I am just curious, in what scenario, one would like to migrate from torch7 to Pytorch (other than one like to work on python)?
More efficient memory consumption, and if you use nngraph in torch7, constructing computational graph is much easier and simpler in PyTorch.
Built-in autograd and RNN modules convinced me.
Anecdotally, a fairly complex seq2seq+attention model took me about a week to build in Torch. The majority of that was figuring out backpropagation. The same model took 3 weeks (and counting, I gave up) in TensorFlow. In PyTorch it took about an hour.
Thanks a lot. This seems more compelling reason, infact the Pytorch seems more intuitive than Torch containers.