Skip-Thoughts.torch is a lightweight porting of skip-thought pretrained models from Theano to Torch7 using the rnn library of Element-Research and npy4th.
- python3 (python2 not tested)
- torch
- numpy
$ git clone --recursive http://github.com/Cadene/skip-thoughts.torch
It uses the nn.GRU layer from torch with the cudnn backend. It is the fastest implementation, but the dropout is sampled after each time-step in the cudnn implementation... (equals bad regularization)
It uses the nn.GRUCell layer from torch with the cudnn backend. It is slightly slower than UniSkip, however the dropout is sampled once for all time-steps in a sequence (good regularization).
It uses a custom GRU layer with a torch backend. It is at least two times slower than UniSkip, however the dropout is sampled once for all time-steps for each Linear (best regularization).
Equivalent to UniSkip, but with a bi-sequential GRU.
import torch
from torch.autograd import Variable
import sys
sys.path.append('skip-thoughts.torch/pytorch')
from skipthoughts import UniSkip
dir_st = 'data/skip-thoughts'
vocab = ['robots', 'are', 'very', 'cool', '<eos>', 'BiDiBu']
uniskip = UniSkip(dir_st, vocab)
input = Variable(torch.LongTensor([
[1,2,3,4,0], # robots are very cool 0
[6,2,3,4,5] # bidibu are very cool <eos>
])) # <eos> token is optional
print(input.size()) # batch_size x seq_len
output_seq2vec = uniskip(input, lengths=[4,5])
print(output_seq2vec.size()) # batch_size x 2400
output_seq2seq = uniskip(input)
print(output_seq2seq.size()) # batch_size x seq_len x 2400$ luarocks install tds # for the hashmap
$ luarocks install rnn # for the rnn utils
$ luarocks install --server=http://luarocks.org/dev skipthoughts
The skip-thoughts package enables you to download the pretrained torch7 hashmaps and GRUs, and also to cleanly set the pretrained skip-thoughts models. In fact, the initial vocabulary is made of 930,913 words (including the vocabulary of word2vec). That is why, it is preferable to create a nn.LookupTableMaskZero in order to map your smaller vocabulary to their corresponding vectors in an efficient and "fine-tunable" way. See an example bellow:
st = require 'skipthoughts'
vocab = {'skipthoughts', 'are', 'cool'}
inputs = torch.Tensor{{1,2,3}} -- batch x seq
-- Download and load pretrained models on the fly
uni_skip = st.createUniSkip(vocab, 'data')
print(uni_skip:forward(inputs):size()) -- batch x 2400For further examples please refer to torch/example.lua or torch/test.lua
The uni-skip model is made of:
- a hashmap which, just as word2vec, map a word (from a dictionnary of 930,913 words) to its corresponding vector (620 dimensions),
- a GRU which takes as input the latter vector and process the final skip-thought vector (2400 dimensions).
The bi-skip model is made of:
- a different hashmap (but same dictionnary),
- a first GRU (forward) which takes a vector (620 dimensions) and output a vector (1200 dimensions),
- a second GRU (backward) which takes the same vector (620 dimensions) and output a vector (1200 dimensions). The final skip-thought vector is the result of the concatenation of the two vectors (2400 dimensions).
The combine-skip model outputs the concatenation of both models output vectors (4800 dimensions).
Finally, once those pretrained models are set to take as input a sequence of words (notably by using the nn.Sequencer and nn.BiSequencer), they can be used to compute a sequence of features of the same size (seq2seq) or a features vector (seq2vec).
We provide a new GRU called GRUST for "Gated Recurrent Unit for Skip-Toughts".
In fact, the authors of skip-thoughts models did not use the same implementation as in the rnn library.
The implementation of GRUST corresponds to the following algorithm:
z[t] = σ(W[x->z]x[t] + W[s->z]s[t−1] + b[1->z]) (1)
r[t] = σ(W[x->r]x[t] + W[s->r]s[t−1] + b[1->r]) (2)
h[t] = tanh(W[x->h]x[t] + r[t] .* W[hr->c](s[t−1]) + b[1->h]) (3)
s[t] = (1-z[t])h[t] + z[t]s[t-1] (4)
(with .* the element wise product)
Note: It is also the same implementation of GRU from pytorch.
Whereas, the implementation of GRU from the rnn package corresponds to the following algorithm:
h[t] = tanh(W[x->h]x[t] + W[hr->c](s[t−1] .* r[t]) + b[1->h]) (3)
(with .* the element wise product)
We provide a new layer for the bi-skip and (thus) combine-skip models. In fact, the backward GRU may recieve inputs with right zero padding instead of the usual left zero padding. Thus, the MaskZeroCopy layer will copy the last outputs of the backward GRU when it sees vectors of zero, instead of replacing the actual content by zero (usual behaviour of MaskZero).
-- The input is "hello world" but our full model takes batchs of size 3 only
-- thus we need to add a 0 on the left (left zero padding).
-- The ouput of the GRU forward must be the result of precessing hello and then word (= features(hello,word)).
-- Wheras the output of the GRU backward must be the result of processing word, then hello (= features(word,hello)).
input = {0, hello, world}
reverse_input = {world, hello, 0}
-- GRU forward in bi-skip model
-- without MaskZero the final output will be features(0,hello,world)
GRU_fw:forward(input) = {features(0), features(0,hello), features(0,hello,world)}
-- with MaskZero the final output will be features(hello,world)
GRU_fw:forward(input) = {0, features(hello), features(hello,world)}
-- GRU backward in bi-skip model
-- without MaskZero the final output will be features(world,hello,0)
GRU_bw:forward(reverse_input) = {features(world), features(world,hello), features(world,hello,0)}
-- with MaskZero the final output will be 0
GRU_bw:forward(reverse_input) = {features(world), features(world,hello), 0}
-- with MaskZeroCopy the final output will be features(word,hello)
GRU_bw:forward(reverse_input) = {features(world), features(word,hello), features(word,hello)}
Lua/Torch7 and Python2.
$ luarocks install tds
$ luarocks install rnn
$ luarocks install npy4th
$ pip install numpy
$ pip install theano
$ git clone https://github.com/Cadene/skip-thoughts.torch.git
$ cd skip-thoughts.torch
$ git submodule update --init --recursive # download my fork in theano/skip-thoughts
Create uni_hashmap.t7 and bi_hashmap.t7 (both of type tds.Hash) in data/final:
$ th torch/create_hashmaps.lua -dirname data
Create uni_gru.t7, bi_gru_fwd.t7 and bi_gru_bwd.t7 (every three of type GRUST) in data/final:
$ th torch/create_grus.lua -dirname data
$ th torch/test.lua -dirname data
Beside the whole deep learning community, we would like to specifically thanks:
- the authors of the original paper and implementation,
- the authors of DPPnet who first propose a porting,
- the authors of Multi Modal Residual Learning who also propose a 4A5A porting.