A10_project

import os
import time
import functools
import pandas as pd
import tensorflow as tf
import matplotlib.pyplot as plt

from tensorflow.keras.preprocessing.image import ImageDataGenerator
from sklearn.model_selection import train_test_split
from transformers import TFViTModel

DATA_PATH = "/content/drive/MyDrive/dataset/Animals_10"

dataset_dir_path = os.path.join(DATA_PATH, "raw-img")
print(dataset_dir_path)

# /content/drive/MyDrive/dataset/Animals_10/raw-img

get_dir_list = os.listdir(dataset_dir_path)
print(get_dir_list)

# ['mucca', 'farfalla', 'gallina', 'cavallo', 'pecora', 'gatto', 'cane', 'scoiattolo', 'ragno', 'elefante']

translate = {"cane": "dog", "cavallo": "horse", "elefante": "elephant", "farfalla": "butterfly", "gallina": "chicken", "gatto": "cat", "mucca": "cow", "pecora": "sheep", "ragno": "spider", "scoiattolo": "squirrel", "dog": "cane", "cavallo": "horse", "elephant" : "elefante", "butterfly": "farfalla", "chicken": "gallina", "cat": "gatto", "cow": "mucca", "spider": "ragno", "squirrel": "scoiattolo"}

get_idx_lable = {}
get_label_idx = []
for idx, lable in enumerate(get_dir_list):
    get_idx_lable[translate[lable]] = idx
    get_label_idx.append(translate[lable])
print(get_idx_lable)
print(get_label_idx)

#{'cow': 0, 'butterfly': 1, 'chicken': 2, 'horse': 3, 'sheep': 4, 'cat': 5, 'dog': 6, 'squirrel': 7, 'spider': 8, 'elephant': 9}
#['cow', 'butterfly', 'chicken', 'horse', 'sheep', 'cat', 'dog', 'squirrel', 'spider', 'elephant']

get_all_datalist, get_all_label = [], []
for get_dir in get_dir_list:
    get_dir_path = os.path.join(dataset_dir_path, get_dir)
    get_file_list = os.listdir(get_dir_path)
    for get_file in get_file_list:
        get_file_path = os.path.join(get_dir_path, get_file)
        get_all_datalist.append(get_file_path)
        translate_lable = translate[get_file_path.split("/")[-2]]
        get_all_label.append(translate_lable)

print(get_all_datalist)
print(set(get_all_label))

#['/content/drive/MyDrive/dataset/Animals_10/raw-img/mucca/OIP-6XBZIAF5HUK7s1CfQZmR2gHaES.jpeg', 
#{'butterfly', 'dog', 'squirrel', 'cat', 'sheep', 'elephant', 'horse', 'chicken', 'cow', 'spider'}

train_data, test_data, train_label, test_label = train_test_split(get_all_datalist, get_all_label, test_size=0.2, random_state=42)

print(len(train_data), len(train_label))
print(len(test_data), len(test_label))

#21055 21055
#5264 5264

train_df = pd.DataFrame({
    'data': train_data,
    'label': train_label
})

test_df = pd.DataFrame({
    'data': test_data,
    'label': test_label
})

print(train_df.head())
print(test_df.head())

                                                data      label
0  /content/drive/MyDrive/dataset/Animals_10/raw-...        dog
1  /content/drive/MyDrive/dataset/Animals_10/raw-...      sheep
2  /content/drive/MyDrive/dataset/Animals_10/raw-...    chicken
3  /content/drive/MyDrive/dataset/Animals_10/raw-...      sheep
4  /content/drive/MyDrive/dataset/Animals_10/raw-...  butterfly
                                                data  label
0  /content/drive/MyDrive/dataset/Animals_10/raw-...  horse
1  /content/drive/MyDrive/dataset/Animals_10/raw-...    dog
2  /content/drive/MyDrive/dataset/Animals_10/raw-...  sheep
3  /content/drive/MyDrive/dataset/Animals_10/raw-...  horse
4  /content/drive/MyDrive/dataset/Animals_10/raw-...    cow

batch_size = 128
image_size = 224

train_idg = ImageDataGenerator(rescale=1./255,
                               rotation_range=0.1,
                               width_shift_range=0.1,
                               height_shift_range=0.1,
                               zoom_range=0.1,
                               horizontal_flip=True,
                               vertical_flip=False)

test_idg = ImageDataGenerator(rescale=1./255)

train_generator = train_idg.flow_from_dataframe(dataframe=train_df,
                                                directory=None,
                                                x_col="data",
                                                y_col="label",
                                                class_mode="categorical",
                                                target_size=(224,224),
                                                batch_size=batch_size)

test_generator = test_idg.flow_from_dataframe(dataframe=test_df,
                                              directory=None,
                                              x_col="data",
                                              y_col="label",
                                              class_mode="categorical",
                                              target_size=(224,224),
                                              batch_size=batch_size)

#Found 21055 validated image filenames belonging to 10 classes.
#Found 5264 validated image filenames belonging to 10 classes.

