Skip to content

Instantly share code, notes, and snippets.

@ffmpbgrnn

ffmpbgrnn/cpu.log

Last active August 29, 2015 14:25
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 ffmpbgrnn/842e1910f216b1e00e27 to your computer and use it in GitHub Desktop.
Save ffmpbgrnn/842e1910f216b1e00e27 to your computer and use it in GitHub Desktop.
Download ZIP
theano profile
Raw
cpu.log
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:328
Time in 8 calls to Function.__call__: 1.477107e+02s
Time in Function.fn.__call__: 1.474529e+02s (99.825%)
Time in thunks: 1.470501e+02s (99.553%)
Total compile time: 1.377320e+02s
Number of Apply nodes: 525
Theano Optimizer time: 2.186309e+01s
Theano validate time: 2.525887e-01s
Theano Linker time (includes C, CUDA code generation/compiling): 1.157349e+02s
Import time 3.757732e-01s
Time in all call to theano.grad() 4.098411e-01s
Class
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Class name>
64.7% 64.7% 95.128s 9.91e-01s C 96 12 theano.tensor.blas.Dot22
11.5% 76.2% 16.857s 1.05e+00s Py 16 2 theano.scan_module.scan_op.Scan
9.5% 85.6% 13.954s 1.74e+00s C 8 1 theano.tensor.nnet.nnet.Softmax
7.9% 93.6% 11.631s 5.37e-03s C 2168 271 theano.tensor.elemwise.Elemwise
2.1% 95.7% 3.089s 3.86e-01s C 8 1 theano.tensor.blas.Dot22Scalar
1.6% 97.3% 2.389s 3.32e-02s C 72 9 theano.tensor.elemwise.Sum
1.4% 98.7% 2.029s 2.54e-01s Py 8 1 theano.tensor.subtensor.AdvancedIncSubtensor
0.4% 99.1% 0.618s 7.72e-02s C 8 1 theano.tensor.nnet.nnet.SoftmaxGrad
0.3% 99.4% 0.454s 1.89e-02s Py 24 3 theano.tensor.subtensor.AdvancedSubtensor
0.3% 99.7% 0.453s 1.95e-03s C 232 29 theano.tensor.basic.Reshape
0.2% 99.9% 0.278s 2.32e-03s C 120 15 theano.tensor.basic.Alloc
0.1% 100.0% 0.164s 2.28e-03s C 72 9 theano.tensor.subtensor.IncSubtensor
0.0% 100.0% 0.002s 2.15e-04s Py 8 1 theano.tensor.basic.Nonzero
0.0% 100.0% 0.001s 2.26e-06s C 424 53 theano.compile.ops.Shape_i
0.0% 100.0% 0.001s 3.05e-06s C 264 33 theano.tensor.elemwise.DimShuffle
0.0% 100.0% 0.001s 4.06e-06s C 184 23 theano.tensor.opt.MakeVector
0.0% 100.0% 0.001s 3.71e-06s C 200 25 theano.tensor.subtensor.Subtensor
0.0% 100.0% 0.000s 1.16e-06s C 248 31 theano.tensor.basic.ScalarFromTensor
0.0% 100.0% 0.000s 2.26e-05s C 8 1 theano.tensor.basic.Join
0.0% 100.0% 0.000s 2.40e-06s C 24 3 theano.tensor.basic.Flatten
... (remaining 1 Classes account for 0.00%(0.00s) of the runtime)
Ops
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Op name>
64.7% 64.7% 95.128s 9.91e-01s C 96 12 Dot22
9.5% 74.2% 13.954s 1.74e+00s C 8 1 Softmax
7.1% 81.3% 10.468s 1.31e+00s Py 8 1 forall_inplace,cpu,grad_of_scan_fn}
5.4% 86.7% 7.876s 9.85e-01s C 8 1 Elemwise{Composite{(i0 * log((i1 / i2)))}}
4.3% 91.0% 6.388s 7.99e-01s Py 8 1 forall_inplace,cpu,scan_fn}
2.1% 93.1% 3.089s 3.86e-01s C 8 1 Dot22Scalar
1.4% 94.5% 2.029s 2.54e-01s Py 8 1 AdvancedIncSubtensor{inplace=False, set_instead_of_inc=False}
1.3% 95.8% 1.907s 4.77e-02s C 40 5 Sum{axis=[0, 1], acc_dtype=float64}
1.2% 97.0% 1.788s 2.24e-01s C 8 1 Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((i3 * i4 * i5) / i6) + i7))}}
0.4% 97.4% 0.618s 7.72e-02s C 8 1 SoftmaxGrad
0.4% 97.8% 0.557s 1.51e-03s C 368 46 Elemwise{add,no_inplace}
0.3% 98.1% 0.482s 2.01e-02s C 24 3 Sum{axis=[1], acc_dtype=float64}
0.3% 98.4% 0.454s 1.89e-02s Py 24 3 AdvancedSubtensor
0.3% 98.7% 0.453s 2.83e-03s C 160 20 Reshape{2}
0.3% 99.0% 0.418s 5.22e-03s C 80 10 Elemwise{mul,no_inplace}
0.2% 99.2% 0.318s 3.98e-02s C 8 1 Elemwise{clip,no_inplace}
0.2% 99.4% 0.278s 2.32e-03s C 120 15 Alloc
0.2% 99.6% 0.262s 8.19e-03s C 32 4 Elemwise{Composite{(i0 - ((i1 * i2) / sqrt((i3 + i4 + i5))))}}[(0, 0)]
0.1% 99.7% 0.206s 1.84e-03s C 112 14 Elemwise{Composite{(i0 * sqr(i1))}}
0.1% 99.8% 0.101s 3.15e-03s C 32 4 IncSubtensor{Inc;:int64:}
... (remaining 98 Ops account for 0.19%(0.27s) of the runtime)
Apply
------
<% time> <sum %> <apply time> <time per call> <#call> <id> <Apply name>
11.8% 11.8% 17.349s 2.17e+00s 8 92 Dot22(Reshape{2}.0, Reshape{2}.0)
9.5% 21.3% 13.954s 1.74e+00s 8 397 Softmax(Reshape{2}.0)
9.1% 30.4% 13.341s 1.67e+00s 8 489 Dot22(InplaceDimShuffle{1,0}.0, Reshape{2}.0)
7.7% 38.1% 11.309s 1.41e+00s 8 93 Dot22(Reshape{2}.0, Reshape{2}.0)
7.1% 45.2% 10.468s 1.31e+00s 8 446 forall_inplace,cpu,grad_of_scan_fn}(Elemwise{Composite{minimum(minimum(minimum(minimum(i0, i1), i2), i3), i4)}}.0, Elemwise{mul,no_inplace}.0, Elemwise{mul,no_inplace}.0, InplaceDimShuffle{0,2,1}.0, Subtensor{int64:int64:int64}.0, Subtensor{int64:int64:int64}.0, Subtensor{int64:int64:int64}.0, Subtensor{int64:int64:int64}.0, Subtensor{int64:int64:int64}.0, Subtensor{int64:int64:int64}.0, Subtensor{::int64}.0, Alloc.0, Alloc.0, Alloc.0, Alloc.0, Elem
6.6% 51.8% 9.739s 1.22e+00s 8 94 Dot22(Reshape{2}.0, Reshape{2}.0)
6.5% 58.2% 9.491s 1.19e+00s 8 91 Dot22(Reshape{2}.0, Reshape{2}.0)
5.6% 63.9% 8.252s 1.03e+00s 8 491 Dot22(InplaceDimShuffle{1,0}.0, Reshape{2}.0)
5.6% 69.5% 8.230s 1.03e+00s 8 433 Dot22(InplaceDimShuffle{1,0}.0, Reshape{2}.0)
5.4% 74.8% 7.876s 9.85e-01s 8 411 Elemwise{Composite{(i0 * log((i1 / i2)))}}(AdvancedSubtensor.0, Elemwise{clip,no_inplace}.0, InplaceDimShuffle{0,x}.0)
4.3% 79.2% 6.388s 7.99e-01s 8 337 forall_inplace,cpu,scan_fn}(Elemwise{Composite{minimum(minimum(minimum(minimum(i0, i1), i2), i3), i4)}}.0, Subtensor{int64:int64:int8}.0, Subtensor{int64:int64:int8}.0, Subtensor{int64:int64:int8}.0, Subtensor{int64:int64:int8}.0, Subtensor{int64:int64:int8}.0, Alloc.0, Alloc.0, <TensorType(float32, matrix)>, <TensorType(float32, matrix)>, <TensorType(float32, matrix)>, <TensorType(float32, matrix)>)
4.0% 83.2% 5.945s 7.43e-01s 8 485 Dot22(InplaceDimShuffle{1,0}.0, Reshape{2}.0)
3.0% 86.2% 4.401s 5.50e-01s 8 493 Dot22(InplaceDimShuffle{1,0}.0, Reshape{2}.0)
2.1% 88.3% 3.094s 3.87e-01s 8 486 Dot22(Flatten{2}.0, Reshape{2}.0)
2.1% 90.4% 3.089s 3.86e-01s 8 487 Dot22Scalar(Flatten{2}.0, Reshape{2}.0, TensorConstant{0.0010000000475})
1.4% 91.8% 2.092s 2.62e-01s 8 434 Dot22(Reshape{2}.0, InplaceDimShuffle{1,0}.0)
1.4% 93.2% 2.029s 2.54e-01s 8 425 AdvancedIncSubtensor{inplace=False, set_instead_of_inc=False}(Alloc.0, Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((i3 * i4 * i5) / i6) + i7))}}.0, Subtensor{int64}.0, Subtensor{int64}.0)
1.3% 94.5% 1.885s 2.36e-01s 8 370 Dot22(Reshape{2}.0, Reshape{2}.0)
1.2% 95.7% 1.788s 2.24e-01s 8 424 Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((i3 * i4 * i5) / i6) + i7))}}(AdvancedSubtensor.0, TensorConstant{(1, 1) of 1e-07}, TensorConstant{(1, 1) of 1.0}, TensorConstant{(1, 1) of -1.0}, InplaceDimShuffle{0,x}.0, AdvancedSubtensor.0, Elemwise{clip,no_inplace}.0, Elemwise{true_div,no_inplace}.0)
1.1% 96.8% 1.561s 1.95e-01s 8 430 Sum{axis=[0, 1], acc_dtype=float64}(InplaceDimShuffle{1,0,2}.0)
... (remaining 505 Apply instances account for 3.24%(4.77s) of the runtime)
Scan Op profiling ( scan_fn )
==================
Message: None
Time in 8 calls of the op (for a total of 240 steps) 6.386277e+00s
Total time spent in calling the VM 6.356155e+00s (99.528%)
Total overhead (computing slices..) 3.012228e-02s (0.472%)
Time in all call to theano.grad() 4.098411e-01s
Class
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Class name>
97.9% 97.9% 6.218s 6.48e-03s C 960 4 theano.tensor.blas.Gemm
2.1% 100.0% 0.135s 1.88e-04s C 720 3 theano.tensor.elemwise.Elemwise
... (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>
97.9% 97.9% 6.218s 6.48e-03s C 960 4 Gemm{no_inplace}
1.3% 99.2% 0.085s 3.52e-04s C 240 1 Elemwise{Composite{((i0 * i1 * i2) + (i3 * i4))}}
0.8% 100.0% 0.051s 1.06e-04s C 480 2 Elemwise{mul,no_inplace}
... (remaining 0 Ops account for 0.00%(0.00s) of the runtime)
Apply
------
<% time> <sum %> <apply time> <time per call> <#call> <id> <Apply name>
