Kaggle比赛——HistopathologicCancerDetection代码开源（Resnet50模型为基础）

       Kaggle比赛 Histopathologic Cancer Detection 代码开源。
　　模型使用了Resnet50&＃xff0c;修改最后几层网络结构重新训练&＃xff0c;并且使用了五折交叉验证取平均值来提高精度。其中的一些trick在代码中已经标注。后面考虑修改网络结构&＃xff0c;损失函数来进一步提高精度。更多比赛代码可查看我的github 大黄大黄大黄 。

import os
import datetime
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split
from keras.preprocessing.image import ImageDataGeneratorfrom keras.applications.resnet50 import ResNet50
from keras.models import Model
from keras.layers import Dense, Flatten, Dropout, Activation, BatchNormalization, GlobalAveragePooling2D, GlobalMaxPooling2D, Concatenate, Input
from keras.optimizers import Adam
from keras.callbacks import EarlyStopping, ReduceLROnPlateau, TensorBoard, ModelCheckpoint
from keras.utils.vis_utils import plot_model
from sklearn.cross_validation import KFold# **本地用matplotlib绘图可以&＃xff0c;但是在Linux服务器运行代码绘图的时候会报错&＃xff0c;需要添加这行代码**
plt.switch_backend(&＃39;agg&＃39;)os.environ["CUDA_VISIBLE_DEVICES"] &＃61; &＃39;2,3,4&＃39;train_dir &＃61; &＃39;./dataset/train/&＃39;IMG_SIZE &＃61; (96, 96)
IN_SHAPE &＃61; (*IMG_SIZE, 3)dropout_rate &＃61; 0.5def load_data(train, valid):train_datagen &＃61; ImageDataGenerator(preprocessing_function&＃61;lambda x: (x - x.mean()) / x.std() if x.std() > 0 else x,horizontal_flip&＃61;True,vertical_flip&＃61;True)valid_datagen &＃61; ImageDataGenerator(preprocessing_function&＃61;lambda x: (x - x.mean()) / x.std() if x.std() > 0 else x)train_generator &＃61; train_datagen.flow_from_dataframe(dataframe&＃61;train,directory&＃61;train_dir,x_col&＃61;&＃39;id&＃39;,y_col&＃61;&＃39;label&＃39;,has_ext&＃61;False,# subset&＃61;&＃39;training&＃39;,batch_size&＃61;32,seed&＃61;2018,shuffle&＃61;True,class_mode&＃61;&＃39;binary&＃39;,target_size&＃61;(96, 96))valid_generator &＃61; valid_datagen.flow_from_dataframe(dataframe&＃61;valid,directory&＃61;train_dir,x_col&＃61;&＃39;id&＃39;,y_col&＃61;&＃39;label&＃39;,has_ext&＃61;False,# subset&＃61;&＃39;validation&＃39;,batch_size&＃61;32,seed&＃61;2018,shuffle&＃61;False,class_mode&＃61;&＃39;binary&＃39;,target_size&＃61;(96, 96))return train_generator, valid_generatordef model_bulid():inputs &＃61; Input(IN_SHAPE)conv_base &＃61; ResNet50(weights&＃61;&＃39;imagenet&＃39;,include_top&＃61;False,input_shape&＃61;IN_SHAPE)x &＃61; conv_base(inputs)out &＃61; Flatten()(x)# **如果换成256验证集效果很差&＃xff0c;我也不知道为什么**out &＃61; Dense(512)(out)out &＃61; BatchNormalization()(out)out &＃61; Activation(activation&＃61;&＃39;relu&＃39;)(out)out &＃61; Dropout(dropout_rate)(out)out &＃61; Dense(1, activation&＃61;"sigmoid", name&＃61;"3_")(out)model &＃61; Model(inputs, out)conv_base.Trainable &＃61; Trueset_trainable &＃61; Falsefor layer in conv_base.layers:if layer.name &＃61;&＃61; &＃39;res5a_branch2a&＃39;:set_trainable &＃61; Trueif set_trainable:layer.trainable &＃61; Trueelse:layer.trainable &＃61; Falsemodel.compile(Adam(0.001),loss&＃61;"binary_crossentropy", metrics&＃61;["accuracy"])model.summary()plot_model(model, to_file&＃61;&＃39;NetStruct.png&＃39;, show_shapes&＃61;True)return