Skip to content

Instantly share code, notes, and snippets.

@shackenberg

shackenberg/minimal.rnn.theano.py

Last active December 22, 2015 05:19
Show Gist options
  • Star 0 You must be signed in to star a gist
  • Fork 0 You must be signed in to fork a gist
  • Save shackenberg/6423638 to your computer and use it in GitHub Desktop.
Save shackenberg/6423638 to your computer and use it in GitHub Desktop.
Download ZIP
Raw
minimal.rnn.theano.py
import numpy
import theano
import theano.tensor as TT
# number of hidden units
n = 50
# number of input units
nin = 5
# number of output units
nout = 5
# input (where first dimension is time)
u = TT.matrix()
# target (where first dimension is time)
t = TT.matrix()
# initial hidden state of the RNN
h0 = TT.vector()
# learning rate
lr = TT.scalar()
# recurrent weights as a shared variable
W = theano.shared(numpy.random.uniform(size=(n, n), low=-.01, high=.01))
# input to hidden layer weights
W_in = theano.shared(numpy.random.uniform(size=(nin, n), low=-.01, high=.01))
# hidden to output layer weights
W_out = theano.shared(numpy.random.uniform(size=(n, nout), low=-.01, high=.01))
# recurrent function (using tanh activation function) and linear output
# activation function
def step(u_t, h_tm1, W, W_in, W_out):
h_t = TT.tanh(TT.dot(u_t, W_in) + TT.dot(h_tm1, W))
y_t = TT.dot(h_t, W_out)
return h_t, y_t
# the hidden state `h` for the entire sequence, and the output for the
# entrie sequence `y` (first dimension is always time)
[h, y], _ = theano.scan(step,
sequences=u,
outputs_info=[h0, None],
non_sequences=[W, W_in, W_out])
# error between output and target
error = ((y - t) ** 2).sum()
# gradients on the weights using BPTT
gW, gW_in, gW_out = TT.grad(error, [W, W_in, W_out])
# training function, that computes the error and updates the weights using
# SGD.
fn = theano.function([h0, u, t, lr],
error,
updates={W: W - lr * gW,
W_in: W_in - lr * gW_in,
W_out: W_out - lr * gW_out},
profile=True)
ntime = 1000
data = numpy.random.uniform(size=(ntime, nin))
targets = numpy.round(numpy.random.uniform(size=(ntime, nout)))
init_state = numpy.zeros(n)
learning_rate = 0.1
result = fn(init_state, data,targets, learning_rate)
Raw
profile.report.txt
Function profiling
==================
Message: None
Time in 1 calls to Function.__call__: 7.018185e-02s
Time in Function.fn.__call__: 7.013297e-02s (99.930%)
Time in thunks: 7.007813e-02s (99.852%)
Total compile time: 4.485335e+00s
Theano Optimizer time: 2.545068e+00s
Theano validate time: 2.463958e-01s
Theano Linker time (includes C, CUDA code generation/compiling): 3.004949e-01s
Class
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Class name>
98.3% 98.3% 0.069s 3.44e-02s Py 2 2 <class 'theano.scan_module.scan_op.Scan'>
0.5% 98.8% 0.000s 3.68e-04s C 1 1 <class 'theano.tensor.blas.Gemm'>
0.5% 99.3% 0.000s 1.01e-05s C 35 35 <class 'theano.tensor.elemwise.Elemwise'>
0.3% 99.6% 0.000s 6.64e-05s C 3 3 <class 'theano.tensor.basic.Alloc'>
0.3% 99.9% 0.000s 1.90e-04s C 1 1 <class 'theano.tensor.subtensor.IncSubtensor'>
0.0% 99.9% 0.000s 2.86e-06s C 8 8 <class 'theano.tensor.subtensor.Subtensor'>
0.0% 99.9% 0.000s 1.77e-06s C 9 9 <class 'theano.tensor.elemwise.DimShuffle'>
0.0% 100.0% 0.000s 9.06e-06s Py 1 1 <class 'theano.tensor.basic.Rebroadcast'>
0.0% 100.0% 0.000s 9.06e-06s C 1 1 <class 'theano.tensor.elemwise.Sum'>
0.0% 100.0% 0.000s 9.88e-07s C 7 7 <class 'theano.tensor.opt.Shape_i'>
0.0% 100.0% 0.000s 7.68e-07s C 9 9 <class 'theano.tensor.basic.ScalarFromTensor'>
0.0% 100.0% 0.000s 2.03e-06s C 2 2 <class 'theano.tensor.opt.MakeVector'>
0.0% 100.0% 0.000s 4.77e-07s C 2 2 <class 'theano.tensor.basic.Reshape'>
... (remaining 0 Classes account for 0.00%(0.00s) of the runtime)
Ops
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Op name>
73.3% 73.3% 0.051s 5.14e-02s Py 1 1 forall_inplace,cpu,grad_of_scan_fn}
