def train_discriminator():
input_size = 1
n_hidden = 3
n_units = 256
out_size = 1
out_sigmoid = True
batch_size = 512
n_epochs = 800
lr = 1e-2
# create model, training criterion and optimizer.
criterion = ValueFunction()
model = MLP2(input_size, n_hidden, n_units, out_size, out_sigmoid).to(device)
print(model)
optimizer = torch.optim.SGD(model.parameters(), lr=lr)
# get iterators for distributions
f1 = iter(samplers.distribution4(batch_size))
f0 = iter(samplers.distribution3(batch_size))
losses = train(model, f1, f0, optimizer, criterion, device, n_epochs)
# visualise loss.
plt.figure()
plt.plot(losses)
# plt.show()
return model
def gan_eval():
epochs = 1000
batch_size = 1000
hidden_size = 25
n_hidden = 2
input_dim = 1
f0 = samplers.distribution3(batch_size)
f1 = samplers.distribution4(batch_size)
D = Discriminator(input_dim, hidden_size, n_hidden)
train(D, f0, f1, GAN, batch_size=batch_size, epochs=epochs)
xx = torch.randn(10000)
f = lambda x: torch.tanh(x * 2 + 1) + x * 0.75
d = lambda x: (1 - torch.tanh(x * 2 + 1)**2) * 2 + 0.75
plt.hist(f(xx), 100, alpha=0.5, density=1)
plt.hist(xx, 100, alpha=0.5, density=1)
plt.xlim(-5, 5)
plt.savefig('histogram')
plt.close()
xx = np.linspace(-5, 5, 1000)
N = lambda x: np.exp(-x**2 / 2.) / ((2 * np.pi)**0.5)
f0_x_tensor = Variable(
torch.from_numpy(np.float32(xx.reshape(batch_size, input_dim))))
D_x = D(f0_x_tensor)
f1_est = (N(f0_x_tensor) * D_x) / (1 - D_x)
# Plot the discriminator output.
r = -D_x.detach().numpy()
plt.figure(figsize=(8, 4))
plt.subplot(1, 2, 1)
plt.plot(xx, r)
plt.title(r'$D(x)$')
estimate = f1_est.detach().numpy()
# Plot the density.
plt.subplot(1, 2, 2)
plt.plot(xx, estimate)
plt.plot(
f(torch.from_numpy(xx)).numpy(),
d(torch.from_numpy(xx)).numpy()**(-1) * N(xx))
plt.legend(['Estimated', 'True'])
plt.title('Estimated vs True')
plt.savefig('Estimated_vs_Exact.png')
plt.close()
# exact
xx = np.linspace(-5, 5, 1000)
N = lambda x: np.exp(-x**2 / 2.) / ((2 * np.pi)**0.5)
plt.plot(
f(torch.from_numpy(xx)).numpy(),
d(torch.from_numpy(xx)).numpy()**(-1) * N(xx))
plt.plot(xx, N(xx))
############### import the sampler ``samplers.distribution4''
############### train a discriminator on distribution4 and standard gaussian
############### estimate the density of distribution4
#######--- INSERT YOUR CODE BELOW ---#######
dist4 = samplers.distribution4(batch_size=256)
dist3 = samplers.distribution3(batch_size=256)
D = train('JSD', dist4, dist3, args.use_cuda, args.setup)
#dist1 = samplers.distribution4(batch_size=1)
#dist0 = samplers.distribution3(batch_size=1)
data_points_D = []
data_points_d4 = []
print(xx)
for x in xx:
if args.use_cuda:
y = D(torch.from_numpy(np.array([N(x)])).float().cuda())
else:
y = D(torch.from_numpy(np.array([N(x)])).float())
data_points_D.append(y)
data_points_d4.extend(N(x) * y / (1 - y))
xx = np.linspace(-5, 5, 1000)
def N(x):
return np.exp(-x**2 / 2.) / ((2 * np.pi)**0.5)
plt.plot(
f(torch.from_numpy(xx)).numpy(),
d(torch.from_numpy(xx)).numpy()**(-1) * N(xx))
plt.plot(xx, N(xx))
batch_size = 512
m_minibatch = 100
p_iter = iter(distribution3(batch_size))
fo = p_iter
q_iter = iter(distribution4(batch_size))
f1 = q_iter
Discrim, jsd = js_divergence(f1, fo, m_minibatch)
Discrim = Discrim(torch.Tensor(xx).unsqueeze(dim=1))
r = Discrim.detach().numpy().reshape(-1)
plt.figure(figsize=(8, 4))
plt.subplot(1, 2, 1)
plt.plot(xx, r)
plt.title(r'$D(x)$')
# estimate the density of distribution4 (on xx) using the discriminator;
estimate = N(xx) * r / (1 - r)
# Create model
discriminator = sq()
discriminator.add(Dense(units=128, activation='relu', input_dim=1))
