def temporal_model_step(self, mu_q_t, logscale_q_t, h, attribute=None):
"""
generate z_t autoregressively
"""
# mix with temporal info
h_mu, h_scale = self.h_process(h, h)
mu = torch.cat([mu_q_t, h_mu], dim=-1)
logscale = torch.cat([logscale_q_t, h_scale], dim=-1)
mu_z_t, logscale_z_t = self.psi_dy(mu, logscale)
scale_z_t = logscale_z_t.exp()
# final posterior distribution with rnn information
posterior_t = Independent(Normal(mu_z_t, scale_z_t), 1)
posterior_sample_t = posterior_t.rsample()
# prior distribution with rnn information
if not attribute is None:
mixed_h_mu = torch.cat([h_mu, attribute], dim=-1)
mixed_h_scale = torch.cat([h_scale, attribute], dim=-1)
mu_p_t, logscale_p_t = self.psi_p(mixed_h_mu, mixed_h_scale)
else:
mu_p_t, logscale_p_t = self.psi_p(h_mu, h_scale)
# scale_p_t = logscale_p_t.exp()
scale_p_t = torch.ones_like(logscale_p_t)
prior_t = Independent(Normal(mu_p_t, scale_p_t), 1)
kl_div_t = torch.mean(kl_divergence(posterior_t, prior_t))
return posterior_sample_t, kl_div_t, mu_z_t, scale_z_t
def forward( # type: ignore
self,
batch: Batch,
state: Optional[Union[dict, Batch, np.ndarray]] = None,
input: str = "obs",
**kwargs: Any,
) -> Batch:
obs = batch[input]
logits, h = self.actor(obs, state=state, info=batch.info)
assert isinstance(logits, tuple)
dist = Independent(Normal(*logits), 1)
if self._deterministic_eval and not self.training:
x = logits[0]
else:
x = dist.rsample()
y = torch.tanh(x)
act = y * self._action_scale + self._action_bias
y = self._action_scale * (1 - y.pow(2)) + self.__eps
log_prob = dist.log_prob(x).unsqueeze(-1)
log_prob = log_prob - torch.log(y).sum(-1, keepdim=True)
if self._noise is not None and self.training and not self.updating:
act += to_torch_as(self._noise(act.shape), act)
act = act.clamp(self._range[0], self._range[1])
return Batch(
logits=logits, act=act, state=h, dist=dist, log_prob=log_prob)
def latent(self, conditioning):
"""
z_enc: [batch, frames, latent_dims]
attributes: [batch, frames, attribute_dims]
"""
z_enc = conditioning['z']
attributes = conditioning['attributes']
batch_size, n_frames, _encdims = z_enc.shape
if len(attributes.shape) < 3:
# expand along frame dimension
attributes = attributes.unsqueeze(1).expand(-1, n_frames, -1)
mu_q, logscale_q = self.psi_q(z_enc, z_enc)
# mix with latent
mu_a, logscale_a = self.attribute_latent(mu_q, logscale_q, attributes)
# feed into temporal model
h_0 = torch.rand(1, batch_size, self.latent_dims).to(
z_enc.device) * 0.01
temp_q = self.temporal_latent_model(mu_q, logscale_q, h_0)
# posterior distribution
mu_z, scale_z = self.mix_with_temp(mu_a, logscale_a, temp_q)
posterior = Independent(Normal(mu_z, scale_z), 1)
posterior_sample = posterior.rsample()
# prior
output = torch.cat([temp_q, attributes], dim=-1)
mu, scale = self.psi_p(output, output)
scale = scale.exp()
prior = Independent(Normal(mu, scale), 1)
# sum over latent dim mean over batch
kl_div = torch.mean(kl_divergence(posterior, prior))
return posterior_sample, kl_div, [mu_z, scale_z]
def forward( # type: ignore
self,
batch: Batch,
state: Optional[Union[dict, Batch, np.ndarray]] = None,
input: str = "obs",
**kwargs: Any,
) -> Batch:
obs = batch[input]
logits, h = self.actor(obs, state=state, info=batch.info)
assert isinstance(logits, tuple)
dist = Independent(Normal(*logits), 1)
if self._deterministic_eval and not self.training:
act = logits[0]
else:
act = dist.rsample()
log_prob = dist.log_prob(act).unsqueeze(-1)
# apply correction for Tanh squashing when computing logprob from Gaussian
# You can check out the original SAC paper (arXiv 1801.01290): Eq 21.
# in appendix C to get some understanding of this equation.
if self.action_scaling and self.action_space is not None:
action_scale = to_torch_as(
(self.action_space.high - self.action_space.low) / 2.0, act)
else:
action_scale = 1.0 # type: ignore
squashed_action = torch.tanh(act)
log_prob = log_prob - torch.log(action_scale *
(1 - squashed_action.pow(2)) +
self.__eps).sum(-1, keepdim=True)
return Batch(logits=logits,
act=squashed_action,
state=h,
dist=dist,
log_prob=log_prob)
def forward( # type: ignore
self,
batch: Batch,
state: Optional[Union[dict, Batch, np.ndarray]] = None,
input: str = "obs",
**kwargs: Any,
) -> Batch:
obs = batch[input]
logits, h = self.actor(obs, state=state, info=batch.info)
assert isinstance(logits, tuple)
dist = Independent(Normal(*logits), 1)
if self._deterministic_eval and not self.training:
x = logits[0]
else:
x = dist.rsample()
y = torch.tanh(x)
act = y * self._action_scale + self._action_bias
# __eps is used to avoid log of zero/negative number.
y = self._action_scale * (1 - y.pow(2)) + self.__eps
# Compute logprob from Gaussian, and then apply correction for Tanh squashing.
# You can check out the original SAC paper (arXiv 1801.01290): Eq 21.
# in appendix C to get some understanding of this equation.
log_prob = dist.log_prob(x).unsqueeze(-1)
log_prob = log_prob - torch.log(y).sum(-1, keepdim=True)
return Batch(logits=logits,
act=act,
state=h,
dist=dist,
log_prob=log_prob)
def forward(self, state, eval=False, with_log_prob=False):
x = F.relu(self.fc1(state))
x = F.relu(self.fc2(x))
mu = self.fc3(x)
log_sigma = self.fc4(x)
# clip value of log_sigma, as was done in Haarnoja's implementation of SAC:
# https://github.com/haarnoja/sac.git
log_sigma = torch.clamp(log_sigma, -20.0, 2.0)
sigma = torch.exp(log_sigma)
distribution = Independent(Normal(mu, sigma), 1)
if not eval:
# use rsample() instead of sample(), as sample() does not allow back-propagation through params
u = distribution.rsample()
if with_log_prob:
log_prob = distribution.log_prob(u)
log_prob -= 2.0 * torch.sum(
(np.log(2.0) + 0.5 * np.log(self.ctrl_range) - u -
F.softplus(-2.0 * u)),
dim=1)
else:
log_prob = None
else:
u = mu
log_prob = None
# apply tanh so that the resulting action lies in (-1, 1)^D
a = self.ctrl_range * torch.tanh(u)
return a, log_prob
def generate(self,
synth,
h_0,
f0_hz,
enc_frame_setting='fine',
n_samples=16000):
"""
synth: synth to generate audio
h_0: initial state of RNN [batch, latent_dims]
f0_hz: f0 conditioning of synth [batch, f0_n_frames, 1]
enc_frame_setting: fft/hop size
n_samples: output audio length in samples
"""
h = h_0
n_fft, hop_length = get_window_hop(enc_frame_setting)
n_frames = math.ceil((n_samples - n_fft) / hop_length) + 1
f0_hz = resample_frames(f0_hz,
n_frames) # needs to have same dimension as z
params_list = []
z = torch.zeros(h_0.shape[0], n_frames, self.latent_dims).to(h.device)
for t in range(n_frames):
# prior distribution with rnn information
mu_p_t, scale_p_t = self.get_prior(h)
prior_t = Independent(Normal(mu_p_t, scale_p_t), 1)
prior_sample_t = prior_t.rsample()
h = self.temporal(prior_sample_t, h)
z[:, t, :] = prior_sample_t
cond = {}
cond['z'] = z
cond['f0_hz'] = f0_hz
y_params = self.decode(cond)
params = synth.fill_params(y_params, cond)
resyn_audio, outputs = synth(params, n_samples)
return params, resyn_audio
def trajectory(self, current_state):
'''
Maybe this implementation doesn't utilize GPUs very well, but I have no clue or not.
