def _back_prop(self, z, a, X_train):
delta = self.loss.delta(X_train, a[self.n_layers])
dw = coo_matrix(self.w[self.n_layers - 1])
# compute backpropagation updates
sparseoperations.backpropagation_updates_Cython(
a[self.n_layers - 1], delta, dw.row, dw.col, dw.data
) # If you have problems with Cython please use the backpropagation_updates_Numpy method by uncommenting the line below and commenting the one above. Please note that the running time will be much higher
# If you have problems with Cython please use the backpropagation_updates_Numpy method by uncommenting the line below and commenting the one above. Please note that the running time will be much higher
#backpropagation_updates_Numpy(a[self.n_layers - 1], delta, dw.row, dw.col, dw.data)
update_params = {self.n_layers - 1: (dw.tocsr(), delta)}
for i in reversed(range(2, self.n_layers)):
delta = (delta
@ self.w[i].transpose()) * self.activations[i].prime(z[i])
dw = coo_matrix(self.w[i - 1])
# compute backpropagation updates
sparseoperations.backpropagation_updates_Cython(
a[i - 1], delta, dw.row, dw.col, dw.data
) # If you have problems with Cython please use the backpropagation_updates_Numpy method by uncommenting the line below and commenting the one above. Please note that the running time will be much higher
# If you have problems with Cython please use the backpropagation_updates_Numpy method by uncommenting the line below and commenting the one above. Please note that the running time will be much higher
#backpropagation_updates_Numpy(a[i - 1], delta, dw.row, dw.col, dw.data)
update_params[i - 1] = (dw.tocsr(), delta)
for k, v in update_params.items():
self._update_w_b(k, v[0], v[1])
def _back_prop(self, z, a, y_true):
"""
The input dicts keys represent the layers of the net.
a = { 1: x,
2: f(w1(x) + b1)
3: f(w2(a2) + b2)
4: f(w3(a3) + b3)
5: f(w4(a4) + b4)
}
:param z: (dict) w(x) + b
:param a: (dict) f(z)
:param y_true: (array) One hot encoded truth vector.
:return:
"""
# Determine partial derivative and delta for the output layer.
# delta output layer
delta = self.loss.delta(y_true, a[self.n_layers])
dw = coo_matrix(self.w[self.n_layers - 1])
# compute backpropagation updates
sparseoperations.backpropagation_updates_Cython(
a[self.n_layers - 1], delta, dw.row, dw.col, dw.data)
# If you have problems with Cython please use the backpropagation_updates_Numpy method by uncommenting the line below and commenting the one above.
# Please note that the running time will be much higher
# backpropagation_updates_Numpy(a[self.n_layers - 1], delta, dw.row, dw.col, dw.data)
update_params = {self.n_layers - 1: (dw.tocsr(), delta)}
# In case of three layer net will iterate over i = 2 and i = 1
# Determine partial derivative and delta for the rest of the layers.
# Each iteration requires the delta from the previous layer, propagating backwards.
for i in reversed(range(2, self.n_layers)):
delta = (delta
@ self.w[i].transpose()) * self.activations[i].prime(z[i])
dw = coo_matrix(self.w[i - 1])
# compute backpropagation updates
sparseoperations.backpropagation_updates_Cython(
a[i - 1], delta, dw.row, dw.col, dw.data)
# If you have problems with Cython please use the backpropagation_updates_Numpy method by uncommenting the line below and commenting the one above.
# Please note that the running time will be much higher
# backpropagation_updates_Numpy(a[i - 1], delta, dw.row, dw.col, dw.data)
update_params[i - 1] = (dw.tocsr(), delta)
for k, v in update_params.items():
self._update_w_b(k, v[0], v[1])