33.6% 33.6% 2.132s 8.88e-03s 240 2 Gemm{no_inplace}(<TensorType(float32, matrix)>, TensorConstant{1.0}, Elemwise{mul,no_inplace}.0, <TensorType(float32, matrix)>, TensorConstant{1.0})
23.1% 56.6% 1.467s 6.11e-03s 240 1 Gemm{no_inplace}(<TensorType(float32, matrix)>, TensorConstant{1.0}, Elemwise{mul,no_inplace}.0, <TensorType(float32, matrix)>, TensorConstant{1.0})
20.8% 77.5% 1.322s 5.51e-03s 240 4 Gemm{no_inplace}(<TensorType(float32, matrix)>, TensorConstant{1.0}, Elemwise{mul,no_inplace}.0, <TensorType(float32, matrix)>, TensorConstant{1.0})
20.4% 97.9% 1.297s 5.40e-03s 240 3 Gemm{no_inplace}(<TensorType(float32, matrix)>, TensorConstant{1.0}, Elemwise{mul,no_inplace}.0, <TensorType(float32, matrix)>, TensorConstant{1.0})
1.3% 99.2% 0.085s 3.52e-04s 240 5 Elemwise{Composite{((i0 * i1 * i2) + (i3 * i4))}}(Gemm{no_inplace}.0, <TensorType(int8, col)>, <TensorType(float32, matrix)>, Gemm{no_inplace}.0, Gemm{no_inplace}.0)
0.5% 99.8% 0.035s 1.45e-04s 240 0 Elemwise{mul,no_inplace}(<TensorType(int8, col)>, <TensorType(float32, matrix)>)
0.2% 100.0% 0.016s 6.60e-05s 240 6 Elemwise{mul,no_inplace}(Gemm{no_inplace}.0, Elemwise{Composite{((i0 * i1 * i2) + (i3 * i4))}}.0)
... (remaining 0 Apply instances account for 0.00%(0.00s) of the runtime)
Scan Op profiling ( grad_of_scan_fn )
==================
Message: None
Time in 8 calls of the op (for a total of 240 steps) 1.046483e+01s
Total time spent in calling the VM 1.024789e+01s (97.927%)
Total overhead (computing slices..) 2.169311e-01s (2.073%)
Time in all call to theano.grad() 4.098411e-01s
Class
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Class name>
60.8% 60.8% 6.226s 2.36e-03s C 2640 11 theano.tensor.blas.Gemm
32.6% 93.5% 3.340s 2.32e-03s C 1440 6 theano.tensor.blas.Dot22
6.5% 100.0% 0.667s 1.85e-04s C 3600 15 theano.tensor.elemwise.Elemwise
... (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>
37.8% 37.8% 3.867s 2.30e-03s C 1680 7 Gemm{inplace}
32.6% 70.4% 3.340s 2.32e-03s C 1440 6 Dot22
23.0% 93.5% 2.359s 2.46e-03s C 960 4 Gemm{no_inplace}
3.2% 96.6% 0.323s 2.24e-04s C 1440 6 Elemwise{add,no_inplace}
1.1% 97.7% 0.111s 4.64e-04s C 240 1 Elemwise{Composite{(i0 + (i1 * i2 * i3 * i4) + (i1 * i5 * i2))}}
0.9% 98.7% 0.096s 1.00e-04s C 960 4 Elemwise{mul}
0.9% 99.5% 0.088s 3.65e-04s C 240 1 Elemwise{Composite{(i0 + ((i1 + i2) * i3) + (i4 * i3))}}
0.3% 99.8% 0.029s 6.00e-05s C 480 2 Elemwise{Mul}[(0, 1)]
0.2% 100.0% 0.020s 8.42e-05s C 240 1 Elemwise{Mul}[(0, 2)]
... (remaining 0 Ops account for 0.00%(0.00s) of the runtime)
Apply
------
<% time> <sum %> <apply time> <time per call> <#call> <id> <Apply name>
9.2% 9.2% 0.938s 3.91e-03s 240 26 Gemm{inplace}(Dot22.0, TensorConstant{1.0}, <TensorType(float32, matrix)>, Elemwise{Mul}[(0, 1)].0, TensorConstant{1.0})
7.3% 16.4% 0.744s 3.10e-03s 240 18 Gemm{inplace}(Dot22.0, TensorConstant{1.0}, Elemwise{mul}.0, <TensorType(float32, matrix)>, TensorConstant{1.0})
7.1% 23.6% 0.731s 3.04e-03s 240 17 Dot22(Elemwise{mul}.0, <TensorType(float32, matrix)>)
7.0% 30.6% 0.720s 3.00e-03s 240 0 Gemm{no_inplace}(<TensorType(float32, matrix)>, TensorConstant{1.0}, <TensorType(float32, matrix)>, <TensorType(float32, matrix)>, TensorConstant{1.0})
6.1% 36.7% 0.626s 2.61e-03s 240 19 Dot22(<TensorType(float32, matrix)>, Elemwise{Mul}[(0, 2)].0)
5.7% 42.4% 0.584s 2.43e-03s 240 24 Gemm{inplace}(Dot22.0, TensorConstant{1.0}, <TensorType(float32, matrix)>, Elemwise{Mul}[(0, 1)].0, TensorConstant{1.0})
5.7% 48.1% 0.581s 2.42e-03s 240 10 Dot22(Elemwise{mul}.0, <TensorType(float32, matrix)>)
5.5% 53.6% 0.562s 2.34e-03s 240 2 Gemm{no_inplace}(<TensorType(float32, matrix)>, TensorConstant{1.0}, <TensorType(float32, matrix)>, <TensorType(float32, matrix)>, TensorConstant{1.0})
5.5% 59.1% 0.561s 2.34e-03s 240 1 Gemm{no_inplace}(<TensorType(float32, matrix)>, TensorConstant{1.0}, <TensorType(float32, matrix)>, <TensorType(float32, matrix)>, TensorConstant{1.0})
5.2% 64.3% 0.529s 2.20e-03s 240 14 Dot22(<TensorType(float32, matrix)>, Elemwise{mul}.0)
5.0% 69.3% 0.515s 2.15e-03s 240 4 Gemm{no_inplace}(<TensorType(float32, matrix)>, TensorConstant{1.0}, <TensorType(float32, matrix)>, <TensorType(float32, matrix)>, TensorConstant{1.0})
4.4% 73.7% 0.446s 1.86e-03s 240 25 Gemm{inplace}(Dot22.0, TensorConstant{1.0}, Elemwise{Mul}[(0, 2)].0, <TensorType(float32, matrix)>, TensorConstant{1.0})
4.3% 78.0% 0.440s 1.83e-03s 240 16 Dot22(<TensorType(float32, matrix)>, Elemwise{mul}.0)
4.2% 82.2% 0.434s 1.81e-03s 240 9 Dot22(Elemwise{mul}.0, <TensorType(float32, matrix)>)
4.0% 86.2% 0.411s 1.71e-03s 240 27 Gemm{inplace}(Gemm{inplace}.0, TensorConstant{1.0}, Elemwise{Mul}[(0, 1)].0, <TensorType(float32, matrix)>, TensorConstant{1.0})
3.8% 90.0% 0.391s 1.63e-03s 240 22 Gemm{inplace}(Dot22.0, TensorConstant{1.0}, <TensorType(float32, matrix)>, Elemwise{mul}.0, TensorConstant{1.0})
3.4% 93.5% 0.352s 1.47e-03s 240 21 Gemm{inplace}(Dot22.0, TensorConstant{1.0}, Elemwise{Mul}[(0, 1)].0, <TensorType(float32, matrix)>, TensorConstant{1.0})
1.1% 94.6% 0.111s 4.64e-04s 240 7 Elemwise{Composite{(i0 + (i1 * i2 * i3 * i4) + (i1 * i5 * i2))}}(<TensorType(float32, matrix)>, <TensorType(int8, col)>, Gemm{no_inplace}.0, <TensorType(float32, matrix)>, Gemm{no_inplace}.0, <TensorType(float32, matrix)>)
0.9% 95.4% 0.088s 3.65e-04s 240 31 Elemwise{Composite{(i0 + ((i1 + i2) * i3) + (i4 * i3))}}(<TensorType(float32, matrix)>, Gemm{inplace}.0, Gemm{inplace}.0, <TensorType(int8, col)>, Gemm{inplace}.0)
0.7% 96.2% 0.075s 3.12e-04s 240 30 Elemwise{add,no_inplace}(<TensorType(float32, matrix)>, Gemm{inplace}.0)
... (remaining 12 Apply instances account for 3.85%(0.39s) of the runtime)
Function profiling
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:330
Time in 0 calls to Function.__call__: 0.000000e+00s
Total compile time: 2.309456e+01s
Number of Apply nodes: 536
Theano Optimizer time: 1.634476e+01s
Theano validate time: 2.549255e-01s
Theano Linker time (includes C, CUDA code generation/compiling): 6.595885e+00s
Import time 7.425475e-02s
Time in all call to theano.grad() 4.098411e-01s
Time in all call to theano.grad() 4.098411e-01s
Time in all call to theano.grad() 4.098411e-01s
Function profiling
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:332
Time in 0 calls to Function.__call__: 0.000000e+00s
Total compile time: 8.350207e+00s
Number of Apply nodes: 134
Theano Optimizer time: 1.283817e+00s
Theano validate time: 5.622768e-02s
Theano Linker time (includes C, CUDA code generation/compiling): 7.014123e+00s
Import time 3.218102e-02s
Time in all call to theano.grad() 4.098411e-01s
Time in all call to theano.grad() 4.098411e-01s
Function profiling
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:334
Time in 0 calls to Function.__call__: 0.000000e+00s
Total compile time: 5.702252e+00s
Number of Apply nodes: 152
Theano Optimizer time: 1.817622e+00s
Theano validate time: 7.331610e-02s
Theano Linker time (includes C, CUDA code generation/compiling): 3.821592e+00s
Import time 1.338577e-02s
Time in all call to theano.grad() 4.098411e-01s
Time in all call to theano.grad() 4.098411e-01s
Function profiling
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:336
Time in 0 calls to Function.__call__: 0.000000e+00s
Total compile time: 1.727933e+00s
Number of Apply nodes: 163
Theano Optimizer time: 1.428595e+00s
Theano validate time: 7.478404e-02s
Theano Linker time (includes C, CUDA code generation/compiling): 2.332819e-01s
Import time 0.000000e+00s
Time in all call to theano.grad() 4.098411e-01s
Time in all call to theano.grad() 4.098411e-01s
Function profiling
==================
Message: Sum of all(5) printed profiles at exit excluding Scan op profile.
