Hi, Thanks for developing this package. I find it very neat and flexible and would like to use it for my research. I noticed that in the paper "Resampling Base Distributions of Normalizing Flows", the bpd of your glow can reach 3.2~3.3, which is comparable to the original paper.

I was wondering if it is possible to share your training scripts and details to train glow on cifar10 to achieve the above bpd. The current example notebook is too sketchy and only results in bpd of 3.8. Thanks very much!

Without it being a buffer, when you load a model and sample from it, the flow thinks it is the first run through so overwrites all the trained ActNorm parameters.

Fixes the deprecation warning documented in https://github.com/VincentStimper/normalizing-flows/issues/12.

Sanity check: Running this before the change:

import normflows as nf
import torch

torch.manual_seed(42)

flow = nf.NormalizingFlow(
    nf.distributions.DiagGaussian(1, trainable=False),
    [
        nf.flows.AutoregressiveRationalQuadraticSpline(1, 1, 1),
        nf.flows.LULinearPermute(1)
    ]
)

with torch.no_grad():
    samples_flow, _ = flow.sample(4)

print(samples_flow)

gives:

tensor([[0.4528],
        [0.6410],
        [0.5200],
        [0.5567]])

After the change, the output stays the same.

Running the following minimal example:

import normflows as nf
import torch

torch.manual_seed(42)

flow = nf.NormalizingFlow(
    nf.distributions.DiagGaussian(1, trainable=False),
    [
        nf.flows.AutoregressiveRationalQuadraticSpline(1, 1, 1),
        nf.flows.LULinearPermute(1)
    ]
)

with torch.no_grad():
    samples_flow, _ = flow.sample(4)

print(samples_flow)

raises a UserWarning about an upcoming deprecation:

/Users/timothy/Desktop/normalizing-flows/normflows/flows/mixing.py:437: UserWarning: torch.triangular_solve is deprecated in favor of torch.linalg.solve_triangular and will be removed in a future PyTorch release.
torch.linalg.solve_triangular has its arguments reversed and does not return a copy of one of the inputs.
X = torch.triangular_solve(B, A).solution
should be replaced with
X = torch.linalg.solve_triangular(A, B). (Triggered internally at  /Users/runner/work/pytorch/pytorch/pytorch/aten/src/ATen/native/BatchLinearAlgebra.cpp:2189.)
  outputs, _ = torch.triangular_solve(

I will submit a PR shortly that fixes the issue 🙂

Added mkdocs structure, and refactored the docstrings (and applied black)

to install the dependencies for documentation building:

pip install -e ".[docs]"

To see the doc:

mkdocs serve

This starts a live server. Modifications of the documentation are rendered live (excluded the modications to docstrings)

To build the docs:

mkdocs build

this will create the site folder (including index.html)

To expend the docs:

Markdown files can be added in the docs folder, then the "nav" section of the mkdocs.yml file has to be updated, e.g.

nav:
  - about: index.md
  - API: references.md
  - my other page: mymarkdown.md

• good to know: markdown can be used in the docstrings.
• apparently, deploying the documentation online on github after built is as simple as calling mkdocs gh-deploy (I did not try it yet)

I still need to do:

• continuous build on github (documentation is rebuilt and deployed at each merge into master)
• a correction pass on the docstrings (I updated them, but did not check them one by one yet)
• the layout is not so nice (especially for the API), needs to be improved
• apparently mkdocs allows to display jupyter notebooks, I need to dig

Add the option to clip the gradient of the target log prob within HMC. For some target distributions, the log prob may have some very large gradients which can cause numerical instability - the gradient clipping can help with this.

This commit made three changes to the original repo:

1. Fixes the warning regarding the 'is' keyword:

/home/donglin/Github/normalizing-flows/normflow/nets.py:45: SyntaxWarning: "is" with a literal. Did you mean "=="?
  if output_fn is "sigmoid":
/home/donglin/Github/normalizing-flows/normflow/nets.py:47: SyntaxWarning: "is" with a literal. Did you mean "=="?
  elif output_fn is "relu":
/home/donglin/Github/normalizing-flows/normflow/nets.py:49: SyntaxWarning: "is" with a literal. Did you mean "=="?
  elif output_fn is "tanh":
/home/donglin/Github/normalizing-flows/normflow/nets.py:51: SyntaxWarning: "is" with a literal. Did you mean "=="?
  elif output_fn is "clampexp":

2. Fixes the warning regarding 'torch.qr':

/home/donglin/Github/normalizing-flows/normflow/flows.py:616: UserWarning: torch.qr is deprecated in favor of torch.linalg.qr and will be removed in a future PyTorch release.
The boolean parameter 'some' has been replaced with a string parameter 'mode'.
Q, R = torch.qr(A, some)
should be replaced with
Q, R = torch.linalg.qr(A, 'reduced' if some else 'complete') (Triggered internally at  /opt/conda/conda-bld/pytorch_1623448278899/work/aten/src/ATen/native/BatchLinearAlgebra.cpp:1940.)
  Q = torch.qr(torch.randn(self.num_channels, self.num_channels))[0]

3. Eliminated the "nf.util.ToDevice" call during the data pre-processing in glow.ipynb

Hi, when doing experiments, I'd suggest doing some other tutorials, for example for the ClassCondFlow, as while trying to one on my own, I keep encountering this error.

I can't help but to wonder why the NormalizingFlow class use the flows' inverse method when computing forward_kl, but, on the contrary, when using the NormalizingFlowVAE, it uses the flows' forward method.

This way, when trying to fit MNIST with NormalizingFlow, when training and passing a batch of say (64, 784) images I get the following error:

     34 for i in range(len(self.flows) - 1, -1, -1):
     35     z, log_det = self.flows[i].inverse(z)
---> 36     log_q += log_det
     37 log_q += self.q0.log_prob(z)
     38 return -torch.mean(log_q)

RuntimeError: output with shape [64] doesn't match the broadcast shape [1, 64]

Any help/suggestion?

In NormalizingFlowVAE class in core.py, this line says that the encoder outputs log_q:

z, log_q = self.q0(x, num_samples=num_samples)

Suppose that, as in this example, the encoder Gaussian (q0) is parameterized by an MLP. Looking at distributions.encoder.py source code, the forward method of NNDiagGaussian class says that it outputs log_p:

return z, log_p

Inconsistency or not?

Thank you for a repo that's easy to handle with a normalizing flow of one's choice!

I would like to implement a normalizing flow that optimizes multiple target distributions at once depending on the context I would provide to it. Yet, currently, afai, no context can be providided in the .forward_kld method of the NormalizingFlow class.

Would be great if that's added!

Cheers,

Yves

Hi, this PR added some new interface to get better control during sampling, e.g. to repeatedly generate on same latent code when training the model. Please check if it is useful as a merge:)