Final output looks like:
[(s_0, a_0, r_0), ..., (s_L, a_L, r_l)]
'''
output_history = []
while True:
mu, sigma = self.forward(current_state)
distribution = Independent(Normal(mu, sigma), 1)
picked_action = distribution.rsample()
action = picked_action.detach()
#print(action)
new_state, reward = self.env.state_and_reward(
current_state, action
) #Get the reward and the new state that the action in the environment resulted in. None if action caused death. TODO build in environment
#Attempting this
output_history.append(
(current_state, action, reward, distribution.log_prob(action)))
if new_state is None: #essentially, you died or finished your trajectory
break
else:
current_state = new_state
return output_history
class TanhNormal(Distribution):
"""Copied from Kaixhi"""
def __init__(self, loc, scale):
super().__init__()
self.normal = Independent(Normal(loc, scale), 1)
def sample(self):
return torch.tanh(self.normal.sample())
# samples with re-parametrization trick (differentiable)
def rsample(self):
return torch.tanh(self.normal.rsample())
# Calculates log probability of value using the change-of-variables technique
# (uses log1p = log(1 + x) for extra numerical stability)
def log_prob(self, value):
inv_value = (torch.log1p(value) - torch.log1p(-value)) / 2 # artanh(y)
# log p(f^-1(y)) + log |det(J(f^-1(y)))|
return self.normal.log_prob(inv_value) - torch.log1p(-value.pow(2) +
1e-6).sum(dim=1)
@property
def mean(self):
return torch.tanh(self.normal.mean)
def get_std(self):
return self.normal.stddev
def forward(self, x):
kl = torch.zeros(1).to(device)
z = 0.
# Unet encoder result
x_enc = self.unet_encoder(x)
# VAE regularisation
if not self.unet:
emb = self.vae_encoder(x)
# Split encoder outputs into a mean and variance vector
mu, log_var = torch.chunk(emb, 2, dim=1)
# Make sure that the log variance is positive
log_var = softplus(log_var)
sigma = torch.exp(log_var / 2)
# Instantiate a diagonal Gaussian with mean=mu, std=sigma
# This is the approximate latent distribution q(z|x)
posterior = Independent(Normal(loc=mu, scale=sigma), 1)
z = posterior.rsample()
# Instantiate a standard Gaussian with mean=mu_0, std=sigma_0
# This is the prior distribution p(z)
prior = Independent(Normal(loc=self.mu_0, scale=self.sigma_0), 1)
# Estimate the KLD between q(z|x)|| p(z)
kl = KLD(posterior, prior).sum()
# Outputs for MSE
xHat = self.decoder(x_enc, z)
return kl, xHat
class IndependentNormal(Distribution):
arg_constraints = {'loc': constraints.real, 'scale': constraints.positive}
support = constraints.positive
has_rsample = True
def __init__(self, loc, scale, validate_args=None):
self.base_dist = Independent(Normal(loc=loc,
scale=scale,
validate_args=validate_args),
len(loc.shape) - 1,
validate_args=validate_args)
super(IndependentNormal, self).__init__(self.base_dist.batch_shape,
self.base_dist.event_shape,
validate_args=validate_args)
def log_prob(self, value):
return self.base_dist.log_prob(value)
@property
def mean(self):
return self.base_dist.mean
@property
def variance(self):
return self.base_dist.variance
def sample(self, sample_shape=torch.Size()):
return self.base_dist.sample(sample_shape)
def rsample(self, sample_shape=torch.Size()):
return self.base_dist.rsample(sample_shape)
def entropy(self):
entropy = self.base_dist.entropy()
return entropy
def generate(self,
synth,
h_0,
f0_hz,
attributes,
enc_frame_setting='fine',
n_samples=16000):
"""
