def rsample(self, sample_shape=torch.Size([])):
a_sampler = D.OneHotCategorical(probs=torch.ones(len(self.a_domain)))
probs = {}
x_a_dists = {}
for a in self.a_domain:
probs[a] = {}
for x in self.x_support:
probs[a][x] = math.exp(self.log_prob(x, a))
normalise = sum(probs[a].values())
for x in self.x_support:
probs[a][x] = probs[a][x] / normalise
x_a_dists[a] = EmpiricalDistribution(None, domain=self.x_support, probs=probs[a])
self.test = EmpiricalDistribution(None, domain=self.x_support, probs=probs[a])
a_vals = a_sampler.sample_n(sample_shape.numel())
a_counts = torch.sum(a_vals, axis=0)
x_samples = []
a_samples = []
for a_c, a_vals in zip(a_counts, self.a_domain):
a_c = int(a_c)
x_samples.append(x_a_dists[a_vals].sample_n(a_c))
a_samples.append(torch.Tensor([a_vals] * a_c))
x_samples = torch.cat(x_samples).view(*sample_shape, -1)
a_samples = torch.cat(a_samples).view(*sample_shape, -1)
return x_samples, a_samples
test_a = torch.Tensor(test_df.loc[:, sensitive].values)
test_a = torch.cat([test_a, 1 - test_a], axis=1)
test_sample = list(zip(test_x, test_a))
#%%
# Fair Density setup
x_support = all_binaries(DOMAIN, threshold=[1, 2])
a_domain = [tuple(int(v) for v in t) for t in torch.eye(train_a.shape[1])]
train_x_cond_a = []
for i in range(train_a.shape[1]):
a_mask = train_a[:, i] > 0
train_x_cond_a.append(
EmpiricalDistribution(train_x[a_mask],
domain=x_support,
baseline=0))
#train_x_cond_a.append(EmpiricalDistribution([], domain=x_support, baseline=1))
save_dir = 'dt_compas_small_{}_{}/fold_{}'.format(SENSITIVE_ATTRIBUTE, TAU,
f_idx)
if not os.path.exists(save_dir):
os.makedirs(save_dir)
#%%
boost = BoostDensityDT(TAU, train_sample, test_sample, train_x_cond_a,
x_support, a_domain)
#%%
boost.init_boost(num_iter=50, max_depth=8)
## nn.Linear(20, 20),
# nn.ReLU(),
# nn.Linear(20, 1),
nn.Hardtanh(min_val=-math.log(2), max_val=math.log(2))
)
#%%
# Fair Density setup
x_support = all_binaries(DOMAIN, threshold=[1,2])
a_domain = [tuple(int(v) for v in t) for t in torch.eye(train_a.shape[1])]
train_x_cond_a = []
for i in range(train_a.shape[1]):
a_mask = train_a[:, i] > 0
train_x_cond_a.append(EmpiricalDistribution(train_x[a_mask], domain=x_support, baseline=1))
#train_x_cond_a.append(EmpiricalDistribution([], domain=x_support, baseline=1))
#%%
boost = BoostDensity(TAU, train_sample, test_sample, train_x_cond_a, x_support, a_domain, model)
#%%
boost.init_boost(optimiser_gen=optim.Adam, batch_size=128)
#%%
start = time.time()
boost.boost()
final = time.time()
results[f_idx] = {
'fold': f_idx,