######################################################################
-parser = argparse.ArgumentParser(description="Tiny LeNet-like auto-encoder.")
+parser = argparse.ArgumentParser(
+ description="Very simple implementation of a VAE for teaching."
+)
+
+parser.add_argument("--nb_epochs", type=int, default=100)
-parser.add_argument("--nb_epochs", type=int, default=25)
+parser.add_argument("--learning_rate", type=float, default=2e-4)
parser.add_argument("--batch_size", type=int, default=100)
######################################################################
-def sample_gaussian(mu, log_var):
+def sample_gaussian(param):
+ mu, log_var = param
std = log_var.mul(0.5).exp()
return torch.randn(mu.size(), device=mu.device) * std + mu
-def log_p_gaussian(x, mu, log_var):
+def log_p_gaussian(x, param):
+ mu, log_var = param
var = log_var.exp()
return (
(-0.5 * ((x - mu).pow(2) / var) - 0.5 * log_var - 0.5 * math.log(2 * math.pi))
)
-def dkl_gaussians(mu_a, log_var_a, mu_b, log_var_b):
- mu_a, log_var_a = mu_a.flatten(1), log_var_a.flatten(1)
- mu_b, log_var_b = mu_b.flatten(1), log_var_b.flatten(1)
+def dkl_gaussians(param_a, param_b):
+ mean_a, log_var_a = param_a[0].flatten(1), param_a[1].flatten(1)
+ mean_b, log_var_b = param_b[0].flatten(1), param_b[1].flatten(1)
var_a = log_var_a.exp()
var_b = log_var_b.exp()
return 0.5 * (
- log_var_b - log_var_a - 1 + (mu_a - mu_b).pow(2) / var_b + var_a / var_b
+ log_var_b - log_var_a - 1 + (mean_a - mean_b).pow(2) / var_b + var_a / var_b
).sum(1)
def forward(self, z):
output = self.model(z.view(z.size(0), -1, 1, 1))
mu, log_var = output[:, 0:1], output[:, 1:2]
+ log_var.flatten(1)[...] = log_var.flatten(1)[:, :1]
return mu, log_var
optimizer = optim.Adam(
itertools.chain(model_p_X_given_z.parameters(), model_q_Z_given_x.parameters()),
- lr=4e-4,
+ lr=args.learning_rate,
)
model_p_X_given_z.to(device)
######################################################################
-mu_p_Z = train_input.new_zeros(1, args.latent_dim)
-log_var_p_Z = mu_p_Z
+zeros = train_input.new_zeros(1, args.latent_dim)
+
+param_p_Z = zeros, zeros
for epoch in range(args.nb_epochs):
acc_loss = 0
for x in train_input.split(args.batch_size):
- mu_q_Z_given_x, log_var_q_Z_given_x = model_q_Z_given_x(x)
- z = sample_gaussian(mu_q_Z_given_x, log_var_q_Z_given_x)
- mu_p_X_given_z, log_var_p_X_given_z = model_p_X_given_z(z)
+ param_q_Z_given_x = model_q_Z_given_x(x)
+ z = sample_gaussian(param_q_Z_given_x)
+ param_p_X_given_z = model_p_X_given_z(z)
+ log_p_x_given_z = log_p_gaussian(x, param_p_X_given_z)
if args.no_dkl:
- log_q_z_given_x = log_p_gaussian(z, mu_q_Z_given_x, log_var_q_Z_given_x)
- log_p_x_z = log_p_gaussian(
- x, mu_p_X_given_z, log_var_p_X_given_z
- ) + log_p_gaussian(z, mu_p_Z, log_var_p_Z)
+ log_q_z_given_x = log_p_gaussian(z, param_q_Z_given_x)
+ log_p_z = log_p_gaussian(z, param_p_Z)
+ log_p_x_z = log_p_x_given_z + log_p_x_z
loss = -(log_p_x_z - log_q_z_given_x).mean()
else:
- log_p_x_given_z = log_p_gaussian(x, mu_p_X_given_z, log_var_p_X_given_z)
- dkl_q_Z_given_x_from_p_Z = dkl_gaussians(
- mu_q_Z_given_x, log_var_q_Z_given_x, mu_p_Z, log_var_p_Z
- )
+ dkl_q_Z_given_x_from_p_Z = dkl_gaussians(param_q_Z_given_x, param_p_Z)
loss = (-log_p_x_given_z + dkl_q_Z_given_x_from_p_Z).mean()
optimizer.zero_grad()
# Save the same images after encoding / decoding
-mu_q_Z_given_x, log_var_q_Z_given_x = model_q_Z_given_x(x)
-z = sample_gaussian(mu_q_Z_given_x, log_var_q_Z_given_x)
-mu_p_X_given_z, log_var_p_X_given_z = model_p_X_given_z(z)
-x = sample_gaussian(mu_p_X_given_z, log_var_p_X_given_z)
+param_q_Z_given_x = model_q_Z_given_x(x)
+z = sample_gaussian(param_q_Z_given_x)
+param_p_X_given_z = model_p_X_given_z(z)
+x = sample_gaussian(param_p_X_given_z)
save_image(x, "output.png")
+save_image(param_p_X_given_z[0], "output_mean.png")
# Generate a bunch of images
-z = sample_gaussian(mu_p_Z.expand(x.size(0), -1), log_var_p_Z.expand(x.size(0), -1))
-mu_p_X_given_z, log_var_p_X_given_z = model_p_X_given_z(z)
-x = sample_gaussian(mu_p_X_given_z, log_var_p_X_given_z)
+z = sample_gaussian(
+ (param_p_Z[0].expand(x.size(0), -1), param_p_Z[1].expand(x.size(0), -1))
+)
+param_p_X_given_z = model_p_X_given_z(z)
+x = sample_gaussian(param_p_X_given_z)
save_image(x, "synth.png")
+save_image(param_p_X_given_z[0], "synth_mean.png")
######################################################################