synth: synth to generate audio
h_0: initial seed of RNN [batch, latent_dims]
f0_hz: f0 conditioning of synth [batch, f0_n_frames, 1]
attributes: attributes [batch, n_frames, attribute_size]
enc_frame_setting: fft/hop size
n_samples: output audio length in samples
"""
if len(h_0.shape) == 2:
h = h_0[None, :, :] # 1, batch, latent_dims
else:
h = h_0
n_fft, hop_length = get_window_hop(enc_frame_setting)
n_frames = math.ceil((n_samples - n_fft) / hop_length) + 1
f0_hz = resample_frames(f0_hz,
n_frames) # needs to have same dimension as z
params_list = []
for i in range(n_frames):
cond = {}
output = torch.cat([h.permute(1, 0, 2), attributes], dim=-1)
mu, logscale = self.psi_p(output, output)
scale = logscale.exp()
prior = Independent(Normal(mu, scale), 1)
prior_sample = prior.rsample()
cond['z'] = prior_sample
cond['f0_hz'] = f0_hz[:, i, :].unsqueeze(1)
cond['f0_scaled'] = hz_to_midi(cond['f0_hz']) / 127.0
# generate x
y = self.decode(cond)
params = synth.fill_params(y, cond)
params_list.append(params)
x_tilde, _outputs = synth(
params, n_samples=n_fft) # write exactly one frame
cond['audio'] = x_tilde
# encode
cond = self.encoder(cond)
z_enc = cond['z']
# get psi_q
mu, logscale = self.psi_q(z_enc, z_enc)
psi = torch.cat([mu, logscale], dim=-1)
# temporal model
temp_q, h = self.temporal_q(psi, h) # one off
param_names = params_list[0].keys()
final_params = {}
for pn in param_names:
#cat over frames
final_params[pn] = torch.cat([par[pn] for par in params_list],
dim=1)
final_audio, _outputs = synth(final_params, n_samples=n_samples)
return final_params, final_audio
def forward(self, obs, act=None, deterministic=False):
# Optionally pass in an action to get the log_prob of that action
mu = self.mu_layer(obs)
std = torch.exp(self.log_std_layer)
pi = Independent(Normal(mu, std), 1)
if act is None:
act = pi.mean if deterministic else pi.rsample()
log_prob = pi.log_prob(act)
return pi, act, log_prob
def generate(self,
synth,
h_0,
f0_hz,
enc_frame_setting='fine',
n_samples=16000):
"""
synth: synth to generate audio
h_0: initial state of RNN [batch, latent_dims]
f0_hz: f0 conditioning of synth [batch, f0_n_frames, 1]
enc_frame_setting: fft/hop size
n_samples: output audio length in samples
"""
h = h_0
n_fft, hop_length = get_window_hop(enc_frame_setting)
n_frames = math.ceil((n_samples - n_fft) / hop_length) + 1
f0_hz = resample_frames(f0_hz,
n_frames) # needs to have same dimension as z
params_list = []
z = torch.zeros(h_0.shape[0], n_frames, self.latent_dims).to(h.device)
for t in range(n_frames):
h_mu, h_scale = self.h_process(h, h)
mu_t, logscale_t = self.psi_p(h_mu,
h_scale) # [batch, latent_size]
scale_t = logscale_t.exp()
prior_t = Independent(Normal(mu_t, scale_t), 1)
prior_sample_t = prior_t.rsample()
cond = {}
z[:, t, :] = prior_sample_t
cond['z'] = prior_sample_t.unsqueeze(1)
cond['f0_hz'] = f0_hz[:, t, :].unsqueeze(1)
cond['f0_scaled'] = hz_to_midi(cond['f0_hz']) / 127.0
# generate x
y = self.decode(cond)
params = synth.fill_params(y, cond)
params_list.append(params)
x_tilde, _outputs = synth(
params, n_samples=n_fft) # write exactly one frame
cond['audio'] = x_tilde
# encode
cond = self.encoder(cond)
z_enc = cond['z'].squeeze(1)
# get psi_q
mu, logscale = self.psi_q(z_enc, z_enc)
rnn_input = torch.cat([mu, logscale, prior_sample_t], dim=-1)
# temporal model
h = self.temporal_q(rnn_input, h) # one off
cond = {}
cond['z'] = z
cond['f0_hz'] = f0_hz