Time in 8 calls to Function.__call__: 1.477107e+02s
Time in Function.fn.__call__: 1.474529e+02s (99.825%)
Time in thunks: 1.470501e+02s (99.553%)
Total compile time: 1.766070e+02s
Number of Apply nodes: 525
Theano Optimizer time: 4.273788e+01s
Theano validate time: 7.118421e-01s
Theano Linker time (includes C, CUDA code generation/compiling): 1.333998e+02s
Import time 4.955947e-01s
Time in all call to theano.grad() 4.098411e-01s
Class
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Class name>
64.7% 64.7% 95.128s 9.91e-01s C 96 12 theano.tensor.blas.Dot22
11.5% 76.2% 16.857s 1.05e+00s Py 16 2 theano.scan_module.scan_op.Scan
9.5% 85.6% 13.954s 1.74e+00s C 8 1 theano.tensor.nnet.nnet.Softmax
7.9% 93.6% 11.631s 5.37e-03s C 2168 271 theano.tensor.elemwise.Elemwise
2.1% 95.7% 3.089s 3.86e-01s C 8 1 theano.tensor.blas.Dot22Scalar
1.6% 97.3% 2.389s 3.32e-02s C 72 9 theano.tensor.elemwise.Sum
1.4% 98.7% 2.029s 2.54e-01s Py 8 1 theano.tensor.subtensor.AdvancedIncSubtensor
0.4% 99.1% 0.618s 7.72e-02s C 8 1 theano.tensor.nnet.nnet.SoftmaxGrad
0.3% 99.4% 0.454s 1.89e-02s Py 24 3 theano.tensor.subtensor.AdvancedSubtensor
0.3% 99.7% 0.453s 1.95e-03s C 232 29 theano.tensor.basic.Reshape
0.2% 99.9% 0.278s 2.32e-03s C 120 15 theano.tensor.basic.Alloc
0.1% 100.0% 0.164s 2.28e-03s C 72 9 theano.tensor.subtensor.IncSubtensor
0.0% 100.0% 0.002s 2.15e-04s Py 8 1 theano.tensor.basic.Nonzero
0.0% 100.0% 0.001s 2.26e-06s C 424 53 theano.compile.ops.Shape_i
0.0% 100.0% 0.001s 3.05e-06s C 264 33 theano.tensor.elemwise.DimShuffle
0.0% 100.0% 0.001s 4.06e-06s C 184 23 theano.tensor.opt.MakeVector
0.0% 100.0% 0.001s 3.71e-06s C 200 25 theano.tensor.subtensor.Subtensor
0.0% 100.0% 0.000s 1.16e-06s C 248 31 theano.tensor.basic.ScalarFromTensor
0.0% 100.0% 0.000s 2.26e-05s C 8 1 theano.tensor.basic.Join
0.0% 100.0% 0.000s 2.40e-06s C 24 3 theano.tensor.basic.Flatten
... (remaining 1 Classes account for 0.00%(0.00s) of the runtime)
Ops
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Op name>
64.7% 64.7% 95.128s 9.91e-01s C 96 12 Dot22
9.5% 74.2% 13.954s 1.74e+00s C 8 1 Softmax
7.1% 81.3% 10.468s 1.31e+00s Py 8 1 forall_inplace,cpu,grad_of_scan_fn}
5.4% 86.7% 7.876s 9.85e-01s C 8 1 Elemwise{Composite{(i0 * log((i1 / i2)))}}
4.3% 91.0% 6.388s 7.99e-01s Py 8 1 forall_inplace,cpu,scan_fn}
2.1% 93.1% 3.089s 3.86e-01s C 8 1 Dot22Scalar
1.4% 94.5% 2.029s 2.54e-01s Py 8 1 AdvancedIncSubtensor{inplace=False, set_instead_of_inc=False}
1.3% 95.8% 1.907s 4.77e-02s C 40 5 Sum{axis=[0, 1], acc_dtype=float64}
1.2% 97.0% 1.788s 2.24e-01s C 8 1 Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((i3 * i4 * i5) / i6) + i7))}}
0.4% 97.4% 0.618s 7.72e-02s C 8 1 SoftmaxGrad
0.4% 97.8% 0.557s 1.51e-03s C 368 46 Elemwise{add,no_inplace}
0.3% 98.1% 0.482s 2.01e-02s C 24 3 Sum{axis=[1], acc_dtype=float64}
0.3% 98.4% 0.454s 1.89e-02s Py 24 3 AdvancedSubtensor
0.3% 98.7% 0.453s 2.83e-03s C 160 20 Reshape{2}
0.3% 99.0% 0.418s 5.22e-03s C 80 10 Elemwise{mul,no_inplace}
0.2% 99.2% 0.318s 3.98e-02s C 8 1 Elemwise{clip,no_inplace}
0.2% 99.4% 0.278s 2.32e-03s C 120 15 Alloc
0.2% 99.6% 0.262s 8.19e-03s C 32 4 Elemwise{Composite{(i0 - ((i1 * i2) / sqrt((i3 + i4 + i5))))}}[(0, 0)]
0.1% 99.7% 0.206s 1.84e-03s C 112 14 Elemwise{Composite{(i0 * sqr(i1))}}
0.1% 99.8% 0.101s 3.15e-03s C 32 4 IncSubtensor{Inc;:int64:}
... (remaining 98 Ops account for 0.19%(0.27s) of the runtime)
Apply
------
<% time> <sum %> <apply time> <time per call> <#call> <id> <Apply name>
11.8% 11.8% 17.349s 2.17e+00s 8 92 Dot22(Reshape{2}.0, Reshape{2}.0)
9.5% 21.3% 13.954s 1.74e+00s 8 397 Softmax(Reshape{2}.0)
9.1% 30.4% 13.341s 1.67e+00s 8 489 Dot22(InplaceDimShuffle{1,0}.0, Reshape{2}.0)
7.7% 38.1% 11.309s 1.41e+00s 8 93 Dot22(Reshape{2}.0, Reshape{2}.0)
7.1% 45.2% 10.468s 1.31e+00s 8 446 forall_inplace,cpu,grad_of_scan_fn}(Elemwise{Composite{minimum(minimum(minimum(minimum(i0, i1), i2), i3), i4)}}.0, Elemwise{mul,no_inplace}.0, Elemwise{mul,no_inplace}.0, InplaceDimShuffle{0,2,1}.0, Subtensor{int64:int64:int64}.0, Subtensor{int64:int64:int64}.0, Subtensor{int64:int64:int64}.0, Subtensor{int64:int64:int64}.0, Subtensor{int64:int64:int64}.0, Subtensor{int64:int64:int64}.0, Subtensor{::int64}.0, Alloc.0, Alloc.0, Alloc.0, Alloc.0, Elem
6.6% 51.8% 9.739s 1.22e+00s 8 94 Dot22(Reshape{2}.0, Reshape{2}.0)
6.5% 58.2% 9.491s 1.19e+00s 8 91 Dot22(Reshape{2}.0, Reshape{2}.0)
5.6% 63.9% 8.252s 1.03e+00s 8 491 Dot22(InplaceDimShuffle{1,0}.0, Reshape{2}.0)
5.6% 69.5% 8.230s 1.03e+00s 8 433 Dot22(InplaceDimShuffle{1,0}.0, Reshape{2}.0)
5.4% 74.8% 7.876s 9.85e-01s 8 411 Elemwise{Composite{(i0 * log((i1 / i2)))}}(AdvancedSubtensor.0, Elemwise{clip,no_inplace}.0, InplaceDimShuffle{0,x}.0)
4.3% 79.2% 6.388s 7.99e-01s 8 337 forall_inplace,cpu,scan_fn}(Elemwise{Composite{minimum(minimum(minimum(minimum(i0, i1), i2), i3), i4)}}.0, Subtensor{int64:int64:int8}.0, Subtensor{int64:int64:int8}.0, Subtensor{int64:int64:int8}.0, Subtensor{int64:int64:int8}.0, Subtensor{int64:int64:int8}.0, Alloc.0, Alloc.0, <TensorType(float32, matrix)>, <TensorType(float32, matrix)>, <TensorType(float32, matrix)>, <TensorType(float32, matrix)>)
4.0% 83.2% 5.945s 7.43e-01s 8 485 Dot22(InplaceDimShuffle{1,0}.0, Reshape{2}.0)
3.0% 86.2% 4.401s 5.50e-01s 8 493 Dot22(InplaceDimShuffle{1,0}.0, Reshape{2}.0)
2.1% 88.3% 3.094s 3.87e-01s 8 486 Dot22(Flatten{2}.0, Reshape{2}.0)
2.1% 90.4% 3.089s 3.86e-01s 8 487 Dot22Scalar(Flatten{2}.0, Reshape{2}.0, TensorConstant{0.0010000000475})
1.4% 91.8% 2.092s 2.62e-01s 8 434 Dot22(Reshape{2}.0, InplaceDimShuffle{1,0}.0)
1.4% 93.2% 2.029s 2.54e-01s 8 425 AdvancedIncSubtensor{inplace=False, set_instead_of_inc=False}(Alloc.0, Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((i3 * i4 * i5) / i6) + i7))}}.0, Subtensor{int64}.0, Subtensor{int64}.0)
1.3% 94.5% 1.885s 2.36e-01s 8 370 Dot22(Reshape{2}.0, Reshape{2}.0)
1.2% 95.7% 1.788s 2.24e-01s 8 424 Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((i3 * i4 * i5) / i6) + i7))}}(AdvancedSubtensor.0, TensorConstant{(1, 1) of 1e-07}, TensorConstant{(1, 1) of 1.0}, TensorConstant{(1, 1) of -1.0}, InplaceDimShuffle{0,x}.0, AdvancedSubtensor.0, Elemwise{clip,no_inplace}.0, Elemwise{true_div,no_inplace}.0)
1.1% 96.8% 1.561s 1.95e-01s 8 430 Sum{axis=[0, 1], acc_dtype=float64}(InplaceDimShuffle{1,0,2}.0)
... (remaining 505 Apply instances account for 3.24%(4.77s) of the runtime)
training time 148.004755974
Raw
linear.log
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:328
Time in 8 calls to Function.__call__: 1.782110e+01s
Time in Function.fn.__call__: 1.747024e+01s (98.031%)
Time in thunks: 1.711180e+01s (96.020%)
Total compile time: 2.021389e+01s
Number of Apply nodes: 530
Theano Optimizer time: 1.830422e+01s
Theano validate time: 2.977979e-01s
Theano Linker time (includes C, CUDA code generation/compiling): 1.739421e+00s
Import time 1.584115e-01s
Time in all call to theano.grad() 4.073255e-01s
Class
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Class name>
60.1% 60.1% 10.281s 6.15e-03s C 1672 209 theano.tensor.elemwise.Elemwise
12.1% 72.2% 2.077s 2.60e-01s Py 8 1 theano.tensor.subtensor.AdvancedIncSubtensor
5.1% 77.3% 0.874s 5.47e-02s Py 16 2 theano.scan_module.scan_op.Scan
4.2% 81.6% 0.725s 7.56e-03s C 96 12 theano.sandbox.cuda.blas.GpuDot22
3.7% 85.3% 0.634s 2.64e-02s C 24 3 theano.sandbox.cuda.basic_ops.HostFromGpu
3.6% 88.8% 0.610s 7.63e-02s C 8 1 theano.tensor.nnet.nnet.SoftmaxGrad
2.8% 91.6% 0.482s 1.51e-02s C 32 4 theano.tensor.elemwise.Sum
2.6% 94.2% 0.444s 1.85e-02s Py 24 3 theano.tensor.subtensor.AdvancedSubtensor
2.4% 96.7% 0.417s 1.74e-02s C 24 3 theano.sandbox.cuda.basic_ops.GpuFromHost
1.0% 97.7% 0.168s 1.05e-02s C 16 2 theano.tensor.basic.Alloc
0.7% 98.4% 0.124s 5.76e-04s C 216 27 theano.sandbox.cuda.basic_ops.GpuReshape
0.7% 99.1% 0.124s 2.45e-04s C 504 63 theano.sandbox.cuda.basic_ops.GpuElemwise
0.3% 99.4% 0.055s 1.38e-03s C 40 5 theano.sandbox.cuda.basic_ops.GpuCAReduce
0.2% 99.6% 0.028s 3.87e-04s C 72 9 theano.sandbox.cuda.basic_ops.GpuIncSubtensor
0.2% 99.8% 0.027s 2.62e-04s C 104 13 theano.sandbox.cuda.basic_ops.GpuAlloc
0.1% 99.8% 0.015s 1.92e-03s C 8 1 theano.sandbox.cuda.nnet.GpuSoftmax
0.1% 99.9% 0.012s 1.47e-03s C 8 1 theano.sandbox.cuda.blas.GpuDot22Scalar
0.1% 100.0% 0.009s 1.16e-03s Py 8 1 theano.sandbox.cuda.basic_ops.GpuFlatten
0.0% 100.0% 0.002s 2.42e-04s Py 8 1 theano.tensor.basic.Nonzero
0.0% 100.0% 0.001s 2.13e-06s C 416 52 theano.compile.ops.Shape_i
... (remaining 10 Classes account for 0.02%(0.00s) of the runtime)
Ops
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Op name>
46.1% 46.1% 7.889s 9.86e-01s C 8 1 Elemwise{Composite{(i0 * log((i1 / i2)))}}
12.1% 58.2% 2.077s 2.60e-01s Py 8 1 AdvancedIncSubtensor{inplace=False, set_instead_of_inc=False}
10.4% 68.7% 1.783s 2.23e-01s C 8 1 Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((i3 * i4 * i5) / i6) + i7))}}
4.2% 72.9% 0.725s 7.56e-03s C 96 12 GpuDot22
4.1% 77.0% 0.704s 8.80e-02s Py 8 1 forall_inplace,gpu,grad_of_scan_fn}
3.7% 80.7% 0.634s 2.64e-02s C 24 3 HostFromGpu
3.6% 84.3% 0.610s 7.63e-02s C 8 1 SoftmaxGrad
2.8% 87.1% 0.482s 2.01e-02s C 24 3 Sum{axis=[1], acc_dtype=float64}
2.6% 89.7% 0.444s 1.85e-02s Py 24 3 AdvancedSubtensor
2.4% 92.1% 0.417s 1.74e-02s C 24 3 GpuFromHost
1.8% 94.0% 0.315s 3.94e-02s C 8 1 Elemwise{clip,no_inplace}
1.7% 95.7% 0.290s 1.21e-02s C 24 3 Elemwise{mul,no_inplace}
1.0% 96.7% 0.170s 2.13e-02s Py 8 1 forall_inplace,gpu,scan_fn}
