Skip to content

fix syntax error #5

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
rikrd wants to merge 3 commits into
base: master
Choose a base branch
from
Open

fix syntax error #5

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
9a249a9
fix syntax error
Jun 20, 2019
13764ab
Add gitignore
Jun 21, 2019
c2328aa
add a sampler toy example
Jun 24, 2019
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion torchkit/flows.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -41,7 +41,7 @@ def sample(self, n=1, context=None, **kwargs):
if self.gpu:
spl = spl.cuda()
lgd = lgd.cuda()
context = context.gpu()
context = context.cuda()

return self.forward((spl, lgd, context))

Expand Down
14 changes: 8 additions & 6 deletions torchkit/model_iaf_toy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,7 +31,7 @@ def __init__(self, target_energy):
# flows.IAF(2, 128, 1, 3),
# flows.FlipFlow(1),
# flows.IAF(2, 128, 1, 3))
self.mdl = flows.IAF_DSF(2, 64, 1, 4, realify=nn_.softplus,
self.mdl = flows.IAF_DSF(2, 64, 1, 4, #realify=nn_.softplus,
num_ds_dim=5, num_ds_layers=2)

self.optim = optim.Adam(self.mdl.parameters(), lr=0.0005,
Expand All @@ -41,7 +41,7 @@ def __init__(self, target_energy):

def train(self):

total = 10000
total = 1000

for it in range(total):

Expand All @@ -56,7 +56,7 @@ def train(self):

if ((it + 1) % 1000) == 0:
print 'Iteration: [%4d/%4d] loss: %.8f' % \
(it+1, total, loss.data[0])
(it+1, total, loss.item())



Expand All @@ -76,20 +76,22 @@ def train(self):
X = np.concatenate((xx.reshape(n**2,1),yy.reshape(n**2,1)),1)
X = X.astype('float32')
X = Variable(torch.from_numpy(X))
Z = ef(X).data.numpy().reshape(n,n)
Z = ef(X).data.cpu().numpy().reshape(n,n)
ax.pcolormesh(xx,yy,np.exp(Z))
ax.axis('off')
plt.xlim((-10,10))
plt.ylim((-10,10))

ax = fig.add_subplot(1,2,2)
data = mdl.mdl.sample(n**2)[0].data.numpy()
data = mdl.mdl.sample(n**2)[0].data.cpu().numpy()
XX = data[:,0]
YY = data[:,1]
plot = ax.hist2d(XX,YY,200,range=np.array([(-10, 10), (-10, 10)]))
#plot = ax.hist2d(XX,YY,200,range=np.array([(-10, 10), (-10, 10)]))
plt.scatter(XX, YY, s=1, c='r', alpha=0.01)
plt.xlim((-10,10))
plt.ylim((-10,10))
plt.axis('off')

plt.show()


64 changes: 43 additions & 21 deletions torchkit/model_maf_toy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,7 +23,7 @@

class model(object):

def __init__(self, sampler, n=64):
def __init__(self, sampler, n=64, cuda=False):
# self.mdl = nn_.SequentialFlow(
# flows.IAF(2, 64, 1, 2),
# flows.FlipFlow(1),
Expand All @@ -36,10 +36,10 @@ def __init__(self, sampler, n=64):
# flows.IAF(2, 64, 1, 2),
# flows.FlipFlow(1),
# flows.IAF(2, 64, 1, 2))
# self.mdl = flows.IAF_DDSF(2, 64, 1, 3,
# num_ds_dim=2, num_ds_layers=2)
self.mdl = flows.IAF_DSF(2, 64, 1, 3,
num_ds_dim=4)
self.mdl = flows.IAF_DDSF(2, 16, 1, 2,
num_ds_dim=16, num_ds_layers=2)
# self.mdl = flows.IAF_DSF(2, 64, 1, 3,
# num_ds_dim=4)

