# Using CNN to Predict Bone age ###### tags: `python` `tensorflow` `NN` `Keras` ## VGG16 ### Import library ```python # Import data & preprocessing import tensorflow as tf import keras import numpy as np # linear algebra import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv) import matplotlib.pyplot as plt # showing and rendering figures # io related from skimage.io import imread import os from glob import glob # not needed in Kaggle, but required in Jupyter %matplotlib inline ``` ### Import dataset ```python # Import data base_bone_dir = os.path.join('.') print (base_bone_dir) age_df = pd.read_csv(os.path.join(base_bone_dir, 'boneage-training-dataset.csv')) ion for now boneage_mean = 0 boneage_div = 1.0# 製作需要的欄位 age_df['path'] = age_df['id'].map(lambda x: os.path.join(base_bone_dir, 'resize', '{}.png'.format(x))) age_df['exists'] = age_df['path'].map(os.path.exists) print(age_df['exists'].sum(), 'images found of', age_df.shape[0], 'total') age_df['gender'] = age_df['male'].map(lambda x: 'male' if x else 'female') boneage_mean = age_df['boneage'].mean() boneage_div = 2*age_df['boneage'].std() # we don't want normalizat age_df['boneage_zscore'] = age_df['boneage'].map(lambda x: (x-boneage_mean)/boneage_div) age_df.dropna(inplace = True) age_df['male']= age_df['male'].map(lambda x: 1 if x else 0) age_df[['boneage', 'male', 'boneage_zscore']].hist(figsize = (10, 5)) age_df['boneage_category'] = pd.cut(age_df['boneage'], 10) ``` ![](https://i.imgur.com/RFGcEPX.png) ```python age_df['boneage_category'] = pd.cut(age_df['boneage'], 10) new_age_df = age_df.groupby(['boneage_category', 'male']).apply(lambda x: x.sample(400, replace = True) ).reset_index(drop = True) print('New Data Size:', new_age_df.shape[0], 'Old Size:', age_df.shape[0]) new_age_df[['boneage', 'male']].hist(figsize = (10, 5)) ``` New Data Size: 8000 Old Size: 12611 array([[<matplotlib.axes._subplots.AxesSubplot object at 0x00000134C390E948>, <matplotlib.axes._subplots.AxesSubplot object at 0x00000134C3C0BB88>]], dtype=object) ![](https://i.imgur.com/lZuLTyX.png) ```python # 將Training & validation 以3:1比例分開 from sklearn.model_selection import train_test_split raw_train_df, valid_df = train_test_split(age_df, test_size = 0.25, random_state = 2018, stratify = age_df['boneage_category']) print('train', raw_train_df.shape[0], 'validation', valid_df.shape[0]) ``` train 9458 validation 3153 ```python # 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) print('New Data Size:', train_df.shape[0], 'Old Size:', raw_train_df.shape[0]) train_df[['boneage', 'male']].hist(figsize = (10, 5)) # 圖檔預處理 from keras.preprocessing.image import ImageDataGenerator from keras.applications.vgg16 import preprocess_input IMG_SIZE = (384, 384) # slightly smaller than vgg16 normally expects core_idg = ImageDataGenerator(samplewise_center=False, samplewise_std_normalization=False, horizontal_flip = True, vertical_flip = False, height_shift_range = 0.15, width_shift_range = 0.15, rotation_range = 5, shear_range = 0.01, fill_mode = 'nearest', zoom_range=0.25, preprocessing_function = preprocess_input) ``` New Data Size: 10000 Old Size: 9458 ```python def flow_from_dataframe(img_data_gen, in_df, path_col, y_col, **dflow_args): base_dir = os.path.dirname(in_df[path_col].values[0]) print('## Ignore next message from keras, values are replaced anyways') df_gen = img_data_gen.flow_from_directory(base_dir, class_mode = 'sparse', **dflow_args) df_gen.filenames = in_df[path_col].values df_gen.filepaths.extend(df_gen.filenames) df_gen.classes = np.stack(in_df[y_col].values) df_gen.samples = in_df.shape[0] df_gen.n = in_df.shape[0] df_gen._set_index_array() df_gen.directory = '' # since we have the full path print('Reinserting dataframe: {} images'.format(in_df.shape[0])) return df_gen train_gen = flow_from_dataframe(core_idg, train_df, path_col = 'path', y_col = 'boneage_zscore', target_size = IMG_SIZE, color_mode = 'rgb', batch_size = 32) valid_gen = flow_from_dataframe(core_idg, valid_df, path_col = 'path', y_col = 'boneage_zscore', target_size = IMG_SIZE, color_mode = 'rgb', batch_size = 256) # we can use much larger batches for evaluation # used a fixed dataset for evaluating the algorithm test_X, test_Y = next(flow_from_dataframe(core_idg, valid_df, path_col = 'path', y_col = 'boneage_zscore', target_size = IMG_SIZE, color_mode = 'rgb', batch_size = 1024)) # one big batch ``` ## Ignore next message from keras, values are replaced anyways Found 0 images belonging to 0 classes. Reinserting dataframe: 10000 images ## Ignore next message from keras, values are replaced anyways Found 0 images belonging to 0 classes. Reinserting dataframe: 3153 images ## Ignore next message from keras, values are replaced anyways Found 0 images belonging to 0 classes. Reinserting dataframe: 3153 images ```python # 確認載入的圖檔 t_x, t_y = next(train_gen) fig, m_axs = plt.subplots(2, 4, figsize = (16, 8)) for (c_x, c_y, c_ax) in zip(t_x, t_y, m_axs.flatten()): c_ax.imshow(c_x[:,:,0], cmap = 'bone', vmin = -127, vmax = 127) c_ax.set_title('%2.0f months' % (c_y*boneage_div+boneage_mean)) c_ax.axis('off') ``` ![](https://i.imgur.com/GKFwzC6.png) ```python # VGG16參數設定 from keras.applications.vgg16 import VGG16 from keras.layers import GlobalAveragePooling2D, Dense, Dropout, Flatten, Input, Conv2D, multiply, LocallyConnected2D, Lambda from keras.models import Model in_lay = Input(t_x.shape[1:]) base_pretrained_model = VGG16(input_shape = t_x.shape[1:], include_top = False, weights = 'imagenet') base_pretrained_model.trainable = False pt_depth = base_pretrained_model.get_output_shape_at(0)[-1] pt_features = base_pretrained_model(in_lay) from keras.layers import BatchNormalization bn_features = BatchNormalization()(pt_features) # here we do an attention mechanism to turn pixels in the GAP on an off attn_layer = Conv2D(64, kernel_size = (1,1), padding = 'same', activation = 'relu')(bn_features) attn_layer = Conv2D(16, kernel_size = (1,1), padding = 'same', activation = 'relu')(attn_layer) attn_layer = LocallyConnected2D(1, kernel_size = (1,1), padding = 'valid', activation = 'sigmoid')(attn_layer) # fan it out to all of the channels up_c2_w = np.ones((1, 1, 1, pt_depth)) up_c2 = Conv2D(pt_depth, kernel_size = (1,1), padding = 'same', activation = 'linear', use_bias = False, weights = [up_c2_w]) up_c2.trainable = False attn_layer = up_c2(attn_layer) mask_features = multiply([attn_layer, bn_features]) gap_features = GlobalAveragePooling2D()(mask_features) gap_mask = GlobalAveragePooling2D()(attn_layer) # to account for missing values from the attention model gap = Lambda(lambda x: x[0]/x[1], name = 'RescaleGAP')([gap_features, gap_mask]) gap_dr = Dropout(0.5)(gap) dr_steps = Dropout(0.25)(Dense(1024, activation = 'elu')(gap_dr)) out_layer = Dense(1, activation = 'linear')(dr_steps) # linear is what 16bit did bone_age_model = Model(inputs = [in_lay], outputs = [out_layer]) # MAE from keras.metrics import mean_absolute_error def mae_months(in_gt, in_pred): return mean_absolute_error(boneage_div*in_gt, boneage_div*in_pred) bone_age_model.compile(optimizer = 'adam', loss = 'mse', metrics = [mae_months]) bone_age_model.summary() ``` Downloading data from https://github.com/fchollet/deep-learning-models/releases/download/v0.1/vgg16_weights_tf_dim_ordering_tf_kernels_notop.h5 58892288/58889256 [==============================] - 32s 1us/step Model: "model_1" __________________________________________________________________________________________________ Layer (type) Output Shape Param # Connected to ================================================================================================== input_1 (InputLayer) (None, 384, 384, 3) 0 __________________________________________________________________________________________________ vgg16 (Model) (None, 