def eval_hpl(ae_sd_path, zgen_sd_path, zdis_sd_path):
"""
Evaluate on held-out set
Setting *_sd_path=None will evaluate random intialization for that net
"""
device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
if not os.path.isdir(DATASET_PATH):
os.mkdir(DATASET_PATH)
mnist_test_loader = get_mnist_test_data(store_location=DATASET_PATH)
test_batch, _ = next(
iter(mnist_test_loader)) # Used for visual checkpoints of progress
z_size = 128
autoencoder = AE(input_shape=784, z_size=z_size).to(device)
if ae_sd_path:
autoencoder.load_state_dict(
torch.load(os.path.join(MODEL_PATH, ae_sd_path),
map_location=device))
zgen = LatentGeneratorMLP(z_size, 32).to(device)
if zgen_sd_path:
zgen.load_state_dict(
torch.load(os.path.join(MODEL_PATH, zgen_sd_path),
map_location=device))
zdis = LatentDiscriminatorMLP(z_size, 32).to(device)
if zdis_sd_path:
zdis.load_state_dict(
torch.load(os.path.join(MODEL_PATH, zdis_sd_path),
map_location=device))
print(f"Evaluating HPL on MNIST digits held-out set")
autoencoder.eval()
zgen.eval()
zdis.eval()
Dloss = 0
Gloss = 0
for image_tensors, _ in tqdm(mnist_test_loader):
ones_target = torch.ones(len(image_tensors), 1).to(device)
zeros_target = torch.zeros(len(image_tensors), 1).to(device)
# Compute discriminator loss with real and fake latent codes
input_features = image_tensors.view(-1, 784).to(device)
input_features = (input_features * 2) - 1
real_codes = autoencoder.encode(input_features)
real_codes = real_codes.detach()
real_scores = zdis(real_codes)
real_loss = F.binary_cross_entropy_with_logits(
real_scores, 0.9 * ones_target) # Smoothed "real" label
prior = (torch.rand(len(image_tensors), z_size).to(device) * 2) - 1
fake_codes = zgen(prior)
fake_scores = zdis(fake_codes.detach())
fake_loss = F.binary_cross_entropy_with_logits(fake_scores,
zeros_target)
zdis_loss = real_loss + fake_loss
Dloss += zdis_loss.item()
# Compute generator loss for maximizing fooling of zdis
prior = (torch.rand(len(image_tensors), z_size).to(device) * 2) - 1
fake_codes = zgen(prior)
fake_scores = zdis(fake_codes)
zgen_loss = F.binary_cross_entropy_with_logits(fake_scores,
ones_target)
Gloss += zgen_loss.item()
Dloss /= len(mnist_test_loader)
Gloss /= len(mnist_test_loader)
print(f"Held-out D-loss = {Dloss:.6f}, G-loss = {Gloss:.6f}")
# Display reconstructions on first batch
prior = (torch.rand(len(test_batch), z_size).to(device) * 2) - 1
testZ = zgen(prior)
testX = (autoencoder.decode(testZ) + 1) * 0.5
save_images(os.path.join(RESULT_PATH, "HPL-grid-eval"), testX)
def train_gan(epochs):
device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
if not os.path.isdir(DATASET_PATH):
os.mkdir(DATASET_PATH)
mnist_train_loader = get_mnist_train_data(store_location=DATASET_PATH)
mnist_test_loader = get_mnist_test_data(store_location=DATASET_PATH)
test_batch, _ = next(iter(mnist_test_loader)) # Used for visual checkpoints of progress