y_params = self.decode(cond)
params = synth.fill_params(y_params, cond)
resyn_audio, outputs = synth(params, n_samples)
return params, resyn_audio
def act(self, obs, deterministic=False):
action_mean = self.forward(obs)
normal = Normal(action_mean, torch.exp(self.log_scale))
dist = Independent(normal, 1)
if deterministic:
action = action_mean
else:
action = dist.rsample()
action_logprobs = dist.log_prob(torch.squeeze(action))
return action, action_logprobs
def latent(self, conditioning):
z_enc = conditioning['z']
batch_size, num_frames, _ = z_enc.shape
mu = self.linear_mu(z_enc)
log_var = self.linear_logvar(z_enc)
eps = torch.randn_like(mu).detach().to(mu.device)
posterior = Independent(Normal(mu, log_var.exp().sqrt()), 1)
posterior_sample = posterior.rsample()
# Compute KL divergence
prior = Independent(Normal(torch.zeros_like(mu), torch.ones_like(log_var)), 1)
kl_div = torch.mean(kl_divergence(posterior, prior))
return posterior_sample, kl_div, [mu, log_var.exp().sqrt()]
def temporal_model_step(self, z_enc_t, h, attribute=None):
"""
generate z_t autoregressively
"""
# mix with temporal info
mu_z_t, scale_z_t = self.get_posterior(h, z_enc_t)
scale_z_t = scale_z_t + 1e-4 # minimum
# final posterior distribution with rnn information
posterior_t = Independent(Normal(mu_z_t, scale_z_t), 1)
posterior_sample_t = posterior_t.rsample()
# prior distribution with rnn information
mu_p_t, scale_p_t = self.get_prior(h)
scale_p_t = scale_p_t + 1e-4 # minimum
prior_t = Independent(Normal(mu_p_t, scale_p_t), 1)
kl_div_t = torch.mean(kl_divergence(posterior_t, prior_t))
return posterior_sample_t, kl_div_t, mu_z_t, scale_z_t
def forward(self, obs, deterministic=False):
mu = self.mu_layer(obs)
log_std = self.log_std_layer(obs)
std = torch.clamp(log_std, LOG_STD_MIN, LOG_STD_MAX).exp()
# Pre-squash distribution and sample
pi_distribution = Independent(Normal(mu, std), 1)
act = mu if deterministic else pi_distribution.rsample()
log_prob = pi_distribution.log_prob(act)
squashed_action = torch.tanh(act)
log_prob = log_prob - torch.log((1 - squashed_action.pow(2)) +
self.__eps).sum(axis=-1)
return squashed_action, log_prob
def _mcvi_forward(self, x):
if self.certain or not self.deterministic:
x_mean = x if not isinstance(x, tuple) else x[0]
x_var = x_mean * x_mean
else:
x_mean = x[0]
x_var = x[1]
W_var = self._get_var(self.W_logvar)
bias_var = self._get_var(self.bias_logvar)
z_mean = F.conv2d(x_mean, self.W, self.bias, self.stride, self.padding)
z_var = F.conv2d(x_var, W_var, bias_var, self.stride, self.padding)
dst = Independent(Normal(z_mean, z_var), 1)
sample = dst.rsample()
return sample, None
class VAE(nn.Module):
def __init__(self, encoder, decoder, device=None):
super().__init__()
self.encoder = encoder
self.decoder = decoder
if device is None:
self.device = torch.device(
"cuda:0" if torch.cuda.is_available() else "cpu")
else:
self.device = device
def forward(self, x=None):
bs = x.size(0)
ls = self.encoder.latent_dims
mu, sigma = self.encoder(x)
self.pz = Independent(Normal(loc=torch.zeros(bs, ls).to(self.device),
scale=torch.ones(bs, ls).to(self.device)),
reinterpreted_batch_ndims=1)
self.qz_x = Independent(Normal(loc=mu, scale=torch.exp(sigma)),
reinterpreted_batch_ndims=1)
self.z = self.qz_x.rsample()
decoded = self.decoder(self.z)
return decoded
def compute_loss(self, x, y, scale_kl=False):
px_z = Independent(ContinuousBernoulli(logits=y),
