global_rewards = []
obs_gotten = None
while frame_count < max_frame:
cache = {'obs': [], 'acts': [], 'rews': [], 'dones': []}
probs_cache = {'mu': [], 'sig': []}
for _ in range(n_steps):
obs = envs.reset() if obs_gotten is None else obs_gotten
obs_in = torch.FloatTensor(obs).to(device)
mu, sig = actor(obs_in)
with torch.no_grad():
a = Normal(mu, sig).sample()
a.clamp_(-2.0 + 1e-7, 2.0 - 1e-7)
obs_gotten, rews, dones, _ = envs.step(a)
for i in range(num_envs):
rewards[i][-1] += rews[i]
if dones[i]:
global_rewards.append(rewards[i][-1])
rewards[i].append(0.)
cache['obs'].append(obs)
cache['acts'].append(a)
cache['rews'].append(rews * 0.1)
cache['dones'].append(dones)
probs_cache['mu'].append(mu)