z_size = 128
gen = PixelGeneratorMLP(z_size, 32, 784).to(device)
gen_opt = optim.Adam(gen.parameters(), lr=1e-3)
dis = PixelDiscriminatorMLP(784, 32, 1).to(device)
dis_opt = optim.Adam(dis.parameters(), lr=1e-3)
print(gen)
print(dis)
with torch.no_grad():
testZ = (torch.rand(len(test_batch), z_size).to(device) * 2) - 1
testX = (gen(testZ) + 1) * 0.5
save_images(os.path.join(RESULT_PATH, "GAN-grid-0"), testX)
print(f"Training GAN for {epochs} epochs on MNIST digits")
for epoch in range(epochs):
Dloss = 0
Gloss = 0
for image_tensors, _ in tqdm(mnist_train_loader):
ones_target = torch.ones(len(image_tensors), 1).to(device)
zeros_target = torch.zeros(len(image_tensors), 1).to(device)
# Compute discriminator loss with real and fake latent codes
real_images = image_tensors.view(-1, 784).to(device)
real_images = (real_images * 2) - 1
# real_images += torch.normal(torch.zeros(real_images.shape), torch.ones(real_images.shape) * (1. / (4. * 255.))).to(device) # add a bit of noise
# real_images = torch.clamp(real_images, -1, 1)
real_scores = dis(real_images)
real_loss = F.binary_cross_entropy_with_logits(real_scores, 0.9 * ones_target) # Smoothed "real" label
prior = (torch.rand(len(image_tensors), z_size).to(device) * 2) - 1
fake_images = gen(prior)
fake_scores = dis(fake_images.detach())
fake_loss = F.binary_cross_entropy_with_logits(fake_scores, zeros_target)
dis_loss = real_loss + fake_loss
Dloss += dis_loss.item()
dis_opt.zero_grad()
dis_loss.backward()
dis_opt.step()
# Compute generator loss for maximizing fooling of dis
prior = (torch.rand(len(image_tensors), z_size).to(device) * 2) - 1
fake_images = gen(prior)
fake_scores = dis(fake_images)
gen_loss = F.binary_cross_entropy_with_logits(fake_scores, ones_target)
Gloss += gen_loss.item()
# print(gen_loss.item())
gen_opt.zero_grad()
gen_loss.backward()
gen_opt.step()
Dloss /= len(mnist_train_loader)
Gloss /= len(mnist_train_loader)
with torch.no_grad():
testZ = (torch.rand(len(test_batch), z_size).to(device) * 2) - 1
testX = (gen(testZ) + 1) * 0.5
meanX, stdX = testX.mean(), testX.std()
print("epoch : {}/{}, D-loss = {:.6f}, G-loss = {:.6f}, mean-X = {:.3f}, std-X = {:.3f}".format(epoch + 1, epochs, Dloss, Gloss, meanX, stdX))
save_images(os.path.join(RESULT_PATH, f"GAN-grid-{epoch+1}"), testX)
if not os.path.isdir(MODEL_PATH):
os.mkdir(MODEL_PATH)
torch.save(gen.state_dict(), os.path.join(MODEL_PATH, f"gen{epochs}.pt"))
torch.save(dis.state_dict(), os.path.join(MODEL_PATH, f"dis{epochs}.pt"))
def train_hpl_from_GAN(epochs, gen_sd_path):
"""
Trains latent-space generator which attempts to map the prior Z distribution to
match the unconditional latent distribution of a GAN generator.
Starts with a pre-trained generator (which is frozen and unchanged).
Learns pixel-space encoder for producing the "real" Z vectors for the latent discriminator's loss.
