classification for peta

README.MD

Use PETA Dataset for Gender Detection

Steps

1. Download dataset from here
2. Uncompress the dataset to PETA, it should contain 10 subfolders like this

drwxrwxr-x 3 dola dola 4096 Oct 20  2014 3DPeS
drwxrwxr-x 3 dola dola 4096 Oct 20  2014 CAVIAR4REID
drwxrwxr-x 3 dola dola 4096 Oct 20  2014 CUHK
drwxrwxr-x 3 dola dola 4096 Oct 20  2014 GRID
drwxrwxr-x 3 dola dola 4096 Oct 20  2014 i-LID
drwxrwxr-x 3 dola dola 4096 Oct 20  2014 MIT
drwxrwxr-x 3 dola dola 4096 Oct 20  2014 PRID
drwxrwxr-x 3 dola dola 4096 Oct 20  2014 SARC3D
drwxrwxr-x 3 dola dola 4096 Oct 20  2014 TownCentre
drwxrwxr-x 3 dola dola 4096 Oct 20  2014 VIPeR

3. organize the photos into two subfolders male and female by running

python generate_dataset.py --peta_dir ./PETA --train_dir ./train

4. train with tensorflow.hub Inception v3

python ~/ws/hub/examples/image_retraining/retrain.py \
	--image_dir train \
	--output_graph model/gender_graph.pb \
	--output_labels model/gender_labels.txt \
	--how_many_training_steps 4000 \

5. freeze the model with strip_unused and quantization

python ~/ws/tensorflow/tensorflow/python/tools/strip_unused.py \
	--input_graph=./model/gender_graph.pb \
	--output_graph=./model/stripped_gender_graph.pb \
	--input_node_names=Placeholder \
	--output_node_names=final_result \
	--input_binary=true

and

python ~/ws/tensorflow/tensorflow/tools/quantization/quantize_graph.py \
	--input=model/stripped_gender_graph.pb \
	--output_node_names=final_result \
	--output=model/quantized_stripped_gender_graph.pb \
	--mode=weights

6. deploy to the mobile tf_classify app and run by changing the following in ClassifierActivity.java

private static final int INPUT_SIZE = 299; 
private static final int IMAGE_MEAN = 128; 
private static final float IMAGE_STD = 128; 
private static final String INPUT_NAME = "Placeholder"; 
private static final String OUTPUT_NAME = "final_result"; 
private static final String MODEL_FILE = "file:///android_asset/quantized_stripped_gender_graph.pb"; 
private static final String LABEL_FILE = 
"file:///android_asset/dog_retrained_labels.txt";  

generate_dataset.py

from __future__ import absolute_import
from __future__ import print_function
from __future__ import division

import os
import pandas as pd
from argparse import ArgumentParser
import shutil
import re

ARGV = None

def main():
	# parsing
	parser = ArgumentParser()
	parser.add_argument(
		'--peta_dir',
		required=True,
		type=str,
		help='where peta data is located, including 10 subfolders'
	)
	parser.add_argument(
		'--train_dir',
		required=True,
		type=str,
		help='where training data (male/female folders) are generated'
	)
	
	ARGV, _ = parser.parse_known_args()
	print(ARGV)
	PETA_DIR = os.path.abspath(ARGV.peta_dir)
	TRAIN_DIR = os.path.abspath(ARGV.train_dir)
	# scaffolding
	if os.path.exists(TRAIN_DIR):
		shutil.rmtree(TRAIN_DIR)
	os.makedirs(TRAIN_DIR)
	os.makedirs(os.path.join(TRAIN_DIR, 'male'))
	os.makedirs(os.path.join(TRAIN_DIR, 'female'))
	# generating
	for subdir in os.listdir(PETA_DIR):
		if not os.path.isdir(os.path.join(PETA_DIR, subdir)): continue
		print('Generating for {0}'.format(subdir))
		subdir = os.path.join(PETA_DIR, subdir, 'archive')
		lines = open(os.path.join(subdir, 'Label.txt')).readlines()
		labels = []
		for line in lines:
			line = line.strip()
			if line:
				try:
					label = (re.findall(r'^\d+', line)[0], re.findall(r'Male|Female', line)[0])
					labels.append(label)
				except:
					print('ERROR', line)

		male_head = '|'.join(set([head for head, gender in labels if gender == 'Male']))
		female_head = '|'.join(set([head for head, gender in labels if gender == 'Female']))

		for f in os.listdir(subdir):
			if f.endswith('txt'): continue
			src_f = os.path.join(subdir, f)
			if re.match('^'+male_head, f):
				dst_f = os.path.join(TRAIN_DIR, 'male', f)
			elif re.match('^'+female_head, f):
				dst_f = os.path.join(TRAIN_DIR, 'female', f)
			shutil.copy(src_f, dst_f)

if __name__ == '__main__':
	main()