def Hellinger(theta):
# Example of usage of the code provided for answering Q2.5 as well as recommended hyper parameters.
model = q2_model.Critic(2)
optim = torch.optim.SGD(model.parameters(), lr=1e-3)
sampler1 = iter(q2_sampler.distribution1(0, 512))
sampler2 = iter(q2_sampler.distribution1(theta, 512))
lambda_reg_lp = 50 # Recommended hyper parameters for the lipschitz regularizer.
steps = 500
for step in range(steps):
data1 = torch.from_numpy(next(sampler1)).float()
data2 = torch.from_numpy(next(sampler2)).float()
loss = -vf_squared_hellinger(data1, data2, model)
print('Step {} : loss {}'.format(step, loss))
optim.zero_grad()
loss.backward()
optim.step()
data1 = torch.from_numpy(next(sampler1)).float()
data2 = torch.from_numpy(next(sampler2)).float()
return vf_squared_hellinger(data1, data2, model)
t = one - torch.exp(-critic_y)
return torch.mean(one - torch.exp(-critic_x)) + torch.mean(-t / (one - t))
if __name__ == '__main__':
# Example of usage of the code provided for answering Q2.5 as well as recommended hyper parameters.
lambda_reg_lp = 50 # Recommended hyper parameters for the lipschitz regularizer.
sh = []
w = []
wlp = []
thetas = [theta * 0.1 for theta in range(0, 21)]
for theta1 in thetas:
modelsh = q2_model.Critic(2)
#modelw = q2_model.Critic(2)
modelwlp = q2_model.Critic(2)
optimsh = torch.optim.SGD(modelsh.parameters(), lr=1e-3)
#optimw = torch.optim.SGD(modelw.parameters(), lr=1e-3)
optimwlp = torch.optim.SGD(modelwlp.parameters(), lr=1e-3)
sampler1 = iter(q2_sampler.distribution1(0, 512))
sampler2 = iter(q2_sampler.distribution1(theta1, 512))
for i in range(500):
x = torch.Tensor(next(sampler1))
y = torch.Tensor(next(sampler2))
modelsh.zero_grad()
:param critic: (Module) - torch module used to compute the Wasserstein distance
:return: (FloatTensor) - shape: (1,) - Estimate of the Wasserstein distance
"""
return torch.mean(critic(x)) - torch.mean(critic(y))
def vf_squared_hellinger(x, y, critic):
"""
Complete me. DONT MODIFY THE PARAMETERS OF THE FUNCTION. Otherwise, tests might fail.
*** The notation used for the parameters follow the one from Nowazin et al: https://arxiv.org/pdf/1606.00709.pdf
In other word, x are samples from the distribution P and y are samples from the distribution Q. Please note that the Critic is unbounded. ***
:param p: (FloatTensor) - shape: (batchsize x featuresize) - Samples from a distribution p.
:param q: (FloatTensor) - shape: (batchsize x featuresize) - Samples from a distribution q.
:param critic: (Module) - torch module used to compute the Squared Hellinger.
:return: (FloatTensor) - shape: (1,) - Estimate of the Squared Hellinger
"""
return torch.mean(1 - torch.exp(-critic(x))) - torch.mean(
(1 - torch.exp(-critic(y))) / (torch.exp(-critic(y))))
if __name__ == '__main__':
# Example of usage of the code provided for answering Q2.5 as well as recommended hyper parameters.
model = q2_model.Critic(2)
optim = torch.optim.SGD(model.parameters(), lr=1e-3)
sampler1 = iter(q2_sampler.distribution1(0, 512))
theta = 0
sampler2 = iter(q2_sampler.distribution1(theta, 512))
lambda_reg_lp = 50 # Recommended hyper parameters for the lipschitz regularizer.
if __name__ == '__main__':
# Example of usage of the code provided for answering Q2.5 as well as recommended hyper parameters.
thetas = np.arange(0.0, 2.1, 0.1)
device = 'cuda' if torch.cuda.is_available() else 'cpu'
dtype = torch.cuda.FloatTensor if torch.cuda.is_available(
) else torch.FloatTensor
print("Using ", device)
torch.manual_seed(10)
np.random.seed(10)
from matplotlib import pyplot as plt
hellinger_dict = {}
wd_dict = {}
for theta in thetas:
model1 = q2_model.Critic(2).to(device)
optim1 = torch.optim.SGD(model1.parameters(), lr=1e-3)
model2 = q2_model.Critic(2).to(device)
optim2 = torch.optim.SGD(model2.parameters(), lr=1e-3)
lambda_reg_lp = 50 # Recommended hyper parameters for the lipschitz regularizer.
iterations = 2500
for i in range(iterations):
# print("iteration and theta is: ", i, " ", theta)
## data is the same for both the models
sampler1 = iter(q2_sampler.distribution1(0, 512))
sampler2 = iter(q2_sampler.distribution1(theta, 512))
data1 = torch.from_numpy(next(sampler1)).to(device)