discriminator.add(Dense(units=128, activation='relu'))
discriminator.add(Dense(units=128, activation='relu'))
discriminator.add(Dense(units=1, activation='sigmoid'))
discriminator.compile(loss=D_Loss, optimizer=Adam(lr=0.00005))
# Make target dummy for Keras
target_dummy = np.zeros(512 * 2)
#Create samplers
gaussian_dist_gen = samplers.distribution3()
mystery_dist_gen = samplers.distribution4(batch_size=512)
for _ in range(EPOCHS):
#Sample the distributions
x = next(mystery_dist_gen)
y = next(gaussian_dist_gen)
#Train
discriminator.train_on_batch(np.concatenate((x, y)), target_dummy)
############### plotting things
############### (1) plot the output of your trained discriminator
############### (2) plot the estimated density contrasted with the true density
# the binary cross entropy l(y,D(x)) is: -1 * [y log(D(x)) + (1 - y)*log(1 - D(x))]
# if we take y to be zero (distribution q), we optimize -log(1-D(x))
# if we take y to be 1 (distribution p), we optimize -log(D(x))
# we want to minimize this (this is directly equivalent to maximizing our objective function)
criterion = nn.BCEWithLogitsLoss(
) # binary cross entropy with built-in sigmoid
optimizer = optim.SGD(d.parameters(), lr=init_lr)
for epoch in range(num_epochs):
# sample minibatches from the two distributions:
distr1 = sp.distribution1(0, batch_size)
dist1 = iter(distr1)
samples1 = np.squeeze(next(dist1)[:, 0])
t1 = torch.Tensor(samples1).to(device)
distr3 = sp.distribution3(batch_size)
dist3 = iter(distr3)
samples3 = np.squeeze(next(dist3))
t3 = torch.Tensor(samples3).to(device)
d.zero_grad() # gradients to zero
# gradients on 'real' distribution:
out_r = d(t1)
err_r = criterion(out_r, torch.Tensor([1]).to(device))
err_r.backward()
# gradients on 'fake' distribution:
out_f = d(t3)
err_f = criterion(out_f, torch.Tensor([0]).to(device))
err_f.backward()
xx = np.linspace(-5, 5, 1000)
N = lambda x: np.exp(-x**2 / 2.) / ((2 * np.pi)**0.5)
plt.plot(
f(torch.from_numpy(xx)).numpy(),
d(torch.from_numpy(xx)).numpy()**(-1) * N(xx))
plt.plot(xx, N(xx))
############### import the sampler ``samplers.distribution4''
############### train a discriminator on distribution4 and standard gaussian
############### estimate the density of distribution4
#######--- INSERT YOUR CODE BELOW ---#######
from samplers import distribution3, distribution4
from problem1 import MLP_Disc, discriminator_obj, js_obj
f0 = distribution3(2048)
f1 = distribution4(2048)
func = MLP_Disc(dim=1)
_ = func._train(f1, f0, discriminator_obj, lr=0.001, epochs=1500)
# We know from problem 5 that f1 = f0 D*/(1-D*)
############### plotting things
############### (1) plot the output of your trained discriminator
############### (2) plot the estimated density contrasted with the true density
r = func(torch.FloatTensor(xx).view(-1, 1)).detach().numpy(
)[:, 0] # evaluate xx using your discriminator; replace xx with the output
plt.figure(figsize=(8, 4))
plt.subplot(1, 2, 1)
plt.plot(xx, r)
plt.title(r'$D(x)$')
plt.plot(
f(torch.from_numpy(xx)).numpy(),
d(torch.from_numpy(xx)).numpy()**(-1) * N(xx))
plt.plot(xx, N(xx))
plt.savefig("./images/q1_4_1.png")
############### import the sampler ``samplers.distribution4''
############### train a discriminator on distribution4 and standard gaussian
############### estimate the density of distribution4
#######--- INSERT YOUR CODE BELOW ---#######
to_tensor = lambda x: torch.as_tensor(x).float
from samplers import distribution4, distribution3
f_0 = distribution3(batch_size=512) # standard gaussian
f_1 = distribution4(batch_size=512) # modified 'unknown' distribution
print("Training discriminator...")
