新手。。请多多指教~

数据集rsna-bone-age来自kaggle
云盘链接 提取码: n24q
因为上传限制，训练集分成了三个压缩包。解压后照着boneage-test-dataset，把图片放在boneage-training-dataset文件夹里。
讯享网

数据集解压后，boneage-training-dataset文件夹中是：

boneage-training-dataset.csv中是：

第一步，读取数据集

1

import os import pandas as pd print("Reading data...") base_data_dir = ’/input/rsna-bone-age‘ img_dir = os.path.join(base_data_dir , 'boneage-training-dataset/') csv_dir = os.path.join(base_data_dir , 'boneage-training-dataset.csv') boneage_df = pd.read_csv(csv_dir) print (boneage_df) column_headers = list(boneage_df.columns.values) print(column_headers) #列标签 print(boneage_df.sample(3)) #csv中随机抽取3行 

输出：

讯享网(12611, 3) ['id', 'boneage', 'male'] #列标签 id boneage male 11314 11314 96 True 11583 11583 228 True 4583 4583 94 False 

参考：
1.pandas文档 —— pandas.read_csv
2.python pandas读取csv后，获取列标签
——————————————————————————————————

2

获得需要的标签和数据，比如性别（male或female），img文件夹中图片的路径等

boneage_df['path'] = boneage_df['id'].map(lambda x: img_dir+"{}.png".format(x)) boneage_df['gender'] = boneage_df['male'].map(lambda x: "male" if x else "female") boneage_df['exists'] = boneage_df['path'].map(os.path.exists) #判断img_dir中的图片和csv中的数据是否对应 #标准化骨龄数据 mu = boneage_df['boneage'].mean() sigma = boneage_df['boneage'].std() boneage_df['zscore'] = boneage_df['boneage'].map(lambda x: (x-mu)/sigma) boneage_df.dropna(inplace=True) column_headers = list(boneage_df.columns.values) print("column_hander = ",column_headers) #列标签 #print(boneage_df.sample(3)) #csv中随机抽取3行 print("{} images found out of total {} images".format(boneage_df['exists'].sum(),boneage_df.shape[0])) print("Reading complete !!!\n") boneage_df [['boneage','zscore']].hist() 

输出：

讯享网column_hander = ['id', 'boneage', 'male', 'path', 'exists', 'gender', 'zscore'] 12611 images found out of total 12611 images Reading complete !!! 

横轴是骨龄（按月计算），纵轴是样本数量，标准化后，样本的骨龄数据转化为-3~3的分值

参考：
3.pandas中强大的绘制图表功能——DataFrame.hist()
4.3.4.2数据标准化（一） - Z-Score标准化

第二步，准备数据集（训练集/测试集/验证集）

1 划分数据集

from sklearn.model_selection import train_test_split print("Preparing training, testing and validation datasets ...") #按骨龄的梯度分为10个量级 boneage_df['boneage_category'] = pd.cut(boneage_df['boneage'], 10) #划分数据集 raw_train_df, test_df = train_test_split(boneage_df, test_size = 0.2, random_state = 2018,stratify = boneage_df['boneage_category']) raw_train_df, valid_df = train_test_split(raw_train_df,test_size = 0.25,random_state = 2018,stratify = raw_train_df['boneage_category']) raw_train_df[['boneage']].hist(figsize = (10, 5)) #绘图（均衡前） # Training images: 7566 | Validation images: 2522 | Test images: 2523 #Balance the distribution in the training set train_df = raw_train_df.groupby(['boneage_category', 'male']).apply(lambda x: x.sample(500, replace = True)).reset_index(drop=True) train_df[['boneage']].hist(figsize = (10, 5)) #绘图 (均衡后) # Training images: 10000 | Validation images: 2522 | Test images: 2523 train_size = train_df.shape[0] valid_size = valid_df.shape[0] test_size = test_df.shape[0] print("# Training images: {}".format(train_size)) print("# Validation images: {}".format(valid_size)) print("# Test images: {}".format(test_size)) 

