You missed the peephole weights (one from the cell state to each gate unit, i.e....
You can just add dummy targets (e.g., zeros) of the right size (one integer per frame...
I'm using it with CUDA 6.5. You need to apply a small patch in order to get best...
Hi Alex, The outputs of the bidirectional layer are the concatenation of the activations...
The input layer has no weights, thus the learning rate has no effect. Conceptually,...
I cannot reproduce this bug, sorry. If I insert "learningRate": 0 as a member into...
The targetStrings variable is not required, and in fact it has no effect in CURRENNT....
A BLSTM layer consists of 2 LSTM layers of the same size, i.e., a BLSTM of size N...
A BLSTM layer of size N consists of 2 LSTM layers of size N/2. The number of weights...
Hi Scott, It also has to do with the parallel_sequences option, because the more...
Hi gibbi92, Thanks for the code. I think this is very useful to add as another PostOutputLayer...
For a ce layer you need regresssion targets (target probabilities), not class indices....
Should I use ce or multiclass_classification for the post output layer while training...
This is usually a problem with the cache file (by default it is in /tmp). Make sure...
I think these questions are answered in the paper: http://jmlr.org/papers/v16/we...
Hi Tomasz, You are right: the network is able to look ahead for input values, and...
We're not using multiple GPUs here. The mini-batches (data fractions) are processed...
No, that's why it's called maxSeqTagLength - in this case, all tags need to be <...
NetCDF models string arrays as character arrays (rather than a list of arrays), thus...
NetCDF models string arrays as character arrays (rather than a list of arrays), thus...
NetCDF models strings as character arrays (rather than a list of arrays), thus a...
Yes, exactly. The example would look something like inputs = 0, 0, 0, 1, 1, 0, 0,...
Hi Tomasz, the sequences are processed from left to right, and output_time_lag provides...
Hi gibbi, yes, the state of the net is reset when splitting sequences. As long as...
Alex, for language modeling with RNNs you need N+1 inputs and outputs, where N is...
Alex, for language modeling you need N+1 inputs and outputs, where N is the vocabulary...
Alex, for language modeling you need N+1 inputs and outputs, where N is the vocabulary...
DNN layer.
Hi Harlei, glad that you found it useful. For your current problem, I would do the...
Hi Yonatan, sorry for the late reply - good to see that you found the source of the...
Yes, cross-entropy post-output layer should work. Make sure you use a feedforward_logistic...
By the way, the output labels have to be 0, 1, 2, 3, 4 by convention.
This seems like a classical sequence prediction task. I'd suggest the following:...
The bug is caused by some lines in main.cpp that overwrite the json's input size...
It seems that some untested code has crept into the rc code. This will be fixed in...
It is supposed to work without changing the NetCDF file. The JSON you posted is correct,...
Hi, the order in which the LSTM weights are saved in the json weight container is...
Cool. I think the code should be fine as is (without the compiler-level check), since...
Thanks!! I will try this ASAP. Not sure if it's the limit - but since Dmytro reported...
Harlei, the numTimesteps dimension corresponds to the total number of timesteps (50...
This library is all about multivariate regression and classification. You can look...
This library is all about multivariate regression and classification. You can look...
This library is all about multivariate regression and classification. You can look...
This library is all about multivariate regression and classification. You can look...
This library is all about multivariate regression and classification. You can look...
No, it's not possible. If it's audio, then you can use openSMILE (but you have to...
You should probably start a new thread for a new question. Anyway, the procedure...
./currennt --list_devices 1 export CURRENNT_CUDA_DEVICE=<device_id>
Hi Xingyu, the limiting factor is that the state variables (cell and gate activations)...
The sequence truncation is only done for training (might be changed in the future),...
We found that this is a problem with new CUDA versions. Actually CUDA 6.5 is roughly...
This seems very slow. How fast does the example (speech recognition, chime, subsampling)...
Dmytro, I think you should use the truncate_seq option since your sequences are rather...
Currently not but since I received multiple feature requests for this I might consider...
I found the bug, it was due to the targets not being loaded from the nc file in case...
Hi Simon Thank you very much for downloading CURRENNT and your detailed bug report....