def select_actions(self, states, explore_noise=0.0): # CPU array to GPU tensor to CPU array
states = torch.tensor(states, dtype=torch.float32, device=self.device)
actions = self.act(states)
if explore_noise != 0.0:
pis = self.act.actor(states)
actions = Normal(*pis).sample()
actions = actions.tanh()
actions = actions.cpu().data.numpy()
return actions
with torch.cuda.device(3):
# Do it with CUDA if possible.
device = 'cuda' if torch.cuda.is_available() else 'cpu'
if device == 'cuda':
encd.cuda()
decd.cuda()
disc.cuda()
if len(sys.argv) > 1:
for batch in data.batches():
batch = batch.to(device)
with torch.no_grad():
mu, sd = decd(encd(batch))
y = Normal(mu, sd).sample()
np.savetxt(sys.stdout, y.cpu().numpy(), fmt='%.4f')
sys.exit()
lr = 0.001 # The celebrated learning rate
aopt = torch.optim.Adam(encd.parameters(), lr=lr)
bopt = torch.optim.Adam(decd.parameters(), lr=lr)
copt = torch.optim.Adam(disc.parameters(), lr=lr)
asched = torch.optim.lr_scheduler.MultiStepLR(aopt, [800])
bsched = torch.optim.lr_scheduler.MultiStepLR(bopt, [800])
csched = torch.optim.lr_scheduler.MultiStepLR(copt, [800])
# (Binary) cross-entropy loss.
loss_clsf = nn.BCELoss(reduction='mean')
