In [1]:
# windows only hack for graphviz path
import os
for path in os.environ['PATH'].split(os.pathsep):
if path.endswith("Library\\bin"):
os.environ['PATH']+=os.pathsep+os.path.join(path, 'graphviz')
In [2]:
# 設定環境變數來控制 keras, theano
os.environ['KERAS_BACKEND']="tensorflow"
#os.environ['THEANO_FLAGS']="floatX=float32, device=cuda"
In [3]:
import keras
from keras.models import Sequential
from PIL import Image
import numpy as np
In [4]:
import keras.backend as K
# 設定 channels_first 或 channels_last
K.set_image_data_format('channels_last')
In [5]:
base_model = keras.applications.vgg16.VGG16(weights='imagenet', include_top=False, input_shape=(64,64,3))
In [6]:
from keras.layers import GlobalAveragePooling2D, Dense
from keras.models import Model
_ = base_model.get_layer("block5_conv3").output
_ = GlobalAveragePooling2D()(_)
_ = Dense(512, activation='relu')(_)
# 10 個輸出
predictions = Dense(10, activation='softmax')(_)
# 這是我們的 model
model = Model(inputs=base_model.input, outputs=predictions)
# 將原來的模型設定成不可訓練，只訓練上面的兩層 dense layers
for layer in base_model.layers:
layer.trainable = False
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
In [7]:
# 讀取 cifar10 dataset
# 只有 train 和 test 沒有 validation
import tarfile
import pickle
train_X=[]
train_y=[]
tar_gz = "../Week06/cifar-10-python.tar.gz"
with tarfile.open(tar_gz) as tarf:
for i in range(1, 6):
dataset = "cifar-10-batches-py/data_batch_%d"%i
print("load",dataset)
with tarf.extractfile(dataset) as f:
result = pickle.load(f, encoding='latin1')
train_X.extend(result['data'])
train_y.extend(result['labels'])
train_X=np.float32(train_X)
train_y=np.int32(train_y)
dataset = "cifar-10-batches-py/test_batch"
print("load",dataset)
with tarf.extractfile(dataset) as f:
result = pickle.load(f, encoding='latin1')
test_X=np.float32(result['data'])
test_y=np.int32(result['labels'])
train_Y = np.eye(10)[train_y]
test_Y = np.eye(10)[test_y]
train_X = np.moveaxis(train_X.reshape(-1,3,32,32),1,3)
test_X = np.moveaxis(test_X.reshape(-1,3,32,32),1,3)
validation_data = (test_X[:1000], test_Y[:1000])
test_data = (test_X[1000:], test_Y[1000:])
In [8]:
# 改變輸入圖形的大小並且處理
from keras.applications.vgg16 import preprocess_input
def resize_image(arr):
img = Image.fromarray(arr)
img = img.resize((64,64))
img = img.convert("RGB")
return np.array(img)
def preprocess_X(X):
X = X.astype('uint8')
X = np.float32([resize_image(X[i]) for i in range(X.shape[0])])
return preprocess_input(X)
In [9]:
from PIL import Image
from IPython.display import display
def showX(X):
int_X = X.clip(0,255).astype('uint8')
int_X_reshape = int_X.swapaxes(0,1).reshape(32,-1,3)
display(Image.fromarray(int_X_reshape))
# 訓練資料， X 的前 20 筆
showX(train_X[:20])
In [10]:
import numpy as np
def generate(X, Y):
while 1:
idx = np.random.choice(X.shape[0], size=32, replace=False)
_X = preprocess_X(X[idx])
_Y = Y[idx]
yield (_X, _Y)
v_data = (preprocess_X(test_X[:200]), test_Y[:200])
model.fit_generator(generate(train_X, train_Y), steps_per_epoch=500, epochs=10, validation_data=v_data)
Out[10]:
In [11]:
# 實際
predict_y = (model.predict(preprocess_X(test_X))).argmax(axis=1)
test_y = test_Y.argmax(axis=1)
In [12]:
(predict_y == test_y).mean()
Out[12]:
In [13]:
print(predict_y[:30])
showX(test_X[:30])
In [14]:
for n in ["block5_conv1", "block5_conv2", "block5_conv3"]:
model.get_layer(n).trainable = True
from keras.optimizers import Adam
model.compile(optimizer=Adam(lr=0.0001), loss='categorical_crossentropy', metrics=["accuracy"])
In [15]:
model.fit_generator(generate(train_X, train_Y), steps_per_epoch=500, epochs=10, validation_data=v_data)
Out[15]:
In [16]:
predictions = (model.predict(preprocess_X(test_X))).argmax(axis=1)
(predictions == test_y).mean()
Out[16]:
In [ ]: