeError')
        except OSError:
            print('OSError')

if __name__ == "__main__":
    query = urllib2.quote(QUERY)
    step = 20
    num = 50

    url_param_dict = {
        'Query': "'"+QUERY+"'",
        'Market': "'ja-JP'",
    }
    url_param_base = urllib.urlencode(url_param_dict)
    url_param_base = url_param_base + "&$format=json&$top=%d&$skip="%(num)

    for skip in range(0, num*step, num):
        url_param = url_param_base + str(skip)
        url = BING_URL + url_param
        print url

        response = requests.get(url,
                                auth = (MS_ACCTKEY, MS_ACCTKEY),
                                headers = {'User-Agent': 'My API Robot'})
        response = response.json()

        urllist = [item['MediaUrl'] for item in response['d']['results']]
        download_urllist(urllist)

Execute it with the following command.

$ python get_images.py

MS_ACCTKEY is the primary account key of Azure Market Place, QUERY is the character string you want to search, and ʻOUTPUT_DIR_PATH` is the storage directory of the retrieved file, so select it as appropriate. With this, you can get about 1,000 images for 50 x 20 pages. You did it!

4. Cut out the face part of the learning image

I wanted to use the previous cutout_face.py as it is, but since some of the images collected from the web are irregular, I added exception handling when cropping. By the way, I didn't like the fact that the cut out image file name contained Japanese, so I try to use the serial number as the file name.

cutout_talent_face.py


#!/usr/bin/env python
# -*- coding: UTF-8 -*-

import numpy
import os
import cv2

CASCADE_PATH = '/usr/local/opt/opencv/share/OpenCV/haarcascades/haarcascade_frontalface_alt.xml'
INPUT_DIR_PATH = './talent/'
OUTPUT_DIR_PATH = './talent_cutout/'
OUTPUT_FILE_FMT = '%s%d_%d%s'

count = 1
files = os.listdir(INPUT_DIR_PATH)
for file in files:
    input_image_path = INPUT_DIR_PATH + file

    #File reading
    image = cv2.imread(input_image_path)
    #Grayscale conversion
    try:
        image_gray = cv2.cvtColor(image, cv2.cv.CV_BGR2GRAY)
    except cv2.error:
        continue

    #Acquire the features of the cascade classifier
    cascade = cv2.CascadeClassifier(CASCADE_PATH)

    #Execution of object recognition (face recognition)
    facerect = cascade.detectMultiScale(image_gray, scaleFactor=1.1, minNeighbors=1, minSize=(1, 1))

    if len(facerect) > 0:
        #Saving recognition results
        i = 1
        for rect in facerect:
            print rect
            x = rect[0]
            y = rect[1]
            w = rect[2]
            h = rect[3]

            path, ext = os.path.splitext(os.path.basename(file))
            output_image_path = OUTPUT_FILE_FMT % (OUTPUT_DIR_PATH, count, i, ext)
            i += 1
            try:
                cv2.imwrite(output_image_path, image[y:y+h, x:x+w])
            except cv2.error:
                print file
                continue

    count += 1

Usage is similar to cutout_face.py.

5. Create a discriminator by learning 4. with Python + Chainer

I referred to the following. Chainer's NIN allows you to deep-learn and recognize your own image set --- shi3z's long diary Can Osomatsu-san's six children be identified by deep learning? ~ Implementation ~ --bohemia diary

Clone the source.

[work_dir]$ git clone https://github.com/shi3z/chainer_imagenet_tools.git

The ImageNet source will also be cloned.

[work_dir]$ git clone https://github.com/pfnet/chainer.git
[work_dir]$ cd chainer
[chainer]$ git checkout -b 1.4.1 refs/tags/v1.4.1
[chainer]$ cd ..

