def __init__(self, inputs):
self.control_force_spacing = F_max/self.num_of_control_resolution_oneside
super(DQN_measurement, self).__init__()
self.inputs=inputs
self.Batch_Normalize = False
k1 = 13; s1 = 5
k2 = 11; s2 = 4
k3 = 9; s3 = 4
filters = (32, 64, 64)
self.conv1=nn.Conv1d(2, filters[0], kernel_size = k1, stride = s1, padding = 0, bias = not self.Batch_Normalize)
if self.Batch_Normalize: self.bn1=layers.BatchRenorm1d(filters[0])
self.conv2=nn.Conv1d(filters[0], filters[1], kernel_size = k2, stride = s2, padding = 0, bias = not self.Batch_Normalize)
if self.Batch_Normalize: self.bn2=layers.BatchRenorm1d(filters[1])
self.conv3=nn.Conv1d(filters[1], filters[2], kernel_size = k3, stride = s3, padding = 0, bias = not self.Batch_Normalize)
if self.Batch_Normalize: self.bn3=layers.BatchRenorm1d(filters[2])
if not self.Batch_Normalize:
for conv in [self.conv1, self.conv2, self.conv3]:
conv.bias.data.zero_()
def calculate_next_layer_dim(number, k, s):
return (number-(k-1)+(s-1)) // s
number = calculate_next_layer_dim(inputs, k1, s1)
number = calculate_next_layer_dim(number, k2, s2)
number = calculate_next_layer_dim(number, k3, s3)
self.fc1=nn.Linear(number * filters[2], 256)
self.fc21=layers.FactorizedNoisy(256, 256)
self.fc31=layers.FactorizedNoisy(256, self.num_of_control_resolution_oneside*2 + 1) # left and right means doubling, 1 extra for no control
self.fc22=layers.Linear_weight_normalize(256, 128)
self.fc32=layers.Linear_weight_normalize(128, 1) # and 1 for mean prediction
print(' + Number of params: {}'.format(sum([p.data.nelement() for p in self.parameters()])))
def __init__(self, data_length, noisy_layers = 2):
self.control_force_spacing = F_max/self.num_of_control_resolution_oneside
super(direct_DQN, self).__init__()
self.inputs={'data_length':data_length, 'noisy_layers':noisy_layers}
self.fc1=layers.Linear_weight_normalize(data_length, 512)
self.fc2=layers.Linear_weight_normalize(512, 512)
if noisy_layers >= 2:
self.fc31=layers.FactorizedNoisy(512, 256)
else:
self.fc31=layers.Linear_weight_normalize(512, 256)
self.fc32=layers.Linear_weight_normalize(512, 128)
if noisy_layers >= 1:
self.fc41=layers.FactorizedNoisy(256, self.num_of_control_resolution_oneside*2 + 1) # left and right, and 1 extra for no control
else:
self.fc41=layers.Linear_weight_normalize(256, self.num_of_control_resolution_oneside*2 + 1)
self.fc42=layers.Linear_weight_normalize(128, 1) # mean prediction
print(' + Number of params: {}'.format(sum([p.data.nelement() for p in self.parameters()])))