def process_batch(engine, batch):
optimizer.zero_grad()
loss_v = common.calc_loss_dqn(
batch, net, tgt_net.target_model, gamma=params.gamma, device=device
)
loss_v.backward()
optimizer.step()
if engine.state.iteration % params.target_net_sync == 0:
tgt_net.sync()
return {
"loss": loss_v.item(),
"epsilon": batch_generator.epsilon,
}
def process_batch(engine_, batch):
optimizer.zero_grad()
loss_v = common.calc_loss_dqn(
batch, net, tgt_net.target_model, gamma=params.gamma ** args.n, device=device
)
loss_v.backward()
optimizer.step()
epsilon_tracker.frame(engine_.state.iteration)
if engine_.state.iteration % params.target_net_sync == 0:
tgt_net.sync()
return {
"loss": loss_v.item(),
"epsilon": selector.epsilon,
}
def process_batch(engine, batch):
optimizer.zero_grad()
loss_v = common.calc_loss_dqn(batch,
net,
tgt_net.target_model,
gamma=params.gamma,
device=device)
loss_v.backward()
optimizer.step()
if engine.state.iteration % params.target_net_sync == 0:
tgt_net.sync()
if engine.state.iteration % NOISY_SNR_EVERY_ITERS == 0:
for layer_idx, sigma_l2 in enumerate(net.noisy_layers_sigma_snr()):
engine.state.metrics[f"snr_{layer_idx+1}"] = sigma_l2
return {
"loss": loss_v.item(),
}
def process_batch(engine, batch):
optimizer.zero_grad()
loss_v = common.calc_loss_dqn(
batch, net, tgt_net.target_model, gamma=params.gamma, device=device
)
loss_v.backward()
optimizer.step()
epsilon_tracker.frame(engine.state.iteration)
if engine.state.iteration % params.target_net_sync == 0:
tgt_net.sync()
if engine.state.iteration % EVAL_EVERY_FRAME == 0:
eval_states = getattr(engine.state, "eval_states", None)
if eval_states is None:
eval_states = buffer.sample(STATES_TO_EVALUATE)
eval_states = [np.array(transition.state, copy=False) for transition in eval_states]
eval_states = np.array(eval_states, copy=False)
engine.state.eval_states = eval_states
evaluate_states(eval_states, net, device, engine)
return {
"loss": loss_v.item(),
"epsilon": selector.epsilon,
}