"""
device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
if not os.path.isdir(DATASET_PATH):
os.mkdir(DATASET_PATH)
mnist_train_loader = get_mnist_train_data(store_location=DATASET_PATH)
mnist_test_loader = get_mnist_test_data(store_location=DATASET_PATH)
test_batch, _ = next(
iter(mnist_test_loader)) # Used for visual checkpoints of progress
z_size = 128
# Pixel-space GANs
xgen = PixelGeneratorMLP(z_size, 32, 784).to(device)
if gen_sd_path:
xgen.load_state_dict(
torch.load(os.path.join(MODEL_PATH, gen_sd_path),
map_location=device))
xgen.eval(
) # stops dropout, otherwise it may be harder for encoder to learn mapping from X back to Z
encoder = Encoder(input_shape=784, z_size=z_size).to(device)
encoder_opt = optim.Adam(encoder.parameters(), lr=1e-3)
# Latent-space GANs
zgen = LatentGeneratorMLP(z_size, 32).to(device)
zgen_opt = optim.Adam(zgen.parameters(), lr=1e-3)
zdis = LatentDiscriminatorMLP(z_size, 32).to(device)
zdis_opt = optim.Adam(zdis.parameters(), lr=1e-3)
print(zgen)
print(zdis)
with torch.no_grad():
prior = (torch.rand(len(test_batch), z_size).to(device) * 2) - 1
testZ = zgen(prior)
testX = (xgen(testZ) + 1) * 0.5
save_images(os.path.join(RESULT_PATH, f"HPL-grid-0"), testX)
print(f"Training HPL transfer mapping for {epochs} epochs on MNIST digits")
for epoch in range(epochs):
Eloss = 0
Dloss = 0
Gloss = 0
for image_tensors, _ in tqdm(mnist_train_loader):
ones_target = torch.ones(len(image_tensors), 1).to(device)
zeros_target = torch.zeros(len(image_tensors), 1).to(device)
# Compute discriminator loss with real and fake latent codes
input_features = image_tensors.view(-1, 784).to(device)
input_features = (input_features * 2) - 1
real_codes = encoder(input_features)
real_codes = real_codes.detach()
real_scores = zdis(real_codes)
real_loss = F.binary_cross_entropy_with_logits(
real_scores, 0.9 * ones_target) # Smoothed "real" label
prior = (torch.rand(len(image_tensors), z_size).to(device) * 2) - 1
fake_codes = zgen(prior)
fake_scores = zdis(fake_codes.detach())
fake_loss = F.binary_cross_entropy_with_logits(
fake_scores, zeros_target)
zdis_loss = real_loss + fake_loss
Dloss += zdis_loss.item()
zdis_opt.zero_grad()
zdis_loss.backward()
zdis_opt.step()
# Compute generator loss for maximizing fooling of zdis
prior = (torch.rand(len(image_tensors), z_size).to(device) * 2) - 1
fake_codes = zgen(prior)
fake_scores = zdis(fake_codes)
zgen_loss = F.binary_cross_entropy_with_logits(
fake_scores, ones_target)
Gloss += zgen_loss.item()
zgen_opt.zero_grad()
zgen_loss.backward()
zgen_opt.step()
# Compute encoder loss for matching Z input to pixel-space generator
prior = (torch.rand(len(image_tensors), z_size).to(device) * 2) - 1
fake_images = xgen(prior)
prior_recon = encoder(fake_images.detach())
encoder_loss = F.mse_loss(prior_recon, prior)
Eloss += encoder_loss.item()
encoder_opt.zero_grad()
encoder_loss.backward()
encoder_opt.step()
Dloss /= len(mnist_train_loader)
Gloss /= len(mnist_train_loader)
Eloss /= len(mnist_train_loader)
with torch.no_grad():
prior = (torch.rand(len(test_batch), z_size).to(device) * 2) - 1
testZ = zgen(prior)
testX = (xgen(testZ) + 1) * 0.5
meanZ, stdZ = testZ.mean(), testZ.std()
print(
"epoch : {}/{}, D-loss = {:.6f}, G-loss = {:.6f}, E-loss = {:.6f}, mean-Z = {:.3f}, std-Z = {:.3f}"
.format(epoch + 1, epochs, Dloss, Gloss, Eloss, meanZ, stdZ))
save_images(os.path.join(RESULT_PATH, f"HPL-grid-{epoch+1}"), testX)
if not os.path.isdir(MODEL_PATH):
os.mkdir(MODEL_PATH)
torch.save(zgen.state_dict(), os.path.join(MODEL_PATH, f"zgen{epochs}.pt"))
torch.save(zdis.state_dict(), os.path.join(MODEL_PATH, f"zdis{epochs}.pt"))
torch.save(encoder.state_dict(), os.path.join(MODEL_PATH,
f"enc{epochs}.pt"))
def train_hpl(epochs, state_dict_path_AE):
"""
Trains latent-space generator which attempts to map the prior Z distribution to
match the latent distribution of an AutoEncoder.
