def evel_model(G, save_path, name, data1, data2, data3, r_data, distance):
no_graph = False
nz = 2
z = torch.randn(10000, nz).cuda()
label = torch.zeros(10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data1_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data2_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(10000).long().cuda() + 1 # torch.LongTensor(10000).random_(2).cuda()#
data3_g = G(z=z, label=label).squeeze().cpu().detach()
df1 = pd.DataFrame()
df2 = pd.DataFrame()
df1['score_{0}'.format(0)] = data1_g.numpy()
df1['score_{0}'.format(1)] = data2_g.numpy()
df1['score_{0}'.format(2)] = data3_g.numpy()
g_data = torch.cat([data1_g, data2_g, data3_g], dim=0).numpy()
np.save(save_path + '/%s_data'%name, g_data)
df2['score_{0}'.format(2)] = g_data
if not no_graph:
fig, ax = plt.subplots(1, 1)
for s in df1.columns:
df1[s].plot(kind='kde')
for s in df2.columns:
df2[s].plot(style='--', kind='kde')
plt.xlim((-4, 9 + distance * 2))
ax.legend(["Class_0", "Class_1", "Class_2", "Marginal"])
# plt.title(name)
fig.savefig(save_path + '/%s.eps'%name)
mean0_0, var0_0 = polynomial_mmd(np.expand_dims(data1_g.numpy(), axis=1), np.expand_dims(data1.cpu().numpy(),axis=1))
mean0_1, var0_1 = polynomial_mmd(np.expand_dims(data2_g.numpy(), axis=1),
np.expand_dims(data2.cpu().numpy(), axis=1))
mean0_2, var0_2 = polynomial_mmd(np.expand_dims(data3_g.numpy(), axis=1),
np.expand_dims(data3.cpu().numpy(), axis=1))
mean0, var0 = polynomial_mmd(np.expand_dims(g_data, axis=1), np.expand_dims(r_data, axis=1))
with open(save_path + f'/results.txt', 'a+') as f:
f.write(f'{name}:\n')
f.write(f'{mean0_0}, {var0_0}\n')
f.write(f'{mean0_1}, {var0_1}\n')
f.write(f'{mean0_2}, {var0_2}\n')
f.write(f'{mean0}, {var0}\n')
return (mean0_0, var0_0), (mean0_1, var0_1), (mean0_2, var0_2), (mean0, var0)
def multi_results(distance):
# time.sleep(distance*3)
nz = 2
G = G_guassian(nz=nz, num_classes=3).cuda()
D = D_guassian(num_classes=3).cuda()
optg = optim.Adam(G.parameters(), lr=0.002,
betas=(0.5, 0.999))
optd = optim.Adam(D.parameters(), lr=0.002,
betas=(0.5, 0.999))
distance = (distance+2)/2
if os.path.exists(os.path.join('MOG','1D', str(distance) + '_1D')):
pass
else:
os.makedirs(os.path.join('MOG','1D', str(distance) + '_1D'))
save_path = os.path.join('MOG','1D', str(distance) + '_1D')
data1 = torch.randn(128000).cuda()
data2 = torch.randn(128000).cuda() * 2 + distance
data3 = torch.randn(128000).cuda() * 3 + distance * 2
df1 = pd.DataFrame()
df2 = pd.DataFrame()
df1['score_{0}'.format(0)] = data1.cpu().numpy()
df1['score_{0}'.format(1)] = data2.cpu().numpy()
df1['score_{0}'.format(2)] = data3.cpu().numpy()
r_data = torch.cat([data1, data2, data3], dim=0).cpu().numpy()
df2['score_{0}'.format(2)] = r_data
np.save(save_path+'/o_data',r_data)
fig, ax = plt.subplots(1, 1)
for s in df1.columns:
df1[s].plot(kind='kde')
for s in df2.columns:
df2[s].plot(style='--',kind='kde')
plt.xlim((-4, 9 + distance * 2))
ax.legend(["Class_0", "Class_1","Class_2","Marginal"])
plt.title('Original')
fig.savefig(save_path + '/original.eps')
train(data1,data2,data3,G,D,optd,optg,AC=True,MI=True)
z = torch.randn(10000, nz).cuda()
label = torch.zeros(10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data1_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data2_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(10000).long().cuda() + 1 # torch.LongTensor(10000).random_(2).cuda()#
data3_g = G(z=z, label=label).squeeze().cpu().detach()
