ai-di / brancher Goto Github PK

Missing import? Typo?

flake8 testing of https://github.com/AI-DI/Brancher on Python 3.7.1

$ flake8 . --count --select=E9,F63,F72,F82 --show-source --statistics

./brancher/functions.py:135:12: F821 undefined name 'matmul'
    return matmul(weight, input) + bias
           ^
1     F821 undefined name 'matmul'
1

E901,E999,F821,F822,F823 are the "showstopper" flake8 issues that can halt the runtime with a SyntaxError, NameError, etc. These 5 are different from most other flake8 issues which are merely "style violations" -- useful for readability but they do not effect runtime safety.

• F821: undefined name name
• F822: undefined name name in __all__
• F823: local variable name referenced before assignment
• E901: SyntaxError or IndentationError
• E999: SyntaxError -- failed to compile a file into an Abstract Syntax Tree

The variables.py module contains some functions that have a for_gradient parameter. I assume it is for calculating gradients using the results of the functions but I cannot find what actually happens in the code when this variable is set to True. Only estimate_log_model_evidence seems to use for_gradient.

Seems like lt, le, gt, ge, eq, ne are all useful operations in probabilistic models. I tried out adding these to the Variable & PartialLink classes but maybe there's more to it than that. Are these operations that are doable?

I have been trying to fix multiple runs of the exact same program so that it is repeatable and having no luck. I have been using the following method calls:

def fix_seed():
    seed_value = 1337

    np.random.seed(seed_value) 
    torch.manual_seed(seed_value) 
    random.seed(seed_value)
    torch.cuda.manual_seed(seed_value)
    torch.cuda.manual_seed_all(seed_value) 
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False

I have tried setting cfg.set_device to gpu and to cpu and in both cases I cannot reproduce Brancher calls. My actual program just does an inference and then samples from the posterior. I was under the impression that if I fix all of these random seeds before a run and input is the exact same, runs should be reproducible. So far I have not been able to achieve that.

The perform_inference method appears to always show a tqdm progress bar. Is there a way of disabling this?

In many cases, the probability model built by Brancher might be just a part of a model, which is usually built by pytorch (e.g., ResNet). How to combine them together in a simple/easy way?

Thank you for this interesting package. Can we use GPU to accelerate the computation? Like "model.cuda()" in pytorch?

Hi, I am looking to implement a probabilistic model in Brancher that requires a truncated normal distribution. In pymc3 that can be achieved with https://docs.pymc.io/api/bounds.html as well as an explicit TruncatedNormal class. Is there an equivalent or some sort of work-around in Brancher?
Thanks

The init function of StandardVariables gets called twice because it gets called in new and when new is done.

The following code with scale=0.01

x = NormalVariable(loc=0., scale=0.01, name="mu")
model = ProbabilisticModel([x])
sample = model.get_sample(2)
print(sample)

returns this:

   mu_scale  mu_loc        mu
0 -4.600166     0.0 -0.030332
1 -4.600166     0.0 -0.008252

Value of the mu_scale has been changed. Is this expected behavior since I have declared the random variable x with the scale value scale=0.01, but got something like log(0.01) = -4.600166.

Was trying to test out this model and couldn't find this distribution.

Hi :),
setting the negative value of the scale parameter doesn't throw an exception and allows user to even sample from that model. There is only a RuntimeWarning: "invalid value encountered in log return np.log(np.exp(y - self.lower_bound) - 1)".

x = NormalVariable(loc=0., scale=-10., name="mu")
model = ProbabilisticModel([x])

If one try to sample from this model, output will contain NaN but maybe it would be better to throw an exception for the invalid parameters of the model.

sample = model.get_sample(2)
print(sample)

I am currently working on the documentation. Progress on documentation:

• variables.py
• distributions.py
• utilities.py
• standard_variables.py
• config.py
• functions.py
• geometric_ranges.py
• gradient_estimators.py
• inference.py
• modules.py
• optimizers.py
• pandas_interface.py
• particle_inference_tools.py
• stochastic_processes.py
• transformations.py
• visualizations.py