1.0% 97.6% 0.168s 1.05e-02s C 16 2 Alloc
0.7% 98.4% 0.124s 8.16e-04s C 152 19 GpuReshape{2}
0.3% 98.7% 0.055s 1.38e-03s C 40 5 GpuCAReduce{add}{1,1,0}
0.2% 98.9% 0.030s 3.46e-04s C 88 11 GpuElemwise{Add}[(0, 0)]
0.2% 99.0% 0.027s 2.78e-04s C 96 12 GpuAlloc{memset_0=True}
0.1% 99.2% 0.025s 2.41e-04s C 104 13 GpuElemwise{Composite{(i0 * sqr(i1))},no_inplace}
0.1% 99.3% 0.022s 3.85e-04s C 56 7 GpuElemwise{Composite{(i0 - ((i1 * i2) / sqrt((i3 + i4 + i5))))},no_inplace}
... (remaining 107 Ops account for 0.71%(0.12s) of the runtime)
Apply
------
<% time> <sum %> <apply time> <time per call> <#call> <id> <Apply name>
46.1% 46.1% 7.889s 9.86e-01s 8 418 Elemwise{Composite{(i0 * log((i1 / i2)))}}(AdvancedSubtensor.0, Elemwise{clip,no_inplace}.0, InplaceDimShuffle{0,x}.0)
12.1% 58.2% 2.077s 2.60e-01s 8 421 AdvancedIncSubtensor{inplace=False, set_instead_of_inc=False}(Alloc.0, Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((i3 * i4 * i5) / i6) + i7))}}.0, Subtensor{int64}.0, Subtensor{int64}.0)
10.4% 68.7% 1.783s 2.23e-01s 8 419 Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((i3 * i4 * i5) / i6) + i7))}}(AdvancedSubtensor.0, TensorConstant{(1, 1) of 1e-07}, TensorConstant{(1, 1) of 1.0}, TensorConstant{(1, 1) of -1.0}, InplaceDimShuffle{0,x}.0, AdvancedSubtensor.0, Elemwise{clip,no_inplace}.0, Elemwise{true_div,no_inplace}.0)
4.1% 72.8% 0.704s 8.80e-02s 8 451 forall_inplace,gpu,grad_of_scan_fn}(Elemwise{Composite{minimum(minimum(minimum(minimum(i0, i1), i2), i3), i4)}}.0, GpuElemwise{Mul}[(0, 1)].0, GpuElemwise{mul,no_inplace}.0, GpuDimShuffle{0,2,1}.0, GpuSubtensor{int64:int64:int64}.0, GpuSubtensor{int64:int64:int64}.0, GpuSubtensor{int64:int64:int64}.0, GpuSubtensor{int64:int64:int64}.0, Subtensor{int64:int64:int64}.0, GpuSubtensor{int64:int64:int64}.0, GpuSubtensor{::int64}.0, GpuAlloc{memset_0=True}
3.6% 76.3% 0.610s 7.63e-02s 8 425 SoftmaxGrad(Reshape{2}.0, HostFromGpu.0)
1.9% 78.2% 0.319s 3.99e-02s 8 406 HostFromGpu(GpuReshape{3}.0)
1.8% 80.0% 0.315s 3.94e-02s 8 414 Elemwise{clip,no_inplace}(AdvancedSubtensor.0, TensorConstant{(1, 1) of 1e-07}, TensorConstant{(1, 1) of 1.0})
1.8% 81.9% 0.314s 3.93e-02s 8 401 HostFromGpu(GpuSoftmax.0)
1.7% 83.6% 0.290s 3.63e-02s 8 150 Elemwise{mul,no_inplace}(InplaceDimShuffle{0,x}.0, AdvancedSubtensor.0)
1.3% 84.9% 0.228s 2.84e-02s 8 75 AdvancedSubtensor(y, Subtensor{int64}.0, Subtensor{int64}.0)
1.3% 86.2% 0.216s 2.70e-02s 8 412 AdvancedSubtensor(HostFromGpu.0, Subtensor{int64}.0, Subtensor{int64}.0)
1.2% 87.4% 0.209s 2.61e-02s 8 6 GpuFromHost(<TensorType(float32, 3D)>)
1.2% 88.6% 0.208s 2.60e-02s 8 427 GpuFromHost(SoftmaxGrad.0)
1.0% 89.6% 0.170s 2.13e-02s 8 352 forall_inplace,gpu,scan_fn}(Elemwise{Composite{minimum(minimum(minimum(minimum(i0, i1), i2), i3), i4)}}.0, GpuSubtensor{int64:int64:int8}.0, GpuSubtensor{int64:int64:int8}.0, GpuSubtensor{int64:int64:int8}.0, GpuSubtensor{int64:int64:int8}.0, Subtensor{int64:int64:int8}.0, GpuAlloc{memset_0=True}.0, GpuAlloc{memset_0=True}.0, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(
1.0% 90.6% 0.167s 2.09e-02s 8 413 Alloc(TensorConstant{(1, 1, 1) of 0.0}, Shape_i{0}.0, Shape_i{1}.0, Shape_i{2}.0)
0.9% 91.5% 0.161s 2.01e-02s 8 420 Sum{axis=[1], acc_dtype=float64}(Elemwise{Composite{(i0 * log((i1 / i2)))}}.0)
0.9% 92.5% 0.161s 2.01e-02s 8 173 Sum{axis=[1], acc_dtype=float64}(Elemwise{mul,no_inplace}.0)
0.9% 93.4% 0.161s 2.01e-02s 8 415 Sum{axis=[1], acc_dtype=float64}(Elemwise{clip,no_inplace}.0)
0.4% 93.8% 0.066s 8.26e-03s 8 94 GpuDot22(GpuReshape{2}.0, GpuReshape{2}.0)
0.4% 94.2% 0.065s 8.16e-03s 8 435 GpuDot22(GpuReshape{2}.0, GpuDimShuffle{1,0}.0)
... (remaining 510 Apply instances account for 5.84%(1.00s) of the runtime)
Scan Op profiling ( scan_fn )
==================
Message: None
Time in 8 calls of the op (for a total of 240 steps) 1.655595e-01s
Total time spent in calling the VM 1.615922e-01s (97.604%)
Total overhead (computing slices..) 3.967285e-03s (2.396%)
Time in all call to theano.grad() 4.073255e-01s
Class
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Class name>
81.1% 81.1% 0.129s 1.35e-04s C 960 4 theano.sandbox.cuda.blas.GpuGemm
13.8% 95.0% 0.022s 3.06e-05s C 720 3 theano.sandbox.cuda.basic_ops.GpuElemwise
4.9% 99.8% 0.008s 3.23e-05s C 240 1 theano.sandbox.cuda.basic_ops.GpuFromHost
0.2% 100.0% 0.000s 1.12e-06s C 240 1 theano.tensor.elemwise.Elemwise
... (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>
81.1% 81.1% 0.129s 1.35e-04s C 960 4 GpuGemm{no_inplace}
8.6% 89.7% 0.014s 2.84e-05s C 480 2 GpuElemwise{mul,no_inplace}
5.3% 95.0% 0.008s 3.50e-05s C 240 1 GpuElemwise{Composite{((i0 * (i1 * i2)) + (i3 * i4))},no_inplace}
4.9% 99.8% 0.008s 3.23e-05s C 240 1 GpuFromHost
0.2% 100.0% 0.000s 1.12e-06s C 240 1 Elemwise{Cast{float32}}
... (remaining 0 Ops account for 0.00%(0.00s) of the runtime)
Apply
------
<% time> <sum %> <apply time> <time per call> <#call> <id> <Apply name>
20.7% 20.7% 0.033s 1.38e-04s 240 3 GpuGemm{no_inplace}(<CudaNdarrayType(float32, matrix)>, TensorConstant{1.0}, GpuElemwise{mul,no_inplace}.0, <CudaNdarrayType(float32, matrix)>, TensorConstant{1.0})
20.5% 41.2% 0.033s 1.36e-04s 240 6 GpuGemm{no_inplace}(<CudaNdarrayType(float32, matrix)>, TensorConstant{1.0}, GpuElemwise{mul,no_inplace}.0, <CudaNdarrayType(float32, matrix)>, TensorConstant{1.0})
20.0% 61.2% 0.032s 1.33e-04s 240 5 GpuGemm{no_inplace}(<CudaNdarrayType(float32, matrix)>, TensorConstant{1.0}, GpuElemwise{mul,no_inplace}.0, <CudaNdarrayType(float32, matrix)>, TensorConstant{1.0})
19.9% 81.1% 0.032s 1.32e-04s 240 4 GpuGemm{no_inplace}(<CudaNdarrayType(float32, matrix)>, TensorConstant{1.0}, GpuElemwise{mul,no_inplace}.0, <CudaNdarrayType(float32, matrix)>, TensorConstant{1.0})
5.3% 86.4% 0.008s 3.50e-05s 240 7 GpuElemwise{Composite{((i0 * (i1 * i2)) + (i3 * i4))},no_inplace}(GpuGemm{no_inplace}.0, GpuFromHost.0, <CudaNdarrayType(float32, matrix)>, GpuGemm{no_inplace}.0, GpuGemm{no_inplace}.0)
5.2% 91.6% 0.008s 3.45e-05s 240 2 GpuElemwise{mul,no_inplace}(GpuFromHost.0, <CudaNdarrayType(float32, matrix)>)
4.9% 96.5% 0.008s 3.23e-05s 240 1 GpuFromHost(Elemwise{Cast{float32}}.0)
3.4% 99.8% 0.005s 2.23e-05s 240 8 GpuElemwise{mul,no_inplace}(GpuGemm{no_inplace}.0, GpuElemwise{Composite{((i0 * (i1 * i2)) + (i3 * i4))},no_inplace}.0)
0.2% 100.0% 0.000s 1.12e-06s 240 0 Elemwise{Cast{float32}}(<TensorType(int8, col)>)
... (remaining 0 Apply instances account for 0.00%(0.00s) of the runtime)
Scan Op profiling ( grad_of_scan_fn )
==================
Message: None
Time in 8 calls of the op (for a total of 240 steps) 6.921008e-01s
Total time spent in calling the VM 6.207879e-01s (89.696%)
Total overhead (computing slices..) 7.131290e-02s (10.304%)
Time in all call to theano.grad() 4.073255e-01s
Class
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Class name>
53.6% 53.6% 0.329s 1.25e-04s C 2640 11 theano.sandbox.cuda.blas.GpuGemm
24.6% 78.2% 0.151s 1.05e-04s C 1440 6 theano.sandbox.cuda.blas.GpuDot22
20.2% 98.4% 0.124s 3.24e-05s C 3840 16 theano.sandbox.cuda.basic_ops.GpuElemwise
1.5% 99.9% 0.009s 3.94e-05s C 240 1 theano.sandbox.cuda.basic_ops.GpuFromHost
0.1% 100.0% 0.000s 1.36e-06s C 240 1 theano.tensor.elemwise.Elemwise
... (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>
32.2% 32.2% 0.198s 1.18e-04s C 1680 7 GpuGemm{inplace}
24.6% 56.8% 0.151s 1.05e-04s C 1440 6 GpuDot22
21.4% 78.2% 0.131s 1.37e-04s C 960 4 GpuGemm{no_inplace}
8.6% 86.8% 0.053s 3.67e-05s C 1440 6 GpuElemwise{add,no_inplace}
4.2% 91.0% 0.026s 2.69e-05s C 960 4 GpuElemwise{mul,no_inplace}
2.8% 93.7% 0.017s 2.36e-05s C 720 3 GpuElemwise{Mul}[(0, 1)]
1.9% 95.6% 0.012s 4.86e-05s C 240 1 GpuElemwise{Composite{((((i0 + i1) * i2) + (i3 * i2)) + i4)},no_inplace}
1.9% 97.5% 0.012s 4.85e-05s C 240 1 GpuElemwise{Composite{((((i0 * i1) * i2) + ((i3 * i1) * i2)) + i4)},no_inplace}
1.5% 99.1% 0.009s 3.94e-05s C 240 1 GpuFromHost
0.9% 99.9% 0.005s 2.25e-05s C 240 1 GpuElemwise{Mul}[(0, 0)]
0.1% 100.0% 0.000s 1.36e-06s C 240 1 Elemwise{Cast{float32}}
... (remaining 0 Ops account for 0.00%(0.00s) of the runtime)
Apply
------
<% time> <sum %> <apply time> <time per call> <#call> <id> <Apply name>
5.6% 5.6% 0.034s 1.43e-04s 240 4 GpuGemm{no_inplace}(<CudaNdarrayType(float32, matrix)>, TensorConstant{1.0}, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(float32, matrix)>, TensorConstant{1.0})
5.3% 10.9% 0.033s 1.36e-04s 240 5 GpuGemm{no_inplace}(<CudaNdarrayType(float32, matrix)>, TensorConstant{1.0}, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(float32, matrix)>, TensorConstant{1.0})
5.3% 16.2% 0.032s 1.35e-04s 240 2 GpuGemm{no_inplace}(<CudaNdarrayType(float32, matrix)>, TensorConstant{1.0}, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(float32, matrix)>, TensorConstant{1.0})
5.2% 21.4% 0.032s 1.33e-04s 240 1 GpuGemm{no_inplace}(<CudaNdarrayType(float32, matrix)>, TensorConstant{1.0}, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(float32, matrix)>, TensorConstant{1.0})
4.7% 26.1% 0.029s 1.21e-04s 240 28 GpuGemm{inplace}(GpuDot22.0, TensorConstant{1.0}, GpuElemwise{Mul}[(0, 0)].0, <CudaNdarrayType(float32, matrix)>, TensorConstant{1.0})
4.7% 30.8% 0.029s 1.21e-04s 240 18 GpuGemm{inplace}(GpuDot22.0, TensorConstant{1.0}, GpuElemwise{mul,no_inplace}.0, <CudaNdarrayType(float32, matrix)>, TensorConstant{1.0})
4.7% 35.5% 0.029s 1.20e-04s 240 24 GpuGemm{inplace}(GpuDot22.0, TensorConstant{1.0}, GpuElemwise{Mul}[(0, 1)].0, <CudaNdarrayType(float32, matrix)>, TensorConstant{1.0})
