# Load RNN
rnn_dir = join(args.originallogdir, 'mdrnn')
rnn_file = join(rnn_dir, 'best.tar')
assert exists(rnn_file), 'No trained MDNRNN in the originallogdir...'
mdrnn = MDRNN(LSIZE, ASIZE, RSIZE, 5)
mdrnn.to(device)
mdrnn_optimizer = torch.optim.RMSprop(mdrnn.parameters(), lr=1e-3, alpha=.9)
mdrnn_scheduler = ReduceLROnPlateau(mdrnn_optimizer,
'min',
factor=0.5,
patience=5)
rnn_state = torch.load(rnn_file, map_location={'cuda:0': str(device)})
print("Loading MDRNN at epoch {} "
"with test error {}".format(rnn_state["epoch"], rnn_state["precision"]))
mdrnn.load_state_dict(rnn_state["state_dict"])
def collation_fn(rollouts):
rollout_items = [[], [], [], [], []]
for rollout in rollouts:
for i in range(len(rollout)):
rollout_items[i].append(torch.Tensor(rollout[i]))
for i in range(len(rollout_items)):
rollout_items[i] = pack_sequence(
sorted(rollout_items[i], key=len, reverse=True))
return tuple(rollout_items)
# Data Loading
transform = lambda x: np.transpose(x, (0, 3, 1, 2)) / 255
class RolloutGenerator(object):
""" Utility to generate rollouts.
Encapsulate everything that is needed to generate rollouts in the TRUE ENV
using a controller with previously trained VAE and MDRNN.
:attr vae: VAE model loaded from mdir/vae
:attr mdrnn: MDRNN model loaded from mdir/mdrnn
:attr controller: Controller, either loaded from mdir/ctrl or randomly
initialized
:attr env: instance of the CarRacing-v0 gym environment
:attr device: device used to run VAE, MDRNN and Controller
:attr time_limit: rollouts have a maximum of time_limit timesteps
"""
def __init__(self, mdir, device, time_limit, explorer=False):
""" Build vae, rnn, controller and environment. """
self.explorer = explorer
# Load controllers
vae_file, rnn_file, ctrl_file = \
[join(mdir, m, 'best.tar') for m in ['vae', 'mdrnn', 'ctrl']]
if self.explorer:
ctrl_file = join(mdir, 'exp', 'best.tar')
assert exists(vae_file) and exists(rnn_file),\
"Either vae or mdrnn is untrained."
vae_state, rnn_state = [
torch.load(fname, map_location={'cuda:0': str(device)})
for fname in (vae_file, rnn_file)
]
for m, s in (('VAE', vae_state), ('MDRNN', rnn_state)):
print("Loading {} at epoch {} "
"with test loss {}".format(m, s['epoch'], s['precision']))
self.vae = VAE(3, LSIZE).to(device)
self.vae.load_state_dict(vae_state['state_dict'])
# MDRNNCell
self.mdrnn = MDRNNCell(LSIZE, ASIZE, RSIZE, 5).to(device)
self.mdrnn.load_state_dict(
{k.strip('_l0'): v
for k, v in rnn_state['state_dict'].items()})
self.controller = Controller(LSIZE, RSIZE, ASIZE).to(device)
# load controller if it was previously saved
if exists(ctrl_file):
ctrl_state = torch.load(ctrl_file,
map_location={'cuda:0': str(device)})
print("Loading Controller with reward {}".format(
ctrl_state['reward']))
self.controller.load_state_dict(ctrl_state['state_dict'])
self.env = gym.make('CarRacing-v0')
self.device = device
self.time_limit = time_limit
self.mdrnn_notcell = MDRNN(LSIZE, ASIZE, RSIZE, 5)
self.mdrnn_notcell.to(device)
self.mdrnn_notcell.load_state_dict(rnn_state['state_dict'])
#####$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$
# VERY LAZY. Copied from the other trainmdrnn file
# from trainmdrnn import get_loss, to_latent
def to_latent(self, obs, next_obs):
""" Transform observations to latent space.
