def flowBuilder(n,numFlow,innerBuilder=None,typeLayer=3,relax=False,shift=False):
nn = n*2
op = source.Gaussian([nn]).to(torch.float64)
if innerBuilder is None:
raise Exception("innerBuilder is None")
if relax:
f3 = flow.DiagScaling(nn,initValue=0.1*np.random.randn(nn),fix=[0]*n+[0]*n,shift=shift)
else:
f3 = flow.DiagScaling(nn,initValue=0.1*np.random.randn(nn),fix=[0]*n+[1]*n,shift=shift)
layers=[f3]
if typeLayer == 0:
layers.append(flow.Symplectic(nn))
else:
for d in range(numFlow):
if typeLayer == 3:
layers.append(flow.PointTransformation(innerBuilder(n)))
layers.append(flow.Symplectic(nn))
elif typeLayer ==2:
layers.append(flow.Symplectic(nn))
elif typeLayer ==1:
layers.append(flow.PointTransformation(innerBuilder(n)))
elif typeLayer!=0:
raise Exception("No such type")
return flow.FlowNet(layers,op).double()
def extractPPrior(flowCon):
layers = []
_diag = flowCon.layerList[0]
nn = _diag.shift.shape[0] // 2
_op = source.Gaussian([nn]).to(torch.float64)
assert _diag.shift.sum() == 0
assert _diag.fix.sum() == nn
layers.append(
flow.DiagScaling(nn, initValue=_diag.elements.clone().detach()[nn:]))
return flow.FlowNet(layers, _op).to(torch.float64)
def extractFlow(flowCon):
from copy import deepcopy
layers = []
_op = deepcopy(flowCon.prior)
_rnvp = deepcopy(flowCon.layerList[1].flow)
_diag = flowCon.layerList[0]
nn = _diag.shift.shape[0]//2
layers.append(flow.DiagScaling(nn,initValue=_diag.elements.clone().detach()[:nn]))
layers.append(_rnvp)
return flow.FlowNet(layers,_op).double()
def extractFlow(flowCon):
from copy import deepcopy
layers = []
_rnvp = deepcopy(flowCon.layerList[1].flow)
_diag = flowCon.layerList[0]
nn = _diag.shift.shape[0] // 2
_op = source.Gaussian([nn]).to(torch.float64)
assert _diag.shift.sum() == 0
assert _diag.fix.sum() == nn
layers.append(
flow.DiagScaling(nn, initValue=_diag.elements.clone().detach()[:nn]))
layers.append(_rnvp)
return flow.FlowNet(layers, _op).double()
def test_saveload():
p = source.Gaussian([4])
f1 = flow.Scaling(4, [2, 3])
maskList = []
for n in range(4):
if n % 2 == 0:
b = torch.zeros(1, 4)
i = torch.randperm(b.numel()).narrow(0, 0, b.numel() // 2)
b.zero_()[:, i] = 1
b = b.reshape(1, 4)
else:
b = 1 - b
maskList.append(b)
maskList = torch.cat(maskList, 0).to(torch.float32)
f2 = flow.NICE(maskList, [
utils.SimpleMLPreshape([4, 32, 32, 4],
[nn.ELU(), nn.ELU(), None]) for _ in range(4)
])
f = flow.FlowNet([f1, f2], p)
f1 = flow.Scaling(4)
maskList = []
for n in range(4):
if n % 2 == 0:
b = torch.zeros(1, 4)
i = torch.randperm(b.numel()).narrow(0, 0, b.numel() // 2)
b.zero_()[:, i] = 1
b = b.reshape(1, 4)
else:
b = 1 - b
maskList.append(b)
maskList = torch.cat(maskList, 0).to(torch.float32)
f2 = flow.NICE(maskList, [
utils.SimpleMLPreshape([4, 32, 32, 4],
[nn.ELU(), nn.ELU(), None]) for _ in range(4)
])
blankf = flow.FlowNet([f1, f2], p)
saveload(f, blankf)
def buildSource(f):
return flow.FlowNet([f.layerList[0]], f.prior).double()