4.7% 40.2% 0.029s 1.19e-04s 240 30 GpuGemm{inplace}(GpuGemm{inplace}.0, TensorConstant{1.0}, GpuElemwise{Mul}[(0, 1)].0, <CudaNdarrayType(float32, matrix)>, TensorConstant{1.0})
4.6% 44.7% 0.028s 1.17e-04s 240 10 GpuDot22(GpuElemwise{mul,no_inplace}.0, <CudaNdarrayType(float32, matrix)>)
4.5% 49.2% 0.028s 1.15e-04s 240 21 GpuDot22(GpuElemwise{mul,no_inplace}.0, <CudaNdarrayType(float32, matrix)>)
4.5% 53.7% 0.028s 1.15e-04s 240 25 GpuGemm{inplace}(GpuDot22.0, TensorConstant{1.0}, <CudaNdarrayType(float32, matrix)>, GpuElemwise{mul,no_inplace}.0, TensorConstant{1.0})
4.5% 58.2% 0.027s 1.15e-04s 240 29 GpuGemm{inplace}(GpuDot22.0, TensorConstant{1.0}, <CudaNdarrayType(float32, matrix)>, GpuElemwise{Mul}[(0, 1)].0, TensorConstant{1.0})
4.5% 62.7% 0.027s 1.14e-04s 240 27 GpuGemm{inplace}(GpuDot22.0, TensorConstant{1.0}, <CudaNdarrayType(float32, matrix)>, GpuElemwise{Mul}[(0, 1)].0, TensorConstant{1.0})
4.4% 67.1% 0.027s 1.14e-04s 240 9 GpuDot22(GpuElemwise{mul,no_inplace}.0, <CudaNdarrayType(float32, matrix)>)
3.7% 70.8% 0.023s 9.53e-05s 240 17 GpuDot22(<CudaNdarrayType(float32, matrix)>, GpuElemwise{mul,no_inplace}.0)
3.7% 74.5% 0.023s 9.42e-05s 240 22 GpuDot22(<CudaNdarrayType(float32, matrix)>, GpuElemwise{Mul}[(0, 0)].0)
3.7% 78.2% 0.023s 9.40e-05s 240 20 GpuDot22(<CudaNdarrayType(float32, matrix)>, GpuElemwise{mul,no_inplace}.0)
2.0% 80.2% 0.012s 5.16e-05s 240 31 GpuElemwise{add,no_inplace}(<CudaNdarrayType(float32, matrix)>, GpuGemm{inplace}.0)
2.0% 82.2% 0.012s 5.06e-05s 240 32 GpuElemwise{add,no_inplace}(<CudaNdarrayType(float32, matrix)>, GpuGemm{inplace}.0)
2.0% 84.1% 0.012s 5.06e-05s 240 33 GpuElemwise{add,no_inplace}(<CudaNdarrayType(float32, matrix)>, GpuGemm{inplace}.0)
... (remaining 15 Apply instances account for 15.86%(0.10s) of the runtime)
Function profiling
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:330
Time in 0 calls to Function.__call__: 0.000000e+00s
Total compile time: 2.060873e+01s
Number of Apply nodes: 541
Theano Optimizer time: 1.904841e+01s
Theano validate time: 8.013768e-01s
Theano Linker time (includes C, CUDA code generation/compiling): 1.383735e+00s
Import time 1.165748e-02s
Time in all call to theano.grad() 4.073255e-01s
Time in all call to theano.grad() 4.073255e-01s
Time in all call to theano.grad() 4.073255e-01s
Function profiling
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:332
Time in 0 calls to Function.__call__: 0.000000e+00s
Total compile time: 1.925702e+00s
Number of Apply nodes: 137
Theano Optimizer time: 1.554818e+00s
Theano validate time: 5.621839e-02s
Theano Linker time (includes C, CUDA code generation/compiling): 3.086841e-01s
Import time 7.719040e-03s
Time in all call to theano.grad() 4.073255e-01s
Time in all call to theano.grad() 4.073255e-01s
Function profiling
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:334
Time in 0 calls to Function.__call__: 0.000000e+00s
Total compile time: 2.623655e+00s
Number of Apply nodes: 155
Theano Optimizer time: 2.232972e+00s
Theano validate time: 7.306647e-02s
Theano Linker time (includes C, CUDA code generation/compiling): 3.238871e-01s
Import time 5.110025e-03s
Time in all call to theano.grad() 4.073255e-01s
Time in all call to theano.grad() 4.073255e-01s
Function profiling
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:336
Time in 0 calls to Function.__call__: 0.000000e+00s
Total compile time: 2.094579e+00s
Number of Apply nodes: 166
Theano Optimizer time: 1.699915e+00s
Theano validate time: 7.587528e-02s
Theano Linker time (includes C, CUDA code generation/compiling): 3.269272e-01s
Import time 0.000000e+00s
Time in all call to theano.grad() 4.073255e-01s
Time in all call to theano.grad() 4.073255e-01s
Function profiling
==================
Message: Sum of all(5) printed profiles at exit excluding Scan op profile.
Time in 8 calls to Function.__call__: 1.782110e+01s
Time in Function.fn.__call__: 1.747024e+01s (98.031%)
Time in thunks: 1.711180e+01s (96.020%)
Total compile time: 4.746655e+01s
Number of Apply nodes: 530
Theano Optimizer time: 4.284033e+01s
Theano validate time: 1.304335e+00s
Theano Linker time (includes C, CUDA code generation/compiling): 4.082654e+00s
Import time 1.828980e-01s
Time in all call to theano.grad() 4.073255e-01s
Class
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Class name>
60.1% 60.1% 10.281s 6.15e-03s C 1672 209 theano.tensor.elemwise.Elemwise
12.1% 72.2% 2.077s 2.60e-01s Py 8 1 theano.tensor.subtensor.AdvancedIncSubtensor
5.1% 77.3% 0.874s 5.47e-02s Py 16 2 theano.scan_module.scan_op.Scan
4.2% 81.6% 0.725s 7.56e-03s C 96 12 theano.sandbox.cuda.blas.GpuDot22
3.7% 85.3% 0.634s 2.64e-02s C 24 3 theano.sandbox.cuda.basic_ops.HostFromGpu
3.6% 88.8% 0.610s 7.63e-02s C 8 1 theano.tensor.nnet.nnet.SoftmaxGrad
2.8% 91.6% 0.482s 1.51e-02s C 32 4 theano.tensor.elemwise.Sum
2.6% 94.2% 0.444s 1.85e-02s Py 24 3 theano.tensor.subtensor.AdvancedSubtensor
2.4% 96.7% 0.417s 1.74e-02s C 24 3 theano.sandbox.cuda.basic_ops.GpuFromHost
1.0% 97.7% 0.168s 1.05e-02s C 16 2 theano.tensor.basic.Alloc
0.7% 98.4% 0.124s 5.76e-04s C 216 27 theano.sandbox.cuda.basic_ops.GpuReshape
0.7% 99.1% 0.124s 2.45e-04s C 504 63 theano.sandbox.cuda.basic_ops.GpuElemwise
0.3% 99.4% 0.055s 1.38e-03s C 40 5 theano.sandbox.cuda.basic_ops.GpuCAReduce
0.2% 99.6% 0.028s 3.87e-04s C 72 9 theano.sandbox.cuda.basic_ops.GpuIncSubtensor
0.2% 99.8% 0.027s 2.62e-04s C 104 13 theano.sandbox.cuda.basic_ops.GpuAlloc
0.1% 99.8% 0.015s 1.92e-03s C 8 1 theano.sandbox.cuda.nnet.GpuSoftmax
0.1% 99.9% 0.012s 1.47e-03s C 8 1 theano.sandbox.cuda.blas.GpuDot22Scalar
0.1% 100.0% 0.009s 1.16e-03s Py 8 1 theano.sandbox.cuda.basic_ops.GpuFlatten
0.0% 100.0% 0.002s 2.42e-04s Py 8 1 theano.tensor.basic.Nonzero
0.0% 100.0% 0.001s 2.13e-06s C 416 52 theano.compile.ops.Shape_i
... (remaining 10 Classes account for 0.02%(0.00s) of the runtime)
Ops
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Op name>
46.1% 46.1% 7.889s 9.86e-01s C 8 1 Elemwise{Composite{(i0 * log((i1 / i2)))}}
12.1% 58.2% 2.077s 2.60e-01s Py 8 1 AdvancedIncSubtensor{inplace=False, set_instead_of_inc=False}
10.4% 68.7% 1.783s 2.23e-01s C 8 1 Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((i3 * i4 * i5) / i6) + i7))}}
4.2% 72.9% 0.725s 7.56e-03s C 96 12 GpuDot22
4.1% 77.0% 0.704s 8.80e-02s Py 8 1 forall_inplace,gpu,grad_of_scan_fn}
3.7% 80.7% 0.634s 2.64e-02s C 24 3 HostFromGpu
3.6% 84.3% 0.610s 7.63e-02s C 8 1 SoftmaxGrad
2.8% 87.1% 0.482s 2.01e-02s C 24 3 Sum{axis=[1], acc_dtype=float64}
2.6% 89.7% 0.444s 1.85e-02s Py 24 3 AdvancedSubtensor
2.4% 92.1% 0.417s 1.74e-02s C 24 3 GpuFromHost
1.8% 94.0% 0.315s 3.94e-02s C 8 1 Elemwise{clip,no_inplace}
1.7% 95.7% 0.290s 1.21e-02s C 24 3 Elemwise{mul,no_inplace}
1.0% 96.7% 0.170s 2.13e-02s Py 8 1 forall_inplace,gpu,scan_fn}
1.0% 97.6% 0.168s 1.05e-02s C 16 2 Alloc
0.7% 98.4% 0.124s 8.16e-04s C 152 19 GpuReshape{2}
0.3% 98.7% 0.055s 1.38e-03s C 40 5 GpuCAReduce{add}{1,1,0}
0.2% 98.9% 0.030s 3.46e-04s C 88 11 GpuElemwise{Add}[(0, 0)]
0.2% 99.0% 0.027s 2.78e-04s C 96 12 GpuAlloc{memset_0=True}
0.1% 99.2% 0.025s 2.41e-04s C 104 13 GpuElemwise{Composite{(i0 * sqr(i1))},no_inplace}
0.1% 99.3% 0.022s 3.85e-04s C 56 7 GpuElemwise{Composite{(i0 - ((i1 * i2) / sqrt((i3 + i4 + i5))))},no_inplace}
... (remaining 107 Ops account for 0.71%(0.12s) of the runtime)
Apply
------
<% time> <sum %> <apply time> <time per call> <#call> <id> <Apply name>
46.1% 46.1% 7.889s 9.86e-01s 8 418 Elemwise{Composite{(i0 * log((i1 / i2)))}}(AdvancedSubtensor.0, Elemwise{clip,no_inplace}.0, InplaceDimShuffle{0,x}.0)
12.1% 58.2% 2.077s 2.60e-01s 8 421 AdvancedIncSubtensor{inplace=False, set_instead_of_inc=False}(Alloc.0, Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((i3 * i4 * i5) / i6) + i7))}}.0, Subtensor{int64}.0, Subtensor{int64}.0)
10.4% 68.7% 1.783s 2.23e-01s 8 419 Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((i3 * i4 * i5) / i6) + i7))}}(AdvancedSubtensor.0, TensorConstant{(1, 1) of 1e-07}, TensorConstant{(1, 1) of 1.0}, TensorConstant{(1, 1) of -1.0}, InplaceDimShuffle{0,x}.0, AdvancedSubtensor.0, Elemwise{clip,no_inplace}.0, Elemwise{true_div,no_inplace}.0)
4.1% 72.8% 0.704s 8.80e-02s 8 451 forall_inplace,gpu,grad_of_scan_fn}(Elemwise{Composite{minimum(minimum(minimum(minimum(i0, i1), i2), i3), i4)}}.0, GpuElemwise{Mul}[(0, 1)].0, GpuElemwise{mul,no_inplace}.0, GpuDimShuffle{0,2,1}.0, GpuSubtensor{int64:int64:int64}.0, GpuSubtensor{int64:int64:int64}.0, GpuSubtensor{int64:int64:int64}.0, GpuSubtensor{int64:int64:int64}.0, Subtensor{int64:int64:int64}.0, GpuSubtensor{int64:int64:int64}.0, GpuSubtensor{::int64}.0, GpuAlloc{memset_0=True}
3.6% 76.3% 0.610s 7.63e-02s 8 425 SoftmaxGrad(Reshape{2}.0, HostFromGpu.0)
1.9% 78.2% 0.319s 3.99e-02s 8 406 HostFromGpu(GpuReshape{3}.0)
1.8% 80.0% 0.315s 3.94e-02s 8 414 Elemwise{clip,no_inplace}(AdvancedSubtensor.0, TensorConstant{(1, 1) of 1e-07}, TensorConstant{(1, 1) of 1.0})
1.8% 81.9% 0.314s 3.93e-02s 8 401 HostFromGpu(GpuSoftmax.0)
1.7% 83.6% 0.290s 3.63e-02s 8 150 Elemwise{mul,no_inplace}(InplaceDimShuffle{0,x}.0, AdvancedSubtensor.0)
1.3% 84.9% 0.228s 2.84e-02s 8 75 AdvancedSubtensor(y, Subtensor{int64}.0, Subtensor{int64}.0)
1.3% 86.2% 0.216s 2.70e-02s 8 412 AdvancedSubtensor(HostFromGpu.0, Subtensor{int64}.0, Subtensor{int64}.0)
1.2% 87.4% 0.209s 2.61e-02s 8 6 GpuFromHost(<TensorType(float32, 3D)>)
1.2% 88.6% 0.208s 2.60e-02s 8 427 GpuFromHost(SoftmaxGrad.0)
