def __init__(self, nn_params):
super(ImitativePolicy, self).__init__()
# Store the parameters:
self.hidden_w = nn_params['hid_width']
self.depth = nn_params['hid_depth']
self.n_in_input = nn_params['dx']
self.n_out = nn_params['du']
self.activation = nn_params['activation']
self.d = nn_params['dropout']
self.loss_fnc = nn.MSELoss()
# super(ImitativePolicy, self).__init__()
# Takes objects from the training parameters
layers = []
layers.append(nn.Linear(self.n_in_input,
self.hidden_w)) # input layer
layers.append(self.activation)
layers.append(nn.Dropout(p=self.d))
for d in range(self.depth):
# add modules
# input layer
layers.append(nn.Linear(self.hidden_w, self.hidden_w))
layers.append(self.activation)
layers.append(nn.Dropout(p=self.d))
# output layer
layers.append(nn.Linear(self.hidden_w, self.n_out))
self.features = nn.Sequential(*layers)
# Need to scale the state variables again etc
# inputs state, output an action (PWMs)
self.scalarX = MinMaxScaler(feature_range=(-1, 1))
self.scalarU = MinMaxScaler(feature_range=(-1, 1))
def __init__(self, nn_params):
super(GeneralNN, self).__init__()
"""
Simpler implementation of my other neural net class. After parameter tuning, now just keep the structure and change it if needed. Note that the data passed into this network is only that which is used.
"""
# Store the parameters:
self.prob = nn_params['bayesian_flag']
self.hidden_w = nn_params['hid_width']
self.depth = nn_params['hid_depth']
self.n_in_input = nn_params['du']
self.n_in_state = nn_params['dx']
self.n_in = self.n_in_input + self.n_in_state
self.n_out = nn_params['dt']
self.activation = nn_params['activation']
self.d = nn_params['dropout']
self.split_flag = nn_params['split_flag']
self.E = 0 # clarify that these models are not ensembles
# Can store with a helper function for when re-loading and figuring out what was trained on
self.state_list = []
self.input_list = []
self.change_state_list = []
self.scalarX = StandardScaler()# MinMaxScaler(feature_range=(-1,1))#StandardScaler()# RobustScaler()
self.scalarU = MinMaxScaler(feature_range=(-1,1))
self.scalardX = MinMaxScaler(feature_range=(-1,1)) #StandardScaler() #MinMaxScaler(feature_range=(-1,1))#StandardScaler() # RobustScaler(quantile_range=(25.0, 90.0))
# Sets loss function
if self.prob:
# INIT max/minlogvar if PNN
self.max_logvar = torch.nn.Parameter(torch.tensor(1*np.ones([1, self.n_out]),dtype=torch.float, requires_grad=True))
self.min_logvar = torch.nn.Parameter(torch.tensor(-1*np.ones([1, self.n_out]),dtype=torch.float, requires_grad=True))
self.loss_fnc = PNNLoss_Gaussian()
self.n_out *= 2
else:
self.loss_fnc = nn.MSELoss()
# If using split model, initiate here:
if self.split_flag:
self.features = nn.Sequential(
SplitModel(self.n_in, self.n_out,
prob = self.prob,
width = self.hidden_w,
activation = self.activation,
dropout = self.d))
else:
# create object nicely
layers = []
layers.append(('dynm_input_lin', nn.Linear(
self.n_in, self.hidden_w))) # input layer
layers.append(('dynm_input_act', self.activation))
# layers.append(nn.Dropout(p=self.d))
for d in range(self.depth):
# add modules
# input layer
layers.append(
('dynm_lin_'+str(d), nn.Linear(self.hidden_w, self.hidden_w)))
layers.append(('dynm_act_'+str(d), self.activation))
# layers.append(nn.Dropout(p=self.d))
# output layer
layers.append(('dynm_out_lin', nn.Linear(self.hidden_w, self.n_out)))
# print(*layers)
self.features = nn.Sequential(OrderedDict([*layers]))
temp_data_test = DataCSV_All('data/tempAMAL_test.csv', number_classes,
sequence_length)
data = DataLoader(temp_data_train,
batch_size=batch_size,
shuffle=True,
drop_last=True)
data_test = DataLoader(temp_data_test,
batch_size=1,
shuffle=False,
drop_last=True)
data_sizes = temp_data_train.data.size()
latent_size = 10
model = RNN(1, latent_size, 1)
loss = nn.MSELoss()
optim = torch.optim.Adam(model.parameters(), lr=10**-3)
iterations = 1000
#GPU
model.to(device)
loss.to(device)
writer = SummaryWriter()
for i in range(iterations):
train_loss = 0
nt = 0
for x in data:
def __init__(self, nn_params):
super(GeneralNN, self).__init__()
# Store the parameters:
self.prob = nn_params['bayesian_flag']
self.hidden_w = nn_params['hid_width']
self.depth = nn_params['hid_depth']
self.n_in_input = nn_params['du']
self.n_in_state = nn_params['dx']
self.n_in = self.n_in_input + self.n_in_state
self.n_out = nn_params['dt']
self.activation = nn_params['activation']
self.d = nn_params['dropout']
self.split_flag = nn_params['split_flag']
self.epsilon = nn_params['epsilon']
self.E = 0
# Can store with a helper function for when re-loading and figuring out what was trained on
self.state_list = []
self.input_list = []
self.change_state_list = []
self.scalarX = StandardScaler(
) # MinMaxScaler(feature_range=(-1,1))#StandardScaler()# RobustScaler()
self.scalarU = MinMaxScaler(feature_range=(-1, 1))
self.scalardX = MinMaxScaler(
feature_range=(-1, 1)
) #StandardScaler() #MinMaxScaler(feature_range=(-1,1))#StandardScaler() # RobustScaler(quantile_range=(25.0, 90.0))
# Sets loss function
if self.prob:
# INIT max/minlogvar if PNN
self.max_logvar = torch.nn.Parameter(
torch.tensor(1 * np.ones([1, self.n_out]),
dtype=torch.float,
requires_grad=True))
self.min_logvar = torch.nn.Parameter(
torch.tensor(-1 * np.ones([1, self.n_out]),
dtype=torch.float,
requires_grad=True))
self.loss_fnc = PNNLoss_Gaussian()
self.n_out *= 2
else:
self.loss_fnc = nn.MSELoss()
layers = []
layers.append(
('dynm_input_lin', nn.Linear(self.n_in,
self.hidden_w))) # input layer
layers.append(('dynm_input_act', self.activation))
layers.append(('dynm_input_dropout', nn.Dropout(p=self.d)))
for d in range(self.depth):
layers.append(
('dynm_lin_' + str(d), nn.Linear(self.hidden_w,
self.hidden_w)))
layers.append(('dynm_act_' + str(d), self.activation))
layers.append(('dynm_dropout_' + str(d), nn.Dropout(p=self.d)))
layers.append(('dynm_out_lin', nn.Linear(self.hidden_w, self.n_out)))
self.features = nn.Sequential(OrderedDict([*layers]))
