Skip to content

Instantly share code, notes, and snippets.

@mmmikael

mmmikael/fcn_bench_profiling.txt

Created October 9, 2015 12:41
Show Gist options
  • Save mmmikael/73b62a911748a33f17ef to your computer and use it in GitHub Desktop.
Save mmmikael/73b62a911748a33f17ef to your computer and use it in GitHub Desktop.
Download ZIP
fcn keras profiliing
Raw
fcn_bench_profiling.txt
Function profiling
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:399
Time in 1 calls to Function.__call__: 1.027354e+01s
Time in Function.fn.__call__: 1.026697e+01s (99.936%)
Time in thunks: 1.017123e+01s (99.004%)
Total compile time: 1.421726e+00s
Number of Apply nodes: 140
Theano Optimizer time: 1.223497e+00s
Theano validate time: 6.331444e-02s
Theano Linker time (includes C, CUDA code generation/compiling): 1.472740e-01s
Import time 4.217982e-02s
Time in all call to theano.grad() 5.796099e-02s
Time since theano import 22.311s
Class
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Class name>
27.1% 27.1% 2.752s 9.17e-01s C 3 3 theano.sandbox.cuda.dnn.GpuDnnConvGradW
25.5% 52.6% 2.595s 2.36e-01s C 11 11 theano.tensor.elemwise.Elemwise
16.1% 68.6% 1.635s 1.63e+00s Py 1 1 theano.tensor.subtensor.AdvancedIncSubtensor
5.9% 74.5% 0.600s 2.00e-01s Py 3 3 theano.tensor.subtensor.AdvancedSubtensor
5.8% 80.3% 0.585s 2.93e-01s C 2 2 theano.sandbox.cuda.dnn.GpuDnnConvGradI
3.8% 84.1% 0.384s 9.59e-02s C 4 4 theano.tensor.elemwise.Sum
3.6% 87.6% 0.362s 3.62e-01s C 1 1 theano.sandbox.cuda.basic_ops.HostFromGpu
2.8% 90.4% 0.287s 1.06e-02s C 27 27 theano.sandbox.cuda.basic_ops.GpuElemwise
2.3% 92.7% 0.232s 7.74e-02s C 3 3 theano.sandbox.cuda.dnn.GpuDnnConv
2.0% 94.7% 0.202s 2.02e-01s Py 1 1 theano.tensor.basic.Nonzero
1.4% 96.1% 0.145s 1.45e-01s C 1 1 theano.sandbox.cuda.dnn.GpuDnnSoftmax
1.2% 97.3% 0.118s 9.86e-03s C 12 12 theano.sandbox.cuda.basic_ops.GpuContiguous
1.0% 98.3% 0.102s 1.02e-01s C 1 1 theano.sandbox.cuda.dnn.GpuDnnSoftmaxGrad
0.8% 99.1% 0.081s 4.06e-02s C 2 2 theano.sandbox.cuda.basic_ops.GpuFromHost
0.5% 99.6% 0.054s 5.45e-02s C 1 1 theano.tensor.basic.Alloc
0.3% 100.0% 0.033s 1.12e-02s C 3 3 theano.sandbox.cuda.basic_ops.GpuCAReduce
0.0% 100.0% 0.003s 3.82e-04s C 8 8 theano.sandbox.cuda.basic_ops.GpuAllocEmpty
0.0% 100.0% 0.000s 3.38e-06s C 12 12 theano.sandbox.cuda.basic_ops.GpuDimShuffle
0.0% 100.0% 0.000s 1.55e-06s C 22 22 theano.compile.ops.Shape_i
0.0% 100.0% 0.000s 2.57e-06s C 9 9 theano.tensor.opt.MakeVector
... (remaining 4 Classes account for 0.00%(0.00s) of the runtime)
Ops
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Op name>
27.1% 27.1% 2.752s 9.17e-01s C 3 3 GpuDnnConvGradW{algo='none', inplace=True}
17.1% 44.1% 1.736s 1.74e+00s C 1 1 Elemwise{Composite{(i0 * log((i1 / i2)))}}[(0, 0)]
16.1% 60.2% 1.635s 1.63e+00s Py 1 1 AdvancedIncSubtensor{inplace=False, set_instead_of_inc=False}
5.9% 66.1% 0.600s 2.00e-01s Py 3 3 AdvancedSubtensor
5.8% 71.9% 0.585s 2.93e-01s C 2 2 GpuDnnConvGradI{algo='none', inplace=True}
4.4% 76.3% 0.451s 4.51e-01s C 1 1 Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((-i3) / i4) + i5))}}[(0, 0)]
3.8% 80.1% 0.390s 3.90e-01s C 1 1 Elemwise{clip,no_inplace}
3.6% 83.7% 0.362s 3.62e-01s C 1 1 HostFromGpu
2.6% 86.3% 0.265s 1.33e-01s C 2 2 Sum{axis=[1], acc_dtype=float64}
2.3% 88.6% 0.232s 7.74e-02s C 3 3 GpuDnnConv{algo='small', inplace=True}
2.0% 90.6% 0.202s 2.02e-01s Py 1 1 Nonzero
1.4% 92.0% 0.145s 1.45e-01s C 1 1 GpuDnnSoftmax{tensor_format='bc01', mode='channel', algo='accurate'}
1.3% 93.3% 0.132s 4.40e-02s C 3 3 GpuElemwise{Composite{((i0 * i1) + (i0 * i1 * sgn(i2)))}}[(0, 1)]
