def apply_to_model(self, model, reference_weights, percentage):
initial_weights_vector = utils.weights_to_vector(reference_weights)
current_weights_vector = utils.weights_to_vector(
utils.get_model_weights(model))
diff_vector = np.abs(initial_weights_vector - current_weights_vector)
mask_vector = self._ranker(diff_vector, percentage)
last_idx = 0
layer_masks = []
reinitialized_layers = []
for layer in model.layers:
if len(layer.get_weights()) > 0:
kernel, bias = layer.get_weights()[0], layer.get_weights()[1]
kernel_mask = mask_vector[last_idx:last_idx +
kernel.size].reshape(kernel.shape)
last_idx += kernel.size
bias_mask = mask_vector[last_idx:last_idx + bias.size].reshape(
bias.shape)
last_idx += bias.size
kernel_reinitialized = self._kernel_initializer(kernel.shape)
bias_reinitialized = self._bias_initializer(bias.shape)
new_kernel = np.copy(kernel)
new_bias = np.copy(bias)
new_kernel[kernel_mask] = kernel_reinitialized[kernel_mask]
new_bias[bias_mask] = bias_reinitialized[bias_mask]
layer.set_weights((new_kernel, new_bias))
layer_masks.append([kernel_mask, bias_mask])
reinitialized_layers.append(
[kernel_reinitialized, bias_reinitialized])
else:
layer_masks.append(None)
reinitialized_layers.append(None)
return layer_masks, reinitialized_layers
def apply_to_model(self, model, reference_weights, percentage):
weights_vector = utils.weights_to_vector(
utils.get_model_weights(model))
mask_vector = self._ranker(weights_vector, percentage)
last_idx = 0
layer_masks = []
reinitialized_layers = []
for i in range(len(model.layers)):
layer = model.layers[i]
if len(layer.get_weights()) > 0:
kernel, bias = layer.get_weights()[0], layer.get_weights()[1]
kernel_mask = mask_vector[last_idx:last_idx +
kernel.size].reshape(kernel.shape)
last_idx += kernel.size
bias_mask = mask_vector[last_idx:last_idx + bias.size].reshape(
bias.shape)
last_idx += bias.size
kernel_reinitialized = self._kernel_initializer(kernel.shape)
bias_reinitialized = self._bias_initializer(bias.shape)
new_kernel = np.copy(reference_weights[i][0])
new_bias = np.copy(reference_weights[i][1])
new_kernel[kernel_mask] = kernel_reinitialized[kernel_mask]
new_bias[bias_mask] = bias_reinitialized[bias_mask]
layer.set_weights((new_kernel, new_bias))
layer_masks.append([kernel_mask, bias_mask])
reinitialized_layers.append(
[kernel_reinitialized, bias_reinitialized])
else:
layer_masks.append(None)
reinitialized_layers.append(None)
def __init__(self, model, reinitializer, mask=None, percentage=None):
super().__init__(model, reinitializer, mask, percentage)
self._reference_weights = utils.get_model_weights(self._model)
def reinitialize(self, percentage=None):
if percentage is None:
percentage = self._percentage
self._reinitializer.apply_to_model(self._model,
self._reference_weights, percentage)
self._reference_weights = utils.get_model_weights(self._model)