Download and unzip Caltech 101

[work_dir]$ cd chainer_imagenet_tools
[chainer_imagenet_tools]$ wget http://www.vision.caltech.edu/Image_Datasets/Caltech101/101_ObjectCategories.tar.gz
[chainer_imagenet_tools]$ tar xzvf 101_ObjectCategories.tar.gz

In the 101_ObjectCategories directory, put the directory containing the image cut out in 4. Also, if there are many directories, it will take time to learn, so some directories have been deleted. The structure of the 101_ObjectCategories directory is as follows. The panda is cute.

drwxr-xr-x  472 talent_cutout
drwxr-xr-x  438 Faces_easy
drwxr-xr-x  438 Faces
drwxr-xr-x   41 panda

Create train.txt, test.txt, label.txt.

[chainer_imagenet_tools]$ python make_train_data.py 101_ObjectCategories

Set the image size to 256 x 256.

[chainer_imagenet_tools]$ python crop.py images/ images/

Generate mean.npy.

[chainer_imagenet_tools]$ python ../chainer/examples/imagenet/compute_mean.py train.txt

Start learning!

[chainer_imagenet_tools]$ python ../chainer/examples/imagenet/train_imagenet.py -g -1 -E 20000 train.txt test.txt 2>&1 | tee log

...

train 3315 updates (106080 samples) time: 14:22:12.501801 (2.05054842326 images/sec)epoch 110

It doesn't end even after 14 hours (´; ω; `)

next time

That's all for this time. (I will study and come back again.) Continue to the final episode (scheduled to be published later this year)! !!

Postscript

solved! → First Deep Learning ~ Solution ~ --Qiita

Recommended Posts

First Deep Learning ~ Struggle ~
First Deep Learning ~ Preparation ~
First Deep Learning ~ Solution ~
Deep Learning
Deep Learning Memorandum
Start Deep learning
Python Deep Learning
Deep learning × Python
Python: Deep Learning Practices
Deep learning / activation functions
Deep Learning from scratch
Deep learning 1 Practice of deep learning
Deep learning / cross entropy
[AI] Deep Metric Learning
I tried deep learning
Python: Deep Learning Tuning
Deep learning large-scale technology
First deep learning in C #-Imitating implementation in Python-
Deep learning / softmax function
Deep Learning from scratch 1-3 chapters
Try deep learning with TensorFlow
Deep Learning Gaiden ~ GPU Programming ~
<Course> Deep Learning: Day2 CNN
Deep learning image recognition 1 theory
Deep running 2 Tuning of deep learning
Reinforcement learning 6 First Chainer RL
Deep learning / LSTM scratch code
Rabbit Challenge Deep Learning 1Day
<Course> Deep Learning: Day1 NN
Deep Kernel Learning with Pyro
Try Deep Learning with FPGA
Deep learning for compound formation?
Introducing Udacity Deep Learning Nanodegree
Subjects> Deep Learning: Day3 RNN
Introduction to Deep Learning ~ Learning Rules ~
Rabbit Challenge Deep Learning 2Day
Reinforcement learning 4 CartPole first step
Deep Reinforcement Learning 1 Introduction to Reinforcement Learning
Deep reinforcement learning 2 Implementation of reinforcement learning
Generate Pokemon with Deep Learning
Introduction to Deep Learning ~ Backpropagation ~
Deep Learning Model Lightening Library Distiller
Deep Learning / Deep Learning from Zero 2 Chapter 4 Memo
Try Deep Learning with FPGA-Select Cucumbers
Cat breed identification with deep learning
Deep Learning / Deep Learning from Zero Chapter 3 Memo
Make ASCII art with deep learning
Deep Learning / Deep Learning from Zero 2 Chapter 5 Memo
Implement Deep Learning / VAE (Variational Autoencoder)
Introduction to Deep Learning ~ Function Approximation ~
Try deep learning with TensorFlow Part 2
Deep learning from scratch (cost calculation)
About Deep Learning (DNN) Project Management
Deep learning to start without GPU
Solve three-dimensional PDEs with deep learning.
Introduction to Deep Learning ~ Coding Preparation ~
Organize machine learning and deep learning platforms
Deep learning learned by implementation 1 (regression)
Deep Learning / Deep Learning from Zero 2 Chapter 7 Memo
Deep Learning / Deep Learning from Zero 2 Chapter 8 Memo
Microsoft's Deep Learning Library "CNTK" Tutorial