df1 = pd.DataFrame()
df2 = pd.DataFrame()
df1['score_{0}'.format(0)] = data1_g.numpy()
df1['score_{0}'.format(1)] = data2_g.numpy()
df1['score_{0}'.format(2)] = data3_g.numpy()
g_data = torch.cat([data1_g, data2_g, data3_g], dim=0).numpy()
np.save(save_path + '/twin_ac_data', g_data)
df2['score_{0}'.format(2)] = g_data
fig, ax = plt.subplots(1, 1)
for s in df1.columns:
df1[s].plot(kind='kde')
for s in df2.columns:
df2[s].plot(style='--',kind='kde')
plt.xlim((-4, 9 + distance * 2))
ax.legend(["Class_0", "Class_1", "Class_2", "Marginal"])
plt.title('TAC')
fig.savefig(save_path + '/twin_ac.eps')
mean0_0,var0_0 = polynomial_mmd(np.expand_dims(data1_g.numpy(), axis=1), np.expand_dims(data1.cpu().numpy(),axis=1))
mean0_1, var0_1 = polynomial_mmd(np.expand_dims(data2_g.numpy(), axis=1),
np.expand_dims(data2.cpu().numpy(), axis=1))
mean0_2, var0_2 = polynomial_mmd(np.expand_dims(data3_g.numpy(), axis=1),
np.expand_dims(data3.cpu().numpy(), axis=1))
mean0, var0 = polynomial_mmd(np.expand_dims(g_data, axis=1), np.expand_dims(r_data, axis=1))
# ac
G = G_guassian(nz=nz, num_classes=3).cuda()
D = D_guassian(num_classes=3).cuda()
optg = optim.Adam(G.parameters(), lr=0.002,
betas=(0.5, 0.999))
optd = optim.Adam(D.parameters(), lr=0.002,
betas=(0.5, 0.999))
for _ in range(20):
for i in range(1000):
#####D step
for _ in range(1):
data = torch.cat(
[data1[128 * i:128 * i + 128], data2[128 * i:128 * i + 128], data3[128 * i:128 * i + 128]],
dim=0).unsqueeze(dim=1)
label = torch.cat([torch.ones(128).cuda().long() * 0, torch.ones(128).cuda().long() * 1,
torch.ones(128).cuda().long() * 2], dim=0)
###D
d_real, c, _ = D(data)
z = torch.randn(256, nz).cuda()
fake_label = torch.LongTensor(256).random_(3).cuda()
fake_data = G(z, label=fake_label)
d_fake, _, mi = D(fake_data)
D_loss = F.binary_cross_entropy(d_real, torch.ones(384).cuda()) \
+ F.binary_cross_entropy(d_fake, torch.zeros(256).cuda()) \
+ F.cross_entropy(c, label)
optd.zero_grad()
D_loss.backward()
optd.step()
#####G step
if i % 10 == 0:
z = torch.randn(256, nz).cuda()
fake_label = torch.LongTensor(256).random_(3).cuda()
fake_data = G(z, label=fake_label)
d_fake, c, mi = D(fake_data)
G_loss = F.binary_cross_entropy(d_fake, torch.ones(256).cuda()) + F.cross_entropy(c,fake_label)
optg.zero_grad()
G_loss.backward()
optg.step()
z = torch.randn(10000, nz).cuda()
label = torch.zeros(10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data1_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data2_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(10000).long().cuda() + 1 # torch.LongTensor(10000).random_(2).cuda()#
data3_g = G(z=z, label=label).squeeze().cpu().detach()
df1 = pd.DataFrame()
df2 = pd.DataFrame()
df1['score_{0}'.format(0)] = data1_g.numpy()
df1['score_{0}'.format(1)] = data2_g.numpy()
df1['score_{0}'.format(2)] = data3_g.numpy()
g_data = torch.cat([data1_g, data2_g, data3_g], dim=0).numpy()
np.save(save_path + '/ac_data', g_data)
df2['score_{0}'.format(2)] = g_data
fig, ax = plt.subplots(1, 1)
for s in df1.columns:
df1[s].plot(kind='kde')
for s in df2.columns:
df2[s].plot(style='--',kind='kde')
plt.xlim((-4, 9 + distance * 2))
ax.legend(["Class_0", "Class_1", "Class_2", "Marginal"])
plt.title('AC')
fig.savefig(save_path + '/ac.eps')
mean1_0, var1_0 = polynomial_mmd(np.expand_dims(data1_g.numpy(), axis=1),
np.expand_dims(data1.cpu().numpy(), axis=1))
mean1_1, var1_1 = polynomial_mmd(np.expand_dims(data2_g.numpy(), axis=1),
