resnet_50.py

#!/usr/bin/env python
# coding: utf-8

# In[1]:


import pandas as pd
import numpy as np
import glob
 
import os
import zipfile

import tensorflow as tf

from keras_preprocessing.image import ImageDataGenerator
from sklearn.model_selection import train_test_split

from tensorflow.keras.models import Model, Sequential, load_model
from tensorflow.keras.layers import Dense, Dropout, BatchNormalization, Input
from tensorflow.keras.optimizers import Adam

from tensorflow.keras.models import load_model


# In[2]:


# 다운로드 받은 이미지 압축 파일 해제
# print(os.listdir("./drive/MyDrive"))
local_zip = './train.zip'

zip_ref = zipfile.ZipFile(local_zip, 'r')
zip_ref.extractall('./res/train')
zip_ref.close()

local_zip = './test.zip'

zip_ref = zipfile.ZipFile(local_zip, 'r')
zip_ref.extractall('./res/test')
zip_ref.close()


# In[3]:


base_dir = "./res/"

train_dir = "train/train/"

train_class = ['dog', 'elephant', 'giraffe', 'guitar', 'horse', 'house', 'person']


test_dir = "test/test/0"


# In[4]:


train_one_hot = []

for path in range(len(train_class)):
    class2label = [''] + [0] * len(train_class)
    for img in os.listdir(base_dir + train_dir + train_class[path]):
        class2label[0] = (train_dir + train_class[path] + "/" + img)[6:]
        class2label[path + 1] = 1
        train_one_hot.append(class2label)

        
train_answer = pd.DataFrame(train_one_hot, columns = ["path", "dogs", 'elephant', 'giraffe', 'guitar', 'horse' , 'house', 'person'])
train_answer.to_csv('./train_answer.csv', index=False)

data = pd.read_csv("./train_answer.csv")
columns = data.columns

datagen=ImageDataGenerator(
      rescale = 1/255,
      rotation_range=40,
      width_shift_range=0.2,
      height_shift_range=0.2,
      shear_range=0.2,
      zoom_range=0.2,
      horizontal_flip=True,
      fill_mode='nearest')
train_generator=datagen.flow_from_dataframe(
                                            dataframe=data,
                                            directory='./res/train/',
                                            x_col="path",
                                            y_col=columns[1:],
                                            batch_size=48,
                                            shuffle=False,
                                            class_mode="raw",
                                            target_size=(224,224))


# In[5]:


res_net = tf.keras.applications.ResNet50V2(
    include_top=True, weights='imagenet',input_shape=(224,224,3), pooling="max")
res_net.trainable = True

pretrained_data = res_net.predict(train_generator, verbose=1)

x_train, x_valid, y_train, y_valid = train_test_split(pretrained_data, data.iloc[:,1:], test_size=0.2, random_state=42)


# In[ ]:


x_train, x_valid, y_train, y_valid = train_test_split(pretrained_data, data.iloc[:,1:], test_size=0.2, random_state=42)
x_train


# In[6]:


model = Sequential(
    [Input(1000,),
    BatchNormalization(),
    
    Dense(512,  activation='relu'),
#     BatchNormalization(),
    
    Dense(256,  activation='relu'),
#     BatchNormalization(),
    
    Dense(128,  activation='relu'),
#     BatchNormalization(),
    
    Dense(7,  activation='softmax')
    ])
model.summary()
 
mc = tf.keras.callbacks.ModelCheckpoint('res_model.h5', monitor='val_loss', mode='min', verbose=1, save_best_only=True)


# In[7]:


model.compile(loss='categorical_crossentropy', optimizer=Adam(0.001), metrics=['accuracy']) # 최적화 함수 학습률 1e-4에서 0.001로 변경
model.fit(x_train, y_train, epochs=30, batch_size=48, validation_data=(x_valid, y_valid),callbacks=[mc])


# In[8]:


loaded_model = load_model('res_model.h5')
 
final_model = Sequential([res_net, model])
final_model.summary()


# In[9]:


test_df = pd.DataFrame()
test = glob.glob("./res/test/test/0/*.jpg")
test_dir = []
for img in test:
    test_dir.append(img[11:])
test_df["path"] = test_dir
 
test_datagen=ImageDataGenerator(rescale = 1/255)
test_generator = test_datagen.flow_from_dataframe(  dataframe=test_df[:],
                                                    directory='./res/test/',
                                                    x_col="path",
                                                    y_col=columns[0],
                                                    batch_size=48,
                                                    shuffle=False,
                                                    class_mode="raw",
                                                    target_size=(224,224))


# In[10]:


pred = final_model.predict(test_generator, verbose=1)


# In[11]:


answer = np.array([y.argmax() for y in pred])
test_df = pd.read_csv("./test_answer_sample_.csv")
test_df.iloc[:,1] = answer
test_df.to_csv('res_net_50v2.csv', index=False)