Starts with a pre-trained AutoEncoder (which is frozen and unchanged).
"""
device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")
if not os.path.isdir(DATASET_PATH):
os.mkdir(DATASET_PATH)
mnist_train_loader = get_mnist_train_data(store_location=DATASET_PATH)
mnist_test_loader = get_mnist_test_data(store_location=DATASET_PATH)
test_batch, _ = next(
iter(mnist_test_loader)) # Used for visual checkpoints of progress
z_size = 128
autoencoder = AE(input_shape=784, z_size=z_size).to(device)
if state_dict_path_AE:
autoencoder.load_state_dict(
torch.load(os.path.join(MODEL_PATH, state_dict_path_AE),
map_location=device))
zgen = LatentGeneratorMLP(z_size, 32).to(device)
zgen_opt = optim.Adam(zgen.parameters(), lr=1e-3)
zdis = LatentDiscriminatorMLP(z_size, 32).to(device)
zdis_opt = optim.Adam(zdis.parameters(), lr=1e-3)
print(zgen)
print(zdis)
with torch.no_grad():
prior = (torch.rand(len(test_batch), z_size).to(device) * 2) - 1
testZ = zgen(prior)
testX = (autoencoder.decode(testZ) + 1) * 0.5
save_images(os.path.join(RESULT_PATH, "HPL-grid-0"), testX)
print(f"Training HPL transfer mapping for {epochs} epochs on MNIST digits")
for epoch in range(epochs):
Dloss = 0
Gloss = 0
for image_tensors, _ in tqdm(mnist_train_loader):
ones_target = torch.ones(len(image_tensors), 1).to(device)
zeros_target = torch.zeros(len(image_tensors), 1).to(device)
# Compute discriminator loss with real and fake latent codes
input_features = image_tensors.view(-1, 784).to(device)
input_features = (input_features * 2) - 1
real_codes = autoencoder.encode(input_features)
real_codes = real_codes.detach()
real_scores = zdis(real_codes)
real_loss = F.binary_cross_entropy_with_logits(
real_scores, 0.9 * ones_target) # Smoothed "real" label
prior = (torch.rand(len(image_tensors), z_size).to(device) * 2) - 1
fake_codes = zgen(prior)
fake_scores = zdis(fake_codes.detach())
fake_loss = F.binary_cross_entropy_with_logits(
fake_scores, zeros_target)
zdis_loss = real_loss + fake_loss
Dloss += zdis_loss.item()
zdis_opt.zero_grad()
zdis_loss.backward()
zdis_opt.step()
# Compute generator loss for maximizing fooling of zdis
prior = (torch.rand(len(image_tensors), z_size).to(device) * 2) - 1
fake_codes = zgen(prior)
fake_scores = zdis(fake_codes)
zgen_loss = F.binary_cross_entropy_with_logits(
fake_scores, ones_target)
Gloss += zgen_loss.item()
zgen_opt.zero_grad()
zgen_loss.backward()
zgen_opt.step()
Dloss /= len(mnist_train_loader)
Gloss /= len(mnist_train_loader)
with torch.no_grad():
prior = (torch.rand(len(test_batch), z_size).to(device) * 2) - 1
testZ = zgen(prior)
testX = (autoencoder.decode(testZ) + 1) * 0.5
meanZ, stdZ = testZ.mean(), testZ.std()
print(
"epoch : {}/{}, D-loss = {:.6f}, G-loss = {:.6f}, mean-Z = {:.3f}, std-Z = {:.3f}"
.format(epoch + 1, epochs, Dloss, Gloss, meanZ, stdZ))
save_images(os.path.join(RESULT_PATH, f"HPL-grid-{epoch+1}"), testX)
if not os.path.isdir(MODEL_PATH):
os.mkdir(MODEL_PATH)
torch.save(zgen.state_dict(), os.path.join(MODEL_PATH, f"zgen{epochs}.pt"))
torch.save(zdis.state_dict(), os.path.join(MODEL_PATH, f"zdis{epochs}.pt"))