np.expand_dims(data2.cpu().numpy(), axis=1))
mean1_2, var1_2 = polynomial_mmd(np.expand_dims(data3_g.numpy(), axis=1),
np.expand_dims(data3.cpu().numpy(), axis=1))
mean1, var1 = polynomial_mmd(np.expand_dims(g_data, axis=1), np.expand_dims(r_data, axis=1))
####projection
G = G_guassian(nz=nz, num_classes=3).cuda()
D = D_guassian(num_classes=3,AC=False).cuda()
optg = optim.Adam(G.parameters(), lr=0.002,
betas=(0.5, 0.999))
optd = optim.Adam(D.parameters(), lr=0.002,
betas=(0.5, 0.999))
for _ in range(20):
for i in range(1000):
#####D step
for _ in range(1):
data = torch.cat(
[data1[128 * i:128 * i + 128], data2[128 * i:128 * i + 128], data3[128 * i:128 * i + 128]],
dim=0).unsqueeze(dim=1)
label = torch.cat([torch.ones(128).cuda().long() * 0, torch.ones(128).cuda().long() * 1,
torch.ones(128).cuda().long() * 2], dim=0)
###D
d_real, c, _ = D(data, label)
z = torch.randn(256, nz).cuda()
fake_label = torch.LongTensor(256).random_(3).cuda()
fake_data = G(z, label=fake_label)
d_fake, _, mi = D(fake_data, fake_label)
D_loss = F.binary_cross_entropy(d_real, torch.ones(384).cuda()) \
+ F.binary_cross_entropy(d_fake, torch.zeros(256).cuda())
optd.zero_grad()
D_loss.backward()
optd.step()
#####G step
if i % 10 == 0:
z = torch.randn(256, nz).cuda()
fake_label = torch.LongTensor(256).random_(3).cuda()
fake_data = G(z, label=fake_label)
d_fake, c, mi = D(fake_data, fake_label)
G_loss = F.binary_cross_entropy(d_fake, torch.ones(256).cuda())
optg.zero_grad()
G_loss.backward()
optg.step()
z = torch.randn(10000, nz).cuda()
label = torch.zeros(10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data1_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data2_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(10000).long().cuda() + 1 # torch.LongTensor(10000).random_(2).cuda()#
data3_g = G(z=z, label=label).squeeze().cpu().detach()
df1 = pd.DataFrame()
df2 = pd.DataFrame()
df1['score_{0}'.format(0)] = data1_g.numpy()
df1['score_{0}'.format(1)] = data2_g.numpy()
df1['score_{0}'.format(2)] = data3_g.numpy()
g_data = torch.cat([data1_g, data2_g, data3_g], dim=0).numpy()
np.save(save_path + '/projection_data', g_data)
df2['score_{0}'.format(2)] = g_data
fig, ax = plt.subplots(1, 1)
for s in df1.columns:
df1[s].plot(kind='kde')
for s in df2.columns:
df2[s].plot(style='--', kind='kde')
plt.xlim((-4, 9 + distance * 2))
ax.legend(["Class_0", "Class_1", "Class_2", "Marginal"])
plt.title('Projection')
fig.savefig(save_path + '/projection.eps')
mean2_0, var2_0 = polynomial_mmd(np.expand_dims(data1_g.numpy(), axis=1),
np.expand_dims(data1.cpu().numpy(), axis=1))
mean2_1, var2_1 = polynomial_mmd(np.expand_dims(data2_g.numpy(), axis=1),
np.expand_dims(data2.cpu().numpy(), axis=1))
mean2_2, var2_2 = polynomial_mmd(np.expand_dims(data3_g.numpy(), axis=1),
np.expand_dims(data3.cpu().numpy(), axis=1))
mean2, var2 = polynomial_mmd(np.expand_dims(g_data, axis=1), np.expand_dims(r_data, axis=1))
result = [str(mean0_0) + ',' + str(var0_0),
str(mean0_1) + ',' + str(var0_1),
str(mean0_2) + ',' + str(var0_2),
str(mean0) + ',' + str(var0),
str(mean1_0) + ',' + str(var1_0),
str(mean1_1) + ',' + str(var1_1),
str(mean1_2) + ',' + str(var1_2),
str(mean1) + ',' + str(var1),
str(mean2_0) + ',' + str(var2_0),
str(mean2_1) + ',' + str(var2_1),
str(mean2_2) + ',' + str(var2_2),
str(mean2) + ',' + str(var2)]
file = open(save_path + '/result.text', 'w')
for content in result:
file.write(content + '\n')
def multi_results(distance):
