def init_param(self, init_values: to.Tensor = None, **kwargs):
if init_values is None:
self.shared.init_param(None, **kwargs)
init_param(self.head_1, **kwargs)
init_param(self.head_2, **kwargs)
else:
self.param_values = init_values
def init_param(self, init_values: to.Tensor = None, **kwargs):
if init_values is None:
# Initialize the layers using default initialization
init_param(self.rnn_layers, **kwargs)
init_param(self.output_layer, **kwargs)
else:
self.param_values = init_values
def init_param(self, init_values: to.Tensor = None, **kwargs):
# See RNNPolicyBase
if init_values is None:
# Initialize the layers using default initialization
init_param(self.rnn_layers, **kwargs)
init_param(self.output_layer, **kwargs)
else:
cp.vector_to_parameters(init_values, self.parameters())
def init_param(self, init_values: to.Tensor = None, **kwargs):
if init_values is None:
if isinstance(self.net, FNN):
# Forward to the FNN's custom initialization function (handles dropout)
self.net.init_param(init_values, **kwargs)
else:
# Initialize using default initialization
init_param(self.net, **kwargs)
else:
self.param_values = init_values
def init_param(self, init_values: to.Tensor = None, **kwargs):
if init_values is None:
# Initialize common layers
init_param(self.obs_layer, **kwargs)
self.resting_level.data = to.randn_like(self.resting_level.data)
# Initialize time constant and potentials if learnable
if self.tau_learnable:
self._log_tau.data = self._log_tau_init
if self.potential_init_learnable:
self._potentials_init.data = to.randn(self.hidden_size)
else:
self.param_values = init_values
def init_param(self, init_values: to.Tensor = None, **kwargs):
super().init_param(init_values, **kwargs)
if init_values is None:
# Initialize layers
init_param(self.prev_act_layer, **kwargs)
if kwargs.get("sigmoid_nlin", False):
self.prev_act_layer.weight.data.fill_(-0.5) # inhibit others
for i in range(self.prev_act_layer.weight.data.shape[0]):
self.prev_act_layer.weight.data[i, i] = 1.0 # excite self
init_param(self.pot_to_act_layer, **kwargs)
# Initialize cubic decay and capacity if learnable
if self.potentials_dot_fcn == pd_cubic and self.kappa_learnable:
self._log_kappa.data = self._log_kappa_init
elif self.potentials_dot_fcn in [pd_capacity_21, pd_capacity_21_abs, pd_capacity_32, pd_capacity_32_abs]:
if self.capacity_learnable:
self._log_capacity.data = self._log_capacity_init
else:
self.param_values = init_values
def init_param(self, init_values: to.Tensor = None, **kwargs):
if init_values is None:
# Initialize RNN layers
init_param(self.obs_layer, **kwargs)
init_param(self.prev_act_layer, **kwargs)
if kwargs.get('sigmoid_nlin', False):
self.prev_act_layer.weight.data.fill_(-0.5) # inhibit others
for i in range(self.prev_act_layer.weight.data.shape[0]):
self.prev_act_layer.weight.data[i, i] = 1. # excite self
init_param(self.nonlin_layer, **kwargs)
# Initialize time constant if modifiable
if self.tau_learnable:
self._log_tau.data = self._log_tau_init
# Initialize cubic decay if modifiable
if self.potentials_dot_fcn == pd_cubic:
if self.kappa_learnable:
self._log_kappa.data = self._log_kappa_init
# Initialize capacity if modifiable
elif self.potentials_dot_fcn in [
pd_capacity_21, pd_capacity_21_abs, pd_capacity_32,
pd_capacity_32_abs
]:
if self.capacity_learnable:
self._log_capacity.data = self._log_capacity_init
else:
self.param_values = init_values
def init_param(self, init_values: to.Tensor = None, **kwargs):
if init_values is None:
# Initialize the layers using default initialization
init_param(self.shared, **kwargs)
init_param(self.head_1, **kwargs)
init_param(self.head_2, **kwargs)
else:
self.param_values = init_values
def init_param(self, init_values: to.Tensor = None, **kwargs):
if init_values is None:
# Initialize RNN layers
init_param(self.obs_layer, **kwargs)
# self.obs_layer.weight.data /= 100.
# self.obs_layer.bias.data /= 100.
self.resting_level.data = to.randn_like(self.resting_level.data)
init_param(self.conv_layer, **kwargs)
# init_param(self.post_conv_layer, **kwargs)
init_param(self.nonlin_layer, **kwargs)
init_param(self.act_layer, **kwargs)
# Initialize time constant if modifiable
if self.tau_learnable:
self._log_tau.data = self._log_tau_init
else:
self.param_values = init_values
def init_param(self, init_values: to.Tensor = None, **kwargs):
super().init_param(init_values, **kwargs)
if init_values is None:
# Initialize layers
init_param(self.conv_layer, **kwargs)
# init_param(self.post_conv_layer, **kwargs)
init_param(self.pot_to_activ, **kwargs)
init_param(self.act_layer, **kwargs)
else:
self.param_values = init_values
def init_param(self, init_values: to.Tensor = None, **kwargs):
"""
Initialize the network's parameters. By default the parameters are initialized randomly.
:param init_values: Tensor of fixed initial network parameter values
"""
if init_values is None:
# Initialize hidden layers
for i, layer in enumerate(self.hidden_layers):
if self.dropout == 0:
# If there is no dropout, initialize weights and biases for every layer
init_param(layer, **kwargs)
elif self.dropout > 0 and i%2 == 0:
# If there is dropout, omit the initialization for the dropout layers
init_param(layer, **kwargs)
# Initialize output layer
init_param(self.output_layer, **kwargs)
else:
self.param_values = init_values
def init_param(self, init_values: to.Tensor = None, **kwargs):
if init_values is None:
# Initialize the layer using default initialization
init_param(self.net, **kwargs)
else:
self.param_values = init_values # ignore the IntelliJ warning
print(
f'ptwise_conv_layer weights shape: {ptwise_conv_layer.weight.shape}'
)
else:
# Standard way
conv_layer = nn.Conv1d(in_channels,
out_channels,
kernel_size,
stride=1,
padding=padding,
dilation=1,
groups=1,
bias=False,
padding_mode=padding_mode)
init_param(conv_layer, bell=use_custom_bell_init)
print(f'conv_layer weights shape: {conv_layer.weight.shape}')
# A ramp filter
if hand_coded_filter:
conv_layer.weight.data = to.linspace(0, 1, kernel_size).repeat(
2, 1).unsqueeze(0)
# Mirrored weighs
elif use_custom_mirr_layer:
conv_layer = MirrConv1d(in_channels,
out_channels,
kernel_size,
stride=1,
padding=padding,
dilation=1,
def init_param(self, init_values: to.Tensor = None, **kwargs):
if init_values is None:
init_param(self.conds, **kwargs)
else:
self.param_values = init_values # ignore the IntelliJ warning