reinterpreted_batch_ndims=3)
px = px_z.log_prob(x)
kl = self.pz.log_prob(self.z) - self.qz_x.log_prob(self.z)
if scale_kl:
kl = kl * scale_kl
loss = -(px + kl).mean()
return loss, kl.mean().item(), px.mean().item()
def rmse(self, input, target):
return torch.sqrt(F.mse_loss(input, target))
def _loss_vae(self, x):
batch_size = x.size(0)
encoder_output = self.encoder(x)
pz = Independent(Normal(loc=torch.zeros(batch_size, self.latent_dim).to(self.device),
scale=torch.ones(batch_size, self.latent_dim).to(self.device)),
reinterpreted_batch_ndims=1)
qz_x = Independent(Normal(loc=encoder_output[:, :self.latent_dim],
scale=torch.exp(encoder_output[:, self.latent_dim:])),
reinterpreted_batch_ndims=1)
z = qz_x.rsample()
decoder_output = self.decoder(z)
px_z = Independent(Bernoulli(logits=decoder_output),
reinterpreted_batch_ndims=1)
loss = -(px_z.log_prob(x) + pz.log_prob(z) - qz_x.log_prob(z)).mean()
return loss, decoder_output
def generate(self,
synth,
h_0,
f0_hz,
attributes,
enc_frame_setting='fine',
n_samples=16000):
"""
synth: synth to generate audio
h_0: initial state of RNN [batch, latent_dims]
f0_hz: f0 conditioning of synth [batch, f0_n_frames, 1]
attributes: attributes [batch, attribute_size] or [batch, n_frames, attribute_size]
enc_frame_setting: fft/hop size
n_samples: output audio length in samples
"""
n_fft, hop_length = get_window_hop(enc_frame_setting)
n_frames = math.ceil((n_samples - n_fft) / hop_length) + 1
f0_hz = resample_frames(f0_hz, n_frames).to(
h_0.device) # needs to have same dimension as z
params_list = []
z = torch.zeros(h_0.shape[0], n_frames,
self.latent_dims).to(h_0.device)
if len(attributes.shape) == 2:
attributes = attributes[:, None, :].expand(-1, n_frames, -1)
# set up initial prior with attributes
z_t = torch.zeros(h_0.shape[0], self.latent_dims).to(h_0.device)
rnn_input = torch.cat([z_t, attributes[:, 0, :]], dim=-1)
h = self.temporal(rnn_input, h_0)
for t in range(n_frames):
# prior distribution with rnn information
mu_p_t, scale_p_t = self.get_prior(h)
prior_t = Independent(Normal(mu_p_t, scale_p_t), 1)
z_t = prior_t.rsample()
rnn_input = torch.cat([z_t, attributes[:, t, :]], dim=-1)
h = self.temporal(rnn_input, h)
z[:, t, :] = z_t
cond = {}
cond['z'] = z
cond['f0_hz'] = f0_hz
y_params = self.decode(cond)
params = synth.fill_params(y_params, cond)
resyn_audio, outputs = synth(params, n_samples)
return params, resyn_audio
def loss_vae_normal(x, encoder, decoder):
batch_size = x.size(0)
encoder_output = encoder(x)
d = encoder_output.shape[1] // 2
pz_loc = F.sigmoid(torch.zeros(batch_size, d).to(device))
pz_scale = torch.ones(batch_size, d).to(device)
pz = Independent(Normal(loc=pz_loc, scale=pz_scale),
reinterpreted_batch_ndims=1)
qz_x_loc = encoder_output[:, :d]
qz_x_log_scale = encoder_output[:, d:]
qz_x = Independent(Normal(loc=qz_x_loc, scale=qz_x_log_scale**2),
reinterpreted_batch_ndims=1)
z = qz_x.rsample()
decoder_output = decoder(z)
optimal_sigma_observed = ((x - decoder_output)**2).mean(
[0, 1, 2, 3], keepdim=True).sqrt()
px_z = Independent(Normal(loc=decoder_output,
scale=optimal_sigma_observed),
reinterpreted_batch_ndims=3)
elbo = (px_z.log_prob(x) - kl_divergence(qz_x, pz)).mean()
return -elbo, decoder_output
class IndependentRescaledBeta(Distribution):
arg_constraints = {
'concentration1': constraints.positive,
'concentration0': constraints.positive
}
support = constraints.interval(-1., 1.)
