Multiple Instance Learning (MIL) для классификации наборов гистологических изображений со светлоклеточным раком почки

from tensorflow import keras from tensorflow.keras import layers from keras.layers import Flatten from keras.layers import Input, Dense, Layer, Dropout, Conv2D, MaxPooling2D, Flatten, multiply from MILAttentionLayer import MILAttentionLayer  def SimpleModel(instance_shape,bag_size):     """ Create Keras model for Multiply Instance Learning     Parameters     -------------------     instance_shape (tuple) - shape of 1 instance in the bag     bag_size (int) - size of the bag     Returns     -------------------      keras.Model     """     # Extract features from inputs.     inputs, embeddings = [], []     conv1_1 = Conv2D(16, kernel_size=(2,2), activation='relu')      conv1_2 = Conv2D(16, kernel_size=(2,2), activation='relu')       mpool_1 = MaxPooling2D((2,2))      conv2_1 = Conv2D(32, kernel_size=(2,2),   activation='relu')       conv2_2 = Conv2D(32, kernel_size=(2,2),activation='relu')      mpool_2 = MaxPooling2D((2,2))      fc0 = Dense(512, activation='relu', name='fc0')      fc1 = Dense(512, activation='relu', name='fc1')      fc2 = Dense(256, activation= 'relu',  name='fc2')          for _ in range(bag_size):         inp = layers.Input(instance_shape)         inputs.append(inp)         x = conv1_1(inp)         x = conv1_2(x)         x = mpool_1(x)          x = conv2_1(x)         x = conv2_2(x)         x = mpool_2(x)          x = Flatten()(x)         x = fc0(x)         x = Dropout(0.5)(x)         x = fc1(x)         x = Dropout(0.5)(x)         x = fc2(x)         x = Dropout(0.2)(x)                  embeddings.append(x)      # Аttention layer.     alpha = MILAttentionLayer(         weight_params_dim=1024,         kernel_regularizer=keras.regularizers.l2(0),# previous - 0.01         use_gated=True,          name="alpha",     )(embeddings)      # Multiply attention weights with the input layers.     multiply_layers = [         layers.multiply([alpha[i], embeddings[i]]) for i in range(len(alpha))     ]      # Concatenate layers.     concat = layers.concatenate(multiply_layers, axis=1)      # Classification output node.     output = layers.Dense(2, activation = 'softmax')(concat)      return keras.Model(inputs, output) 

import tensorflow as tf from CustomDataGenerator import CustomDataGenerator from SimpleModel import SimpleModel  def train_model (train_df, validation_df, model_save_path):     """     Train SimpleModel          Parameters     -------------------     train_df (pandas DataFrame) - DataFrame with the training data. X (bag of instances) - list of images paths. y -label     validation_df (pandas DataFrame) - DataFrame with the validation data. X (bag of instances) - list of images paths. y -label     model_save_path (str) - path for model saving     Returns     -------------------     """             # create generator of the training and validation data     train_generator = CustomDataGenerator(df = train_df, shuffle = True, augmentations = True )     validation_generator = CustomDataGenerator (df = validation_df, shuffle = False, augmentations = False )          # Callbacks     model_checkpoint = tf.keras.callbacks.ModelCheckpoint(         model_save_path,         monitor="val_loss",         verbose=1,         mode="min",         save_best_only=True,         save_weights_only= False)               es = tf.keras.callbacks.EarlyStopping(         monitor="val_loss",         patience=10,         verbose=1,         mode="min")          # optimizer     opt = tf.keras.optimizers.Adam(learning_rate=1e-3, decay=0.0005, beta_1=0.9, beta_2=0.999)          # create and compile model     model = SimpleModel(bag_size = 40, instance_shape = (256, 256, 3) )     model.compile(optimizer = opt,      loss='categorical_crossentropy', metrics=["accuracy",tf.keras.metrics.AUC(name = 'AUC'),                                                         tf.keras.metrics.AUC(curve = 'PR',name = 'PR_AUC'),                                                          tf.keras.metrics.Precision(name = 'Precision', class_id = 1),                                                         tf.keras.metrics.Recall(name = 'Recall',class_id = 1)])     # model fitting     model.fit(         train_generator,         validation_data = validation_generator ,         epochs=100,         batch_size= 1,         callbacks=[model_checkpoint,es],          verbose=1)