:args obs: 5D torch tensor (BSIZE, SEQ_LEN, ASIZE, SIZE, SIZE)
:args next_obs: 5D torch tensor (BSIZE, SEQ_LEN, ASIZE, SIZE, SIZE)
:returns: (latent_obs, latent_next_obs)
- latent_obs: 4D torch tensor (BSIZE, SEQ_LEN, LSIZE)
- next_latent_obs: 4D torch tensor (BSIZE, SEQ_LEN, LSIZE)
"""
with torch.no_grad():
obs, next_obs = [
f.upsample(x.view(-1, 3, SIZE, SIZE),
size=RED_SIZE,
mode='bilinear',
align_corners=True) for x in (obs, next_obs)
]
(obs_mu, obs_logsigma), (next_obs_mu, next_obs_logsigma) = [
self.vae(x)[1:] for x in (obs, next_obs)
]
SEQ_LEN = 1
latent_obs, latent_next_obs = [
(x_mu + x_logsigma.exp() * torch.randn_like(x_mu)).view(
BSIZE, SEQ_LEN, LSIZE)
for x_mu, x_logsigma in [(
obs_mu, obs_logsigma), (next_obs_mu, next_obs_logsigma)]
]
return latent_obs, latent_next_obs
def mdrnn_exp_reward(self, latent_obs, action, reward, latent_next_obs,
hidden):
""" # REMOVE TERMINAL
Compute losses.
The loss that is computed is:
(GMMLoss(latent_next_obs, GMMPredicted) + MSE(reward, predicted_reward) +
BCE(terminal, logit_terminal)) / (LSIZE + 2)
The LSIZE + 2 factor is here to counteract the fact that the GMMLoss scales
approximately linearily with LSIZE. All losses are averaged both on the
batch and the sequence dimensions (the two first dimensions).
:args latent_obs: (BSIZE, SEQ_LEN, LSIZE) torch tensor
:args action: (BSIZE, SEQ_LEN, ASIZE) torch tensor
:args reward: (BSIZE, SEQ_LEN) torch tensor
:args latent_next_obs: (BSIZE, SEQ_LEN, LSIZE) torch tensor
:returns: dictionary of losses, containing the gmm, the mse, the bce and
the averaged loss.
"""
mus, sigmas, logpi, rs, ds, next_hidden = self.mdrnn(
action, latent_obs, hidden)
gmm = gmm_loss(latent_next_obs, mus, sigmas, logpi)
# bce = f.binary_cross_entropy_with_logits(ds, terminal)
mse = f.mse_loss(rs, reward)
loss = (gmm + mse) / (LSIZE + 2)
return loss.squeeze().cpu().numpy()
# def recon_error_reward(self, obs, hidden, obs_new):
# print('recon_error_reward')
# """Find out how good the reconstruction was.
# Encoding the vae to get mu and the controller action is deterministic, so its fine to be duplicated
# ??? maybe remove this and the above function because of unnecessary duplication
# """
# # obs_new = torch.from_numpy(np.moveaxis(obs_new, 2, 0).copy()).unsqueeze(0).to(self.device).type(torch.cuda.FloatTensor)
# # obs = obs.to(self.device).type(torch.cuda.FloatTensor)
# _, latent_mu, _ = self.vae(obs)
# action = self.controller(latent_mu, hidden[0])
# mus, sigmas, logpi, r, d, next_hidden = self.mdrnn(action, latent_mu, hidden)
# print('mus.size()', mus.size())
# print('sigmas.size()', sigmas.size())
# print('logpi.size()', logpi.size())
# print('r.size()', r.size())
# print('d.size()', d.size())
# print('next_hidden.size() [0], [1]', next_hidden[0].size(), next_hidden[1].size())
# recon_x = self.vae.decoder(mus.squeeze()).type(torch.cuda.FloatTensor) # ??? this is just mu, right? Still a bit confused
# print('obs_new.size()', obs_new.size())
# print('recon_x.size()', recon_x.size())
# # reward = -1*((recon_x - obs_new) ** 2).mean()
# reward = -1*F.mse_loss(recon_x, obs_new).item()
def rollout(self, params, render=False):
""" Execute a rollout and return reward
Load :params: into the controller and execute a single rollout. This
is the main API of this class.
:args params: parameters as a single 1D np array
:returns: minus cumulative reward if ctrl mode, cumulative recon_error if exp mode
"""
# copy params into the controller
if params is not None:
load_parameters(params, self.controller)
obs = self.env.reset()