self.optim = optim.Adam(self.mdl.parameters(), lr=0.005,
betas=(0.9, 0.999))
Expand All @@ -50,12 +50,28 @@ def __init__(self, sampler, n=64):
self.context = Variable(torch.FloatTensor(n, 1).zero_()) + 2.0
self.lgd = Variable(torch.FloatTensor(n).zero_())
self.zeros = Variable(torch.FloatTensor(n, 2).zero_())




self.gpu = cuda
if self.gpu:
self.mdl.cuda()

self.context = self.context.cuda()
self.lgd = self.lgd.cuda()
self.zeros = self.zeros.cuda()


def density(self, spl, lgd=None, context=None, zeros=None):
lgd = self.lgd if lgd is None else lgd
context = self.context if context is None else context
zeros = self.zeros if zeros is None else zeros

if self.gpu:
spl = spl.cuda()
lgd = lgd.cuda()
context = context.cuda()
zeros = zeros.cuda()

z, logdet, _ = self.mdl((spl, lgd, context))
losses = - utils.log_normal(z, zeros, zeros+1.0).sum(1) - logdet
return - losses
Expand All @@ -80,7 +96,7 @@ def train(self, total=2000):

if ((it + 1) % 100) == 0:
print 'Iteration: [%4d/%4d] loss: %.8f' % \
(it+1, total, loss.data[0])
(it+1, total, loss.item())

#self.mdl.made.randomize()

Expand Down Expand Up @@ -119,21 +135,21 @@ def rvs(self, n):
grid = np.concatenate([grid[0].reshape(nmodesperdim**2,1),
grid[1].reshape(nmodesperdim**2,1)],1)

#mix = Mixture(
# np.ones(nmodesperdim**2) / float(nmodesperdim**2),
# [multivariate_normal(mean, 1/float(nmodesperdim*np.log(nmodesperdim))) for mean in grid] )
mix = Mixture(
np.ones(nmodesperdim**2) / float(nmodesperdim**2),
[multivariate_normal(mean, 1/float(nmodesperdim*np.log(nmodesperdim))) for mean in grid] )
#mix = Mixture(
# [0.6, 0.4],
# [multivariate_normal((2.0,2.0), 1.0), multivariate_normal((-3.0,-3.0), 0.5)])
mix = Mixture([0.1, 0.3, 0.4, 0.2], [
multivariate_normal([-5., 0]),
multivariate_normal([5., 0]),
multivariate_normal([0, 5.]),
multivariate_normal([0, -5.])])
# mix = Mixture([0.1, 0.3, 0.4, 0.2], [
# multivariate_normal([-5., 0]),
# multivariate_normal([5., 0]),
# multivariate_normal([0, 5.]),
# multivariate_normal([0, -5.])])

mdl = model(mix.rvs, n=64)
mdl = model(mix.rvs, n=256, cuda=True)
#input('x')
mdl.train()
mdl.train(2000)


# plot figure
Expand All @@ -160,7 +176,7 @@ def rvs(self, n):


ax = fig.add_subplot(1,2,2)
Z = mdl.density(X, lgd, context, zeros).data.numpy().reshape(n,n)
Z = mdl.density(X, lgd, context, zeros).data.cpu().numpy().reshape(n,n)
ax.pcolormesh(xx,yy,np.exp(Z))
ax.axis('off')
plt.xlim((-10,10))
Expand All @@ -185,14 +201,20 @@ def rvs(self, n):
zeros = Variable(torch.FloatTensor(n**2, 2).zero_())


Z = mdl.density(X, lgd, context, zeros).data.numpy().reshape(n,n)
Z = mdl.density(X, lgd, context, zeros).data.cpu().numpy().reshape(n,n)
ax.pcolormesh(xx,yy,np.exp(Z))

mdl.mdl.eval()
nspl, logdet, context = mdl.mdl.sample(10000)
nspl = nspl.detach().cpu().numpy()
ax.scatter(nspl[:,0], nspl[:,1], s=1, alpha=0.2, c='r')

ax.axis('off')
plt.xlim((-10,10))
plt.ylim((-10,10))



plt.show()



Loading