[KaggleStudy] Keras U-Net starter - LB 0.277

Notebook

• Kaggle
  Keras U-Net starter - LB 0.277

• transcription
  Keras U-Net starter - LB 0.277.ipynb

1. Intro

• U-Net 사용
  • 작은 데이터셋에서도 비교적 잘 작동함

import os
import sys
import random
import warnings

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

from tqdm import tqdm
from itertools import chain
from skimage.io import imread, imshow, imread_collection, concatenate_images
from skimage.transform import resize
from skimage.morphology import label

from tensorflow.keras.models import Model, load_model
from tensorflow.keras.layers import Input
from tensorflow.keras.layers import Dropout, Lambda
from tensorflow.keras.layers import Conv2D, Conv2DTranspose
from tensorflow.keras.layers import MaxPooling2D
from tensorflow.keras.layers import concatenate
from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint
from tensorflow.keras import backend as K

import tensorflow as tf

# Set some parameters
IMG_WIDTH = 128
IMG_HEIGHT = 128
IMG_CHANNELS = 3
TRAIN_PATH = '/content/stage1_train/'
TEST_PATH = '/content/stage1_test/'

warnings.filterwarnings('ignore', category=UserWarning, module='skimage')
seed = 42
random.seed = seed
np.random.seed = seed

# Get train and test IDs
train_ids = next(os.walk(TRAIN_PATH))[1]
test_ids = next(os.walk(TEST_PATH))[1]

2. Get the data

• 이미지 관련 마스크 모두 가져오기
  • train image, test image 다운 샘플링(예측 마스크는 업샘플링)
  • test image 원본 크기 기록

# Get and resize train images and masks
X_train = np.zeros((len(train_ids), IMG_HEIGHT, IMG_WIDTH, IMG_CHANNELS), dtype=np.uint8)
Y_train = np.zeros((len(train_ids), IMG_HEIGHT, IMG_WIDTH, 1), dtype=bool)
print('Getting and resizing train images and masks ... ')
sys.stdout.flush()
for n, id_ in tqdm(enumerate(train_ids), total=len(train_ids)):
    path = TRAIN_PATH + id_
    img = imread(path + '/images/' + id_ + '.png')[:,:,:IMG_CHANNELS]
    img = resize(img, (IMG_HEIGHT, IMG_WIDTH), mode='constant', preserve_range=True)
    X_train[n] = img
    mask = np.zeros((IMG_HEIGHT, IMG_WIDTH, 1), dtype=bool)
    for mask_file in next(os.walk(path + '/masks/'))[2]:
        mask_ = imread(path + '/masks/' + mask_file)
        mask_ = np.expand_dims(resize(mask_, (IMG_HEIGHT, IMG_WIDTH), mode='constant', 
                                      preserve_range=True), axis=-1)
        mask = np.maximum(mask, mask_)
    Y_train[n] = mask

# Get and resize test images
X_test = np.zeros((len(test_ids), IMG_HEIGHT, IMG_WIDTH, IMG_CHANNELS), dtype=np.uint8)
sizes_test = []
print('Getting and resizing test images ... ')
sys.stdout.flush()
for n, id_ in tqdm(enumerate(test_ids), total=len(test_ids)):
    path = TEST_PATH + id_
    img = imread(path + '/images/' + id_ + '.png')[:,:,:IMG_CHANNELS]
    sizes_test.append([img.shape[0], img.shape[1]])
    img = resize(img, (IMG_HEIGHT, IMG_WIDTH), mode='constant', preserve_range=True)
    X_test[n] = img

print('Done!')

• 랜덤으로 image 관련 mask 확인

# Check if training data looks all right
ix = random.randint(0, len(train_ids))
imshow(X_train[ix])
plt.show()
imshow(np.squeeze(Y_train[ix]))
plt.show()

3. Create our Keras metric

• Keras 이용

  • 다양한 교차점에서의 평균 정밀도 정의
  • TenserFlow : 평균 IoU 메트릭
    • 여러 임계값 평균 기본 지원하지 않음 -> 구현해야 함

• 다만, 결과와 LB 결과간 불일치가 크며, 정확하지 않은 것으로 알려짐

# Define IoU metric
miou = tf.keras.metrics.MeanIoU(num_classes=2)

def mean_iou(y_true, y_pred):
    miou.reset_state()
    prec = []
    for t in np.arange(0.5, 1.0, 0.05):
        y_pred_ = tf.cast(y_pred > t, tf.int32)
        miou.update_state(y_true, y_pred_)
        score = miou.result()
        prec.append(score)
    return K.mean(K.stack(prec), axis=0)