12, 12, 512) 14714688 input_1[0][0] __________________________________________________________________________________________________ batch_normalization_1 (BatchNor (None, 12, 12, 512) 2048 vgg16[1][0] __________________________________________________________________________________________________ conv2d_1 (Conv2D) (None, 12, 12, 64) 32832 batch_normalization_1[0][0] __________________________________________________________________________________________________ conv2d_2 (Conv2D) (None, 12, 12, 16) 1040 conv2d_1[0][0] __________________________________________________________________________________________________ locally_connected2d_1 (LocallyC (None, 12, 12, 1) 2448 conv2d_2[0][0] __________________________________________________________________________________________________ conv2d_3 (Conv2D) (None, 12, 12, 512) 512 locally_connected2d_1[0][0] __________________________________________________________________________________________________ multiply_1 (Multiply) (None, 12, 12, 512) 0 conv2d_3[0][0] batch_normalization_1[0][0] __________________________________________________________________________________________________ global_average_pooling2d_1 (Glo (None, 512) 0 multiply_1[0][0] __________________________________________________________________________________________________ global_average_pooling2d_2 (Glo (None, 512) 0 conv2d_3[0][0] __________________________________________________________________________________________________ RescaleGAP (Lambda) (None, 512) 0 global_average_pooling2d_1[0][0] global_average_pooling2d_2[0][0] __________________________________________________________________________________________________ dropout_1 (Dropout) (None, 512) 0 RescaleGAP[0][0] __________________________________________________________________________________________________ dense_1 (Dense) (None, 1024) 525312 dropout_1[0][0] __________________________________________________________________________________________________ dropout_2 (Dropout) (None, 1024) 0 dense_1[0][0] __________________________________________________________________________________________________ dense_2 (Dense) (None, 1) 1025 dropout_2[0][0] ================================================================================================== Total params: 15,279,905 Trainable params: 563,681 Non-trainable params: 14,716,224 __________________________________________________________________________________________________ ```python # VGG16參數設定(?) from keras.callbacks import ModelCheckpoint, LearningRateScheduler, EarlyStopping, ReduceLROnPlateau weight_path="{}_weights.best.hdf5".format('bone_age') 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=5) # probably needs to be more patient, but kaggle time is limited callbacks_list = [checkpoint, early, reduceLROnPlat] ``` C:\Users\Camille_Seknova\anaconda3\envs\tensorflow1\lib\site-packages\keras\callbacks\callbacks.py:998: UserWarning: `epsilon` argument is deprecated and will be removed, use `min_delta` instead. warnings.warn('`epsilon` argument is deprecated and ' ```python # 開始training # train_gen.batch_size = 16 bone_age_model.fit_generator(train_gen, validation_data = (test_X, test_Y), epochs = 15, callbacks = callbacks_list) ``` Epoch 1/15 313/313 [==============================] - 6021s 19s/step - loss: 2887.5958 - mae_months: 37.1822 - val_loss: 601.9203 - val_mae_months: 19.8439 Epoch 00001: val_loss improved from inf to 601.92025, saving model to bone_age_weights.best.hdf5 Epoch 2/15 313/313 [==============================] - 5879s 19s/step - loss: 834.0107 - mae_months: 22.8183 - val_loss: 517.4436 - val_mae_months: 18.1152 Epoch 00002: val_loss improved from 601.92025 to 517.44357, saving model to bone_age_weights.best.hdf5 Epoch 3/15 313/313 [==============================] - 5853s 19s/step - loss: 674.4779 - mae_months: 20.3838 - val_loss: 476.4907 - val_mae_months: 