1.0% 89.6% 0.170s 2.13e-02s 8 352 forall_inplace,gpu,scan_fn}(Elemwise{Composite{minimum(minimum(minimum(minimum(i0, i1), i2), i3), i4)}}.0, GpuSubtensor{int64:int64:int8}.0, GpuSubtensor{int64:int64:int8}.0, GpuSubtensor{int64:int64:int8}.0, GpuSubtensor{int64:int64:int8}.0, Subtensor{int64:int64:int8}.0, GpuAlloc{memset_0=True}.0, GpuAlloc{memset_0=True}.0, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(
1.0% 90.6% 0.167s 2.09e-02s 8 413 Alloc(TensorConstant{(1, 1, 1) of 0.0}, Shape_i{0}.0, Shape_i{1}.0, Shape_i{2}.0)
0.9% 91.5% 0.161s 2.01e-02s 8 420 Sum{axis=[1], acc_dtype=float64}(Elemwise{Composite{(i0 * log((i1 / i2)))}}.0)
0.9% 92.5% 0.161s 2.01e-02s 8 173 Sum{axis=[1], acc_dtype=float64}(Elemwise{mul,no_inplace}.0)
0.9% 93.4% 0.161s 2.01e-02s 8 415 Sum{axis=[1], acc_dtype=float64}(Elemwise{clip,no_inplace}.0)
0.4% 93.8% 0.066s 8.26e-03s 8 94 GpuDot22(GpuReshape{2}.0, GpuReshape{2}.0)
0.4% 94.2% 0.065s 8.16e-03s 8 435 GpuDot22(GpuReshape{2}.0, GpuDimShuffle{1,0}.0)
... (remaining 510 Apply instances account for 5.84%(1.00s) of the runtime)
training time 18.1833021641
Raw
profile.log
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:328
Time in 8 calls to Function.__call__: 1.868278e+01s
Time in Function.fn.__call__: 1.834628e+01s (98.199%)
Time in thunks: 1.794744e+01s (96.064%)
Total compile time: 2.445579e+01s
Number of Apply nodes: 532
Theano Optimizer time: 2.181430e+01s
Theano validate time: 2.990005e-01s
Theano Linker time (includes C, CUDA code generation/compiling): 2.419208e+00s
Import time 1.909029e-01s
Time in all call to theano.grad() 1.037111e+00s
Class
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Class name>
61.6% 61.6% 11.056s 6.61e-03s C 1672 209 theano.tensor.elemwise.Elemwise
11.4% 73.0% 2.045s 2.56e-01s Py 8 1 theano.tensor.subtensor.AdvancedIncSubtensor
5.6% 78.6% 0.999s 6.24e-02s Py 16 2 theano.scan_module.scan_op.Scan
4.2% 82.8% 0.753s 7.84e-03s C 96 12 theano.sandbox.cuda.blas.GpuDot22
3.6% 86.4% 0.647s 2.69e-02s C 24 3 theano.sandbox.cuda.basic_ops.HostFromGpu
3.4% 89.8% 0.613s 7.66e-02s C 8 1 theano.tensor.nnet.nnet.SoftmaxGrad
2.7% 92.5% 0.482s 1.51e-02s C 32 4 theano.tensor.elemwise.Sum
2.4% 94.9% 0.436s 1.82e-02s Py 24 3 theano.tensor.subtensor.AdvancedSubtensor
1.8% 96.7% 0.321s 1.34e-02s C 24 3 theano.sandbox.cuda.basic_ops.GpuFromHost
0.9% 97.6% 0.169s 1.06e-02s C 16 2 theano.tensor.basic.Alloc
0.7% 98.3% 0.129s 5.98e-04s C 216 27 theano.sandbox.cuda.basic_ops.GpuReshape
0.7% 99.1% 0.129s 2.48e-04s C 520 65 theano.sandbox.cuda.basic_ops.GpuElemwise
0.3% 99.4% 0.054s 1.35e-03s C 40 5 theano.sandbox.cuda.basic_ops.GpuCAReduce
0.2% 99.6% 0.044s 6.08e-04s C 72 9 theano.sandbox.cuda.basic_ops.GpuIncSubtensor
0.2% 99.8% 0.028s 2.67e-04s C 104 13 theano.sandbox.cuda.basic_ops.GpuAlloc
0.1% 99.9% 0.016s 1.99e-03s C 8 1 theano.sandbox.cuda.nnet.GpuSoftmax
0.1% 99.9% 0.012s 1.47e-03s C 8 1 theano.sandbox.cuda.blas.GpuDot22Scalar
0.1% 100.0% 0.009s 1.15e-03s Py 8 1 theano.sandbox.cuda.basic_ops.GpuFlatten
0.0% 100.0% 0.002s 2.06e-04s Py 8 1 theano.tensor.basic.Nonzero
0.0% 100.0% 0.001s 1.92e-06s C 416 52 theano.compile.ops.Shape_i
... (remaining 10 Classes account for 0.01%(0.00s) of the runtime)
Ops
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Op name>
48.1% 48.1% 8.638s 1.08e+00s C 8 1 Elemwise{Composite{(i0 * log((i1 / i2)))}}
11.4% 59.5% 2.045s 2.56e-01s Py 8 1 AdvancedIncSubtensor{inplace=False, set_instead_of_inc=False}
10.0% 69.5% 1.796s 2.24e-01s C 8 1 Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((i3 * i4 * i5) / i6) + i7))}}
4.6% 74.1% 0.826s 1.03e-01s Py 8 1 forall_inplace,gpu,grad_of_scan_fn}
4.2% 78.3% 0.753s 7.84e-03s C 96 12 GpuDot22
3.6% 81.9% 0.647s 2.69e-02s C 24 3 HostFromGpu
3.4% 85.3% 0.613s 7.66e-02s C 8 1 SoftmaxGrad
2.7% 88.0% 0.482s 2.01e-02s C 24 3 Sum{axis=[1], acc_dtype=float64}
2.4% 90.5% 0.436s 1.82e-02s Py 24 3 AdvancedSubtensor
1.8% 92.3% 0.321s 1.34e-02s C 24 3 GpuFromHost
1.8% 94.0% 0.317s 3.96e-02s C 8 1 Elemwise{clip,no_inplace}
1.7% 95.7% 0.302s 1.26e-02s C 24 3 Elemwise{mul,no_inplace}
1.0% 96.7% 0.173s 2.17e-02s Py 8 1 forall_inplace,gpu,scan_fn}
0.9% 97.6% 0.169s 1.06e-02s C 16 2 Alloc
0.7% 98.3% 0.129s 8.48e-04s C 152 19 GpuReshape{2}
0.3% 98.6% 0.054s 1.35e-03s C 40 5 GpuCAReduce{add}{1,1,0}
0.2% 98.8% 0.035s 1.11e-03s C 32 4 GpuIncSubtensor{Inc;:int64:}
0.2% 99.0% 0.030s 3.46e-04s C 88 11 GpuElemwise{Add}[(0, 0)]
0.2% 99.1% 0.027s 2.84e-04s C 96 12 GpuAlloc{memset_0=True}
0.1% 99.3% 0.025s 2.39e-04s C 104 13 GpuElemwise{Composite{(i0 * sqr(i1))},no_inplace}
... (remaining 109 Ops account for 0.71%(0.13s) of the runtime)
Apply
------
<% time> <sum %> <apply time> <time per call> <#call> <id> <Apply name>
48.1% 48.1% 8.638s 1.08e+00s 8 424 Elemwise{Composite{(i0 * log((i1 / i2)))}}(AdvancedSubtensor.0, Elemwise{clip,no_inplace}.0, InplaceDimShuffle{0,x}.0)
11.4% 59.5% 2.045s 2.56e-01s 8 427 AdvancedIncSubtensor{inplace=False, set_instead_of_inc=False}(Alloc.0, Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((i3 * i4 * i5) / i6) + i7))}}.0, Subtensor{int64}.0, Subtensor{int64}.0)
10.0% 69.5% 1.796s 2.24e-01s 8 425 Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((i3 * i4 * i5) / i6) + i7))}}(AdvancedSubtensor.0, TensorConstant{(1, 1) of 1e-07}, TensorConstant{(1, 1) of 1.0}, TensorConstant{(1, 1) of -1.0}, InplaceDimShuffle{0,x}.0, AdvancedSubtensor.0, Elemwise{clip,no_inplace}.0, Elemwise{true_div,no_inplace}.0)
4.6% 74.1% 0.826s 1.03e-01s 8 453 forall_inplace,gpu,grad_of_scan_fn}(Elemwise{Composite{minimum(minimum(minimum(minimum(i0, i1), i2), i3), i4)}}.0, GpuElemwise{mul,no_inplace}.0, GpuElemwise{mul,no_inplace}.0, GpuElemwise{Tanh}[(0, 0)].0, GpuDimShuffle{0,2,1}.0, GpuElemwise{Composite{(i0 - sqr(i1))},no_inplace}.0, GpuSubtensor{int64:int64:int64}.0, GpuSubtensor{int64:int64:int64}.0, GpuSubtensor{int64:int64:int64}.0, GpuSubtensor{int64:int64:int64}.0, Subtensor{int64:int64:int64}.0
3.4% 77.5% 0.613s 7.66e-02s 8 431 SoftmaxGrad(Reshape{2}.0, HostFromGpu.0)
1.8% 79.4% 0.326s 4.07e-02s 8 406 HostFromGpu(GpuReshape{3}.0)
1.8% 81.2% 0.320s 4.01e-02s 8 401 HostFromGpu(GpuSoftmax.0)
1.8% 82.9% 0.317s 3.96e-02s 8 416 Elemwise{clip,no_inplace}(AdvancedSubtensor.0, TensorConstant{(1, 1) of 1e-07}, TensorConstant{(1, 1) of 1.0})
1.7% 84.6% 0.302s 3.77e-02s 8 150 Elemwise{mul,no_inplace}(InplaceDimShuffle{0,x}.0, AdvancedSubtensor.0)
1.2% 85.8% 0.218s 2.73e-02s 8 75 AdvancedSubtensor(y, Subtensor{int64}.0, Subtensor{int64}.0)
1.2% 87.0% 0.218s 2.72e-02s 8 412 AdvancedSubtensor(HostFromGpu.0, Subtensor{int64}.0, Subtensor{int64}.0)
1.0% 88.0% 0.173s 2.17e-02s 8 352 forall_inplace,gpu,scan_fn}(Elemwise{Composite{minimum(minimum(minimum(minimum(i0, i1), i2), i3), i4)}}.0, GpuSubtensor{int64:int64:int8}.0, GpuSubtensor{int64:int64:int8}.0, GpuSubtensor{int64:int64:int8}.0, GpuSubtensor{int64:int64:int8}.0, Subtensor{int64:int64:int8}.0, GpuAlloc{memset_0=True}.0, GpuAlloc{memset_0=True}.0, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(
0.9% 88.9% 0.169s 2.11e-02s 8 413 Alloc(TensorConstant{(1, 1, 1) of 0.0}, Shape_i{0}.0, Shape_i{1}.0, Shape_i{2}.0)
0.9% 89.9% 0.166s 2.07e-02s 8 6 GpuFromHost(<TensorType(float32, 3D)>)
0.9% 90.8% 0.161s 2.01e-02s 8 173 Sum{axis=[1], acc_dtype=float64}(Elemwise{mul,no_inplace}.0)
0.9% 91.6% 0.161s 2.01e-02s 8 426 Sum{axis=[1], acc_dtype=float64}(Elemwise{Composite{(i0 * log((i1 / i2)))}}.0)
0.9% 92.5% 0.161s 2.01e-02s 8 420 Sum{axis=[1], acc_dtype=float64}(Elemwise{clip,no_inplace}.0)
0.9% 93.4% 0.155s 1.93e-02s 8 433 GpuFromHost(SoftmaxGrad.0)
0.5% 93.9% 0.092s 1.15e-02s 8 91 GpuDot22(GpuReshape{2}.0, GpuReshape{2}.0)
0.4% 94.3% 0.066s 8.25e-03s 8 94 GpuDot22(GpuReshape{2}.0, GpuReshape{2}.0)
... (remaining 512 Apply instances account for 5.72%(1.03s) of the runtime)
Scan Op profiling ( scan_fn )
==================
Message: None
Time in 8 calls of the op (for a total of 240 steps) 1.686399e-01s
Total time spent in calling the VM 1.647773e-01s (97.710%)
Total overhead (computing slices..) 3.862619e-03s (2.290%)
Time in all call to theano.grad() 1.037111e+00s
Class
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Class name>
79.9% 79.9% 0.130s 1.35e-04s C 960 4 theano.sandbox.cuda.blas.GpuGemm
15.2% 95.1% 0.025s 3.44e-05s C 720 3 theano.sandbox.cuda.basic_ops.GpuElemwise
4.7% 99.8% 0.008s 3.18e-05s C 240 1 theano.sandbox.cuda.basic_ops.GpuFromHost
0.2% 100.0% 0.000s 1.09e-06s C 240 1 theano.tensor.elemwise.Elemwise
... (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>
79.9% 79.9% 0.130s 1.35e-04s C 960 4 GpuGemm{no_inplace}
5.9% 85.8% 0.010s 4.00e-05s C 240 1 GpuElemwise{Composite{((clip((i0 + i1), i2, i3) * (i4 * i5)) + (clip((i0 + i6), i2, i3) * tanh(i7)))},no_inplace}
5.2% 91.0% 0.008s 3.50e-05s C 240 1 GpuElemwise{mul,no_inplace}
4.7% 95.7% 0.008s 3.18e-05s C 240 1 GpuFromHost
4.2% 99.8% 0.007s 2.81e-05s C 240 1 GpuElemwise{Composite{(clip((i0 + i1), i2, i3) * tanh(i4))},no_inplace}
0.2% 100.0% 0.000s 1.09e-06s C 240 1 Elemwise{Cast{float32}}
... (remaining 0 Ops account for 0.00%(0.00s) of the runtime)
Apply
------
<% time> <sum %> <apply time> <time per call> <#call> <id> <Apply name>
20.6% 20.6% 0.033s 1.39e-04s 240 3 GpuGemm{no_inplace}(<CudaNdarrayType(float32, matrix)>, TensorConstant{1.0}, GpuElemwise{mul,no_inplace}.0, <CudaNdarrayType(float32, matrix)>, TensorConstant{1.0})