time.sleep(distance * 3)
nz = 2
G = G_guassian(nz=nz, num_classes=3).cuda()
D = D_guassian(num_classes=3).cuda()
optg = optim.Adam(G.parameters(), lr=0.002, betas=(0.5, 0.999))
optd = optim.Adam(D.parameters(), lr=0.002, betas=(0.5, 0.999))
distance = (distance + 2) / 2
if os.path.exists(os.path.join('MOG', '2D', str(distance) + '_2D')):
pass
else:
os.makedirs(os.path.join('MOG', '2D', str(distance) + '_2D'))
save_path = os.path.join('MOG', '2D', str(distance) + '_2D')
data1 = torch.randn(128000, 2).cuda()
data2 = torch.randn(128000, 2).cuda() * 2
data2[:, 0] += distance
data3 = torch.randn(128000, 2).cuda() * 3
data3[:, 0] += distance * 2
fig, ax = plt.subplots(1, 1)
sns.kdeplot(data1.cpu().numpy()[:, 0], data1.cpu().numpy()[:, 1])
sns.kdeplot(data2.cpu().numpy()[:, 0], data2.cpu().numpy()[:, 1])
sns.kdeplot(data3.cpu().numpy()[:, 0], data3.cpu().numpy()[:, 1])
fig.legend(["Class_0", "Class_1", "Class_2"])
plt.xlim((-4, 9 + distance * 2))
plt.ylim((-8, 8))
plt.title('Original')
fig.savefig(save_path + '/o_conditional.eps')
r_data = torch.cat([data1, data2, data3], dim=0).cpu().numpy()
np.save(save_path + '/o_data', r_data)
for _ in range(40):
for i in range(1000):
#####D step
for _ in range(1):
data = torch.cat([
data1[128 * i:128 * i + 128], data2[128 * i:128 * i + 128],
data3[128 * i:128 * i + 128]
],
dim=0).unsqueeze(dim=1)
label = torch.cat([
torch.ones(128).cuda().long() * 0,
torch.ones(128).cuda().long() * 1,
torch.ones(128).cuda().long() * 2
],
dim=0)
###D
d_real, c, _ = D(data)
z = torch.randn(256, nz).cuda()
fake_label = torch.LongTensor(256).random_(3).cuda()
fake_data = G(z, label=fake_label)
d_fake, _, mi = D(fake_data)
D_loss = F.binary_cross_entropy(d_real, torch.ones(384).cuda()) \
+ F.binary_cross_entropy(d_fake, torch.zeros(256).cuda()) \
+ F.cross_entropy(c, label) \
+ F.cross_entropy(mi, fake_label)
optd.zero_grad()
D_loss.backward()
optd.step()
#####G step
if i % 10 == 0:
z = torch.randn(256, nz).cuda()
fake_label = torch.LongTensor(256).random_(3).cuda()
fake_data = G(z, label=fake_label)
d_fake, c, mi = D(fake_data)
G_loss = F.binary_cross_entropy(
d_fake,
torch.ones(256).cuda()) + F.cross_entropy(
c, fake_label) - F.cross_entropy(mi, fake_label)
optg.zero_grad()
G_loss.backward()
optg.step()
z = torch.randn(10000, nz).cuda()
label = torch.zeros(
10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data1_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(
10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data2_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(
10000).long().cuda() + 1 # torch.LongTensor(10000).random_(2).cuda()#
data3_g = G(z=z, label=label).squeeze().cpu().detach()
fig, ax = plt.subplots(1, 1)
sns.kdeplot(data1_g.numpy()[:, 0], data1_g.numpy()[:, 1])
sns.kdeplot(data2_g.numpy()[:, 0], data2_g.numpy()[:, 1])
sns.kdeplot(data3_g.numpy()[:, 0], data3_g.numpy()[:, 1])
fig.legend(["Class_0", "Class_1", "Class_2"])
plt.xlim((-4, 9 + distance * 2))
plt.ylim((-8, 8))
plt.title('TAC')
fig.savefig(save_path + '/twin_ac_conditional.eps')
g_data = torch.cat([data1_g, data2_g, data3_g], dim=0).cpu().numpy()
np.save(save_path + '/twin_ac_data', g_data)
mean0_0, var0_0 = polynomial_mmd(data1_g.numpy(), data1.cpu().numpy())
mean0_1, var0_1 = polynomial_mmd(data2_g.numpy(), data2.cpu().numpy())
mean0_2, var0_2 = polynomial_mmd(data3_g.numpy(), data3.cpu().numpy())
mean0, var0 = polynomial_mmd(g_data, r_data)
