I'm planning on setting up tensor flow and working with some sequence to sequence models (https://www.tensorflow.org/tutorials/seq2seq) and setting up a toy translation example between English and French (I've already got TF up and running in docker, just haven't started anything yet because I wanted some advice first). For another case I want to implement though, I was reading through the bucketing and padding section, and my one language (B) it will be a fixed length chunks and I was wondering if I would see too bad of a performance hit if I created separate buckets for each possible tokenized length of A? (Probably end up with about 2 dozen buckets or so and I haven't been able to find anything on performance issues with buckets) I'll be using an approximately 1TB dataset for this project, which is larger than the sample language translation dataset they use in the tutorial.
I'm definitely a newb when it comes to machine learning, I've set up torch-rnn in docker and ran a few toy examples through it (The ones from https://medium.com/@ageitgey/machine-learning-is-fun-part-2-a26a10b68df3 mostly, as well as tiny-shakespeare, the combined works of stephen king and the combined works of robert jordan) and saw some cool output and wanted to try the next logical step of tensor flow, and seq2seq matches one of the problems I really want to play with .
install gentoo
>>61720558
Recently I found this site about Machine Learning for artists (http://ml4a.github.io/), and I'm fairly new to it as well but perhaps you can find some of the resources you need
>>61721685
Gentoo isn't officially supported by CUDA.
>>61720558
You can just use tf.nn.dynamic_rnn and pass in a length tensor if you don't want to deal with ugly bucketing
tensorflow is fun
anyone think the appstores will accept my app?