discriminator = get_optimal_discriminator(f_1,
f_0,
maxsteps=1_000,
threshold=1e-3)
def estimate_density(xx):
d_x = discriminator(xx)
# prevent division by zero:
d_x = np.maximum(d_x, 1e-8)
d_x = np.minimum(d_x, 1 - 1e-8)
self.relu = nn.ReLU()
self.sigmoid = nn.Sigmoid()
def forward(self, x):
x = self.relu(self.map1(x))
x = self.relu(self.map2(x))
return self.sigmoid(self.map3(x))
cuda_available = True if torch.cuda.is_available() else False
#cuda_available = False
if cuda_available:
D.cuda()
Tensor = torch.cuda.FloatTensor if cuda_available else torch.FloatTensor
dist_p = iter(distribution3(batch_size))
dist_q = iter(distribution4(batch_size))
D = Discriminator(input_size=d_input_size,
hidden_size=d_hidden_size,
output_size=d_output_size)
optimizer_D = torch.optim.Adam(D.parameters(), lr=lr)
for epoch in range(n_epochs):
samples_p = next(dist_p)
samples_q = next(dist_q)
unknown_dist = Tensor(samples_q)
std_dist = Tensor(samples_p)
device = torch.device('cpu')
js_model = torch.nn.Sequential(
torch.nn.Linear(1, 64),
torch.nn.ReLU(),
torch.nn.Linear(64, 32),
torch.nn.ReLU(),
torch.nn.Linear(32, 1),
torch.nn.Sigmoid(),
).to(device)
batch_size = 512
learning_rate = 1e-4
criterion = torch.nn.BCELoss()
optimizer_js = torch.optim.Adam(js_model.parameters(), lr=learning_rate)
real_dist = iter(distribution4(batch_size))
fake_dist = iter(distribution3(batch_size))
real_label = torch.full((batch_size, ), 1, device=device)
fake_label = torch.full((batch_size, ), 0, device=device)
for epoch in range(3000):
js_model.zero_grad()
# pdb.set_trace()
# Train on real
real_input = torch.from_numpy(next(real_dist).astype(np.float32))
output_real = js_model(real_input)
errD_real = criterion(output_real, real_label)
errD_real.backward()
# Train on fake
plt.plot(phi, values, 'o-')
if args.loss_type == "JSD":
plt.ylabel("JSD")
plt.xlabel("phi")
plt.title("JSD vs phi")
plt.savefig(directory + '_JSD_phi.png', bbox_inches='tight')
elif args.loss_type == "WD":
plt.ylabel("Wasserstein Distance")
plt.xlabel("Phi")
plt.title("WD vs. Phi")
plt.savefig(directory + '_WD_phi.png', bbox_inches='tight')
plt.clf()
if args.question == 4:
f0_dist = samplers.distribution3(512)
f1_dist = samplers.distribution4(512)
model = Discriminator(1, 50, 512, 0)
for epoch in range(num_epochs):
f1_batch = torch.from_numpy(next(f1_dist))
f0_batch = torch.from_numpy(next(f0_dist))
model.train(f1_batch.type(torch.FloatTensor),
f0_batch.type(torch.FloatTensor))
print("train")
f1_value = []
f1_real = []
f_0_values = []
discriminator_outputs = []
for x in xx:
f_0 = N(x)
x = torch.tensor([[x]])
import numpy as np import matplotlib.pyplot as plt import torch from torch.autograd import Variable import torch.nn as nn import torch.nn.functional as F import torch.optim as optim import samplers as samplers ## the data also come form the distribution cuda = torch.cuda.is_available() f_0 = next(samplers.distribution3(512)) f_1 = next(samplers.distribution4(1)) X_dim = 1 h_dim = 64 # class Net(nn.Module): # def __init__(self, input_dim=1, hidden_size=32, n_hidden=3): # super(Net, self).__init__() # self.D = torch.nn.Sequential( # torch.nn.Linear(X_dim, h_dim), # torch.nn.ReLU(), # torch.nn.Linear(h_dim, h_dim), # torch.nn.ReLU(), # torch.nn.Linear(h_dim, 1), # torch.nn.Sigmoid() # ) # def forward(self, x): # return self.D(x)