# This first render is required !
self.env.render()
hidden = [torch.zeros(1, RSIZE).to(self.device) for _ in range(2)]
cumulative = 0
i = 0
while True:
obs = transform(obs).unsqueeze(0).to(self.device)
# GET ACTION
_, latent_mu, _ = self.vae(obs)
action = self.controller(latent_mu, hidden[0])
_, _, _, _, _, next_hidden = self.mdrnn(action, latent_mu, hidden)
action = action.squeeze().cpu().numpy()
next_obs, reward, done, _ = self.env.step(action)
if self.explorer:
latent_obs, latent_next_obs = self.to_latent(
obs.unsqueeze(0),
transform(next_obs).unsqueeze(0).to(self.device))
action = torch.from_numpy(action).unsqueeze(0)
latent_obs = latent_obs.to(self.device).squeeze().unsqueeze(0)
latent_next_obs = latent_next_obs.to(
self.device).squeeze().unsqueeze(0)
action = action.to(self.device)
reward = torch.from_numpy(np.array(reward)).unsqueeze(0).type(
torch.cuda.FloatTensor)
reward = self.mdrnn_exp_reward(latent_obs, action, reward,
latent_next_obs, hidden)
obs = next_obs
hidden = next_hidden
if render:
self.env.render()
cumulative += reward
if done or i > self.time_limit:
return -cumulative
i += 1
if not exists(rnn_dir):
mkdir(rnn_dir)
mdrnn = MDRNN(LSIZE, ASIZE, RSIZE, 5)
mdrnn.to(device)
optimizer = torch.optim.RMSprop(mdrnn.parameters(), lr=1e-3, alpha=.9)
scheduler = ReduceLROnPlateau(optimizer, 'min', factor=0.5, patience=5)
earlystopping = EarlyStopping('min', patience=30)
if exists(rnn_file) and not args.noreload:
rnn_state = torch.load(rnn_file)
print("Loading MDRNN at epoch {} "
"with test error {}".format(
rnn_state["epoch"], rnn_state["precision"]))
mdrnn.load_state_dict(rnn_state["state_dict"])
optimizer.load_state_dict(rnn_state["optimizer"])
scheduler.load_state_dict(state['scheduler'])
earlystopping.load_state_dict(state['earlystopping'])
# Data Loading
transform = transforms.Lambda(
lambda x: np.transpose(x, (0, 3, 1, 2)) / 255)
train_loader = DataLoader(
RolloutSequenceDataset('datasets/carracing', SEQ_LEN, transform, buffer_size=30),
batch_size=BSIZE, num_workers=8, shuffle=True)
test_loader = DataLoader(
RolloutSequenceDataset('datasets/carracing', SEQ_LEN, transform, train=False, buffer_size=10),
batch_size=BSIZE, num_workers=8)
def train_mdrnn(logdir, traindir, epochs=10, testdir=None):
BSIZE = 80 # maybe should change this back to their initial one of 16
noreload = False #Best model is not reloaded if specified
SEQ_LEN = 32
epochs = int(epochs)
testdir = testdir if testdir else traindir
cuda = torch.cuda.is_available()
torch.manual_seed(123)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Loading VAE
vae_file = join(logdir, 'vae', 'best.tar')
assert exists(vae_file), "No trained VAE in the logdir..."
state = torch.load(vae_file)
print("Loading VAE at epoch {} "
"with test error {}".format(
state['epoch'], state['precision']))
vae = VAE(3, LSIZE).to(device)
vae.load_state_dict(state['state_dict'])
# Loading model
rnn_dir = join(logdir, 'mdrnn')
rnn_file = join(rnn_dir, 'best.tar')
if not exists(rnn_dir):
mkdir(rnn_dir)
mdrnn = MDRNN(LSIZE, ASIZE, RSIZE, 5)
mdrnn.to(device)
optimizer = torch.optim.RMSprop(mdrnn.parameters(), lr=1e-3, alpha=.9)
scheduler = ReduceLROnPlateau(optimizer, 'min', factor=0.5, patience=5)
earlystopping = EarlyStopping('min', patience=30)
if exists(rnn_file) and not noreload:
rnn_state = torch.load(rnn_file)
print("Loading MDRNN at epoch {} "
"with test error {}".format(
rnn_state["epoch"], rnn_state["precision"]))
mdrnn.load_state_dict(rnn_state["state_dict"])
optimizer.load_state_dict(rnn_state["optimizer"])
scheduler.load_state_dict(state['scheduler'])
earlystopping.load_state_dict(state['earlystopping'])
# Data Loading
transform = transforms.Lambda(
lambda x: np.transpose(x, (0, 3, 1, 2)) / 255)
train_loader = DataLoader(
RolloutSequenceDataset(traindir, SEQ_LEN, transform, buffer_size=30),
batch_size=BSIZE, num_workers=8, shuffle=True)
test_loader = DataLoader(
RolloutSequenceDataset(testdir, SEQ_LEN, transform, train=False, buffer_size=10),
batch_size=BSIZE, num_workers=8)
def to_latent(obs, next_obs):
""" Transform observations to latent space.
:args obs: 5D torch tensor (BSIZE, SEQ_LEN, ASIZE, SIZE, SIZE)
:args next_obs: 5D torch tensor (BSIZE, SEQ_LEN, ASIZE, SIZE, SIZE)
:returns: (latent_obs, latent_next_obs)
- latent_obs: 4D torch tensor (BSIZE, SEQ_LEN, LSIZE)
- next_latent_obs: 4D torch tensor (BSIZE, SEQ_LEN, LSIZE)
"""
with torch.no_grad():
obs, next_obs = [
f.upsample(x.view(-1, 3, SIZE, SIZE), size=RED_SIZE,
mode='bilinear', align_corners=True)
for x in (obs, next_obs)]
(obs_mu, obs_logsigma), (next_obs_mu, next_obs_logsigma) = [
vae(x)[1:] for x in (obs, next_obs)]
latent_obs, latent_next_obs = [
(x_mu + x_logsigma.exp() * torch.randn_like(x_mu)).view(BSIZE, SEQ_LEN, LSIZE)
for x_mu, x_logsigma in
[(obs_mu, obs_logsigma), (next_obs_mu, next_obs_logsigma)]]
return latent_obs, latent_next_obs
def get_loss(latent_obs, action, reward, terminal, latent_next_obs):
""" Compute losses.
The loss that is computed is:
(GMMLoss(latent_next_obs, GMMPredicted) + MSE(reward, predicted_reward) +
BCE(terminal, logit_terminal)) / (LSIZE + 2)
The LSIZE + 2 factor is here to counteract the fact that the GMMLoss scales
approximately linearily with LSIZE. All losses are averaged both on the
batch and the sequence dimensions (the two first dimensions).
:args latent_obs: (BSIZE, SEQ_LEN, LSIZE) torch tensor
:args action: (BSIZE, SEQ_LEN, ASIZE) torch tensor
:args reward: (BSIZE, SEQ_LEN) torch tensor
:args latent_next_obs: (BSIZE, SEQ_LEN, LSIZE) torch tensor
:returns: dictionary of losses, containing the gmm, the mse, the bce and
the averaged loss.
"""
latent_obs, action,\
reward, terminal,\
latent_next_obs = [arr.transpose(1, 0)
for arr in [latent_obs, action,
reward, terminal,
latent_next_obs]]
mus, sigmas, logpi, rs, ds = mdrnn(action, latent_obs)
gmm = gmm_loss(latent_next_obs, mus, sigmas, logpi)
bce = f.binary_cross_entropy_with_logits(ds, terminal)
mse = f.mse_loss(rs, reward)
loss = (gmm + bce + mse) / (LSIZE + 2)
return dict(gmm=gmm, bce=bce, mse=mse, loss=loss)
def data_pass(epoch, train): # pylint: disable=too-many-locals
""" One pass through the data """
if train:
mdrnn.train()
loader = train_loader
else:
mdrnn.eval()
loader = test_loader
loader.dataset.load_next_buffer()
cum_loss = 0
cum_gmm = 0
cum_bce = 0
cum_mse = 0
pbar = tqdm(total=len(loader.dataset), desc="Epoch {}".format(epoch))
for i, data in enumerate(loader):
obs, action, reward, terminal, next_obs = [arr.to(device) for arr in data]
# transform obs
latent_obs, latent_next_obs = to_latent(obs, next_obs)
if train:
losses = get_loss(latent_obs, action, reward,
terminal, latent_next_obs)
optimizer.zero_grad()
losses['loss'].backward()
optimizer.step()
else:
with torch.no_grad():
losses = get_loss(latent_obs, action, reward,
terminal, latent_next_obs)
cum_loss += losses['loss'].item()
cum_gmm += losses['gmm'].item()
cum_bce += losses['bce'].item()
cum_mse += losses['mse'].item()
pbar.set_postfix_str("loss={loss:10.6f} bce={bce:10.6f} "
"gmm={gmm:10.6f} mse={mse:10.6f}".format(
loss=cum_loss / (i + 1), bce=cum_bce / (i + 1),
gmm=cum_gmm / LSIZE / (i + 1), mse=cum_mse / (i + 1)))
pbar.update(BSIZE)
pbar.close()
return cum_loss * BSIZE / len(loader.dataset)
train = partial(data_pass, train=True)
test = partial(data_pass, train=False)
for e in range(epochs):
cur_best = None
train(e)
test_loss = test(e)
scheduler.step(test_loss)
earlystopping.step(test_loss)
is_best = not cur_best or test_loss < cur_best
if is_best:
cur_best = test_loss
checkpoint_fname = join(rnn_dir, 'checkpoint.tar')
save_checkpoint({
"state_dict": mdrnn.state_dict(),
"optimizer": optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
'earlystopping': earlystopping.state_dict(),
"precision": test_loss,
"epoch": e}, is_best, checkpoint_fname,
rnn_file)
if earlystopping.stop:
print("End of Training because of early stopping at epoch {}".format(e))
break