# plot_density(generate_data2.cpu().detach().numpy())
# ac
G = G_guassian(nz=nz, num_classes=3).cuda()
D = D_guassian(num_classes=3).cuda()
optg = optim.Adam(G.parameters(), lr=0.002, betas=(0.5, 0.999))
optd = optim.Adam(D.parameters(), lr=0.002, betas=(0.5, 0.999))
for _ in range(40):
for i in range(1000):
#####D step
for _ in range(1):
data = torch.cat([
data1[128 * i:128 * i + 128], data2[128 * i:128 * i + 128],
data3[128 * i:128 * i + 128]
],
dim=0).unsqueeze(dim=1)
label = torch.cat([
torch.ones(128).cuda().long() * 0,
torch.ones(128).cuda().long() * 1,
torch.ones(128).cuda().long() * 2
],
dim=0)
###D
d_real, c, _ = D(data)
z = torch.randn(256, nz).cuda()
fake_label = torch.LongTensor(256).random_(3).cuda()
fake_data = G(z, label=fake_label)
d_fake, _, mi = D(fake_data)
D_loss = F.binary_cross_entropy(d_real, torch.ones(384).cuda()) \
+ F.binary_cross_entropy(d_fake, torch.zeros(256).cuda()) \
+ F.cross_entropy(c, label)
optd.zero_grad()
D_loss.backward()
optd.step()
#####G step
if i % 10 == 0:
z = torch.randn(256, nz).cuda()
fake_label = torch.LongTensor(256).random_(3).cuda()
fake_data = G(z, label=fake_label)
d_fake, c, mi = D(fake_data)
G_loss = F.binary_cross_entropy(
d_fake,
torch.ones(256).cuda()) + F.cross_entropy(c, fake_label)
optg.zero_grad()
G_loss.backward()
optg.step()
z = torch.randn(10000, nz).cuda()
label = torch.zeros(
10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data1_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(
10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data2_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(
10000).long().cuda() + 1 # torch.LongTensor(10000).random_(2).cuda()#
data3_g = G(z=z, label=label).squeeze().cpu().detach()
fig, ax = plt.subplots(1, 1)
sns.kdeplot(data1_g.numpy()[:, 0], data1_g.numpy()[:, 1])
sns.kdeplot(data2_g.numpy()[:, 0], data2_g.numpy()[:, 1])
sns.kdeplot(data3_g.numpy()[:, 0], data3_g.numpy()[:, 1])
fig.legend(["Class_0", "Class_1", "Class_2"])
plt.xlim((-4, 9 + distance * 2))
plt.ylim((-8, 8))
plt.title('AC')
fig.savefig(save_path + '/ac_conditional.eps')
g_data = torch.cat([data1_g, data2_g, data3_g], dim=0).cpu().numpy()
np.save(save_path + '/ac_data', g_data)
mean1_0, var1_0 = polynomial_mmd(data1_g.numpy(), data1.cpu().numpy())
mean1_1, var1_1 = polynomial_mmd(data2_g.numpy(), data2.cpu().numpy())
mean1_2, var1_2 = polynomial_mmd(data3_g.numpy(), data3.cpu().numpy())
mean1, var1 = polynomial_mmd(g_data, r_data)
####projection
G = G_guassian(nz=nz, num_classes=3).cuda()
D = D_guassian(num_classes=3, AC=False).cuda()
optg = optim.Adam(G.parameters(), lr=0.002, betas=(0.5, 0.999))
optd = optim.Adam(D.parameters(), lr=0.002, betas=(0.5, 0.999))
for _ in range(40):
for i in range(1000):
#####D step
for _ in range(1):
data = torch.cat([
data1[128 * i:128 * i + 128], data2[128 * i:128 * i + 128],
data3[128 * i:128 * i + 128]
],
dim=0).unsqueeze(dim=1)
label = torch.cat([
torch.ones(128).cuda().long() * 0,
torch.ones(128).cuda().long() * 1,
torch.ones(128).cuda().long() * 2
],
dim=0)
###D
d_real, c, _ = D(data, label)
z = torch.randn(256, nz).cuda()
fake_label = torch.LongTensor(256).random_(3).cuda()
fake_data = G(z, label=fake_label)
d_fake, _, mi = D(fake_data, fake_label)
D_loss = F.binary_cross_entropy(d_real, torch.ones(384).cuda()) \
+ F.binary_cross_entropy(d_fake, torch.zeros(256).cuda())
optd.zero_grad()
D_loss.backward()
optd.step()
#####G step
if i % 10 == 0:
z = torch.randn(256, nz).cuda()
fake_label = torch.LongTensor(256).random_(3).cuda()
fake_data = G(z, label=fake_label)
d_fake, c, mi = D(fake_data, fake_label)
G_loss = F.binary_cross_entropy(d_fake, torch.ones(256).cuda())
optg.zero_grad()
G_loss.backward()
optg.step()
z = torch.randn(10000, nz).cuda()
label = torch.zeros(
10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data1_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(
10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data2_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(
10000).long().cuda() + 1 # torch.LongTensor(10000).random_(2).cuda()#
data3_g = G(z=z, label=label).squeeze().cpu().detach()
fig, ax = plt.subplots(1, 1)
sns.kdeplot(data1_g.numpy()[:, 0], data1_g.numpy()[:, 1])
sns.kdeplot(data2_g.numpy()[:, 0], data2_g.numpy()[:, 1])
sns.kdeplot(data3_g.numpy()[:, 0], data3_g.numpy()[:, 1])
fig.legend(["Class_0", "Class_1", "Class_2"])
plt.xlim((-4, 9 + distance * 2))
plt.ylim((-8, 8))
fig.savefig(save_path + '/projection_conditional.eps')
g_data = torch.cat([data1_g, data2_g, data3_g], dim=0).cpu().numpy()
np.save(save_path + '/projection_data', g_data)
mean2_0, var2_0 = polynomial_mmd(data1_g.numpy(), data1.cpu().numpy())
mean2_1, var2_1 = polynomial_mmd(data2_g.numpy(), data2.cpu().numpy())
mean2_2, var2_2 = polynomial_mmd(data3_g.numpy(), data3.cpu().numpy())
mean2, var2 = polynomial_mmd(g_data, r_data)
result = [
str(mean0_0) + ',' + str(var0_0),
str(mean0_1) + ',' + str(var0_1),
str(mean0_2) + ',' + str(var0_2),
str(mean0) + ',' + str(var0),
str(mean1_0) + ',' + str(var1_0),
str(mean1_1) + ',' + str(var1_1),
str(mean1_2) + ',' + str(var1_2),
str(mean1) + ',' + str(var1),
str(mean2_0) + ',' + str(var2_0),
str(mean2_1) + ',' + str(var2_1),
str(mean2_2) + ',' + str(var2_2),
str(mean2) + ',' + str(var2)
]
file = open(save_path + '/result.text', 'w')
for content in result:
file.write(content + '\n')
def multi_results(distance, gan_loss='bce', dis_mlp=False, run_id=0, suffix='', no_graph=False):
if not suffix and dis_mlp:
suffix = '_mlp'
# time.sleep(distance*3)
nz = 2
G = G_guassian(nz=nz, num_classes=3).cuda()
D = D_guassian(num_classes=3, AC=True, dis_mlp=dis_mlp).cuda()
optg = optim.Adam(G.parameters(), lr=0.002, betas=(0.5, 0.999))
optd = optim.Adam(D.parameters(), lr=0.002, betas=(0.5, 0.999))
distance = (distance + 2) / 2
if os.path.exists(os.path.join('MOG', '1D', f'{distance}_{gan_loss}{suffix}_{run_id}')):
pass
else:
os.makedirs(os.path.join('MOG', '1D', f'{distance}_{gan_loss}{suffix}_{run_id}'))
save_path = os.path.join('MOG', '1D', f'{distance}_{gan_loss}{suffix}_{run_id}')
data1 = torch.randn(128000).cuda()
data2 = torch.randn(128000).cuda() * 2 + distance
data3 = torch.randn(128000).cuda() * 3 + distance * 2
df1 = pd.DataFrame()
df2 = pd.DataFrame()
df1['score_{0}'.format(0)] = data1.cpu().numpy()
df1['score_{0}'.format(1)] = data2.cpu().numpy()
df1['score_{0}'.format(2)] = data3.cpu().numpy()
r_data = torch.cat([data1, data2, data3], dim=0).cpu().numpy()
df2['score_{0}'.format(2)] = r_data
np.save(save_path+'/o_data', r_data)
if not no_graph:
fig, ax = plt.subplots(1, 1)
for s in df1.columns:
df1[s].plot(kind='kde')
for s in df2.columns:
df2[s].plot(style='--', kind='kde')
plt.xlim((-4, 9 + distance * 2))
ax.legend(["Class_0", "Class_1", "Class_2", "Marginal"])
plt.title('Original')
fig.savefig(save_path + '/original.eps')
train(data1, data2, data3, nz, G, D, optd, optg, AC=True, MI=True, gan_loss=gan_loss)
print('TAC training done.')