输出：

讯享网# Training images: 10000 # Validation images: 2522 # Test images: 2523 

注：

数据均衡过程：

train_df = raw_train_df.groupby(['boneage_category', 'male']).apply(lambda x: x.sample(500, replace = True)).reset_index(drop=True) 

参考：
5.pandas文档 ——pandas.cut
6. .sample method
7.Sample Pandas dataframe based on values in column

2 数据预处理

讯享网from keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img from keras.applications.inception_v3 import InceptionV3, preprocess_input import tensorflow as tf import numpy as np IMG_SIZE = (224,224) BATCH_SIZE_TRAIN = 10 BATCH_SIZE_VAL = 16 SEED = 1234 # 将图像和年龄作为输入，构建数据生成器 def gen_2inputs(imgDatGen, df, batch_size, seed, img_size): gen_img = imgDatGen.flow_from_dataframe(dataframe=df, x_col='path', y_col='zscore', batch_size=batch_size, seed=seed, shuffle=True, class_mode='other', target_size=img_size, color_mode='rgb', drop_duplicates=False) gen_gender = imgDatGen.flow_from_dataframe(dataframe=df, x_col='path', y_col='gender', batch_size=batch_size, seed=seed, shuffle=True, class_mode='other', target_size=img_size, color_mode='rgb', drop_duplicates=False) while True: X1i = gen_img.next() X2i = gen_gender.next() yield [X1i[0], X2i[1]], X1i[1] def test_gen_2inputs(imgDatGen, df, batch_size, img_size): gen_img = imgDatGen.flow_from_dataframe(dataframe=df, x_col='path', y_col='zscore', batch_size=batch_size, shuffle=False, class_mode='other', target_size=img_size, color_mode='rgb', drop_duplicates=False) gen_gender = imgDatGen.flow_from_dataframe(dataframe=df, x_col='path', y_col='gender', batch_size=batch_size, shuffle=False, class_mode='other', target_size=img_size, color_mode='rgb', drop_duplicates=False) while True: X1i = gen_img.next() X2i = gen_gender.next() yield [X1i[0], X2i[1]], X1i[1] 

train_idg = ImageDataGenerator(zoom_range=0.2, fill_mode='nearest', rotation_range=25, width_shift_range=0.25, height_shift_range=0.25, vertical_flip=False, horizontal_flip=True, shear_range = 0.2, samplewise_center=False, samplewise_std_normalization=False) val_idg = ImageDataGenerator(width_shift_range=0.25, height_shift_range=0.25, horizontal_flip=True) test_idg = ImageDataGenerator() train_flow = gen_2inputs(train_idg, train_df, BATCH_SIZE_TRAIN, SEED, IMG_SIZE) valid_flow = gen_2inputs(val_idg, valid_df, BATCH_SIZE_VAL, SEED, IMG_SIZE) test_flow = test_gen_2inputs(test_idg, test_df, 500, IMG_SIZE) # 计算月份平均绝对误差 def mae_months(in_gt, in_pred): return mean_absolute_error(boneage_div * in_gt, boneage_div * in_pred) 

参考：
8.Tutorial on Keras flow_from_dataframe
9.师兄的代码 😄

第三步，构建神经网络

讯享网from keras.utils import plot_model from keras.layers import Input, GlobalAveragePooling2D, Dense, Dropout, Flatten, Concatenate from keras.models import Sequential,Model from keras.metrics import mean_absolute_error from keras.applications.inception_v3 import InceptionV3, preprocess_input print("Compiling deep model ...") IMG_SHAPE = (224, 224,3) # 1、两个输入分别是原始图像和性别输入 img = Input(shape = IMG_SHAPE) gender = Input(shape=(1,)) #2、预训练主干网络（图像） cnn_vec = InceptionV3(input_shape = IMG_SHAPE, include_top = False, weights = 'imagenet')(img) #3、主干网络输出 cnn_vec = GlobalAveragePooling2D()(cnn_vec) cnn_vec = Dropout(0.2)(cnn_vec) #4.性别输入网络 gender_vec = Dense(32,activation='relu')(gender) #5.两个网络拼接 features = Concatenate(axis=-1)([cnn_vec,gender_vec]) dense_layer = Dense(1024, activation = 'relu')(features) dense_layer = Dropout(0.2)(dense_layer) dense_layer = Dense(1024,activation='relu')(dense_layer) dense_layer = Dropout(0.2)(dense_layer) output_layer = Dense(1, activation = 'linear')(dense_layer) # linear is what 16bit did bone_age_model = Model(inputs=[img,gender],outputs=output_layer) def mae_months(in_gt, in_pred): return mean_absolute_error(mu+sigma*in_gt, mu+sigma*in_pred) bone_age_model.compile(optimizer = 'adam', loss = 'mse', metrics = [mae_months]) bone_age_model.summary() print("Model compiled !!!\n") 

第四步，训练模型

from keras.callbacks import ModelCheckpoint, EarlyStopping, ReduceLROnPlateau from sklearn.metrics import mean_absolute_error as sk_mae import matplotlib.pyplot as plt print("Training deep model ...") # 测试步长 EPOCHS = 10 BATCH_SIZE_TEST = len(test_df) // 3 STEP_SIZE_TEST = 3 STEP_SIZE_TRAIN = len(train_df) // BATCH_SIZE_TRAIN STEP_SIZE_VALID = len(valid_df) // BATCH_SIZE_VAL weight_path="bone_age_weights_trainable_inceptionv3_gender_50_epochs_relu_less_dropout_dense.best.hdf5" # saved_model_name checkpoint = ModelCheckpoint(weight_path, monitor='val_loss', verbose=1,save_best_only=True, mode='min', save_weights_only = True) reduceLROnPlat = ReduceLROnPlateau(monitor='val_loss', factor=0.8, patience=10, verbose=1, mode='auto', epsilon=0.0001, cooldown=5, min_lr=0.0001) early = EarlyStopping(monitor="val_loss", mode="min", patience=10) # probably needs to be more patient, but kaggle time is limited callbacks_list = [checkpoint, early, reduceLROnPlat] if not os.path.exists(weight_path): model_history = bone_age_model.fit_generator(generator=train_flow, steps_per_epoch=STEP_SIZE_TRAIN, validation_data=valid_flow, validation_steps=STEP_SIZE_VALID, epochs=EPOCHS, callbacks = callbacks_list) bone_age_model.load_weights(weight_path) print("Training complete !!!\n") # 保存训练的loss loss_history = model_history.history['loss'] history_df = pd.DataFrame.from_dict(model_history.history) history_df.to_csv('loss_history.csv') #Evaluate model on test dataset print("Evaluating model on test data ...\n") print("Preparing testing dataset...") test_X, test_Y = next(test_flow) # one big batch print("Data prepared !!!") pred_Y = mu+sigma*bone_age_model.predict(x=test_X,batch_size=25,verbose=1) test_Y_months = mu+sigma*test_Y print("Mean absolute error on test data: "+str(sk_mae(test_Y_months,pred_Y))) fig, ax1 = plt.subplots(1,1, figsize = (6,6)) ax1.plot(test_Y_months, pred_Y, 'r.', label = 'predictions') ax1.plot(test_Y_months, test_Y_months, 'b-', label = 'actual') ax1.legend() ax1.set_xlabel('Actual Age (Months)') ax1.set_ylabel('Predicted Age (Months)') ord_idx = np.argsort(test_Y) ord_idx = ord_idx[np.linspace(0, len(ord_idx)-1, num=8).astype(int)] # take 8 evenly spaced ones fig, m_axs = plt.subplots(2, 4, figsize = (16, 32)) for (idx, c_ax) in zip(ord_idx, m_axs.flatten()): c_ax.imshow(test_X[0][idx, :,:,0], cmap = 'bone') title = 'Age: %2.1f\nPredicted Age: %2.1f\nGender: ' % (test_Y_months[idx], pred_Y[idx]) if test_X[1][idx]==0: title+="Female\n" else: title+="Male\n" c_ax.set_title(title) c_ax.axis('off') plt.show()