modeldef train_model():nfolds &＃61; 5epoch &＃61; 30random_state &＃61; 2018df_train &＃61; pd.read_csv(&＃39;./dataset/train_labels.csv&＃39;)# df_train &＃61; df_train.sample(n&＃61;20, random_state&＃61;random_state)df_train &＃61; df_train.valueskf &＃61; KFold(len(df_train), n_folds&＃61;nfolds,shuffle&＃61;True, random_state&＃61;random_state)num_fold &＃61; 0for train_index, valid_index in kf:train_df, valid_df &＃61; df_train[train_index], df_train[valid_index]train_df &＃61; pd.DataFrame(train_df, columns&＃61;[&＃39;id&＃39;, &＃39;label&＃39;])valid_df &＃61; pd.DataFrame(valid_df, columns&＃61;[&＃39;id&＃39;, &＃39;label&＃39;])train, valid &＃61; load_data(train_df, valid_df)num_fold &＃43;&＃61; 1print(&＃39;Start KFold number {} from {}&＃39;.format(num_fold, nfolds))print(&＃39;Split train: &＃39;, len(train))print(&＃39;Split valid: &＃39;, len(valid))model &＃61; model_bulid()tensorboard &＃61; TensorBoard(log_dir&＃61;&＃39;./logs_kfold&＃39;, # log 目录# histogram_freq&＃61;1, # 按照何等频率&＃xff08;epoch&＃xff09;来计算直方图&＃xff0c;0为不计算# batch_size&＃61;batch_size, # 用多大量的数据计算直方图write_graph&＃61;True, # 是否存储网络结构图write_grads&＃61;False, # 是否可视化梯度直方图write_images&＃61;False, # 是否可视化参数embeddings_freq&＃61;0,embeddings_layer_names&＃61;None,embeddings_metadata&＃61;None)model_checkpoint &＃61; ModelCheckpoint(&＃39;weights_kfold&＃39;&＃43;&＃39;_&＃39;&＃43;str(num_fold)&＃43;&＃39;.h5&＃39;, monitor&＃61;&＃39;val_loss&＃39;, save_best_only&＃61;True)earlystopper &＃61; EarlyStopping(monitor&＃61;&＃39;val_loss&＃39;, patience&＃61;2, verbose&＃61;1)reducel &＃61; ReduceLROnPlateau(monitor&＃61;&＃39;val_loss&＃39;, patience&＃61;1, verbose&＃61;1, factor&＃61;0.1)history &＃61; model.fit_generator(train,validation_data&＃61;valid,epochs&＃61;epoch,callbacks&＃61;[reducel, earlystopper, model_checkpoint, tensorboard])plt.plot(history.history[&＃39;loss&＃39;], label&＃61;&＃39;train&＃39;)plt.plot(history.history[&＃39;val_loss&＃39;], label&＃61;&＃39;valid&＃39;)plt.title("model loss")plt.ylabel("loss")plt.xlabel("epoch")plt.legend(["train", "valid"], loc&＃61;"upper left")plt.savefig(&＃39;loss_performance&＃39;&＃43;&＃39;_&＃39;&＃43;str(num_fold)&＃43;&＃39;.png&＃39;)plt.clf()plt.plot(history.history[&＃39;acc&＃39;], label&＃61;&＃39;train&＃39;)plt.plot(history.history[&＃39;val_acc&＃39;], label&＃61;&＃39;valid&＃39;)plt.title("model acc")plt.ylabel("acc")plt.xlabel("epoch")plt.legend(["train", "valid"], loc&＃61;"upper left")plt.savefig(&＃39;acc_performance&＃39;&＃43;&＃39;_&＃39;&＃43;str(num_fold)&＃43;&＃39;.png&＃39;)with open(&＃39;logs_kfold.txt&＃39;, &＃39;a&＃43;&＃39;) as f:f.write(str(datetime.datetime.now().strftime("%Y-%m-%d-%H-%M"))&＃43;&＃39;\n&＃39;)f.write(str(num_fold)&＃43;&＃39;\n&＃39;)f.write(str(history.history[&＃39;loss&＃39;])&＃43;&＃39;\n&＃39;)f.write(str(history.history[&＃39;val_loss&＃39;])&＃43;&＃39;\n&＃39;)f.write(str(history.history[&＃39;acc&＃39;])&＃43;&＃39;\n&＃39;)f.write(str(history.history[&＃39;val_acc&＃39;])&＃43;&＃39;\n&＃39;)if __name__ &＃61;&＃61; &＃39;__main__&＃39;:train_model()