plt.hist(xx, 100, alpha=0.5, density=1) plt.xlim(-5,5) # exact xx = np.linspace(-5,5,1000) N = lambda x: np.exp(-x**2/2.)/((2*np.pi)**0.5) plt.plot(f(torch.from_numpy(xx)).numpy(), d(torch.from_numpy(xx)).numpy()**(-1)*N(xx)) plt.plot(xx, N(xx)) ############### import the sampler ``samplers.distribution4'' ############### train a discriminator on distribution4 and standard gaussian ############### estimate the density of distribution4 #######--- INSERT YOUR CODE BELOW ---####### f0 = distribution3(512) f1 = distribution4(512) D = Discriminator(1) train(D, f1, f0, loss_metric='JSD') x = torch.Tensor(xx).reshape(1000, 1) output = D(x) output = output.detach().numpy().reshape(-1) scaling_factor = output / (1 - output) estimated_density = scaling_factor * N(xx) ############### plotting things ############### (1) plot the output of your trained discriminator ############### (2) plot the estimated density contrasted with the true density
def training_loop(LossFct,
x,
distribution=1,
learning_rate=0.0001,
num_epochs=50000):
# Device configuration
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
# Distributions properties
if distribution == 1:
p_gen = samplers.distribution1(0)
q_gen = samplers.distribution1(x)
nb_input = 2
elif distribution == 4:
q_gen = samplers.distribution3(2048)
p_gen = samplers.distribution4(2048)
nb_input = 1
p_gen.send(None)
q_gen.send(None)
D = discriminators.Discriminator(n_input=nb_input).to(device)
# Loss and optimizer
optimizer = torch.optim.SGD(D.parameters(), lr=learning_rate)
# Train the model
trainLoss = []
trainAcc = []
meanLoss = 0
correct = 0
total = 0
log_frequency = 100
for epoch in range(num_epochs):
# exp_lr_scheduler.step()
p = torch.from_numpy(p_gen.send(0)).float().to(device)
q = torch.from_numpy(q_gen.send(x)).float().to(device)
labels_real = torch.ones(p.shape[0]).to(device)
labels_fake = torch.zeros(q.shape[0]).to(device)
# Forward pass
outputs_real = torch.sigmoid(D(p))
outputs_fake = torch.sigmoid(D(q))
predicted_real = (outputs_real.data > 0.5).float().squeeze()
predicted_fake = (outputs_fake.data > 0.5).float().squeeze()
total += 2 * labels_real.size(0)
correct_this_batch = (predicted_real == labels_real).sum().item() + (
predicted_fake == labels_fake).sum().item()
correct += correct_this_batch
loss = LossFct.forward(
outputs_real, outputs_fake, labels_real, labels_fake, p, q, D
) #(torch.log(torch.tensor([2.0])).to(device) + 0.5*criterion(outputs_real, labels_real) + 0.5*criterion(outputs_fake, labels_fake))
meanLoss += loss.item()
# Backward and optimize
optimizer.zero_grad()
loss.backward()
optimizer.step()
if epoch % log_frequency == 0:
print('Epoch [{}/{}]'.format(epoch, num_epochs))
print('Loss: {:.4f}({:.4f}), Acc: {:.3f}({:.3f})'.format(
loss.item(), meanLoss / (epoch + 1),
correct_this_batch * 100 / (2 * labels_real.size(0)),
correct * 100 / total))
trainLoss.append(meanLoss / (epoch + 1))
trainAcc.append(100 * correct / total)
return loss, D
#######--- INSERT YOUR CODE BELOW ---#######
D = Discriminator(1)
optimizer_D = torch.optim.Adam(D.parameters(), lr=setting.lr)
cuda_available = True if torch.cuda.is_available() else False
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
