Refactoring #22
Merged
Refactoring #22
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Closed
|
|
||
| function split_vis(crbm::ConditionalRBM, vis::Mat{Float64}) | ||
| steps=5, sigma=0.01) | ||
| ConditionalRBM(Float64, Bernoulli, Bernoulli, n_vis, n_hid, steps; |
There was a problem hiding this comment.
I think you want this to be ConditionalRBM(Float64, V, H, n_vis, n_hid, ...? If so I can change that when I add my generalization changes.
Rather than assuming the conditioned patterns will be the visible units at previous time steps, we just take a size to condition on and provide a constructor that takes the number of steps as a feature. With the above change the split_vis method has changed and all references to `history` or `hist` have been changed to `cond`. Also, fixed an issue on line 51 which was passing `Bernoulli` instead of V and H parametrics to the more general constructor.
Generalized the conditional rbm.
| end | ||
| return samples | ||
| end | ||
|
|
||
|
|
||
| function sample_hiddens{V,H}(rbm::AbstractRBM{V, H}, vis::Mat{Float64}) | ||
| function sample_hiddens{T,V,H}(rbm::AbstractRBM{T,V,H}, vis::Mat{T}) |
There was a problem hiding this comment.
Might want to accept a vis matrix of typeF<:AbstractFloat and do a convert(Array{T}, vis) if F != T
There was a problem hiding this comment.
If we agree on implicit conversion, I don't see a reason to limit T to be an instance of AbstractFloat - one may be interested in passing matrix of Ints or even Boolean, which is very natural for Bernoulli RBM.
…issues. 1. In the sparsity calculation for conditional rbms `hid_means` takes just the visible input provided by `split_vis(rbm, X)`. 2. The `gemm!` calls in gradient_classic for the conditional weights weren't using the appropriate precision (ie: Float32 w/ sparsity) 3. The `free_energy` function will often produces NaNs do to `log(0)` or lack of precision.
Fixed sparsity calculation for conditional rbm and several precision …
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Main changes:
gradient_classic()andupdate_classic!()are used.update_weights!and a number of functions modifying gradient, including those for weight decay (L1 and L2), sparsity, learning rate, etc. All such functions have the same signature, so one can easily combine them in an arbitrary way.ctx) - a dictionary used for holding configuration and buffers. We don't need to pass 3-5 options + buffers to each function anymore, now they all are combined into a context.TextReporteris used which simply prints current epoch, score and elapsed time).Float64andFloat32. Support for other (real?) data types may be provided given that BLAS functions are defined for them.Two things left outside this refactoring. One is EMF approximation by @sphinxteam which seems like a perfect option for gradient calculation. Corresponding paper has a little bit too much physics for me, but if the team finds it possible to transform their work into this repository, it will be a great advantage for the whole Julia stats/deep learning community.
Another thing that I unsuccessfully tried to borrow is visual monitor from the same fork. Unfortunately, in its current state it's closely bound to TAP free energy, which makes it quite hard to use as is. Yet, it's quite possible I will include something similar later.
Comments and criticism are highly welcome.
CC: @Rory-Finnegan @eric-tramel