17.3899 Epoch 00003: val_loss improved from 517.44357 to 476.49066, saving model to bone_age_weights.best.hdf5 Epoch 4/15 313/313 [==============================] - 5839s 19s/step - loss: 591.0403 - mae_months: 18.9640 - val_loss: 452.4837 - val_mae_months: 16.9138 Epoch 00004: val_loss improved from 476.49066 to 452.48373, saving model to bone_age_weights.best.hdf5 Epoch 5/15 313/313 [==============================] - 5902s 19s/step - loss: 503.1557 - mae_months: 17.5464 - val_loss: 398.2757 - val_mae_months: 15.7191 Epoch 00005: val_loss improved from 452.48373 to 398.27571, saving model to bone_age_weights.best.hdf5 Epoch 6/15 313/313 [==============================] - 5870s 19s/step - loss: 467.0240 - mae_months: 16.8923 - val_loss: 409.5181 - val_mae_months: 15.8301 Epoch 00006: val_loss did not improve from 398.27571 Epoch 7/15 313/313 [==============================] - 5883s 19s/step - loss: 439.9315 - mae_months: 16.3247 - val_loss: 392.8785 - val_mae_months: 15.4780 Epoch 00007: val_loss improved from 398.27571 to 392.87852, saving model to bone_age_weights.best.hdf5 Epoch 8/15 313/313 [==============================] - 5855s 19s/step - loss: 419.6139 - mae_months: 15.9334 - val_loss: 394.2897 - val_mae_months: 15.6384 Epoch 00008: val_loss did not improve from 392.87852 Epoch 9/15 313/313 [==============================] - 5861s 19s/step - loss: 410.1131 - mae_months: 15.7633 - val_loss: 402.8027 - val_mae_months: 15.7428 Epoch 00009: val_loss did not improve from 392.87852 Epoch 10/15 313/313 [==============================] - 5876s 19s/step - loss: 396.9571 - mae_months: 15.4562 - val_loss: 378.6116 - val_mae_months: 15.3340 Epoch 00010: val_loss improved from 392.87852 to 378.61159, saving model to bone_age_weights.best.hdf5 Epoch 11/15 313/313 [==============================] - 6019s 19s/step - loss: 372.0665 - mae_months: 14.9971 - val_loss: 352.4316 - val_mae_months: 14.8172 Epoch 00011: val_loss improved from 378.61159 to 352.43158, saving model to bone_age_weights.best.hdf5 Epoch 12/15 313/313 [==============================] - 6130s 20s/step - loss: 359.2982 - mae_months: 14.8550 - val_loss: 387.3231 - val_mae_months: 15.6298 Epoch 00012: val_loss did not improve from 352.43158 Epoch 13/15 313/313 [==============================] - 6070s 19s/step - loss: 359.9599 - mae_months: 14.8379 - val_loss: 360.8462 - val_mae_months: 15.0052 Epoch 00013: val_loss did not improve from 352.43158 Epoch 14/15 313/313 [==============================] - 6096s 19s/step - loss: 343.5727 - mae_months: 14.5815 - val_loss: 351.4307 - val_mae_months: 14.8820 Epoch 00014: val_loss improved from 352.43158 to 351.43066, saving model to bone_age_weights.best.hdf5 Epoch 15/15 313/313 [==============================] - 6245s 20s/step - loss: 347.8223 - mae_months: 14.5407 - val_loss: 345.8929 - val_mae_months: 14.8902 Epoch 00015: val_loss improved from 351.43066 to 345.89290, saving model to bone_age_weights.best.hdf5 <keras.callbacks.callbacks.History at 0x134cff46c48> ```python # load the best version of the model bone_age_model.load_weights(weight_path) ``` ```python # 開始預測 pred_Y = boneage_div*bone_age_model.predict(test_X, batch_size = 32, verbose = True)+boneage_mean test_Y_months = boneage_div*test_Y+boneage_mean ``` 1024/1024 [==============================] - 578s 564ms/step ```python # 結果放置於 pred_Y 和實際月份 test_Y_months 進行參數比對 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)') ``` Text(0, 0.5, 'Predicted Age (Months)') ![](https://i.imgur.com/1a31fcd.png) ```python from sklearn.metrics import mean_absolute_error, median_absolute_error print(mean_absolute_error(pred_Y, test_Y_months)) print(median_absolute_error(pred_Y, test_Y_months)) ``` 14.890207 12.889206 ## MobieNet * 啟動Anaconda powershell 切到工作目錄下，將影像dataset跟 boneage-training-dataset.csv 都放在同一目錄下 ![](https://i.imgur.com/VLvoIwP.png) * Import data & preprocessing * ```python import tensorflow as tf import keras import numpy as np # linear algebra import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv) import matplotlib.pyplot as plt # showing and rendering figures # io related from skimage.io import imread import os from glob import glob # not needed in Kaggle, but required in Jupyter %matplotlib inline ``` * Import data ```python #base_bone_dir = os.path.join('c:', os.sep, 'Users', 'Elfay', 'ML') base_bone_dir = os.path.join('.') print (base_bone_dir) age_df = pd.read_csv(os.path.join(base_bone_dir,'boneage-training-dataset.csv')) age_df.describe ``` . <index id boneage male 0 1377 180 0 1 1378 12 0 2 1379 94 0 3 1380 120 1 4 1381 82 0 ... ... ... ... [12611 rows x 3 columns]> * 製作需要的欄位 ```python age_df['path'] = age_df['id'].map(lambda x: os.path.join(base_bone_dir, 'resize', '{}.png'.format(x))) age_df['exists'] = age_df['path'].map(os.path.exists) print(age_df['exists'].sum(), 'images found of', age_df.shape[0], 'total') age_df['gender'] = age_df['male'].map(lambda x: 'male' if x else 'female') boneage_mean = age_df['boneage'].mean() boneage_div = 2*age_df['boneage'].std() age_df['boneage_zscore'] = age_df['boneage'].map(lambda x: (x-boneage_mean)/boneage_div) age_df.dropna(inplace = True) age_df.sample(3) ``` 12611 images found of 12611 total * 完成之CSV檔案預覽 <div> <style scoped> .dataframe tbody tr th:only-of-type { vertical-align: middle; } .dataframe tbody tr th { vertical-align: top; } .dataframe thead th { text-align: right; } </style> <table border="1" class="dataframe"> <thead> <tr style="text-align: right;"> <th></th> <th>id</th> <th>boneage</th> <th>male</th> <th>path</th> <th>exists</th> <th>gender</th> <th>boneage_zscore</th> </tr> </thead> <tbody> <tr> <th>2770</th> <td>4650</td> <td>132</td> <td>0</td> <td>.\resize\4650.png</td> <td>True</td> <td>female</td> <td>0.056812</td> </tr> <tr> <th>10121</th> <td>12828</td> <td>120</td> <td>0</td> <td>.\resize\12828.png</td> <td>True</td> <td>female</td> <td>-0.088883</td> </tr> <tr> <th>4176</th> <td>6195</td> <td>162</td> <td>0</td> <td>.\resize\6195.png</td> <td>True</td> <td>female</td> <td>0.421048</td> </tr> </tbody> </table> </div> * 展示資料分佈狀況 ```python age_df[['boneage','male', 'boneage_zscore']].hist(figsize = (10, 5)) ``` array([[<matplotlib.axes._subplots.AxesSubplot object at 0x000000000FE5DE88>, <matplotlib.axes._subplots.AxesSubplot object at 0x00000000100AEA88>], [<matplotlib.axes._subplots.AxesSubplot object at 0x00000000100E9448>, <matplotlib.axes._subplots.AxesSubplot object at 0x000000001011FE08>]], dtype=object) ![](https://i.imgur.com/igFzMRG.png) * 按照boneage & 性別分布取8000筆資料 ```python age_df['boneage_category'] = pd.cut(age_df['boneage'], 10) new_age_df = age_df.groupby(['boneage_category', 'male']).apply(lambda x: x.sample(400, replace = True) ).reset_index(drop = True) print('New Data Size:', new_age_df.shape[0], 'Old Size:', age_df.shape[0]) new_age_df[['boneage', 'male']].hist(figsize = (10, 5)) ``` New Data Size: 8000 Old Size: 12611 ![](https://i.imgur.com/7vO8mhE.png) * 將Training & validation 以3:1比例分開 ```python from sklearn.model_selection import train_test_split train_df, valid_df = train_test_split(new_age_df, test_size = 0.25, random_state = 2018, stratify = new_age_df['boneage_category']) print('train', train_df.shape[0], 'validation', valid_df.shape[0]) ``` train 6000 validation 2000 * 圖檔預處理 ```python from keras.preprocessing.image import ImageDataGenerator IMG_SIZE = (256, 256) core_idg = ImageDataGenerator(samplewise_center=True, samplewise_std_normalization=True, horizontal_flip = True, vertical_flip = False, height_shift_range = 0.1, width_shift_range = 0.1, rotation_range = 10, shear_range = 0.05, fill_mode = 'nearest', zoom_range=0.15) ``` * 載入圖檔 ```python def flow_from_dataframe(img_data_gen, in_df, path_col, y_col, **dflow_args): base_dir = os.path.dirname(in_df[path_col].values[0]) print('## Ignore