has_rsample = True
def __init__(self, concentration1, concentration0, validate_args=None):
self.base_dist = Independent(RescaledBeta(concentration1,
concentration0,
validate_args),
len(concentration1.shape) - 1,
validate_args=validate_args)
super(IndependentRescaledBeta,
self).__init__(self.base_dist.batch_shape,
self.base_dist.event_shape,
validate_args=validate_args)
def log_prob(self, value):
return self.base_dist.log_prob(value)
@property
def mean(self):
return self.base_dist.mean
@property
def variance(self):
return self.base_dist.variance
def sample(self, sample_shape=torch.Size()):
return self.base_dist.sample(sample_shape)
def rsample(self, sample_shape=torch.Size()):
return self.base_dist.rsample(sample_shape)
def entropy(self):
entropy = self.base_dist.entropy()
return entropy
def latent(self, conditioning):
"""
z_enc: [batch, frames, latent_dims]
"""
z_enc = conditioning['z']
batch_size, n_frames, _encdims = z_enc.shape
mu_q, logscale_q = self.psi_q(z_enc, z_enc)
# feed into temporal model
h_0 = torch.randn(1, batch_size, self.latent_dims).to(
z_enc.device) * 0.01
temp_q = self.temporal_latent_model(mu_q, logscale_q, h_0)
# final posterior distribution with rnn information
mu_z, scale_z = self.mix_with_temp(mu_q, logscale_q, temp_q)
posterior = Independent(Normal(mu_z, scale_z), 1)
posterior_sample = posterior.rsample()
# prior distribution with rnn information
mu, scale = self.psi_p(temp_q, temp_q)
scale = scale.exp()
prior = Independent(Normal(mu, scale), 1)
# prior = Independent(Normal(torch.zeros_like(mu), torch.ones_like(scale)), 1)
# sum over latent dim mean over batch
kl_div = torch.mean(kl_divergence(posterior, prior))
return posterior_sample, kl_div, [mu_z, scale_z]
def _zero_mean_forward(self, x):
if self.certain or not self.deterministic:
x_mean = x if not isinstance(x, tuple) else x[0]
x_var = x_mean * x_mean
else:
x_mean = x[0]
x_var = x[1]
W_var = torch.exp(self.log_alpha) * self.weight * self.weight
z_mean = F.conv2d(x_mean, torch.zeros_like(self.weight), self.bias,
self.stride, self.padding)
z_var = F.conv2d(x_var,
W_var,
bias=None,
stride=self.stride,
padding=self.padding)
if self.deterministic:
return z_mean, z_var
else:
dst = Independent(Normal(z_mean, z_var), 1)
sample = dst.rsample()
return sample, None
def loss(self, x):
"""
returns
1. the avergave value of negative ELBO across the minibatch x
2. and the output of the decoder
"""
batch_size = x.size(0)
encoder_output = self.encoder(x)
pz = Independent(Normal(loc=torch.zeros(batch_size,
self.z_dim).to(self.device),
scale=torch.ones(batch_size,
self.z_dim).to(self.device)),
reinterpreted_batch_ndims=1)
qz_x = Independent(Normal(loc=encoder_output[:, :self.z_dim],
scale=torch.exp(
encoder_output[:, self.z_dim:])),
reinterpreted_batch_ndims=1)
z = qz_x.rsample()
decoder_output = self.decoder(z)
px_z = Independent(Bernoulli(logits=decoder_output),
reinterpreted_batch_ndims=1)
loss = -(px_z.log_prob(x) + pz.log_prob(z) - qz_x.log_prob(z)).mean()
return loss, decoder_output
class TorchGaussianMixtureDistribution(TorchDistributionWrapper):
@staticmethod
def required_model_output_shape(action_space, model_config):
return prod((2, model_config['custom_model_config']['num_gaussians']) +
action_space.shape)
def __init__(self, inputs: List[torch.Tensor], model: TorchModelV2):
super(TorchDistributionWrapper, self).__init__(inputs, model)
assert len(inputs.shape) == 2
self.batch_size = self.inputs.shape[0]
self.num_gaussians = model.model_config['custom_model_config'][
'num_gaussians']
self.monte_samples = model.model_config['custom_model_config'][
'monte_samples']
inputs = torch.reshape(self.inputs,
(self.batch_size, 2, self.num_gaussians, -1))
self.action_dim = inputs.shape[-1]
assert not torch.isnan(inputs).any(), "Input nan aborting"
self.means = inputs[:,
0, :, :] # batch_size x num_gaussians x action_dim
self.sigmas = torch.exp(
inputs[:, 1, :, :]) # batch_size x num_gaussians x action_dim
self.cat = Categorical(
torch.ones(self.batch_size,
self.num_gaussians,
device=inputs.device,
requires_grad=False))
self.normals = Independent(Normal(self.means, self.sigmas), 1)
def logp(self, actions: torch.Tensor):
actions = actions.view(
self.batch_size, 1,
-1) # batch_size x 1 (broadcast to num gaussians) x action_dim
mix_lps = self.cat.logits # batch_size x num_gaussians x action_dim
normal_lps = self.normals.log_prob(
actions) # batch_size x num_gaussians x action_dim
assert not torch.isnan(mix_lps).any(), "output nan aborting"
assert not torch.isnan(normal_lps).any(), "output nan aborting"
return torch.logsumexp(mix_lps + normal_lps,
dim=1) # reduce along num gaussians
def deterministic_sample(self) -> torch.Tensor:
self.last_sample = self.means[:,
0, :] # select the mode of the first gaussian
return self.last_sample
def __rsamples(self):
""" Compute samples that can be differentiated through
"""
# Using reparameterization trick i.e. rsample
normal_samples = self.normals.rsample(
(self.monte_samples,
)) # monte_samples x batch_size x num_gaussians x action_dim
cat_samples = self.cat.sample(
(self.monte_samples, )) # monte_samples x batch_size
# First we need to expand cat so that it has the same dimension as normal samples
cat_samples = cat_samples.reshape(self.monte_samples, -1, 1,
1).expand(-1, -1, -1,
self.action_dim)
# We select the normal distribution based on the outputs of
# the categorical distribution
return torch.gather(normal_samples, 2, cat_samples).squeeze(
dim=2) # monte_samples x batch_size x action_dim
def kl(self, q: ActionDistribution) -> torch.Tensor:
""" KL(self || q) estimated with monte carlo sampling
"""
rsamples = self.__rsamples().unbind(0)
log_ratios = torch.stack(
[self.logp(rsample) - q.logp(rsample) for rsample in rsamples])
assert not torch.isnan(log_ratios).any(), "output nan aborting"
return log_ratios.mean(0)
def entropy(self) -> torch.Tensor:
""" H(self) estimated with monte carlo sampling
"""
rsamples = self.__rsamples().unbind(0)
log_ps = torch.stack([-self.logp(rsample) for rsample in rsamples])
assert not torch.isnan(log_ps).any(), "output nan aborting"
return log_ps.mean(0)
def sample(self):
normal_samples = self.normals.sample(
) # batch_size x num_gaussians x action_dim
cat_samples = self.cat.sample() # batch_size
# First we need to expand cat so that it has the same dimension as normal samples
cat_samples = cat_samples.view(-1, 1,
1).expand(-1, -1, self.action_dim)
# We select the normal distribution based on the outputs of
# the categorical distribution
self.last_sample = torch.gather(normal_samples, 1,
cat_samples).squeeze(
dim=1) # batch_size x action_dim
assert len(
self.last_sample.shape) == 2, f"shape, {self.last_sample.shape}"
return self.last_sample