import tensorflow as tf from keras.layers import Input, Conv3D, MaxPooling3D, Dropout, BatchNormalization, Reshape, Dense, ELU, concatenate, add, Lambda, MaxPooling2D, GRU, Masking, advanced_activations from keras.models import Model, save_model from keras.optimizers import Adam from keras import backend as K from tensorflow import reshape, transpose from keras.callbacks import LearningRateScheduler from keras.metrics import binary_crossentropy from keras import activations import numpy as np import math from keras import activations from keras.utils import CustomObjectScope from skimage.transform import resize from tensorflow.python.framework import ops import innvestigate import innvestigate.utils import os import sys sys.path.append('//data/data_wnx3/data_wnx1/_Data/AlzheimersDL/CNN+RNN-2class-1cnn+data/utils') from sepconv3D import SeparableConv3D from augmentation import CustomIterator model_filepath = '//data/data_wnx3/data_wnx1/_Data/AlzheimersDL/CNN+RNN-2class-1cnn+data' ####for 2 class model + RNN### class Parameters(): def __init__ (self, param_dict): self.CNN_w_regularizer = param_dict['CNN_w_regularizer'] self.RNN_w_regularizer = param_dict['RNN_w_regularizer'] self.CNN_batch_size = param_dict['CNN_batch_size'] self.RNN_batch_size = param_dict['RNN_batch_size'] self.CNN_drop_rate = param_dict['CNN_drop_rate'] self.RNN_drop_rate = param_dict['RNN_drop_rate'] self.epochs = param_dict['epochs'] self.gpu = param_dict['gpu'] self.model_filepath = param_dict['model_filepath'] + '/net.h5' self.num_clinical = param_dict['num_clinical'] self.image_shape = param_dict['image_shape'] self.final_layer_size = param_dict['final_layer_size'] self.optimizer = param_dict['optimizer'] class CNN_Net (): def __init__ (self, params): self.params = params self.xls = Input (shape = (self.params.num_clinical,),name='input_xls') self.mri = Input (shape = (self.params.image_shape),name='input_mri') self.jac = Input (shape = (self.params.image_shape),name='input_jac') xalex3D = XAlex3D(w_regularizer = self.params.CNN_w_regularizer, drop_rate = self.params.CNN_drop_rate, final_layer_size=self.params.final_layer_size) with tf.device(self.params.gpu): self.fc_CNN = xalex3D (self.mri, self.jac, self.xls) self.CNNoutput_class = Dense(units = 2, activation = 'softmax', name = 'CNNclass_output') (self.fc_CNN) #use either 1, sigmoid, binarycrossent OR 2, softmax, sparsecategoricalcrossent def train (self, data): train_data, val_data = data train_samples = train_data[0].shape[0] val_samples = len(val_data[0]) data_flow_train = CustomIterator (train_data, batch_size = self.params.CNN_batch_size, shuffle = True) data_flow_val = CustomIterator (val_data, batch_size = self.params.CNN_batch_size, shuffle = True) self.model = Model(inputs = [self.mri,self.jac,self.xls], outputs = [self.CNNoutput_class]) lrate = LearningRateScheduler(step_decay_CNN) callback = [lrate] #optimizer = Adam(lr=1e-5) self.optimizer = self.params.optimizer self.model.compile(optimizer = self.optimizer, loss = 'sparse_categorical_crossentropy', metrics =['acc']) self.model.summary() history = self.model.fit_generator (data_flow_train, steps_per_epoch = train_samples/self.params.CNN_batch_size, epochs = self.params.epochs, callbacks = callback, shuffle = True, #might be being ignored if the input data is a generator?? validation_data = data_flow_val, validation_steps = val_samples/self.params.CNN_batch_size) #Save the model save_model(self.model,'SavedCNNModel') self.model.save_weights('SavedCNNWeights') #get features