# 데이터셋에 저장된 이미지를 출력하는 함수
# 이미지 확인을 위해 사용한다.
def plotImages(images_arr):
    fig, axes = plt.subplots(1, 5, figsize=(20,20))
    axes = axes.flatten()
    for img, ax in zip( images_arr, axes):
        ax.imshow(img)
        ax.axis('off')
    plt.tight_layout()
    plt.show()

augmented_images = [train_generator[0][0][0] for i in range(5)]
plotImages(augmented_images)

base_model = TFViTModel.from_pretrained('google/vit-base-patch16-224-in21k')

pixel_values = tf.keras.layers.Input(shape=(224, 224, 3), dtype='float32')
permute = tf.keras.layers.Permute((3, 1, 2))(pixel_values)

# model layer
vit = base_model.vit(permute)[0]
classifier = tf.keras.layers.Dense(10, name='outputs')(vit[:, 0, :])

# model
model = tf.keras.Model(inputs=pixel_values, outputs=classifier)

# define loss
loss=tf.keras.losses.CategoricalCrossentropy(from_logits=True)

# define metrics
metrics=tf.keras.metrics.CategoricalAccuracy(name="accuracy")

# compile model
model.compile(optimizer=tf.keras.optimizers.Adam(1e-4),
              loss=loss,
              metrics=metrics)

early_stopping_cb = tf.keras.callbacks.EarlyStopping(patience=5,
                                                     monitor='val_loss',
                                                     restore_best_weights=True,
                                                     verbose=1)

train_results = model.fit(
    train_generator,
    steps_per_epoch=len(train_data) // batch_size,
    validation_data=test_generator,
    validation_steps=len(test_data) // batch_size,
    epochs=100,
    callbacks=[early_stopping_cb]
)

Epoch 1/100
WARNING:tensorflow:Gradients do not exist for variables ['tf_vi_t_model/vit/pooler/dense/kernel:0', 'tf_vi_t_model/vit/pooler/dense/bias:0'] when minimizing the loss. If you're using `model.compile()`, did you forget to provide a `loss` argument?
WARNING:tensorflow:Gradients do not exist for variables ['tf_vi_t_model/vit/pooler/dense/kernel:0', 'tf_vi_t_model/vit/pooler/dense/bias:0'] when minimizing the loss. If you're using `model.compile()`, did you forget to provide a `loss` argument?
WARNING:tensorflow:Gradients do not exist for variables ['tf_vi_t_model/vit/pooler/dense/kernel:0', 'tf_vi_t_model/vit/pooler/dense/bias:0'] when minimizing the loss. If you're using `model.compile()`, did you forget to provide a `loss` argument?
WARNING:tensorflow:Gradients do not exist for variables ['tf_vi_t_model/vit/pooler/dense/kernel:0', 'tf_vi_t_model/vit/pooler/dense/bias:0'] when minimizing the loss. If you're using `model.compile()`, did you forget to provide a `loss` argument?
164/164 [==============================] - 20285s 124s/step - loss: 0.2380 - accuracy: 0.9503 - val_loss: 0.0845 - val_accuracy: 0.9779
Epoch 2/100
164/164 [==============================] - 299s 2s/step - loss: 0.0510 - accuracy: 0.9875 - val_loss: 0.0584 - val_accuracy: 0.9849
Epoch 3/100
164/164 [==============================] - 300s 2s/step - loss: 0.0345 - accuracy: 0.9916 - val_loss: 0.0535 - val_accuracy: 0.9844
Epoch 4/100
164/164 [==============================] - 301s 2s/step - loss: 0.0289 - accuracy: 0.9930 - val_loss: 0.0490 - val_accuracy: 0.9878
Epoch 5/100
164/164 [==============================] - 303s 2s/step - loss: 0.0202 - accuracy: 0.9953 - val_loss: 0.0830 - val_accuracy: 0.9783
Epoch 6/100
164/164 [==============================] - 305s 2s/step - loss: 0.0221 - accuracy: 0.9941 - val_loss: 0.0582 - val_accuracy: 0.9836
Epoch 7/100
164/164 [==============================] - 305s 2s/step - loss: 0.0186 - accuracy: 0.9948 - val_loss: 0.0641 - val_accuracy: 0.9846
Epoch 8/100
164/164 [==============================] - 308s 2s/step - loss: 0.0198 - accuracy: 0.9942 - val_loss: 0.0623 - val_accuracy: 0.9844
Epoch 9/100
164/164 [==============================] - ETA: 0s - loss: 0.0149 - accuracy: 0.9955Restoring model weights from the end of the best epoch: 4.
164/164 [==============================] - 309s 2s/step - loss: 0.0149 - accuracy: 0.9955 - val_loss: 0.0733 - val_accuracy: 0.9817
Epoch 9: early stopping