yell / boltzmann-machines Goto Github PK

As far as I can tell from the instructions, this requires nvidia-docker to run with docker, and nvidia-docker does not run on Windows 10 due to lack of GPU passthrough in that OS (seems like it might work on Windows 2016 Server though). But there is no mention of this anywhere. It seems it should work in a non-GPU configuration, which would suffice for learning purposes.

So is nvidia-docker a requirement, or is there some undocumented way to get this to work with docker on Windows 10? Without GPU support?

Continue discussion #7 (comment)

The current lib has a lot of metrics, but in the research is expected try other metrics.

Example:

def my_custom_evaluate_function(metric_name, model, minibatch):
    """Mean of activated units after reconstruction"""
    h_means = model._means_h_given_v(minibatch)
    h0 = self._sample_h_given_v(h_means)
    v_means = model._means_v_given_h(gh)
    v1 = self._sample_v_given_h(v_means)

    with tf.name_scope(metric_name):
        tf.summary.scalar(metric_name, tf.mean(v1, axis=1))  # Maybe axis=1


rbm = BernoulliRBM(n_visible=784, n_hidden=args.n_hidden,
    metrics_config=dict(
        # The default metrics
        msre=True,
        pll=True,
        feg=True,
        train_metrics_every_iter=1000,
        val_metrics_every_epoch=2,
        feg_every_epoch=4,
        n_batches_for_feg=50,
        # New metrics
        my_custom_evaluate=my_custom_evaluate_function
    ),
    verbose=True,
)

utils.py Line 37

# Use // instead /
n_batches = N / batch_size + (N % batch_size > 0)

Python 2: float / int -> int
Python 3: float / int -> float
Python 3: float // int -> int

use tensorflow data batchsize generator

I will add information later

@jordanbCS @SrMouraSilva @hannesdm @yell
What is the sequence of files that needs to be executed to run this repo successfully?
Looking forward to your reply!

@yell @jordanbCS @SrMouraSilva @hannesdm
Hi!
What is fetch_models.sh in the code?What's more,How does it work?
Looking forward to your explanation!Thank you very much!

@SrMouraSilva @jordanbCS @hannesdm @yell
Hi!I'm sorry to bother you again!
I look at the repo,but I got confused!I hope you could make a explainationfor that in detai!
And how were X_aug_mean.npy and X_aug_std.npy generated in notebooks folder?
Looking forward to your reply!

• based on sampling
• based on thresholding (e.g. > 0.5)

The code has unit tests, but the lack of a continuous deploy caused the new version to break in python3

• Travis build
• Coverage
• Readme update

Dear, I want to use your RBM to train for my 3D dataset, I have processed data already and just need some setting so can you guide me a bit to use it and get all the figures and plots,
Am very new to this area therefore need a little guidance to start.

Are the biases for the models stored somehwere?

Right now they cannot be loaded fully from the files.

I am looking for something equivalent to model.predict(X_test) in keras.

res = dbm.reconstruct(X_test)

Traceback (most recent call last):
File "/home/yunus/PycharmProjects/boltz_/boltzmann-machines-clone/examples/dbm_mnist.py", line 402, in
main()
File "/home/yunus/PycharmProjects/boltz_/boltzmann-machines-clone/examples/dbm_mnist.py", line 377, in main
res = dbm.reconstruct(X_test)
File "/home/yunus/PycharmProjects/boltz_/boltzmann-machines-clone/boltzmann_machines/base/tf_model.py", line 28, in wrapped_f
res = f(model, *args, **kwargs)
File "/home/yunus/PycharmProjects/boltz_/boltzmann-machines-clone/boltzmann_machines/dbm.py", line 877, in reconstruct
self._reconstruction = tf.get_collection('reconstruction')[0]
IndexError: list index out of range

tf.get_collection('reconstruction') returns an empty list.

when i run
dummy2 = tf.get_default_graph().get_all_collection_keys() there is no reconstruction entry.

reconstruct is the equivalent i thought but i cant get it to work.
Is there a reconstruct for RBM or how is it called?