033. conver colored image to vector,use cnn model
<![CDATA[
# @
# I want to convert image to vector,
# but with keeping color data
from PIL import Image
import numpy as np
with open("tower.jpg","rb") as file:
opened_file=Image.open(file)
converted_image_to_rgb=opened_file.convert("RGB")
# You reduce image size by reshape
resized_image=converted_image_to_rgb.resize((64,64))
converted_image_to_nparray=np.array(resized_image)
# @
from PIL import Image
import os,glob
import numpy as np
from sklearn.model_selection import train_test_split
# This is folder I want to load
folder_you_want_to_open="./image/101_ObjectCategories"
# These are categories_list which I want to use
categories_list=["chair","camera","butterfly","elephant","flamingo"]
# You define number_of_classes by len(categories_list)
number_of_classes=len(categories_list)
# You define size of image as 64*64
width_of_image=64
height_of_image=64
# Since you will use RGB information,
# you should multiply 3,
# to calculate entire number of element of image vector
number_of_element_of_converted_vector_from_image=width_of_image*height_of_image*3
# You will load image data
X_feature_of_data_list=[]
Y_label_of_data_list=[]
for index,category in enumerate(categories_list):
# You create list for label_list data
label_list=[0 for index in range(number_of_classes)]
label_list[index]=1
# You define image folder
image_directory=folder_you_want_to_open+"/"+category
# glob can load every file in specified folder
loaded_files_from_specified_folder=glob.glob(image_directory+"/*.jpg")
for index,one_file in enumerate(loaded_files_from_specified_folder):
opened_file=Image.open(one_file)
converted_file_to_rgb=opened_file.convert("RGB")
resized_file=converted_file_to_rgb.resize((width_of_image,height_of_image))
filed_represented_in_nparray=np.asarray(resized_file)
X_feature_of_data_list.append(filed_represented_in_nparray)
Y_label_of_data_list.append(label_list)
if index%10==0:
print(index,"\n",data)
X_feature_of_data_nparray=np.array(X_feature_of_data_list)
Y_label_of_data_nparray=np.array(Y_label_of_data_list)
# I separate train dataset and test dataset
x_train_data,x_test_data,y_train_data,y_test_data\
=train_test_split(X_feature_of_data_list,Y_label_of_data_list)
xy_data_set=(x_train_data,x_test_data,y_train_data,y_test_data)
np.save("./image/5obj.npy",xy_data_set)
print("ok,",len(Y_label_of_data_list))
# @
Classifying image by cnn
from keras.models import Sequential
from keras.layers import Convolution2D,MaxPooling2D
from keras.layers import Activation,Dropout,Flatten,Dense
import h5py
from PIL import Image
import numpy as np
import os
# You define categories
categories_list=["chair","camera","butterfly","elephant","flamingo"]
# You define number_of_classes by len(categories_list)
number_of_classes=len(categories_list)
# Since you created image data,
# whose size is 64*64,
# you will disgnate 64,64
width_of_image=64
height_of_image=64
# You load 5obj.npy file,
# which you created from above
x_train_data,x_test_data,y_train_data,y_test_data=np.load("./image/5obj.npy")
# You normalize data,
# because input data should have value ranged from 0 to 1
normalized_x_train_data=x_train_data.astype("float")/256
normalized_x_test_data =x_test_data.astype("float") /256
print("normalized_x_train_data.shape:",normalized_x_train_data.shape)
# You implement cnn model
# Following scheme is ordinal one,
# which can be applied to in almost cases
sequential_object=Sequential()
# You perform filter sliding
sequential_object.add(Convolution2D(32,3,3,
border_mode='same',
input_shape=x_train_data.shape[1:]))
# You perform relu funtion
sequential_object.add(Activation('relu'))
# You perform pooling
sequential_object.add(MaxPooling2D(pool_size=(2,2)))
# You perform dropout
sequential_object.add(Dropout(0.25))
# You perform 2nd convolution layer
sequential_object.add(Convolution2D(64,3,3,border_mode='same'))
sequential_object.add(Activation('relu'))
sequential_object.add(Convolution2D(64,3,3))
sequential_object.add(MaxPooling2D(pool_size=(2,2)))
sequential_object.add(Dropout(0.25))
sequential_object.add(Flatten())
sequential_object.add(Dense(512))
sequential_object.add(Activation('relu'))
sequential_object.add(Dropout(0.5))
sequential_object.add(Dense(number_of_classes))
sequential_object.add(Activation('softmax'))
# You actually run built model
sequential_object.compile(\
loss='binary_crossentropy'\
,optimizer='rmsprop'\
,metrics=['accuracy'])
# You train model
hdf5_file="./image/5obj-sequential_object.hdf5"
if os.path.exists(hdf5_file):
sequential_object.load_weights(hdf5_file)
else:
sequential_object.fit(x_train_data,y_train_data,batch_size=32,nb_epoch=50)
sequential_object.save_weights(hdf5_file)
# You evaluate model
prediction_values=sequential_object.predict(x_test_data)
for index,prediction_value in enumerate(prediction_values):
prediction_value_after_argmax=prediction_value.argmax()
label_value_after_argmax=y_test_data[index].argmax()
one_x_feature_test_data=x_test_data[index]
if label_value_after_argmax==prediction_value_after_argmax:
continue
print("[NG]",categories_list[prediction_value_after_argmax],"!=",categories_list[label_value_after_argmax])
print(prediction_value)
filename_string\
="image/error/"+str(index)+"-"+categories_list[prediction_value_after_argmax]\
+"-ne-"+categories_list[label_value_after_argmax]+".png"
one_x_feature_test_data*=256
opened_file=Image.fromarray(np.uint8(one_x_feature_test_data))
opened_file.save(filename_string)
score=sequential_object.evaluate(x_test_data,y_test_data)
print('loss=',score[0])
print('accuracy=',score[1])
# @
# People who create new models like cnn, rnn are not ordinal developers,
# but researchers in constitute
# Researchers should know every aspect of model in detail
# However, data scientists and ordinal developers,
# who apply these models would be better to focus on applying models,
# to make benefits rather than creating new models
# To apply model, most recommended way is to read related thesis papers,
# which are written by researchers
]]>