4. Build and train our neural network

• U-Net 모델 구축
  • Convolutional Networks for Biomedical Image Segmentation
    • 생의학 이미지 분할을 위한 합성곱 신경망 기반
    • 느슨하게 설정
    • refer: https://github.com/jocicmarko/ultrasound-nerve-segmentation

# Build U-Net model
inputs = Input((IMG_HEIGHT, IMG_WIDTH, IMG_CHANNELS))
s = Lambda(lambda x: x / 255) (inputs)

c1 = Conv2D(16, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (s)
c1 = Dropout(0.1) (c1)
c1 = Conv2D(16, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (c1)
p1 = MaxPooling2D((2, 2)) (c1)

c2 = Conv2D(32, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (p1)
c2 = Dropout(0.1) (c2)
c2 = Conv2D(32, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (c2)
p2 = MaxPooling2D((2, 2)) (c2)

c3 = Conv2D(64, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (p2)
c3 = Dropout(0.2) (c3)
c3 = Conv2D(64, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (c3)
p3 = MaxPooling2D((2, 2)) (c3)

c4 = Conv2D(128, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (p3)
c4 = Dropout(0.2) (c4)
c4 = Conv2D(128, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (c4)
p4 = MaxPooling2D(pool_size=(2, 2)) (c4)

c5 = Conv2D(256, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (p4)
c5 = Dropout(0.3) (c5)
c5 = Conv2D(256, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (c5)

u6 = Conv2DTranspose(128, (2, 2), strides=(2, 2), padding='same') (c5)
u6 = concatenate([u6, c4])
c6 = Conv2D(128, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (u6)
c6 = Dropout(0.2) (c6)
c6 = Conv2D(128, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (c6)

u7 = Conv2DTranspose(64, (2, 2), strides=(2, 2), padding='same') (c6)
u7 = concatenate([u7, c3])
c7 = Conv2D(64, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (u7)
c7 = Dropout(0.2) (c7)
c7 = Conv2D(64, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (c7)

u8 = Conv2DTranspose(32, (2, 2), strides=(2, 2), padding='same') (c7)
u8 = concatenate([u8, c2])
c8 = Conv2D(32, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (u8)
c8 = Dropout(0.1) (c8)
c8 = Conv2D(32, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (c8)

u9 = Conv2DTranspose(16, (2, 2), strides=(2, 2), padding='same') (c8)
u9 = concatenate([u9, c1], axis=3)
c9 = Conv2D(16, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (u9)
c9 = Dropout(0.1) (c9)
c9 = Conv2D(16, (3, 3), activation='elu', kernel_initializer='he_normal', padding='same') (c9)

outputs = Conv2D(1, (1, 1), activation='sigmoid') (c9)

model = Model(inputs=[inputs], outputs=[outputs])
model.compile(optimizer='adam', loss='binary_crossentropy', metrics=[mean_iou])
model.summary()

약 1시간 이상 소요

• Update
  • ELU
  • validation split : 0.1
  • 데이터 양이 작아 -> 배치 크기도 작게 설정
  • checkpoint, early stopping 사용
  • epochs를 50으로 증가

# Fit model
earlystopper = EarlyStopping(patience=5, verbose=1)
checkpointer = ModelCheckpoint('model-dsbowl2018-1.h5.keras', verbose=1, save_best_only=True)
results = model.fit(X_train, Y_train, validation_split=0.1, batch_size=16, epochs=50, 
                    callbacks=[earlystopper, checkpointer])

5. Make predictions

• test set, train set, val set 모두에서의 예측 실행

# Predict on train, val and test
model = load_model('model-dsbowl2018-1.h5.keras', custom_objects={'mean_iou': mean_iou}, safe_mode=False)
preds_train = model.predict(X_train[:int(X_train.shape[0]*0.9)], verbose=1)
preds_val = model.predict(X_train[int(X_train.shape[0]*0.9):], verbose=1)
preds_test = model.predict(X_test, verbose=1)

# Threshold predictions
preds_train_t = (preds_train > 0.5).astype(np.uint8)
preds_val_t = (preds_val > 0.5).astype(np.uint8)
preds_test_t = (preds_test > 0.5).astype(np.uint8)

