implementing LSTM with theano scan, way slower then using loops

Question

I am using Theano/Pylearn2 to implement LSTM model inside my own network. However, I've found that Theano scan is much, much slower than using plain loops. I used the Theano profiler

<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Class name>
  95.4%    95.4%      25.255s       4.31e-02s     Py     586       3   theano.scan_module.scan_op.Scan
   1.8%    97.2%       0.466s       4.72e-05s     C     9864      41   theano.sandbox.cuda.basic_ops.GpuElemwise
   0.8%    97.9%       0.199s       8.75e-05s     C     2276      10   theano.sandbox.cuda.basic_ops.GpuAlloc
   0.7%    98.7%       0.196s       1.14e-04s     C     1724       8   theano.sandbox.cuda.blas.GpuDot22
   0.3%    99.0%       0.087s       1.06e-04s     C      828       3   theano.sandbox.cuda.basic_ops.GpuIncSubtensor
   0.2%    99.2%       0.051s       1.66e-04s     Py     310       2   theano.sandbox.cuda.basic_ops.GpuAdvancedSubtensor1

and the Ops,

<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Op name>
  77.2%    77.2%      20.433s       7.40e-02s     Py     276        1   forall_inplace,gpu,grad_of_lstm__layers}
  18.2%    95.4%       4.822s       1.56e-02s     Py     310        2   forall_inplace,gpu,lstm__layers}

So lots and lots of time are spent on Scan (which is kind of as expected, but I didn't expect it to be soo slow).

The main body of my code is

        def fprop(self, state_below, state_prev = 0, cell_prev = 0):
            if state_prev == None:
              state_prev = self.state_prev;
            if cell_prev == None:
              cell_prev = self.cell_prev;
            i_gate = T.nnet.sigmoid(T.dot(state_below,self.Wi) +
                                                            T.dot(state_prev,self.Ui));
            f_gate = T.nnet.sigmoid(T.dot(state_below,self.Wf) +
                                                            T.dot(state_prev,self.Uf));
            C = T.tanh(T.dot(state_below, self.Wc) +
                               T.dot(state_prev, self.Uc));
            C = i_gate * C + f_gate  * cell_prev;
            o_gate = T.nnet.sigmoid(T.dot(state_below,self.Wo) +
                                                            T.dot(state_prev,self.Uo) +
                                                            T.dot(C, self.Vo));
            h_out = o_gate * T.tanh(C);
            return h_out, C

And I wrote my scan as:

[h,c,out], _ = theano.scan(fn=self.fprop_with_output,
               sequences=[X.T,Y[:,1:].T],
               outputs_info=[dict(initial=h_,taps=[-1]), dict(initial=c_,taps=[-1]),None],n_steps=X.shape[1]-1);

One thing I've noticed is that the type of Theano scan uses Python implementation (?) is that the reason why this is ridiculously slow? or did I do something wrong? Why is Theano python implementation of Scan instead of C's.

(I said using loops is faster, but it's faster at runtime, for large model I didn't manage to compile the version of using loops within reasonable amount of time).

score 2 · Answer 1 · answered Jul 09 '15 at 21:34

2

This was asked a while ago but I had/have the same problem. Answer is that scan is slow on the GPU.

See: https://github.com/Theano/Theano/issues/1168

answered Jul 09 '15 at 21:34

AsianYayaToure

153
1
2
11

xuancong84 · Answer 2 · 2016-07-19T11:48:04.463

1

It takes time for Theano developers to implement scan and gradient-of-scan using C and GPU, because it is much more complicated than other functions. That is why when you profile it, it shows GpuElemwise, GpuGemv, GpuDot22, etc., but you don't see a GpuScan or GpuGradofScan.

Meanwhile, you can only fall back to for loops.

edited Jul 19 '16 at 11:48

answered Jul 19 '16 at 11:43

xuancong84

1,412
16
17

implementing LSTM with theano scan, way slower then using loops

2 Answers2