next message from keras, values are replaced anyways') df_gen = img_data_gen.flow_from_directory(base_dir, class_mode = 'sparse', **dflow_args) df_gen.filenames = in_df[path_col].values df_gen.filepaths.extend(df_gen.filenames) df_gen.classes = np.stack(in_df[y_col].values) df_gen.samples = in_df.shape[0] df_gen.n = in_df.shape[0] df_gen._set_index_array() df_gen.directory = '' # since we have the full path print('Reinserting dataframe: {} images'.format(in_df.shape[0])) return df_gen ``` ```python train_gen = flow_from_dataframe(core_idg, train_df, path_col = 'path', y_col = 'boneage_zscore', target_size = IMG_SIZE, color_mode = 'grayscale', batch_size = 64) valid_gen = flow_from_dataframe(core_idg, valid_df, path_col = 'path', y_col = 'boneage_zscore', target_size = IMG_SIZE, color_mode = 'grayscale', batch_size = 64) # we can use much larger batches for evaluation # used a fixed dataset for evaluating the algorithm test_X, test_Y = next(flow_from_dataframe(core_idg, valid_df, path_col = 'path', y_col = 'boneage_zscore', target_size = IMG_SIZE, color_mode = 'grayscale', batch_size = 500)) # one big batch ``` ## Ignore next message from keras, values are replaced anyways Found 0 images belonging to 0 classes. Reinserting dataframe: 6000 images ## Ignore next message from keras, values are replaced anyways Found 0 images belonging to 0 classes. Reinserting dataframe: 2000 images ## Ignore next message from keras, values are replaced anyways Found 0 images belonging to 0 classes. Reinserting dataframe: 2000 images * 確認載入的圖檔 ```python t_x, t_y = next(train_gen) fig, m_axs = plt.subplots(4, 4, figsize = (16, 16)) for (c_x, c_y, c_ax) in zip(t_x, t_y, m_axs.flatten()): c_ax.imshow(c_x[:,:,0], cmap = 'bone', vmin = -3, vmax = 3) c_ax.set_title('%2.0f months' % (c_y*boneage_div+boneage_mean)) c_ax.axis('off') ``` ![](https://i.imgur.com/hWUsMpO.png) * MobileNet 參數設定 ```python from keras.applications.mobilenet import MobileNet from keras.layers import GlobalAveragePooling2D, Dense, Dropout, Flatten, BatchNormalization from keras.models import Sequential base_mobilenet_model = MobileNet(input_shape = t_x.shape[1:], include_top = False, weights = None) bone_age_model = Sequential() bone_age_model.add(BatchNormalization(input_shape = t_x.shape[1:])) bone_age_model.add(base_mobilenet_model) bone_age_model.add(BatchNormalization()) bone_age_model.add(GlobalAveragePooling2D()) bone_age_model.add(Dropout(0.5)) bone_age_model.add(Dense(1, activation = 'linear' )) # linear is what 16bit did from keras.metrics import mean_absolute_error def mae_months(in_gt, in_pred): return mean_absolute_error(boneage_div*in_gt, boneage_div*in_pred) bone_age_model.compile(optimizer = 'adam', loss = 'mse', metrics = [mae_months]) bone_age_model.summary() ``` Model: "sequential_1" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= batch_normalization_1 (Batch (None, 256, 256, 1) 4 _________________________________________________________________ mobilenet_1.00_256 (Model) (None, 8, 8, 1024) 3228288 _________________________________________________________________ batch_normalization_2 (Batch (None, 8, 8, 1024) 4096 _________________________________________________________________ global_average_pooling2d_1 ( (None, 1024) 0 _________________________________________________________________ dropout_1 (Dropout) (None, 1024) 0 _________________________________________________________________ dense_1 (Dense) (None, 1) 1025 ================================================================= Total params: 3,233,413 Trainable params: 3,209,475 Non-trainable params: 23,938 _________________________________________________________________ * MobileNet 參數設定 ```python from keras.callbacks import ModelCheckpoint, LearningRateScheduler, EarlyStopping, ReduceLROnPlateau weight_path="{}_weights.best.hdf5".format('bone_age') 