def train(self, batches=5000):
losses = np.zeros(batches)
batch_size = 100
loss = nn.MSELoss()
optimiser = optim.SGD(self.actor.parameters(), lr=0.001)
for i in range(batches):
x, u = self.buffer.sample(batch_size)
state = torch.FloatTensor(x)
action = torch.FloatTensor(u)
optimiser.zero_grad()
outputs = self.actor(state)
actor_loss = loss(outputs[:, 0], action)
actor_loss.backward()
optimiser.step()
losses[i] = actor_loss
if i % 500 == 0:
print(f"Batch: {i}: Loss: {actor_loss}")
lib.plot(losses, 100)
self.save()
return losses
def print_update(self, plot_reward=True):
if self.ptr < 5:
return
mean = np.mean(self.rewards[0:self.ptr])
score = self.rewards[-1]
print(f"Run: {self.t_counter} --> Score: {score:.2f} --> Mean: {mean:.2f} ")
if plot_reward:
lib.plot(self.rewards[0:self.ptr], figure_n=2)