cuda_available = True if torch.cuda.is_available() else False
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
if cuda_available:
D.to(device)
Tensor = torch.cuda.FloatTensor if cuda_available else torch.FloatTensor
dist_p = iter(distribution3(setting.samp_per_epoch))
dist_q = iter(distribution4(setting.samp_per_epoch))
for epoch in range(setting.n_epochs):
samples_p = next(dist_p)
samples_q = next(dist_q)
q_dist = Tensor(samples_q)
p_dist = Tensor(samples_p)
optimizer_D.zero_grad()
p_decision = D(p_dist)
q_decision = D(q_dist)
best_list.append(best_loss)
plt.plot(phi_values, best_list, "o")
plt.xlabel('$\phi$')
plt.ylabel('WD')
plt.title('Wasserstein distance per values of $\phi$')
plt.savefig("problem1-3-WD.png")
plt.show()
# Problem 1.4
d_4 = Discriminator(input_size=1)
optimizer = torch.optim.SGD(d_4.parameters(), lr=1e-3)
criterion = JSDLoss2()
f_0 = iter(samplers.distribution3(batch_size=batch_size))
f_1 = iter(samplers.distribution4(batch_size=batch_size))
best_loss = -float('Inf')
for batch in range(10000):
f_0_samples = torch.as_tensor(next(f_0), dtype=torch.float32)
f_1_samples = torch.as_tensor(next(f_1), dtype=torch.float32)
optimizer.zero_grad()
f_0_outputs = d_4(f_0_samples)
f_1_outputs = d_4(f_1_samples)
print((torch.log(1 - f_0_outputs)).mean())
loss = -criterion(f_0_outputs, f_1_outputs)
#######--- INSERT YOUR CODE BELOW ---#######
epoch_count = 25000
unk_disc = Discriminator(input_size=1).cuda()
unk_loss = nn.BCELoss(reduction='mean')
unk_optim = optim.SGD(unk_disc.parameters(), lr=1e-3)
for i in range(epoch_count):
unk_disc.zero_grad()
#get data for both distributions from samplers
real_dist = iter(samplers.distribution4(512))
real_samples = next(real_dist)
real_tensor_samples = torch.tensor(real_samples,
requires_grad=True).float().cuda()
fake_dist = iter(samplers.distribution3(512))
fake_samples = next(fake_dist)
fake_tensor_samples = torch.tensor(fake_samples,
requires_grad=True).float().cuda()
real_targets = torch.ones([len(real_tensor_samples), 1]).cuda()
fake_targets = torch.zeros([len(fake_tensor_samples), 1]).cuda()
real_output = unk_disc(real_tensor_samples)
fake_output = unk_disc(fake_tensor_samples)
loss_real_output = unk_loss(real_output, real_targets)
loss_fake_output = unk_loss(fake_output, fake_targets)
total_output = loss_real_output + loss_fake_output
plt.hist(xx, 100, alpha=0.5, density=1)
plt.xlim(-5, 5)
# exact
xx = np.linspace(-5, 5, 512)
N = lambda x: np.exp(-x**2 / 2.) / ((2 * np.pi)**0.5)
plt.plot(
f(torch.from_numpy(xx)).numpy(),
d(torch.from_numpy(xx)).numpy()**(-1) * N(xx))
plt.plot(xx, N(xx))
############### import the sampler ``samplers.distribution4''
############### train a discriminator on distribution4 and standard gaussian
############### estimate the density of distribution4
#######--- INSERT YOUR CODE BELOW ---#######
f0 = samplers.distribution3()
f1 = samplers.distribution4()
Descr_M = Descr(hidden_size=512,
mini_batch=512,
learning_rate=0.0001,
num_epochs=1000,
print_interval=10)
Descr_M.run_main_loop(f0, f1)
random_input_tensor = Variable(
torch.from_numpy(
np.float32(xx.reshape(Descr_M.minibatch_size, Descr_M.input_size))))
Network_Output = Descr_M.Network(random_input_tensor)
f1_estimated = (1 - Network_Output) / Network_Output * N(random_input_tensor)