# Create list of upsampled test masks
preds_test_upsampled = []
for i in range(len(preds_test)):
    preds_test_upsampled.append(resize(np.squeeze(preds_test[i]), 
                                       (sizes_test[i][0], sizes_test[i][1]), 
                                       mode='constant', preserve_range=True))

• train 데이터에서의 적합

# Perfrom a sanity check on some random training samples
ix = random.randint(0, len(preds_train_t))
imshow(X_train[ix])
plt.show()

imshow(np.squeeze(Y_train[ix]))
plt.show()

imshow(np.squeeze(preds_train_t[ix]))
plt.show()

• validate 데이터에서의 적합

# Perform a sanity check on some random validation samples
ix = random.randint(0, len(preds_val_t))
imshow(X_train[int(X_train.shape[0]*0.9):][ix])
plt.show()

imshow(np.squeeze(Y_train[int(Y_train.shape[0]*0.9):][ix]))
plt.show()

imshow(np.squeeze(preds_val_t[ix]))
plt.show()

6. Encode and submit our results

• Fast Run-Length Encoding (Python) 사용

# Run-length encoding stolen from https://www.kaggle.com/rakhlin/fast-run-length-encoding-python
def rle_encoding(x):
    dots = np.where(x.T.flatten() == 1)[0]
    run_lengths = []
    prev = -2
    for b in dots:
        if (b>prev+1): run_lengths.extend((b + 1, 0))
        run_lengths[-1] += 1
        prev = b
    return run_lengths

def prob_to_rles(x, cutoff=0.5):
    lab_img = label(x > cutoff)
    for i in range(1, lab_img.max() + 1):
        yield rle_encoding(lab_img == i)

• test ID 반복으로 skimage로 식별된 각 개별 마스크 실행 길이 인코딩

new_test_ids = []
rles = []
for n, id_ in enumerate(test_ids):
    rle = list(prob_to_rles(preds_test_upsampled[n]))
    rles.extend(rle)
    new_test_ids.extend([id_] * len(rle))

• submission

# Create submission DataFrame
sub = pd.DataFrame()
sub['ImageId'] = new_test_ids
sub['EncodedPixels'] = pd.Series(rles).apply(lambda x: ' '.join(str(y) for y in x))
sub.to_csv('sub-dsbowl2018-1.csv', index=False)

최종 LB 점수

• (예시 노트북) 0.233점
• 결과가 매우 불규칙(실행 시마다 크게 달라질 수 있음)
• 매개변수 변경, 아키텍쳐 조절을 통해 결과 안정화 & 개선 필요

🚀 troubleshooting

NumPy np.bool 미지원

• NumPy 1.20 버전부터 np.bool이 더 이상 지원되지 않음
  

• np.bool을 bool로 수정

...dtype=bool) # np.bool에서 bool로 수정

TensorFlow 2.x ➡️ tf.to_int32, tf.metrics.mean_iou 미지원

• tf.to_int32 -> tf.cast로 대체
• tf.metrics.mean_iou -> tf.keras.metrics.MeanIoU로 대체

# Before
def mean_iou(y_true, y_pred):
    prec = []
    for t in np.arange(0.5, 1.0, 0.05):
        y_pred_ = tf.to_int32(y_pred > t)
        score, up_opt = tf.metrics.mean_iou(y_true, y_pred_, 2)
        K.get_session().run(tf.local_variables_initializer())
        with tf.control_dependencies([up_opt]):
            score = tf.identity(score)
        prec.append(score)
    return K.mean(K.stack(prec), axis=0)

• 변경 사항
  • MeanIoU를 함수 외부에서 한 번만 초기화하도록 분리
    • miou.reset_state()를 함수 내에서 호출 -> MeanIoU 상태 초기화
    • update_state : 갱신

# After
miou = tf.keras.metrics.MeanIoU(num_classes=2)

def mean_iou(y_true, y_pred):
    miou.reset_state()
    prec = []
    for t in np.arange(0.5, 1.0, 0.05):
        y_pred_ = tf.cast(y_pred > t, tf.int32)
        miou.update_state(y_true, y_pred_)
        score = miou.result()
        prec.append(score)
    return K.mean(K.stack(prec), axis=0)

ValueError: Lambda 레이어 역직렬화 오류

ValueError: The {arg_name} of this Lambda layer is a Python lambda. Deserializing it is unsafe. If you trust the source of the config artifact, you can override this error by passing safe_mode=False to from_config(), or calling keras.config.enable_unsafe_deserialization().

• safe_mode=False로 설정 해제