from last layer featuresModel = Model(inputs = self.model.input, outputs = self.model.layers[-2].output) featuresModel.compile(optimizer = self.params.optimizer, loss = 'sparse_categorical_crossentropy', metrics =['acc']) return history.history, featuresModel def predict (self, data_test): test_mri, test_jac, test_xls, test_labels, test_ptid, test_imageID, test_confid, test_csf = data_test # with open('//data/data_wnx3/data_wnx1/_Data/AlzheimersDL/CNN+RNN-2class/figchecks/test_data.txt', 'w') as testdata: # testdata.write('{}\n'.format(test_data)) # testdata.write('{}\n{}\n{}\n{}\n{}\n'.format(len(test_data),len(test_data[0]),len(test_data[1]),len(test_data[2]),len(test_data[3]))) preds = self.model.predict ([test_mri, test_jac, test_xls]) return preds def evaluate (self, data_test): test_mri, test_jac, test_xls, test_labels, test_ptid, test_imageID, test_confid, test_csf = data_test metrics = self.model.evaluate (x = [test_mri, test_jac, test_xls], y = test_labels, batch_size = self.params.CNN_batch_size) return metrics def load_the_weights (self, SavedWeights): self.model = Model(inputs = [self.mri,self.jac,self.xls], outputs = [self.CNNoutput_class]) self.model.compile(optimizer = self.params.optimizer, loss = 'sparse_categorical_crossentropy', metrics =['acc']) loaded = self.model.load_weights(SavedWeights) return loaded def LRP_heatmap(self, img_data, img_number): #https://github.com/albermax/innvestigate test_mri, test_jac, test_xls, test_labels, test_ptid, test_imageID, test_confid, test_csf = img_data #clear some memory: (these are just pointers anyway?) test_labels=0 test_ptid=0 test_imageID=0 test_confid=0 test_csf=0 print('kill check models130') #create the model without the final softmax layer nosoftmax_model = innvestigate.utils.model_wo_softmax(self.model) print('kill check models135') #create the analyzer analyzer = innvestigate.create_analyzer("lrp.z",nosoftmax_model,disable_model_checks=True) print('kill check models138') ##analyzer = innvestigate.analyzer.LRPZ(nosoftmax_model,disable_model_checks=True) #analyze analysis = analyzer.analyze([[test_mri[img_number]],[test_jac[img_number]],[test_xls[img_number]]]) print('shape of initial LRP heatmap: ', analysis.shape) analysis /= np.max(np.abs(analysis)) analysis = np.squeeze(analysis,0) print('shape of squeezed LRP heatmap: ', analysis.shape) #analysis = np.moveaxis(analysis,0,3) #analysis = resize(analysis,(test_mri[img_number].shape)) #maybe this is actually turning the 1 into 91 instead of moving it to the end. maybe try a channels last adjustment or something?? #analysis = analysis[:,:,:,0,:] print('shape of resized LRP heatmap: ', analysis.shape) return analysis def make_gradcam_heatmap2(self,img_data,img_number): #https://towardsdatascience.com/demystifying-convolutional-neural-networks-using-gradcam-554a85dd4e48 and https://keras.io/examples/vision/grad_cam/ test_mri, test_jac, test_xls, test_labels, test_ptid, test_imageID, test_confid, test_csf = img_data last_conv = self.model.layers[-15] #.get_layer('name') grads = K.gradients(self.model.output[:,1],last_conv.output)[0] #[:,x] corresponds to the class I want I think? So AD = 0 #print('grads: ', grads) print('shape of grads: ', grads.shape) pooled_grads = K.mean(grads,axis=(0,1,2,3)) print('shape of pooled_grads: ', pooled_grads.shape) iterate = K.function([self.model.input[0],self.model.input[1],self.model.input[2]],[pooled_grads,last_conv.output[0]]) print('shape of test_mri[j]: ', test_mri[img_number].shape) pooled_grads_value,conv_layer_output = iterate([[test_mri[img_number]],[test_jac[img_number]],[test_xls[img_number]]]) for i in range(48): #range = size of conv layer units? aka number of filters/channels conv_layer_output[:,:,:,i] *= pooled_grads_value[i] #conv_layer_output[:,:,i] #multiplies feature maps with pooled grads heatmap = np.mean(conv_layer_output,axis=-1) #takes the mean over all the filters/channels to get just one map #for x in range(heatmap.shape[0]): #this chunk applies a relu to keep only the features that have a positive influence on the output map # for y in range(heatmap.shape[1]): # heatmap[x,y] = np.max(heatmap[x,y],0) print('shape of initial heatmap: ', heatmap.shape) heatmap = np.maximum(heatmap,0) #keeps only the positive values (only keep the features that have a positive influence on the output map) heatmap /= np.max(heatmap) #normalizes the heatmap to 0-1 #do I actually want to normalize the same way I normed my images? (x-mean)/std. or min-max? heatmap = resize(heatmap,(test_mri[img_number].shape)) print('shape of resized heatmap: ', heatmap.shape) return heatmap def guided_backprop(self, img_data, img_number): """Guided Backpropagation method for visualizing input saliency.""" #define new model which changes gradient fn for all relu activations acording to Guided Backpropagation if "GuidedBackProp" not in ops._gradient_registry._registry: @ops.RegisterGradient("GuidedBackProp") def _GuidedBackProp(op, grad): dtype = op.inputs[0].dtype return grad * tf.cast(grad > 0., dtype) * \ tf.cast(op.inputs[0] > 0., dtype) g = tf.get_default_graph() with g.gradient_override_map({'Relu': 'GuidedBackProp'}): new_model = self.model test_mri, test_jac, test_xls, test_labels, test_ptid, test_imageID, test_confid, test_csf = img_data layer_output = new_model.layers[-15].output grads = K.gradients(layer_output, [new_model.input[0],new_model.input[1],new_model.input[2]])[0] backprop_fn = K.function([new_model.input[0],new_model.input[1],new_model.input[2], K.learning_phase()], [grads]) grads_val = backprop_fn([[test_mri[img_number]],[test_jac[img_number]],[test_xls[img_number]], 0])[0] print('shape of initial gb: ', grads_val.shape) #grads_val = resize(grads_val,(test_mri[img_number].shape)) grads_val = grads_val[0] print('shape of resized gb: ', grads_val.shape) return grads_val class RNN_Net (): def __init__ (self, params): self.params = params self.fc_CNNt1 = Input (shape = (self.params.final_layer_size,)) #Value corresponds to size of final layer in CNN self.fc_CNNt2 = Input (shape = (self.params.final_layer_size,)) self.fc_CNNt3 = Input (shape = (self.params.final_layer_size,)) #self.fc_CNN = Input (shape = (3,self.params.final_layer_size,)) #for rnn_bgrus_multiTP rnn = rnn_bgrus(drop_rate=self.params.RNN_drop_rate, final_layer_size = self.params.final_layer_size, kernel_regularizer=self.params.RNN_w_regularizer) with tf.device(self.params.gpu): self.fc_RNN = Lambda(rnn, name='rnn')([self.fc_CNNt1,self.fc_CNNt2,self.fc_CNNt3]) #original call without prior masking #self.fc_RNN = Lambda(rnn, name='rnn')(self.fc_CNN) #for multi-TP gru print('Shape of self.fc_RNN: ', self.fc_RNN.shape) self.RNNoutput_class= Dense(units = 2, activation = 'softmax', name = 'RNNclass_output') (self.fc_RNN) #switch to sigmoid from softmax? (for 2 class? for non-sparse categorical?) #back to softmax for multi-class def train (self, data): train_data, train_labels, val_data, val_labels = data #data is now loaded in keeping all scans from same patient aligned across timepoints. #That way train_labels can be just one array which applies to all timepoints (necessary because I only ask for 1 output when I define the model!) print('train data shape: ', train_data[0].shape) print('train labels shape: ', train_labels.shape) # self.fc_CNNt1,self.fc_CNNt2,self.fc_CNNt3, train_dataT1,train_dataT2,train_dataT3 = train_data train_samples = train_data[0].shape[0] # val_samples = len(val_data[0]) self.model = Model(inputs = [self.fc_CNNt1,self.fc_CNNt2,self.fc_CNNt3], outputs = [self.RNNoutput_class]) #self.model = Model(inputs = [self.fc_CNN], outputs = [self.RNNoutput_class]) #for multi-TP gru lrate = LearningRateScheduler(step_decay_RNN) callback = [lrate] #optimizer = Adam(lr=1e-5) self.optimizer = self.params.optimizer self.model.compile(optimizer = self.optimizer, loss = 'sparse_categorical_crossentropy', metrics =['acc']) self.model.summary() history = self.model.fit(x = train_data, y = train_labels, batch_size = self.params.RNN_batch_size, #can change this now ...previously seemed to have to be 1...? Otherwise got a mismatch in my last layer being [1,2] instead of [None,2] epochs = self.params.epochs, callbacks = callback, shuffle = True, verbose = 1, #steps_per_epoch = int(train_samples/self.params.batch_size))#, validation_data = (val_data,val_labels)) #Save the model save_model(self.model,'SavedRNNModel') self.model.save_weights('SavedRNNWeights') return history.history def predict (self, data_test): test_data, test_labels = data_test # with open('//data/data_wnx3/data_wnx1/_Data/AlzheimersDL/CNN+RNN-2class/figchecks/test_data.txt', 'w') as testdata: # print(test_mri) print('shape of test_predsT1 inside predict: ' , test_data[0].shape) # print('test_predsT1 inside predict: ' , test_data[0]) preds = self.model.predict (test_data, batch_size=self.params.RNN_batch_size) return preds def evaluate (self, data_test): # test_mri, test_jac, test_xls, test_labels = data_test test_data, test_labels = data_test metrics = self.model.evaluate (x = test_data, y = test_labels, batch_size = self.params.RNN_batch_size) #self.params.batch_size? return metrics def load_the_weights (self, SavedWeights): self.model = Model(inputs = [self.fc_CNNt1,self.fc_CNNt2,self.fc_CNNt3], outputs = [self.RNNoutput_class]) self.model.compile(optimizer = self.params.optimizer, loss = 'sparse_categorical_crossentropy', metrics =['acc']) loaded = self.model.load_weights(SavedWeights) return loaded def XAlex3D(w_regularizer = None, drop_rate = 0., final_layer_size = 50) : #3D Multi-modal deep learning neural network (refer to fig. 4 for chain graph of architecture) ###Create the CNN architecture def f(mri_volume, mri_volume_jacobian, clinical_inputs): #First conv layers conv1_left = _conv_bn_relu_pool_drop(192, 11, 13, 11, strides = (4, 4, 4), w_regularizer = w_regularizer,drop_rate = drop_rate, pool=True) (mri_volume) #conv1_right = _conv_bn_relu_pool_drop(48, 15, 18, 15, strides = (4, 4, 4), w_regularizer = w_regularizer,drop_rate = drop_rate, pool=True) (mri_volume_jacobian) #Second layer conv2_left =_conv_bn_relu_pool_drop(384, 5, 6, 5, w_regularizer = w_regularizer, drop_rate = drop_rate, pool=True) (conv1_left) #conv2_right =_conv_bn_relu_pool_drop(96, 5, 6, 5, w_regularizer = w_regularizer, drop_rate = drop_rate, pool=True) (conv1_right) #conv2_concat = concatenate([conv2_left, conv2_right], axis = -1) #Third layer #conv3_left =_conv_bn_relu_pool_drop(96, 3, 4, 3, w_regularizer = w_regularizer, drop_rate = drop_rate, pool=True) (conv2_left) #conv3_right =_conv_bn_relu_pool_drop(96, 3, 4, 3, w_regularizer = w_regularizer, drop_rate = drop_rate, pool=True) (conv2_right) #conv3_concat = concatenate([conv2_left, conv2_right], axis = -1) #Introduce Middle Flow (separable convolutions with a residual connection) print('residual shape '+str(conv2_left.shape)) conv_mid_1 = mid_flow (conv2_left, drop_rate, w_regularizer, filters = 384) #changed input to conv2_left from conv2_concat # conv_mid_2 = mid_flow (conv_mid_1, drop_rate, w_regularizer, filters = 192) #Split channels for grouped-style convolution conv_mid_1_1 = Lambda (lambda x:x[:,:,:,:,:192]) (conv_mid_1 ) conv_mid_1_2 = Lambda (lambda x:x[:,:,:,:,192:]) (conv_mid_1 ) conv5_left = _conv_bn_relu_pool_drop (96, 3, 4, 3, w_regularizer = w_regularizer, drop_rate = drop_rate, pool=True) (conv_mid_1_1) conv5_right = _conv_bn_relu_pool_drop (96, 3, 4, 3, w_regularizer = w_regularizer, drop_rate = drop_rate, pool=True) (conv_mid_1_2) conv6_left = _conv_bn_relu_pool_drop (48, 3, 4, 3, w_regularizer = w_regularizer,drop_rate = drop_rate, pool=True) (conv5_left) conv6_right = _conv_bn_relu_pool_drop (48, 3, 4, 3, w_regularizer = w_regularizer,drop_rate = drop_rate, pool=True) (conv5_right) #conv7_left = _conv_bn_relu_pool_drop (16, 3, 4, 3, w_regularizer = w_regularizer,drop_rate = drop_rate, pool=True) (conv6_left) #conv7_right = _conv_bn_relu_pool_drop (16, 3, 4, 3, w_regularizer = w_regularizer,drop_rate = drop_rate, pool=True) (conv6_right) conv6_concat = concatenate([conv6_left, conv6_right], axis = -1) #convExtra = Conv3D(48, (20,30,20), # strides = (1,1,1), kernel_initializer="he_normal", # padding="same", kernel_regularizer = w_regularizer)(conv6_concat) #Flatten 3D conv network representations flat_conv_6 = Reshape((np.prod(K.int_shape(conv6_concat)[1:]),))(conv6_concat) #2-layer Dense network for clinical features vol_fc1 = _fc_bn_relu_drop(64, w_regularizer = w_regularizer, drop_rate = drop_rate)(clinical_inputs) flat_volume = _fc_bn_relu_drop(20, w_regularizer = w_regularizer, drop_rate = drop_rate)(vol_fc1) #Combine image and clinical features embeddings fc1 = _fc_bn_relu_drop (20, w_regularizer, drop_rate = drop_rate, name='final_conv') (flat_conv_6) #fc2 = _fc_bn_relu_drop (40, w_regularizer, drop_rate = drop_rate) (fc1) flat = concatenate([fc1, flat_volume]) #Final 4D embedding fc2 = Dense(units = final_layer_size, activation = 'linear', kernel_regularizer=w_regularizer, name='features') (flat) #was linear activation #fc2 = _fc_bn_relu_drop (final_layer_size, w_regularizer, drop_rate = drop_rate) (flat) #this was the orginal final layer return fc2 return f ###Define pieces of CNN def _fc_bn_relu_drop (units, w_regularizer = None, drop_rate = 0., name = None): #Defines Fully connected block (see fig. 3 in paper) def f(input): fc = Dense(units = units, activation = 'linear', kernel_regularizer=w_regularizer, name = name) (input) #was linear activation fc = BatchNormalization()(fc) fc = ELU()(fc) fc = Dropout (drop_rate) (fc) return fc return f def _conv_bn_relu_pool_drop(filters, height, width, depth, strides=(1, 1, 1), padding = 'same', w_regularizer = None, drop_rate = None, name = None, pool = False): #Defines convolutional block (see fig. 3 in paper) def f(input): conv = Conv3D(filters, (height, width, depth), strides = strides, kernel_initializer="he_normal", padding=padding, kernel_regularizer = w_regularizer, name = name)(input) norm = BatchNormalization()(conv) elu = ELU()(norm) if pool == True: elu = MaxPooling3D(pool_size=3, strides=2, padding = 'same') (elu) return Dropout(drop_rate) (elu) return f def _sepconv_bn_relu_pool_drop (filters, height, width, depth, strides = (1, 1, 1), padding = 'same', depth_multiplier = 1, w_regularizer = None, drop_rate = None, name = None, pool = False): #Defines separable convolutional block (see fig. 3 in paper) def f (input): sep_conv = SeparableConv3D(filters, (height, width, depth), strides = strides, depth_multiplier = depth_multiplier,kernel_initializer="he_normal", padding=padding, kernel_regularizer = w_regularizer, name = name)(input) sep_conv = BatchNormalization()(sep_conv) elu = ELU()(sep_conv) if pool == True: elu = MaxPooling2D(pool_size=3, strides=2, padding = 'same') (elu) return Dropout(drop_rate) (elu) return f def mid_flow (x, drop_rate, w_regularizer, filters): #3 consecutive separable blocks with a residual connection (refer to fig. 4) residual = x x = _sepconv_bn_relu_pool_drop (filters, 3, 3, 3, padding='same', depth_multiplier = 1, drop_rate=drop_rate, w_regularizer = w_regularizer)(x) x = _sepconv_bn_relu_pool_drop (filters, 3, 3, 3, padding='same', depth_multiplier = 1, drop_rate=drop_rate, w_regularizer = w_regularizer)(x) x = _sepconv_bn_relu_pool_drop (filters, 3, 3, 3, padding='same', depth_multiplier = 1, drop_rate=drop_rate, w_regularizer = w_regularizer)(x) # print('x shape '+str(x.shape)) x = add([x, residual]) x = ELU()(x) return x def step_decay_CNN (epoch): #Decaying learning rate function initial_lrate = 4e-4 drop = 0.3 epochs_drop = 10.0 lrate = initial_lrate * math.pow(drop,((1+epoch)/epochs_drop)) return lrate def step_decay_RNN (epoch): #Decaying learning rate function initial_lrate = 2e-3 drop = 0.3 epochs_drop = 10.0 lrate = initial_lrate * math.pow(drop,((1+epoch)/epochs_drop)) return lrate ###Create the RNN def rnn_bgrus (drop_rate,final_layer_size, kernel_regularizer=None, mask=None): def f(inputs): fc2T1 = inputs[0] print ('fc2T1_ogShape: ', fc2T1.shape) print ('fc2T1_ogShape[0]: ', fc2T1.shape[0]) fc2T2 = inputs[1] fc2T3 = inputs[2] batch = K.shape(fc2T1)[0] #just the number of samples in T1 (which is same as T2 and T3) unitsA = 100 unitsB = 100 unitsC = 100 #reshape fc2T1 = tf.reshape(fc2T1,(batch,1,final_layer_size)) #GRU needs input shape: [batch(aka num_samples), timesteps, feature] should first 1 be params.batch_size? fc2T2 = tf.reshape(fc2T2,(batch,1,final_layer_size)) fc2T3 = tf.reshape(fc2T3,(batch,1,final_layer_size)) #Add masking layer to handle missing data fc2T1_mask = Masking(mask_value=-1, input_shape=(1,final_layer_size))(fc2T1) #needs input shape of (samples, timesteps, features) fc2T2_mask = Masking(mask_value=-1, input_shape=(1,final_layer_size))(fc2T2) #should give output shape of (samples,timesteps) fc2T3_mask = Masking(mask_value=-1, input_shape=(1,final_layer_size))(fc2T3) print('fc2T1_masked: ', fc2T1_mask) # first BGRU (a) a_forwardT1 = GRU(unitsA,activation='tanh',dropout=drop_rate,kernel_regularizer=kernel_regularizer)(fc2T1_mask) #output shape = (batch_size, timesteps, units) a_backwardT1 = GRU(unitsA, activation='tanh', go_backwards=True,dropout=drop_rate,kernel_regularizer=kernel_regularizer)(fc2T1_mask) a_forwardT2 = GRU(unitsA,activation='tanh',dropout=drop_rate,kernel_regularizer=kernel_regularizer)(fc2T2_mask) a_backwardT2 = GRU(unitsA, activation='tanh', go_backwards=True,dropout=drop_rate,kernel_regularizer=kernel_regularizer)(fc2T2_mask) a_forwardT3 = GRU(unitsA,activation='tanh',dropout=drop_rate,kernel_regularizer=kernel_regularizer)(fc2T3_mask) a_backwardT3 = GRU(unitsA, activation='tanh', go_backwards=True,dropout=drop_rate,kernel_regularizer=kernel_regularizer)(fc2T3_mask) a_gruT1 = concatenate([a_forwardT1, a_backwardT1], axis=-1) a_gruT2 = concatenate([a_forwardT2, a_backwardT2], axis=-1) a_gruT3 = concatenate([a_forwardT3, a_backwardT3], axis=-1) #reshape a_gruT1 = tf.reshape(a_gruT1,(batch,1,unitsA*2)) #had 1,200,1...why?? a_gruT2 = tf.reshape(a_gruT2,(batch,1,unitsA*2)) a_gruT3 = tf.reshape(a_gruT3,(batch,1,unitsA*2)) # second BGRU (b) b_forwardT1 = GRU(unitsB,activation='tanh',dropout=drop_rate,kernel_regularizer=kernel_regularizer)(a_gruT1) #does this propagate the mask??? Does it need to anymore?? b_backwardT1 = GRU(unitsB, activation='tanh', go_backwards=True,dropout=drop_rate,kernel_regularizer=kernel_regularizer)(a_gruT1) b_forwardT2 = GRU(unitsB,activation='tanh',dropout=drop_rate,kernel_regularizer=kernel_regularizer)(a_gruT2) b_backwardT2 = GRU(unitsB, activation='tanh', go_backwards=True,dropout=drop_rate,kernel_regularizer=kernel_regularizer)(a_gruT2) b_forwardT3 = GRU(unitsB,activation='tanh',dropout=drop_rate,kernel_regularizer=kernel_regularizer)(a_gruT3) b_backwardT3 = GRU(unitsB, activation='tanh', go_backwards=True,dropout=drop_rate,kernel_regularizer=kernel_regularizer)(a_gruT3) b_gruT1 = concatenate([b_forwardT1, b_backwardT1], axis=-1) b_gruT2 = concatenate([b_forwardT2, b_backwardT2], axis=-1) b_gruT3 = concatenate([b_forwardT3, b_backwardT3], axis=-1) #reshape b_gruT1 = tf.reshape(b_gruT1,(batch,1,unitsB*2)) b_gruT2 = tf.reshape(b_gruT2,(batch,1,unitsB*2)) b_gruT3 = tf.reshape(b_gruT3,(batch,1,unitsB*2)) ##ADD a dropout layer or two; or add dropout to GRU (see documentation) # third BGRU (c) c_forwardT1 = GRU(unitsC,activation='tanh',dropout=drop_rate,kernel_regularizer=kernel_regularizer)(b_gruT1) c_backwardT1 = GRU(unitsC, activation='tanh', go_backwards=True,dropout=drop_rate,kernel_regularizer=kernel_regularizer)(b_gruT1) c_forwardT2 = GRU(unitsC,activation='tanh',dropout=drop_rate,kernel_regularizer=kernel_regularizer)(b_gruT2) c_backwardT2 = GRU(unitsC, activation='tanh', go_backwards=True,dropout=drop_rate,kernel_regularizer=kernel_regularizer)(b_gruT2) c_forwardT3 = GRU(unitsC,activation='tanh',dropout=drop_rate,kernel_regularizer=kernel_regularizer)(b_gruT3) c_backwardT3 = GRU(unitsC, activation='tanh', go_backwards=True,dropout=drop_rate,kernel_regularizer=kernel_regularizer)(b_gruT3) c_gruT1 = concatenate([c_forwardT1, c_backwardT1], axis=-1) c_gruT2 = concatenate([c_forwardT2, c_backwardT2], axis=-1) c_gruT3 = concatenate([c_forwardT3, c_backwardT3], axis=-1) #reshape #c_gruT1 = tf.reshape(c_gruT1,(batch,1,unitsC*2)) #c_gruT2 = tf.reshape(c_gruT2,(batch,1,unitsC*2)) #c_gruT3 = tf.reshape(c_gruT3,(batch,1,unitsC*2)) # fourth BGRU (d) #d_forwardT1 = GRU(unitsC,activation='tanh')(c_gruT1) #d_backwardT1 = GRU(unitsC, activation='tanh', go_backwards=True)(c_gruT1) #d_forwardT2 = GRU(unitsC,activation='tanh')(c_gruT2) #d_backwardT2 = GRU(unitsC, activation='tanh', go_backwards=True)(c_gruT2) #d_forwardT3 = GRU(unitsC,activation='tanh')(c_gruT3) #d_backwardT3 = GRU(unitsC, activation='tanh', go_backwards=True)(c_gruT3) #d_gruT1 = concatenate([d_forwardT1, d_backwardT1], axis=-1) #d_gruT2 = concatenate([d_forwardT2, d_backwardT2], axis=-1) #d_gruT3 = concatenate([d_forwardT3, d_backwardT3], axis=-1) #concatenate final BGRU output bgru_total = concatenate([c_gruT1, c_gruT2, c_gruT3], axis=-1) #Fully connected layer rnn_fc1 = Dense(units = 20, activation = 'linear', name = 'RNNfcFinal') (bgru_total) rnn_fc1 = Dropout(drop_rate) (rnn_fc1) return rnn_fc1 return f