Dogs

Dogs

Classification of Dog, Hot Dog and Dog Food.

ImageNet

Brief Description

• 14,197,122 images, 21841 synsets indexed (2018/12/17)
• 334MB for only compressed image URLs

synset is a short for synonym set.

(information science) A set of one or more synonyms that are interchangeable in some context without changing the truth value of the proposition in which they are embedded.

Data

Class	Description
Dog, domestic dog, Canis familiaris	A member of the genus Canis (probably descended from the common wolf) that has been domesticated by man since prehistoric times; occurs in many breeds; "the dog barked all night"
Hotdog, hot dog, red hot	A frankfurter served hot on a bun
Dog food	Food prepared for dogs

The url of images may be invalid

• Size too small
• HTTP Error: 404 not found
• Timeout
• Invalid url character (not ascii)
• Invalid image (can't be opened as array)
• ValueError: Unknown url type

we'll deal with them in data_downloader.py and invalid_image_deleter.py

Dependencies

• resizeimage - pip install python-resize-image

Usage

Download images from ImageNet

python3 data_downloader.py

Delete some invalid images (which can't open)

not sure why it can't be deleted by the same operation logic in data_downloader.py

python3 invalid_image_deleter.py

Train ResNet50

python3 ResNet50model.py

Train VGG16

python3 VGG16model.py

Result

Training after 10 epoch

ResNet50

• Accuracy: 96.77%
• Loss: 0.0797

########################################################################
#      Date:           Thu Dec 27 06:07:09 PST 2018
#    Job ID:           9374.c009
#      User:           u22711
# Resources:           neednodes=1:ppn=2,nodes=1:ppn=2,walltime=06:00:00
########################################################################

Found 1306 images belonging to 3 classes.
Found 434 images belonging to 3 classes.
Found 1740 images belonging to 3 classes.
Epoch 1/10
55/54 [==============================] - 65s 1s/step - loss: 0.3709 - acc: 0.8823 - val_loss: 0.1635 - val_acc: 0.9470
Epoch 2/10
55/54 [==============================] - 62s 1s/step - loss: 0.1748 - acc: 0.9527 - val_loss: 0.1433 - val_acc: 0.9447
Epoch 3/10
55/54 [==============================] - 61s 1s/step - loss: 0.1391 - acc: 0.9568 - val_loss: 0.1248 - val_acc: 0.9539
Epoch 4/10
55/54 [==============================] - 62s 1s/step - loss: 0.1209 - acc: 0.9644 - val_loss: 0.1030 - val_acc: 0.9677
Epoch 5/10
55/54 [==============================] - 62s 1s/step - loss: 0.1134 - acc: 0.9633 - val_loss: 0.0938 - val_acc: 0.9747
Epoch 6/10
55/54 [==============================] - 61s 1s/step - loss: 0.0966 - acc: 0.9667 - val_loss: 0.0993 - val_acc: 0.9724
Epoch 7/10
55/54 [==============================] - 60s 1s/step - loss: 0.0955 - acc: 0.9691 - val_loss: 0.0887 - val_acc: 0.9677
Epoch 8/10
55/54 [==============================] - 61s 1s/step - loss: 0.1035 - acc: 0.9702 - val_loss: 0.0991 - val_acc: 0.9700
Epoch 9/10
55/54 [==============================] - 61s 1s/step - loss: 0.0864 - acc: 0.9750 - val_loss: 0.0909 - val_acc: 0.9700
Epoch 10/10
55/54 [==============================] - 60s 1s/step - loss: 0.0809 - acc: 0.9750 - val_loss: 0.0948 - val_acc: 0.9770
19/18 [===============================] - 15s 788ms/step
loss : 0.07971909043941355
acc : 0.967741926694246
100/100 [==============================] - 22s 221ms/step

########################################################################
# End of output for job 9374.c009
# Date: Thu Dec 27 06:18:18 PST 2018
########################################################################

