def value_objective_grad_func(value_net_flat_params):
set_flat_params(value_net, DOUBLE(value_net_flat_params))
for param in value_net.parameters():
if param.grad is not None:
param.grad.data.fill_(0)
values_pred = value_net(states)
value_loss = nn.MSELoss()(values_pred, returns)
# weight decay
for param in value_net.parameters():
value_loss += param.pow(2).sum() * l2_reg
value_loss.backward() # to get the grad
objective_value_loss_grad = get_flat_grad_params(
value_net).detach().cpu().numpy()
return objective_value_loss_grad
def value_objective_grad_func(value_net_flat_params):
"""
objective function for scipy optimizing
"""
set_flat_params(value_net, FLOAT(value_net_flat_params))
for param in value_net.parameters():
if param.grad is not None:
param.grad.data.fill_(0)
values_pred = value_net(states)
value_loss = nn.MSELoss()(values_pred, returns)
# weight decay
for param in value_net.parameters():
value_loss += param.pow(2).sum() * l2_reg
value_loss.backward() # to get the grad
objective_value_loss_grad = get_flat_grad_params(
value_net).detach().cpu().numpy().astype(np.float64)
return objective_value_loss_grad