İlker Kesen

I forked the benchmark repo and then I tried to install DyNet on our test machine but failed to run examples provided by DyNet. I opened [an issue](https://github.com/clab/dynet/issues/298) about it....

I installed Chainer on our test machine successfully, still working for DyNet. I implemented first model (RNN language model) and I am able to compare it with Chainer and Theano...

I've completed all the benchmark implementations. I will continue with the profiling. Finally, I was able to install DyNet on our test machine; but I couldn't have made the benchmark...

By the way, TF Fold handles minibatching in a different way. They introduced dynamic batching algorithm. They define two different terms, ops and operations. In the above example, embedding a...

I switched from one-hot vectors to indexes in RNN language model and BiLSTM tagger examples. Here is the new results, - RNN lang. model: word_per_sec=15988.6631 - BiLSTM tagger: word_per_sec=585.7604 Right...

We have 30.2239 sentences per second in Tree-structured LSTM example with array indexing right now. Before indexing it was 16 sentences per second. In the paper reported results are 90...

Yes, we have DyNet installed on aitest. Let me do it in this night.

I think I need to think about it again. Apart from Adam optimization, in the simplest case (vanilla SGD) we need to get subembedding array like that, `subembed = embed[batch_inds,:]`...

In RNN language model I've also tried my other idea (concatenate all timesteps' indices, get embeddings from those indices and then take slices like `embed[(t-1)*batchsize+1:t*batchsize-1,:]` per timestep) it didn't give...

Right now all benchmark examples use indexing and here is the result for remaining example, - BiLSTM tagger with chars, word_per_sec=456.1816 (it was word_per_sec=232.5641 previously) I've tried several things on...