20.1% 40.7% 0.033s 1.36e-04s 240 6 GpuGemm{no_inplace}(<CudaNdarrayType(float32, matrix)>, TensorConstant{0.20000000298}, GpuElemwise{mul,no_inplace}.0, <CudaNdarrayType(float32, matrix)>, TensorConstant{0.20000000298})
19.7% 60.4% 0.032s 1.33e-04s 240 5 GpuGemm{no_inplace}(<CudaNdarrayType(float32, matrix)>, TensorConstant{0.20000000298}, GpuElemwise{mul,no_inplace}.0, <CudaNdarrayType(float32, matrix)>, TensorConstant{0.20000000298})
19.5% 79.9% 0.032s 1.32e-04s 240 4 GpuGemm{no_inplace}(<CudaNdarrayType(float32, matrix)>, TensorConstant{0.20000000298}, GpuElemwise{mul,no_inplace}.0, <CudaNdarrayType(float32, matrix)>, TensorConstant{0.20000000298})
5.9% 85.8% 0.010s 4.00e-05s 240 7 GpuElemwise{Composite{((clip((i0 + i1), i2, i3) * (i4 * i5)) + (clip((i0 + i6), i2, i3) * tanh(i7)))},no_inplace}(CudaNdarrayConstant{[[ 0.5]]}, GpuGemm{no_inplace}.0, CudaNdarrayConstant{[[ 0.]]}, CudaNdarrayConstant{[[ 1.]]}, GpuFromHost.0, <CudaNdarrayType(float32, matrix)>, GpuGemm{no_inplace}.0, GpuGemm{no_inplace}.0)
5.2% 91.0% 0.008s 3.50e-05s 240 2 GpuElemwise{mul,no_inplace}(GpuFromHost.0, <CudaNdarrayType(float32, matrix)>)
4.7% 95.7% 0.008s 3.18e-05s 240 1 GpuFromHost(Elemwise{Cast{float32}}.0)
4.2% 99.8% 0.007s 2.81e-05s 240 8 GpuElemwise{Composite{(clip((i0 + i1), i2, i3) * tanh(i4))},no_inplace}(CudaNdarrayConstant{[[ 0.5]]}, GpuGemm{no_inplace}.0, CudaNdarrayConstant{[[ 0.]]}, CudaNdarrayConstant{[[ 1.]]}, GpuElemwise{Composite{((clip((i0 + i1), i2, i3) * (i4 * i5)) + (clip((i0 + i6), i2, i3) * tanh(i7)))},no_inplace}.0)
0.2% 100.0% 0.000s 1.09e-06s 240 0 Elemwise{Cast{float32}}(<TensorType(int8, col)>)
... (remaining 0 Apply instances account for 0.00%(0.00s) of the runtime)
Scan Op profiling ( grad_of_scan_fn )
==================
Message: None
Time in 8 calls of the op (for a total of 240 steps) 8.120770e-01s
Total time spent in calling the VM 6.931341e-01s (85.353%)
Total overhead (computing slices..) 1.189430e-01s (14.647%)
Time in all call to theano.grad() 1.037111e+00s
Class
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Class name>
49.0% 49.0% 0.336s 1.27e-04s C 2640 11 theano.sandbox.cuda.blas.GpuGemm
27.1% 76.1% 0.185s 2.97e-05s C 6240 26 theano.sandbox.cuda.basic_ops.GpuElemwise
22.4% 98.5% 0.154s 1.07e-04s C 1440 6 theano.sandbox.cuda.blas.GpuDot22
1.4% 100.0% 0.010s 4.07e-05s C 240 1 theano.sandbox.cuda.basic_ops.GpuFromHost
0.0% 100.0% 0.000s 1.42e-06s C 240 1 theano.tensor.elemwise.Elemwise
... (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>
29.3% 29.3% 0.201s 1.19e-04s C 1680 7 GpuGemm{inplace}
22.4% 51.7% 0.154s 1.07e-04s C 1440 6 GpuDot22
19.7% 71.4% 0.135s 1.41e-04s C 960 4 GpuGemm{no_inplace}
8.1% 79.5% 0.055s 3.83e-05s C 1440 6 GpuElemwise{add,no_inplace}
2.2% 81.7% 0.015s 2.10e-05s C 720 3 GpuElemwise{Add}[(0, 1)]
2.0% 83.7% 0.014s 2.91e-05s C 480 2 GpuElemwise{mul,no_inplace}
2.0% 85.7% 0.014s 2.83e-05s C 480 2 GpuElemwise{Mul}[(0, 1)]
1.8% 87.5% 0.012s 5.01e-05s C 240 1 GpuElemwise{Composite{((((i0 * clip(i1, i2, i3)) * i4) + ((i5 * clip(i1, i2, i3)) * i4)) + i6)},no_inplace}
1.7% 89.2% 0.012s 2.47e-05s C 480 2 GpuElemwise{Clip}[(0, 0)]
1.7% 90.9% 0.011s 4.78e-05s C 240 1 GpuElemwise{Composite{((((i0 + i1) * i2) + (i3 * i2)) + i4)},no_inplace}
1.5% 92.4% 0.010s 2.12e-05s C 480 2 GpuElemwise{Composite{Cast{float32}(AND(GE(i0, i1), LE(i0, i2)))},no_inplace}
1.4% 93.8% 0.010s 4.07e-05s C 240 1 GpuFromHost
1.1% 94.9% 0.008s 3.15e-05s C 240 1 GpuElemwise{Composite{((i0 * i1) * i2)},no_inplace}
1.1% 96.0% 0.007s 3.08e-05s C 240 1 GpuElemwise{Mul}[(0, 3)]
0.9% 96.9% 0.006s 2.67e-05s C 240 1 GpuElemwise{Composite{((i0 * i1) * i2)}}[(0, 1)]
0.8% 97.7% 0.005s 2.29e-05s C 240 1 GpuElemwise{Composite{((i0 * i1) * i2)}}[(0, 0)]
0.8% 98.5% 0.005s 2.27e-05s C 240 1 GpuElemwise{Tanh}[(0, 0)]
0.8% 99.3% 0.005s 2.15e-05s C 240 1 GpuElemwise{Composite{Cast{float32}(AND(GE(i0, i1), LE(i0, i2)))}}[(0, 0)]
0.7% 100.0% 0.005s 1.94e-05s C 240 1 GpuElemwise{Composite{(i0 - sqr(i1))}}[(0, 1)]
0.0% 100.0% 0.000s 1.42e-06s C 240 1 Elemwise{Cast{float32}}
... (remaining 0 Ops account for 0.00%(0.00s) of the runtime)
Apply
------
<% time> <sum %> <apply time> <time per call> <#call> <id> <Apply name>
5.4% 5.4% 0.037s 1.53e-04s 240 2 GpuGemm{no_inplace}(<CudaNdarrayType(float32, matrix)>, TensorConstant{0.20000000298}, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(float32, matrix)>, TensorConstant{0.20000000298})
5.0% 10.3% 0.034s 1.41e-04s 240 4 GpuGemm{no_inplace}(<CudaNdarrayType(float32, matrix)>, TensorConstant{0.20000000298}, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(float32, matrix)>, TensorConstant{0.20000000298})
4.7% 15.0% 0.032s 1.35e-04s 240 1 GpuGemm{no_inplace}(<CudaNdarrayType(float32, matrix)>, TensorConstant{0.20000000298}, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(float32, matrix)>, TensorConstant{0.20000000298})
4.7% 19.7% 0.032s 1.34e-04s 240 3 GpuGemm{no_inplace}(<CudaNdarrayType(float32, matrix)>, TensorConstant{1.0}, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(float32, matrix)>, TensorConstant{1.0})
4.4% 24.1% 0.030s 1.24e-04s 240 24 GpuGemm{inplace}(GpuDot22.0, TensorConstant{1.0}, GpuElemwise{mul,no_inplace}.0, <CudaNdarrayType(float32, matrix)>, TensorConstant{1.0})
4.3% 28.4% 0.029s 1.23e-04s 240 34 GpuGemm{inplace}(GpuDot22.0, TensorConstant{1.0}, GpuElemwise{Mul}[(0, 1)].0, <CudaNdarrayType(float32, matrix)>, TensorConstant{1.0})
4.3% 32.6% 0.029s 1.22e-04s 240 38 GpuGemm{inplace}(GpuDot22.0, TensorConstant{1.0}, GpuElemwise{Composite{((i0 * i1) * i2)}}[(0, 0)].0, <CudaNdarrayType(float32, matrix)>, TensorConstant{1.0})
4.2% 36.9% 0.029s 1.21e-04s 240 29 GpuDot22(GpuElemwise{Mul}[(0, 1)].0, <CudaNdarrayType(float32, matrix)>)
4.2% 41.1% 0.029s 1.20e-04s 240 41 GpuGemm{inplace}(GpuGemm{inplace}.0, TensorConstant{1.0}, GpuElemwise{Composite{((i0 * i1) * i2)}}[(0, 1)].0, <CudaNdarrayType(float32, matrix)>, TensorConstant{1.0})
4.1% 45.2% 0.028s 1.16e-04s 240 16 GpuDot22(GpuElemwise{mul,no_inplace}.0, <CudaNdarrayType(float32, matrix)>)
4.1% 49.2% 0.028s 1.16e-04s 240 27 GpuDot22(GpuElemwise{Mul}[(0, 3)].0, <CudaNdarrayType(float32, matrix)>)
4.1% 53.3% 0.028s 1.16e-04s 240 33 GpuGemm{inplace}(GpuDot22.0, TensorConstant{1.0}, <CudaNdarrayType(float32, matrix)>, GpuElemwise{Mul}[(0, 1)].0, TensorConstant{1.0})
4.1% 57.3% 0.028s 1.16e-04s 240 31 GpuGemm{inplace}(GpuDot22.0, TensorConstant{1.0}, <CudaNdarrayType(float32, matrix)>, GpuElemwise{Mul}[(0, 1)].0, TensorConstant{1.0})
4.0% 61.4% 0.028s 1.15e-04s 240 42 GpuGemm{inplace}(GpuDot22.0, TensorConstant{1.0}, <CudaNdarrayType(float32, matrix)>, GpuElemwise{Composite{((i0 * i1) * i2)}}[(0, 1)].0, TensorConstant{1.0})
3.4% 64.8% 0.023s 9.66e-05s 240 23 GpuDot22(<CudaNdarrayType(float32, matrix)>, GpuElemwise{mul,no_inplace}.0)
3.3% 68.1% 0.023s 9.52e-05s 240 26 GpuDot22(<CudaNdarrayType(float32, matrix)>, GpuElemwise{Mul}[(0, 3)].0)
3.3% 71.4% 0.023s 9.51e-05s 240 37 GpuDot22(<CudaNdarrayType(float32, matrix)>, GpuElemwise{Composite{((i0 * i1) * i2)}}[(0, 0)].0)
1.8% 73.3% 0.013s 5.24e-05s 240 36 GpuElemwise{add,no_inplace}(<CudaNdarrayType(float32, matrix)>, GpuGemm{inplace}.0)
1.8% 75.1% 0.012s 5.14e-05s 240 44 GpuElemwise{add,no_inplace}(<CudaNdarrayType(float32, matrix)>, GpuGemm{inplace}.0)
1.8% 76.9% 0.012s 5.11e-05s 240 39 GpuElemwise{add,no_inplace}(<CudaNdarrayType(float32, matrix)>, GpuGemm{inplace}.0)
... (remaining 25 Apply instances account for 23.13%(0.16s) of the runtime)
Function profiling
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:330
Time in 0 calls to Function.__call__: 0.000000e+00s
Total compile time: 2.442388e+01s
Number of Apply nodes: 543
Theano Optimizer time: 2.154786e+01s
Theano validate time: 3.120501e-01s
Theano Linker time (includes C, CUDA code generation/compiling): 2.651706e+00s
Import time 1.034188e-02s
Time in all call to theano.grad() 1.037111e+00s
Time in all call to theano.grad() 1.037111e+00s
Time in all call to theano.grad() 1.037111e+00s
Function profiling
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:332
Time in 0 calls to Function.__call__: 0.000000e+00s
Total compile time: 2.318422e+00s
Number of Apply nodes: 137
Theano Optimizer time: 1.716493e+00s
Theano validate time: 5.609584e-02s
Theano Linker time (includes C, CUDA code generation/compiling): 4.905031e-01s
Import time 7.891417e-03s
Time in all call to theano.grad() 1.037111e+00s
Time in all call to theano.grad() 1.037111e+00s
Function profiling
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:334
Time in 0 calls to Function.__call__: 0.000000e+00s
Total compile time: 2.467041e+00s
Number of Apply nodes: 155
Theano Optimizer time: 1.826469e+00s
Theano validate time: 7.342625e-02s
Theano Linker time (includes C, CUDA code generation/compiling): 5.233629e-01s
Import time 5.436182e-03s
Time in all call to theano.grad() 1.037111e+00s
Time in all call to theano.grad() 1.037111e+00s
Function profiling
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:336
Time in 0 calls to Function.__call__: 0.000000e+00s
Total compile time: 3.259403e+00s
Number of Apply nodes: 166
Theano Optimizer time: 2.627834e+00s
Theano validate time: 7.786059e-02s
Theano Linker time (includes C, CUDA code generation/compiling): 5.099900e-01s
Import time 0.000000e+00s
Time in all call to theano.grad() 1.037111e+00s
Time in all call to theano.grad() 1.037111e+00s
Function profiling
==================
Message: Sum of all(5) printed profiles at exit excluding Scan op profile.