def train_adversarial_autoencoder(epochs):
"""
Note "real_codes" and "fake_codes" are reversed compared to HPL
real --> from prior ~ Uniform [0, 1)
fake --> from AAE encoder (generator)
"""
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
if not os.path.isdir(DATASET_PATH):
os.mkdir(DATASET_PATH)
mnist_train_loader = get_mnist_train_data(store_location=DATASET_PATH)
mnist_test_loader = get_mnist_test_data(store_location=DATASET_PATH)
test_batch, _ = next(iter(mnist_test_loader)) # Used for visual checkpoints of progress
autoencoder = AE(input_shape=784).to(device)
ae_opt = optim.Adam(autoencoder.parameters(), lr=1e-3)
zdis = LatentDiscriminatorMLP(128).to(device)
zdis_opt = optim.Adam(zdis.parameters(), lr=1e-3)
with torch.no_grad():
testZ = torch.rand(len(test_batch), 128).to(device) # Directly from random uniform distribution [0, 1)
testX = autoencoder.decode(testZ)
save_images(os.path.join(RESULT_PATH, "AAE-grid-0"), testX)
print(f"Training AAE for {epochs} epochs on MNIST digits")
for epoch in range(epochs):
Dloss = 0
Gloss = 0
AEloss = 0
for image_tensors, _ in tqdm(mnist_train_loader):
ones_target = torch.ones(len(image_tensors), 1).to(device)
zeros_target = torch.zeros(len(image_tensors), 1).to(device)
# Compute discriminator loss with real and fake latent codes
input_features = image_tensors.view(-1, 784).to(device)
fake_codes = autoencoder.encode(input_features)
fake_scores = zdis(fake_codes.detach())
fake_loss = F.binary_cross_entropy(fake_scores, zeros_target)
real_codes = torch.rand(len(image_tensors), 128).to(device)
real_scores = zdis(real_codes.detach())
real_loss = F.binary_cross_entropy(real_scores, 0.9 * ones_target)
zdis_loss = real_loss + fake_loss
Dloss += zdis_loss.item()
zdis_opt.zero_grad()
zdis_loss.backward()
zdis_opt.step()
# Compute AAE loss = reconstruction loss + generator loss for maximizing fooling of zdis
fake_scores = zdis(fake_codes)
gen_loss = F.binary_cross_entropy(fake_scores, ones_target)
Gloss += gen_loss.item()
reconstructions = autoencoder.decode(fake_codes)
recon_loss = F.mse_loss(reconstructions, input_features)
AEloss += recon_loss.item()
ae_loss = gen_loss + recon_loss
ae_opt.zero_grad()
ae_loss.backward()
ae_opt.step()
num_batch = len(mnist_train_loader)
Dloss /= num_batch
Gloss /= num_batch
AEloss /= num_batch
with torch.no_grad():
prior = torch.rand(len(test_batch), 128).to(device) # Directly from random uniform distribution [0, 1)
testX = autoencoder.decode(prior)
test_features = test_batch.view(-1, 784).to(device)
testZ = autoencoder.encode(test_features)
meanZ, stdZ = testZ.mean(), testZ.std()
print("epoch : {}/{}, D-loss = {:.6f}, G-loss = {:.6f}, AE-loss = {:.6f}, mean-Z = {:.3f}, std-Z = {:.3f}".format(epoch + 1, epochs, Dloss, Gloss, AEloss, meanZ, stdZ))
save_images(os.path.join(RESULT_PATH, f"AAE-grid-{epoch+1}"), testX)
if not os.path.isdir(MODEL_PATH):
os.mkdir(MODEL_PATH)
torch.save(autoencoder.state_dict(), os.path.join(MODEL_PATH, f"aae{epochs}.pt"))
torch.save(zdis.state_dict(), os.path.join(MODEL_PATH, f"pzdis{epochs}.pt"))