VGG16

• Accuracy: 93.55%
• Loss: 0.6773

########################################################################
#      Date:           Thu Dec 27 06:07:49 PST 2018
#    Job ID:           9375.c009
#      User:           u22711
# Resources:           neednodes=1:ppn=2,nodes=1:ppn=2,walltime=06:00:00
########################################################################

Found 1306 images belonging to 3 classes.
Found 434 images belonging to 3 classes.
Found 1740 images belonging to 3 classes.
Epoch 1/10
55/54 [==============================] - 48s 878ms/step - loss: 8.8332 - acc: 0.4459 - val_loss: 8.8761 - val_acc: 0.4493
Epoch 2/10
55/54 [==============================] - 44s 806ms/step - loss: 8.8651 - acc: 0.4500 - val_loss: 8.8761 - val_acc: 0.4493
Epoch 3/10
55/54 [==============================] - 45s 823ms/step - loss: 8.8147 - acc: 0.4531 - val_loss: 8.8761 - val_acc: 0.4493
Epoch 4/10
55/54 [==============================] - 45s 822ms/step - loss: 8.8651 - acc: 0.4500 - val_loss: 8.8761 - val_acc: 0.4493
Epoch 5/10
55/54 [==============================] - 45s 823ms/step - loss: 8.8987 - acc: 0.4479 - val_loss: 8.8761 - val_acc: 0.4493
Epoch 6/10
55/54 [==============================] - 45s 820ms/step - loss: 8.8819 - acc: 0.4489 - val_loss: 8.8761 - val_acc: 0.4493
Epoch 7/10
55/54 [==============================] - 45s 823ms/step - loss: 3.8641 - acc: 0.7441 - val_loss: 1.3152 - val_acc: 0.9124
Epoch 8/10
55/54 [==============================] - 46s 831ms/step - loss: 0.8339 - acc: 0.9285 - val_loss: 0.6211 - val_acc: 0.9470
Epoch 9/10
55/54 [==============================] - 46s 836ms/step - loss: 0.5697 - acc: 0.9509 - val_loss: 2.1983 - val_acc: 0.7765
Epoch 10/10
55/54 [==============================] - 46s 836ms/step - loss: 0.6674 - acc: 0.9447 - val_loss: 0.8443 - val_acc: 0.9263

...
19/18 [===============================] - 12s 624ms/step
loss : 0.6773356663387133
acc : 0.9354838665729294

...
100/100 [==============================] - 3s 30ms/step

########################################################################
# End of output for job 9375.c009
# Date: Thu Dec 27 06:15:50 PST 2018
########################################################################

Notes

Download and show an image in python

import urllib
import io
from PIL import Image
a = urllib.request.urlopen('http://static.flickr.com/2611/3680714896_bb5cbc89cb.jpg')
b = io.BytesIO(a.read()) # Seem like this step is redundant, not sure.
c = Image.open(b)
c.show()

Visualize data augmentation

# First make a directory to store the processed data
# mkdir test

# Main
from keras.preprocessing.image import ImageDataGenerator
# Can add more data augmentation trick
gen = ImageDataGenerator(horizontal_flip=True, width_shift_range=0.2, height_shift_range=0.2)
iterator = gen.flow_from_directory('data', save_to_dir='./test')
next(iterator) # Test a batch (default 32)

Links

Other ways to download image from ImageNet

• How to get Images from ImageNet with Python in Google Colaboratory
  • colaboratory notbook
• xkumiyu/imagenet-downloader
• My First Attempt at ImageNet

Data preprocessing

• Kaggle notebook - From image files to Numpy Arrays!

Keras ImageDataGenerator

• Keras Image Preprocessing
• Tutorial on using Keras flow_from_directory and generators
• Tutorial on Keras flow_from_dataframe

Keras Model

• VGG-16 pre-trained model for Keras

Keras Transfer Learning

• Keras Applications
• Kaggle Transfer Learning

Keras Data Augmentation

• Kaggle Data Augmentation