def process_batch(engine, batch):
optimizer.zero_grad()
loss_v = model.calc_loss_dqn(
batch, net, tgt_net.target_model, preproc,
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()
return {
"loss": loss_v.item(),
"epsilon": action_selector.epsilon,
}
def process_batch(engine, batch):
optimizer.zero_grad()
loss_t = model.calc_loss_dqn(batch, prep, tgt_prep.target_model,
net, tgt_net.target_model, GAMMA, device=device)
loss_t.backward()
optimizer.step()
eps = 1 - engine.state.iteration / params.epsilon_steps
agent.epsilon = max(params.epsilon_final, eps)
if engine.state.iteration % params.sync_nets == 0:
tgt_net.sync()
tgt_prep.sync()
return {
"loss": loss_t.item(),
"epsilon": agent.epsilon,
}
def process_batch(engine, batch):
res = {}
optimizer.zero_grad()
loss_v = model.calc_loss_dqn(batch,
net,
tgt_net.target_model,
preproc,
gamma=PARAMS.gamma,
device=device)
loss_v.backward()
optimizer.step()
if epsilon_tracker is not None:
epsilon_tracker.frame(engine.state.iteration)
res['epsilon'] = action_selector.epsilon
if engine.state.iteration % PARAMS.target_net_sync == 0:
tgt_net.sync()
res['loss'] = loss_v.item()
return res
def process_batches(engine, batches):
res = {}
for name, batch, opt, net, tgt_net in zip(
["deer", "tiger"], batches, [deer_optimizer, tiger_optimizer],
[net_deer, net_tiger], [tgt_net_deer, tgt_net_tiger]):
opt.zero_grad()
loss_v = model.calc_loss_dqn(batch,
net,
tgt_net.target_model,
preproc,
gamma=PARAMS.gamma,
device=device)
loss_v.backward()
opt.step()
res[name + "_loss"] = loss_v.item()
if engine.state.iteration % PARAMS.target_net_sync == 0:
tgt_net.sync()
epsilon_tracker.frame(engine.state.iteration)
res['epsilon'] = action_selector.epsilon
return res