z = torch.randn(10000, nz).cuda()
label = torch.zeros(10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data1_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data2_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(10000).long().cuda() + 1 # torch.LongTensor(10000).random_(2).cuda()#
data3_g = G(z=z, label=label).squeeze().cpu().detach()
df1 = pd.DataFrame()
df2 = pd.DataFrame()
df1['score_{0}'.format(0)] = data1_g.numpy()
df1['score_{0}'.format(1)] = data2_g.numpy()
df1['score_{0}'.format(2)] = data3_g.numpy()
g_data = torch.cat([data1_g, data2_g, data3_g], dim=0).numpy()
np.save(save_path + '/twin_ac_data', g_data)
df2['score_{0}'.format(2)] = g_data
if not no_graph:
fig, ax = plt.subplots(1, 1)
for s in df1.columns:
df1[s].plot(kind='kde')
for s in df2.columns:
df2[s].plot(style='--', kind='kde')
plt.xlim((-4, 9 + distance * 2))
ax.legend(["Class_0", "Class_1", "Class_2", "Marginal"])
plt.title('TAC')
fig.savefig(save_path + '/twin_ac.eps')
mean0_0, var0_0 = polynomial_mmd(np.expand_dims(data1_g.numpy(), axis=1), np.expand_dims(data1.cpu().numpy(),axis=1))
mean0_1, var0_1 = polynomial_mmd(np.expand_dims(data2_g.numpy(), axis=1),
np.expand_dims(data2.cpu().numpy(), axis=1))
mean0_2, var0_2 = polynomial_mmd(np.expand_dims(data3_g.numpy(), axis=1),
np.expand_dims(data3.cpu().numpy(), axis=1))
mean0, var0 = polynomial_mmd(np.expand_dims(g_data, axis=1), np.expand_dims(r_data, axis=1))
# ac
G = G_guassian(nz=nz, num_classes=3).cuda()
D = D_guassian(num_classes=3, AC=True, dis_mlp=dis_mlp).cuda()
optg = optim.Adam(G.parameters(), lr=0.002, betas=(0.5, 0.999))
optd = optim.Adam(D.parameters(), lr=0.002, betas=(0.5, 0.999))
train(data1, data2, data3, nz, G, D, optd, optg, AC=True, MI=False, gan_loss=gan_loss)
print('AC training done.')
z = torch.randn(10000, nz).cuda()
label = torch.zeros(10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data1_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data2_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(10000).long().cuda() + 1 # torch.LongTensor(10000).random_(2).cuda()#
data3_g = G(z=z, label=label).squeeze().cpu().detach()
df1 = pd.DataFrame()
df2 = pd.DataFrame()
df1['score_{0}'.format(0)] = data1_g.numpy()
df1['score_{0}'.format(1)] = data2_g.numpy()
df1['score_{0}'.format(2)] = data3_g.numpy()
g_data = torch.cat([data1_g, data2_g, data3_g], dim=0).numpy()
np.save(save_path + '/ac_data', g_data)
df2['score_{0}'.format(2)] = g_data
if not no_graph:
fig, ax = plt.subplots(1, 1)
for s in df1.columns:
df1[s].plot(kind='kde')
for s in df2.columns:
df2[s].plot(style='--', kind='kde')
plt.xlim((-4, 9 + distance * 2))
ax.legend(["Class_0", "Class_1", "Class_2", "Marginal"])
plt.title('AC')
fig.savefig(save_path + '/ac.eps')
mean1_0, var1_0 = polynomial_mmd(np.expand_dims(data1_g.numpy(), axis=1),
np.expand_dims(data1.cpu().numpy(), axis=1))
mean1_1, var1_1 = polynomial_mmd(np.expand_dims(data2_g.numpy(), axis=1),
np.expand_dims(data2.cpu().numpy(), axis=1))
mean1_2, var1_2 = polynomial_mmd(np.expand_dims(data3_g.numpy(), axis=1),
np.expand_dims(data3.cpu().numpy(), axis=1))
mean1, var1 = polynomial_mmd(np.expand_dims(g_data, axis=1), np.expand_dims(r_data, axis=1))
####projection
G = G_guassian(nz=nz, num_classes=3).cuda()
D = D_guassian(num_classes=3, AC=False, dis_mlp=dis_mlp).cuda()
optg = optim.Adam(G.parameters(), lr=0.002, betas=(0.5, 0.999))
optd = optim.Adam(D.parameters(), lr=0.002, betas=(0.5, 0.999))
train(data1, data2, data3, nz, G, D, optd, optg, AC=False, MI=False, gan_loss=gan_loss)
print('Projection training done.')
