If pad_packed_sequence was a chef, would its signature dish be the "Sequence Shuffle"? 🍽️🔀 Source: https://t.co/sN8ljXuvaw #NLP #PyTorchNLP #PyTorch #MachineLearningCommunity #work #pytorch

@_arohan_ I checked two dropout variants (in blue: weight dropout on the BN params, adapted from https://t.co/3mXw4TR083; in purple: spatial dropout after BN layers via torch.nn.Dropout2d) and both perform worse than SAM and SAM-ON.

torchnlp にBatchBucketingSampler があった。少しきになるじっっそうではあるけれど、音声認識につかってみる。 https://t.co/wXBHorsjpq

自然言語処理やるのにpytorchnlp向けのメソッド活用するとうれしいのかなぁ？ 便利ならpytorch覚えようかと考えている

@AlexanderVeysov Attention is slow, b/c it depends on the sequence length. Of course, if you have an end-to-end CUDA kernel, it's gonna be somewhat faster, but I am not sure that this kind of thing is gonna be super slow

Imagine loading pre-trained word2vec and glove in 2020, directly from a library! Glad for torchtext, pytorchnlp, gensim of course and all the new libraries and frameworks for NLP! https://t.co/cfT0WliVfw

@AlexanderVeysov PS: Just in case PytorchNLP is not an official PyTorch sub-package.

Take a look at a dozen options for your next NLP application. https://t.co/qZyGIEPwTq #NLP #SpaCY #PyTorchNLP #OpenNLP