25.0% 98.3% 0.018s 1.75e-02s Py 1 1 forall_inplace,cpu,scan_fn}
0.5% 98.8% 0.000s 3.68e-04s C 1 1 Gemm{inplace}
0.3% 99.1% 0.000s 2.15e-04s C 1 1 Elemwise{Composite{[sub(i0, sqr(i1))]}}
0.3% 99.4% 0.000s 6.64e-05s C 3 3 Alloc
0.3% 99.7% 0.000s 1.90e-04s C 1 1 IncSubtensor{Set;:int64:}
0.0% 99.7% 0.000s 2.91e-05s C 1 1 Elemwise{mul,no_inplace}
0.0% 99.8% 0.000s 7.51e-06s C 2 2 Elemwise{Composite{[sub(i0, mul(i1, i2))]}}[(0, 0)]
0.0% 99.8% 0.000s 9.06e-06s C 1 1 Sum
0.0% 99.8% 0.000s 9.06e-06s Py 1 1 Rebroadcast{0}
0.0% 99.8% 0.000s 4.41e-06s C 2 2 Subtensor{int64:int64:int64}
0.0% 99.8% 0.000s 8.11e-06s C 1 1 Elemwise{Sub{output_types_preference=transfer_type{0}}}[(0, 0)]
0.0% 99.8% 0.000s 7.15e-06s C 1 1 Elemwise{Composite{[Composite{[Composite{[Composite{[Switch(i0, i1, Sw
0.0% 99.8% 0.000s 7.68e-07s C 9 9 ScalarFromTensor
0.0% 99.8% 0.000s 3.46e-06s C 2 2 Elemwise{Composite{[Composite{[Composite{[Composite{[Composite{[Compos
0.0% 99.8% 0.000s 3.46e-06s C 2 2 Elemwise{Composite{[Composite{[Composite{[Switch(i0, i1, maximum(i2, i
0.0% 99.9% 0.000s 3.46e-06s C 2 2 Elemwise{Composite{[Composite{[Composite{[Composite{[Composite{[Compos
0.0% 99.9% 0.000s 6.91e-06s C 1 1 Elemwise{Sqr{output_types_preference=transfer_type{0}}}[(0, 0)]
0.0% 99.9% 0.000s 5.96e-06s C 1 1 Elemwise{Composite{[Switch(i0, i1, minimum(i2, i3))]}}
0.0% 99.9% 0.000s 2.62e-06s C 2 2 Subtensor{::-1}
... (remaining 28 Ops account for 0.12%(0.00s) of the runtime)
Apply
------
<% time> <sum %> <apply time> <time per call> <#call> <id> <Apply name>
73.3% 73.3% 0.051s 5.14e-02s 1 71 forall_inplace,cpu,grad_of_scan_fn}(Shape_i{0}.0, InplaceDimShuffle{0,x,1}.0, Ele
25.0% 98.3% 0.018s 1.75e-02s 1 57 forall_inplace,cpu,scan_fn}(Shape_i{0}.0, Subtensor{int64:int64:int8}.0, IncSubte
0.5% 98.8% 0.000s 3.68e-04s 1 80 Gemm{inplace}(<TensorType(float64, matrix)>, Elemwise{neg,no_inplace}.0, Reshape{
0.3% 99.1% 0.000s 2.15e-04s 1 63 Elemwise{Composite{[sub(i0, sqr(i1))]}}(TensorConstant{(1, 1) of 1.0}, Subtensor{
0.3% 99.4% 0.000s 1.91e-04s 1 18 Alloc(TensorConstant{0.0}, Elemwise{add,no_inplace}.0, Shape_i{0}.0)
0.3% 99.7% 0.000s 1.90e-04s 1 22 IncSubtensor{Set;:int64:}(Alloc.0, Rebroadcast{0}.0, Constant{1})
0.0% 99.7% 0.000s 2.91e-05s 1 65 Elemwise{mul,no_inplace}(TensorConstant{(1, 1) of 2.0}, Subtensor{::-1}.0)
0.0% 99.7% 0.000s 1.31e-05s 1 77 Elemwise{Composite{[sub(i0, mul(i1, i2))]}}[(0, 0)](<TensorType(float64, matrix)>
0.0% 99.8% 0.000s 9.06e-06s 1 70 Sum(Elemwise{Sqr{output_types_preference=transfer_type{0}}}[(0, 0)].0)
0.0% 99.8% 0.000s 9.06e-06s 1 17 Rebroadcast{0}(InplaceDimShuffle{x,0}.0)
0.0% 99.8% 0.000s 8.11e-06s 1 58 Elemwise{Sub{output_types_preference=transfer_type{0}}}[(0, 0)](forall_inplace,cp
0.0% 99.8% 0.000s 7.87e-06s 1 59 Subtensor{int64:int64:int64}(forall_inplace,cpu,scan_fn}.0, ScalarFromTensor.0, S
0.0% 99.8% 0.000s 7.15e-06s 1 45 Elemwise{Composite{[Composite{[Composite{[Composite{[Switch(i0, i1, Switch(i2, i3
0.0% 99.8% 0.000s 6.91e-06s 1 67 Elemwise{Sqr{output_types_preference=transfer_type{0}}}[(0, 0)](Elemwise{Sub{outp
0.0% 99.8% 0.000s 5.96e-06s 1 33 Elemwise{Composite{[Composite{[Composite{[Composite{[Composite{[Composite{[add(Ca
0.0% 99.8% 0.000s 5.96e-06s 1 44 Elemwise{Composite{[Composite{[Composite{[Switch(i0, i1, maximum(i2, i3))]}(i0, i
0.0% 99.8% 0.000s 5.96e-06s 1 28 Elemwise{Composite{[Composite{[Composite{[Composite{[Composite{[Composite{[Switch
0.0% 99.8% 0.000s 5.96e-06s 1 29 Elemwise{Composite{[Switch(i0, i1, minimum(i2, i3))]}}(Elemwise{le,no_inplace}.0,
0.0% 99.9% 0.000s 5.01e-06s 1 48 Elemwise{Composite{[Switch(i0, i1, minimum(i2, i3))]}}[(0, 2)](Elemwise{le,no_inp
0.0% 99.9% 0.000s 5.01e-06s 1 40 Alloc(TensorConstant{0.0}, Elemwise{Composite{[Composite{[Composite{[Composite{[C
... (remaining 61 Apply instances account for 0.14%(0.00s) of the runtime)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment