[Kaggle]Dog Breed Identification

Kaggle 上的比赛，在这里使用了额外数据集Stanford Dogs Dataset。
在Public Leaderboard上排名57/1286（top 5%)



首先使用TensorFlow model zoo中的预训练模型导出图片特征，这里使用了inception_v3 与 resnet_v1_152import sys
sys.path.append('slim')
from tqdm import tqdm
from glob import glob
import numpy as np
import pandas as pd
import os
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
import tensorflow as tf
from nets import inception, resnet_v1
from preprocessing import vgg_preprocessing, inception_preprocessing
from tensorflow.contrib import slim
import h5py

df = pd.read_csv('sample_submission.csv')
synset = list(df.columns[1:])

inception_size = inception.inception_v3.default_image_size
resnet_size = resnet_v1.resnet_v1_152.default_image_size

model_dict = {'InceptionV3':{'model': inception.inception_v3,
'size':inception_size,
'scope':inception.inception_v3_arg_scope(),
'output':'AvgPool_1a',
'numclasses':1001,
'preprocessing':inception_preprocessing,
'ckpt_path':'inception_v3.ckpt'},
'resnet_v1_152':{'model': resnet_v1.resnet_v1_152,
'size':resnet_size,
'scope':resnet_v1.resnet_arg_scope(),
'output':'global_pool',
'numclasses':1000,
'preprocessing':vgg_preprocessing,
'ckpt_path':'resnet_v1_152.ckpt'}}

train_filenames = glob('Images/*/*.jpg')
test_filenames = glob('test/*.jpg')
train_num = len(train_filenames)
test_num = len(test_filenames)
labels = [synset.index(filename.split('/')[2][10:].lower()) for filename in train_filenames]

def get_features(model_name, model):

size = model['size']
feature_model = model['model']
preprocessing = model['preprocessing']

filename = tf.placeholder(tf.string)
file_contents = tf.read_file(filename)
image = tf.image.decode_jpeg(file_contents, channels=3)
processed_image = preprocessing.preprocess_image(image, size, size, is_training=False)
processed_inputs = tf.expand_dims(processed_image, 0)

with slim.arg_scope(model['scope']):
_, end_points = feature_model(processed_inputs, num_classes=model['numclasses'], is_training=False)
feature = tf.squeeze(end_points[model['output']])
init_fn = slim.assign_from_checkpoint_fn(model['ckpt_path'],
slim.get_model_variables(model_name))

train_features = np.zeros([train_num, 2048], dtype=np.float32)
test_features = np.zeros([test_num, 2048], dtype=np.float32)
with tf.Session() as sess:
init_fn(sess)
for i, train_filename in tqdm(enumerate(train_filenames), total=train_num):
train_features[i] = sess.run(feature, feed_dict={filename:train_filename})
for i, test_filename in tqdm(enumerate(test_filenames), total=test_num):
test_features[i] = sess.run(feature, feed_dict={filename:test_filename})

return train_features, test_features

train_features = []
test_features = []
for model_name, model in model_dict.items():
features = get_features(model_name, model)
train_features.append(features[0])
test_features.append(features[1])

train_features = np.concatenate(train_features, axis=1)
test_features = np.concatenate(test_features, axis=1)

with h5py.File('features.h5', 'w') as h:
h.create_dataset('train', data=train_features)
h.create_dataset('test', data=test_features)
h.create_dataset('labels', data=labels)
h.create_dataset('synset', data=synset)将模型导出的特征保存到h5文件中，然后使用keras构建一个简单的网络，进行下一步的训练

import os
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
from glob import glob
import numpy as np
import h5py
from sklearn.utils import shuffle
from keras.models import Model
from keras.utils import to_categorical
from keras.layers import Dropout, Dense, Input, Activation
from keras.layers.normalization import BatchNormalization

with h5py.File('features.h5', 'r') as h:
train = np.array(h['train'])
test = np.array(h['test'])
labels = np.array(h['labels'])
synset = list(h['synset'])

get_filename = lambda x:os.path.splitext(os.path.split(filename)[-1])[0]
test_filenames = [get_filename(filename) for filename in glob('test/*.jpg')]

train, labels = shuffle(train, labels)
labels = to_categorical(labels, num_classes=120)

input_tensor = Input(train.shape[1:])
x = BatchNormalization()(input_tensor)
x = Dense(1024)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Dropout(0.5)(x)
x = Dense(120, activation='softmax')(x)

model = Model(input_tensor, x)

model.compile(optimizer='adadelta',
loss='categorical_crossentropy',
metrics=['accuracy'])

model.fit(train, labels, batch_size=128, epochs=100)
print model.evaluate(train, labels, batch_size=128, verbose=0)
preds = model.predict(test, verbose=1)