z = torch.randn(10000, nz).cuda()
label = torch.zeros(10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data1_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(10000).long().cuda() # torch.LongTensor(10000).random_(2).cuda()#
data2_g = G(z=z, label=label).squeeze().cpu().detach()
z = torch.randn(10000, nz).cuda()
label = torch.ones(10000).long().cuda() + 1 # torch.LongTensor(10000).random_(2).cuda()#
data3_g = G(z=z, label=label).squeeze().cpu().detach()
df1 = pd.DataFrame()
df2 = pd.DataFrame()
df1['score_{0}'.format(0)] = data1_g.numpy()
df1['score_{0}'.format(1)] = data2_g.numpy()
df1['score_{0}'.format(2)] = data3_g.numpy()
g_data = torch.cat([data1_g, data2_g, data3_g], dim=0).numpy()
np.save(save_path + '/projection_data', g_data)
df2['score_{0}'.format(2)] = g_data
if not no_graph:
fig, ax = plt.subplots(1, 1)
for s in df1.columns:
df1[s].plot(kind='kde')
for s in df2.columns:
df2[s].plot(style='--', kind='kde')
plt.xlim((-4, 9 + distance * 2))
ax.legend(["Class_0", "Class_1", "Class_2", "Marginal"])
plt.title('Projection')
fig.savefig(save_path + '/projection.eps')
mean2_0, var2_0 = polynomial_mmd(np.expand_dims(data1_g.numpy(), axis=1),
np.expand_dims(data1.cpu().numpy(), axis=1))
mean2_1, var2_1 = polynomial_mmd(np.expand_dims(data2_g.numpy(), axis=1),
np.expand_dims(data2.cpu().numpy(), axis=1))
mean2_2, var2_2 = polynomial_mmd(np.expand_dims(data3_g.numpy(), axis=1),
np.expand_dims(data3.cpu().numpy(), axis=1))
mean2, var2 = polynomial_mmd(np.expand_dims(g_data, axis=1), np.expand_dims(r_data, axis=1))
result = [str(mean0_0) + ',' + str(var0_0),
str(mean0_1) + ',' + str(var0_1),
str(mean0_2) + ',' + str(var0_2),
str(mean0) + ',' + str(var0),
str(mean1_0) + ',' + str(var1_0),
str(mean1_1) + ',' + str(var1_1),
str(mean1_2) + ',' + str(var1_2),
str(mean1) + ',' + str(var1),
str(mean2_0) + ',' + str(var2_0),
str(mean2_1) + ',' + str(var2_1),
str(mean2_2) + ',' + str(var2_2),
str(mean2) + ',' + str(var2)]
file = open(save_path + f'/result_{gan_loss}.txt', 'w')
for content in result:
file.write(content + '\n')
df2['score_{0}'.format(2)] = g_data
#####Plot generated conditional distribution
fig, ax = plt.subplots(1, 1)
for s in df1.columns:
df1[s].plot(kind='kde')
for s in df2.columns:
df2[s].plot(style='--', kind='kde')
plt.xlim((-4, 9 + distance * 2))
ax.legend(["Class_0", "Class_1", "Class_2", "Marginal"])
plt.title(name)
fig.savefig(save_path + '/' + name + '.eps')
mean0_0, var0_0 = polynomial_mmd(np.expand_dims(data1_g.numpy(), axis=1),
np.expand_dims(data1.cpu().numpy(), axis=1))
mean0_1, var0_1 = polynomial_mmd(np.expand_dims(data2_g.numpy(), axis=1),
np.expand_dims(data2.cpu().numpy(), axis=1))
mean0_2, var0_2 = polynomial_mmd(np.expand_dims(data3_g.numpy(), axis=1),
np.expand_dims(data3.cpu().numpy(), axis=1))
mean0, var0 = polynomial_mmd(np.expand_dims(g_data, axis=1),
np.expand_dims(r_data, axis=1))
print('the distance of class 0 is : ', mean0_0)
print('the distance of class 1 is : ', mean0_1)
print('the distance of class 2 is : ', mean0_2)
print('the distance of marginal is : ', mean0)
with open(args.name, 'w') as f:
f.write(
f'AC:\nclass 0: {mean0_0}\nclass 1: {mean0_1}\nclass 2: {mean0_2}\nmarginal: {mean0}\n'
)