Time in 8 calls to Function.__call__: 1.868278e+01s
Time in Function.fn.__call__: 1.834628e+01s (98.199%)
Time in thunks: 1.794744e+01s (96.064%)
Total compile time: 5.692453e+01s
Number of Apply nodes: 532
Theano Optimizer time: 4.953296e+01s
Theano validate time: 8.184333e-01s
Theano Linker time (includes C, CUDA code generation/compiling): 6.594770e+00s
Import time 2.145724e-01s
Time in all call to theano.grad() 1.037111e+00s
Class
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Class name>
61.6% 61.6% 11.056s 6.61e-03s C 1672 209 theano.tensor.elemwise.Elemwise
11.4% 73.0% 2.045s 2.56e-01s Py 8 1 theano.tensor.subtensor.AdvancedIncSubtensor
5.6% 78.6% 0.999s 6.24e-02s Py 16 2 theano.scan_module.scan_op.Scan
4.2% 82.8% 0.753s 7.84e-03s C 96 12 theano.sandbox.cuda.blas.GpuDot22
3.6% 86.4% 0.647s 2.69e-02s C 24 3 theano.sandbox.cuda.basic_ops.HostFromGpu
3.4% 89.8% 0.613s 7.66e-02s C 8 1 theano.tensor.nnet.nnet.SoftmaxGrad
2.7% 92.5% 0.482s 1.51e-02s C 32 4 theano.tensor.elemwise.Sum
2.4% 94.9% 0.436s 1.82e-02s Py 24 3 theano.tensor.subtensor.AdvancedSubtensor
1.8% 96.7% 0.321s 1.34e-02s C 24 3 theano.sandbox.cuda.basic_ops.GpuFromHost
0.9% 97.6% 0.169s 1.06e-02s C 16 2 theano.tensor.basic.Alloc
0.7% 98.3% 0.129s 5.98e-04s C 216 27 theano.sandbox.cuda.basic_ops.GpuReshape
0.7% 99.1% 0.129s 2.48e-04s C 520 65 theano.sandbox.cuda.basic_ops.GpuElemwise
0.3% 99.4% 0.054s 1.35e-03s C 40 5 theano.sandbox.cuda.basic_ops.GpuCAReduce
0.2% 99.6% 0.044s 6.08e-04s C 72 9 theano.sandbox.cuda.basic_ops.GpuIncSubtensor
0.2% 99.8% 0.028s 2.67e-04s C 104 13 theano.sandbox.cuda.basic_ops.GpuAlloc
0.1% 99.9% 0.016s 1.99e-03s C 8 1 theano.sandbox.cuda.nnet.GpuSoftmax
0.1% 99.9% 0.012s 1.47e-03s C 8 1 theano.sandbox.cuda.blas.GpuDot22Scalar
0.1% 100.0% 0.009s 1.15e-03s Py 8 1 theano.sandbox.cuda.basic_ops.GpuFlatten
0.0% 100.0% 0.002s 2.06e-04s Py 8 1 theano.tensor.basic.Nonzero
0.0% 100.0% 0.001s 1.92e-06s C 416 52 theano.compile.ops.Shape_i
... (remaining 10 Classes account for 0.01%(0.00s) of the runtime)
Ops
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Op name>
48.1% 48.1% 8.638s 1.08e+00s C 8 1 Elemwise{Composite{(i0 * log((i1 / i2)))}}
11.4% 59.5% 2.045s 2.56e-01s Py 8 1 AdvancedIncSubtensor{inplace=False, set_instead_of_inc=False}
10.0% 69.5% 1.796s 2.24e-01s C 8 1 Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((i3 * i4 * i5) / i6) + i7))}}
4.6% 74.1% 0.826s 1.03e-01s Py 8 1 forall_inplace,gpu,grad_of_scan_fn}
4.2% 78.3% 0.753s 7.84e-03s C 96 12 GpuDot22
3.6% 81.9% 0.647s 2.69e-02s C 24 3 HostFromGpu
3.4% 85.3% 0.613s 7.66e-02s C 8 1 SoftmaxGrad
2.7% 88.0% 0.482s 2.01e-02s C 24 3 Sum{axis=[1], acc_dtype=float64}
2.4% 90.5% 0.436s 1.82e-02s Py 24 3 AdvancedSubtensor
1.8% 92.3% 0.321s 1.34e-02s C 24 3 GpuFromHost
1.8% 94.0% 0.317s 3.96e-02s C 8 1 Elemwise{clip,no_inplace}
1.7% 95.7% 0.302s 1.26e-02s C 24 3 Elemwise{mul,no_inplace}
1.0% 96.7% 0.173s 2.17e-02s Py 8 1 forall_inplace,gpu,scan_fn}
0.9% 97.6% 0.169s 1.06e-02s C 16 2 Alloc
0.7% 98.3% 0.129s 8.48e-04s C 152 19 GpuReshape{2}
0.3% 98.6% 0.054s 1.35e-03s C 40 5 GpuCAReduce{add}{1,1,0}
0.2% 98.8% 0.035s 1.11e-03s C 32 4 GpuIncSubtensor{Inc;:int64:}
0.2% 99.0% 0.030s 3.46e-04s C 88 11 GpuElemwise{Add}[(0, 0)]
0.2% 99.1% 0.027s 2.84e-04s C 96 12 GpuAlloc{memset_0=True}
0.1% 99.3% 0.025s 2.39e-04s C 104 13 GpuElemwise{Composite{(i0 * sqr(i1))},no_inplace}
... (remaining 109 Ops account for 0.71%(0.13s) of the runtime)
Apply
------
<% time> <sum %> <apply time> <time per call> <#call> <id> <Apply name>
48.1% 48.1% 8.638s 1.08e+00s 8 424 Elemwise{Composite{(i0 * log((i1 / i2)))}}(AdvancedSubtensor.0, Elemwise{clip,no_inplace}.0, InplaceDimShuffle{0,x}.0)
11.4% 59.5% 2.045s 2.56e-01s 8 427 AdvancedIncSubtensor{inplace=False, set_instead_of_inc=False}(Alloc.0, Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((i3 * i4 * i5) / i6) + i7))}}.0, Subtensor{int64}.0, Subtensor{int64}.0)
10.0% 69.5% 1.796s 2.24e-01s 8 425 Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((i3 * i4 * i5) / i6) + i7))}}(AdvancedSubtensor.0, TensorConstant{(1, 1) of 1e-07}, TensorConstant{(1, 1) of 1.0}, TensorConstant{(1, 1) of -1.0}, InplaceDimShuffle{0,x}.0, AdvancedSubtensor.0, Elemwise{clip,no_inplace}.0, Elemwise{true_div,no_inplace}.0)
4.6% 74.1% 0.826s 1.03e-01s 8 453 forall_inplace,gpu,grad_of_scan_fn}(Elemwise{Composite{minimum(minimum(minimum(minimum(i0, i1), i2), i3), i4)}}.0, GpuElemwise{mul,no_inplace}.0, GpuElemwise{mul,no_inplace}.0, GpuElemwise{Tanh}[(0, 0)].0, GpuDimShuffle{0,2,1}.0, GpuElemwise{Composite{(i0 - sqr(i1))},no_inplace}.0, GpuSubtensor{int64:int64:int64}.0, GpuSubtensor{int64:int64:int64}.0, GpuSubtensor{int64:int64:int64}.0, GpuSubtensor{int64:int64:int64}.0, Subtensor{int64:int64:int64}.0
3.4% 77.5% 0.613s 7.66e-02s 8 431 SoftmaxGrad(Reshape{2}.0, HostFromGpu.0)
1.8% 79.4% 0.326s 4.07e-02s 8 406 HostFromGpu(GpuReshape{3}.0)
1.8% 81.2% 0.320s 4.01e-02s 8 401 HostFromGpu(GpuSoftmax.0)
1.8% 82.9% 0.317s 3.96e-02s 8 416 Elemwise{clip,no_inplace}(AdvancedSubtensor.0, TensorConstant{(1, 1) of 1e-07}, TensorConstant{(1, 1) of 1.0})
1.7% 84.6% 0.302s 3.77e-02s 8 150 Elemwise{mul,no_inplace}(InplaceDimShuffle{0,x}.0, AdvancedSubtensor.0)
1.2% 85.8% 0.218s 2.73e-02s 8 75 AdvancedSubtensor(y, Subtensor{int64}.0, Subtensor{int64}.0)
1.2% 87.0% 0.218s 2.72e-02s 8 412 AdvancedSubtensor(HostFromGpu.0, Subtensor{int64}.0, Subtensor{int64}.0)
1.0% 88.0% 0.173s 2.17e-02s 8 352 forall_inplace,gpu,scan_fn}(Elemwise{Composite{minimum(minimum(minimum(minimum(i0, i1), i2), i3), i4)}}.0, GpuSubtensor{int64:int64:int8}.0, GpuSubtensor{int64:int64:int8}.0, GpuSubtensor{int64:int64:int8}.0, GpuSubtensor{int64:int64:int8}.0, Subtensor{int64:int64:int8}.0, GpuAlloc{memset_0=True}.0, GpuAlloc{memset_0=True}.0, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(float32, matrix)>, <CudaNdarrayType(
0.9% 88.9% 0.169s 2.11e-02s 8 413 Alloc(TensorConstant{(1, 1, 1) of 0.0}, Shape_i{0}.0, Shape_i{1}.0, Shape_i{2}.0)
0.9% 89.9% 0.166s 2.07e-02s 8 6 GpuFromHost(<TensorType(float32, 3D)>)
0.9% 90.8% 0.161s 2.01e-02s 8 173 Sum{axis=[1], acc_dtype=float64}(Elemwise{mul,no_inplace}.0)
0.9% 91.6% 0.161s 2.01e-02s 8 426 Sum{axis=[1], acc_dtype=float64}(Elemwise{Composite{(i0 * log((i1 / i2)))}}.0)
0.9% 92.5% 0.161s 2.01e-02s 8 420 Sum{axis=[1], acc_dtype=float64}(Elemwise{clip,no_inplace}.0)
0.9% 93.4% 0.155s 1.93e-02s 8 433 GpuFromHost(SoftmaxGrad.0)
0.5% 93.9% 0.092s 1.15e-02s 8 91 GpuDot22(GpuReshape{2}.0, GpuReshape{2}.0)
0.4% 94.3% 0.066s 8.25e-03s 8 94 GpuDot22(GpuReshape{2}.0, GpuReshape{2}.0)
... (remaining 512 Apply instances account for 5.72%(1.03s) of the runtime)
training time 19.0255179405
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment