weights_reg=compute_class_weights(Class)
# weights for the segmentation labels
Class=np.loadtxt('/Tmp/cucurulg/datasets/cortical_layers/6layers_segmentation/training_cls.txt')
weights_seg=compute_class_weights(Class)
# Define training functions
print "Defining and compiling training functions"
prediction_seg, prediction_reg = lasagne.layers.get_output([
simple_net_output[0],
simple_net_output[1]])
# Loss function
loss_reg = weighted_crossentropy(prediction_reg, target_var_reg, weight_vector_reg)
loss_reg = loss_reg.mean()
loss_seg = weighted_crossentropy(prediction_seg, target_var_seg, weight_vector_seg)
loss_seg = loss_seg.mean()
loss = loss_reg+loss_seg
# Add regularization
if weight_decay > 0:
weightsl2 = regularize_network_params(
simple_net_output, lasagne.regularization.l2)
loss += weight_decay * weightsl2
# train accuracy
train_acc_reg = accuracy_regions(prediction_reg, target_var_reg)
freq=counts.astype(float)/Class.size
return np.median(freq)/freq
weights=compute_class_weights(Class)
# Define training functions
print "Defining and compiling training functions"
prediction = lasagne.layers.get_output(simple_net_output[0])
# only for DEBUG!!!!
# debug_prediction = theano.function([input_var], prediction)
# Loss function
loss = weighted_crossentropy(prediction, target_var, weight_vector)
loss = loss.mean()
# Add regularization
if weight_decay > 0:
weightsl2 = regularize_network_params(
simple_net_output, lasagne.regularization.l2)
loss += weight_decay * weightsl2
# train accuracy
train_acc = accuracy_regions(prediction, target_var)
# Define the update function
params = lasagne.layers.get_all_params(simple_net_output, trainable=True)
updates = lasagne.updates.adam(loss, params, learning_rate=learn_step)