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=5) # probably needs to be more patient, but kaggle time is limited callbacks_list = [checkpoint, early, reduceLROnPlat] ``` * 開始training，訓練參數置入 bone_age_weights.best.hdf5 ```python train_gen.batch_size = 16 bone_age_model.fit_generator(train_gen, validation_data = (test_X, test_Y), epochs = 10, callbacks = callbacks_list) ``` Epoch 1/10 375/375 [==============================] - 1261s 3s/step - loss: 0.6289 - mae_months: 52.4197 - val_loss: 0.6183 - val_mae_months: 56.1310 Epoch 00001: val_loss improved from inf to 0.61833, saving model to bone_age_weights.best.hdf5 Epoch 2/10 375/375 [==============================] - 1219s 3s/step - loss: 0.3611 - mae_months: 39.8186 - val_loss: 0.4336 - val_mae_months: 46.1007 Epoch 00002: val_loss improved from 0.61833 to 0.43358, saving model to bone_age_weights.best.hdf5 Epoch 3/10 375/375 [==============================] - 1275s 3s/step - loss: 0.2738 - mae_months: 34.3837 - val_loss: 0.2607 - val_mae_months: 32.8493 Epoch 00003: val_loss improved from 0.43358 to 0.26069, saving model to bone_age_weights.best.hdf5 Epoch 4/10 375/375 [==============================] - 1275s 3s/step - loss: 0.2304 - mae_months: 31.2175 - val_loss: 0.4354 - val_mae_months: 45.7760 Epoch 00004: val_loss did not improve from 0.26069 Epoch 5/10 375/375 [==============================] - 1276s 3s/step - loss: 0.1979 - mae_months: 28.9312 - val_loss: 0.1548 - val_mae_months: 24.7830 Epoch 00005: val_loss improved from 0.26069 to 0.15478, saving model to bone_age_weights.best.hdf5 Epoch 6/10 375/375 [==============================] - 1279s 3s/step - loss: 0.1733 - mae_months: 27.0213 - val_loss: 0.1598 - val_mae_months: 25.1762 Epoch 00006: val_loss did not improve from 0.15478 Epoch 7/10 375/375 [==============================] - 1289s 3s/step - loss: 0.1580 - mae_months: 25.6752 - val_loss: 0.1208 - val_mae_months: 22.1466 Epoch 00007: val_loss improved from 0.15478 to 0.12078, saving model to bone_age_weights.best.hdf5 Epoch 8/10 375/375 [==============================] - 1290s 3s/step - loss: 0.1391 - mae_months: 24.2044 - val_loss: 0.1435 - val_mae_months: 24.5172 Epoch 00008: val_loss did not improve from 0.12078 Epoch 9/10 375/375 [==============================] - 1303s 3s/step - loss: 0.1306 - mae_months: 23.4019 - val_loss: 0.0976 - val_mae_months: 19.3900 Epoch 00009: val_loss improved from 0.12078 to 0.09761, saving model to bone_age_weights.best.hdf5 Epoch 10/10 375/375 [==============================] - 1278s 3s/step - loss: 0.1241 - mae_months: 22.9744 - val_loss: 0.0910 - val_mae_months: 19.1049 Epoch 00010: val_loss improved from 0.09761 to 0.09099, saving model to bone_age_weights.best.hdf5 ### 預測資料 * 載入訓練參數 bone_age_weights.best.hdf5 ```python weight_path= bone_age_model.load_weights(weight_path) ``` * 開始預測 ```python pred_Y = boneage_div*bone_age_model.predict(test_X, batch_size = 16, verbose = True)+boneage_mean test_Y_months = boneage_div*test_Y+boneage_mean ``` 500/500 [==============================] - 33s 67ms/step * 結果放置於 pred_Y 和實際月份 test_Y_months 進行參數比對 ```python 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)') ``` * 結果展示 ![](https://i.imgur.com/efkQRmO.png) ```python from sklearn.metrics import mean_absolute_error, median_absolute_error print(mean_absolute_error(pred_Y, test_Y_months)) print(median_absolute_error(pred_Y, test_Y_months)) ``` ### 改影像前處理 ## Reference * [VGG16](https://www.kaggle.com/kmader/attention-on-pretrained-vgg16-for-bone-age) * [InceptionV3](https://www.kaggle.com/kmader/pretrained-inceptionv3-for-bone-age) * [MobileNet](https://www.kaggle.com/kmader/mobilenet-for-bone-age) * [MobileNet Introduction](https://medium.com/@chih.sheng.huang821/%E6%B7%B1%E5%BA%A6%E5%AD%B8%E7%BF%92-mobilenet-depthwise-separable-convolution-f1ed016b3467) * [RestNet50](https://www.kaggle.com/suniliitb96/tutorial-keras-transfer-learning-with-resnet50)