1.2% 94.5% 0.118s 9.86e-03s C 12 12 GpuContiguous
1.2% 95.6% 0.118s 5.92e-02s C 2 2 Sum{acc_dtype=float64}
1.0% 96.6% 0.102s 1.02e-01s C 1 1 GpuDnnSoftmaxGrad{tensor_format='bc01', mode='channel', algo='accurate'}
0.9% 97.5% 0.093s 1.04e-02s C 9 9 GpuElemwise{Add}[(0, 0)]
0.8% 98.3% 0.081s 4.06e-02s C 2 2 GpuFromHost
0.6% 98.9% 0.060s 2.00e-02s C 3 3 GpuElemwise{Composite{(i0 * (i1 + Abs(i1)))},no_inplace}
0.5% 99.5% 0.054s 5.45e-02s C 1 1 Alloc
... (remaining 26 Ops account for 0.54%(0.05s) of the runtime)
Apply
------
<% time> <sum %> <apply time> <time per call> <#call> <id> <Apply name>
19.9% 19.9% 2.023s 2.02e+00s 1 124 GpuDnnConvGradW{algo='none', inplace=True}(GpuContiguous.0, GpuContiguous.0, GpuAllocEmpty.0, GpuDnnConvDesc{border_mode=(1, 1), subsample=(1, 1), conv_mode='conv'}.0, Constant{1.0}, Constant{0.0})
17.1% 37.0% 1.736s 1.74e+00s 1 96 Elemwise{Composite{(i0 * log((i1 / i2)))}}[(0, 0)](AdvancedSubtensor.0, Elemwise{clip,no_inplace}.0, InplaceDimShuffle{0,x}.0)
16.1% 53.0% 1.635s 1.63e+00s 1 97 AdvancedIncSubtensor{inplace=False, set_instead_of_inc=False}(Alloc.0, Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((-i3) / i4) + i5))}}[(0, 0)].0, Subtensor{int64}.0, Subtensor{int64}.0, Subtensor{int64}.0)
4.4% 57.5% 0.451s 4.51e-01s 1 95 Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((-i3) / i4) + i5))}}[(0, 0)](AdvancedSubtensor.0, TensorConstant{(1, 1) of 1e-07}, TensorConstant{(1, 1) of 1.0}, AdvancedSubtensor.0, Elemwise{clip,no_inplace}.0, Elemwise{TrueDiv}[(0, 0)].0)
4.1% 61.6% 0.420s 4.20e-01s 1 113 GpuDnnConvGradW{algo='none', inplace=True}(GpuContiguous.0, GpuContiguous.0, GpuAllocEmpty.0, GpuDnnConvDesc{border_mode=(1, 1), subsample=(1, 1), conv_mode='conv'}.0, Constant{1.0}, Constant{0.0})
3.8% 65.4% 0.390s 3.90e-01s 1 91 Elemwise{clip,no_inplace}(AdvancedSubtensor.0, TensorConstant{(1, 1) of 1e-07}, TensorConstant{(1, 1) of 1.0})
3.6% 69.0% 0.362s 3.62e-01s 1 84 HostFromGpu(GpuReshape{4}.0)
3.5% 72.5% 0.357s 3.57e-01s 1 89 AdvancedSubtensor(HostFromGpu.0, Subtensor{int64}.0, Subtensor{int64}.0, Subtensor{int64}.0)
3.5% 76.0% 0.353s 3.53e-01s 1 123 GpuDnnConvGradI{algo='none', inplace=True}(GpuContiguous.0, GpuContiguous.0, GpuAllocEmpty.0, GpuDnnConvDesc{border_mode=(1, 1), subsample=(1, 1), conv_mode='conv'}.0, Constant{1.0}, Constant{0.0})
3.0% 79.0% 0.309s 3.09e-01s 1 134 GpuDnnConvGradW{algo='none', inplace=True}(GpuContiguous.0, GpuContiguous.0, GpuAllocEmpty.0, GpuDnnConvDesc{border_mode=(1, 1), subsample=(1, 1), conv_mode='conv'}.0, Constant{1.0}, Constant{0.0})
2.3% 81.3% 0.232s 2.32e-01s 1 112 GpuDnnConvGradI{algo='none', inplace=True}(GpuContiguous.0, GpuContiguous.0, GpuAllocEmpty.0, GpuDnnConvDesc{border_mode=(1, 1), subsample=(1, 1), conv_mode='conv'}.0, Constant{1.0}, Constant{0.0})
2.0% 83.3% 0.202s 2.02e-01s 1 23 Nonzero(<TensorType(float32, 4D)>)
1.7% 85.0% 0.173s 1.73e-01s 1 49 AdvancedSubtensor(y, Subtensor{int64}.0, Subtensor{int64}.0, Subtensor{int64}.0)
1.5% 86.5% 0.153s 1.53e-01s 1 92 Sum{axis=[1], acc_dtype=float64}(Elemwise{clip,no_inplace}.0)
1.4% 87.9% 0.145s 1.45e-01s 1 79 GpuDnnSoftmax{tensor_format='bc01', mode='channel', algo='accurate'}(GpuContiguous.0)
1.4% 89.3% 0.145s 1.45e-01s 1 65 GpuDnnConv{algo='small', inplace=True}(GpuContiguous.0, GpuContiguous.0, GpuAllocEmpty.0, GpuDnnConvDesc{border_mode=(1, 1), subsample=(1, 1), conv_mode='conv'}.0, Constant{1.0}, Constant{0.0})
1.2% 90.5% 0.118s 1.18e-01s 1 110 GpuContiguous(GpuElemwise{Composite{((i0 * i1) + (i0 * i1 * sgn(i2)))}}[(0, 1)].0)
1.1% 91.6% 0.114s 1.14e-01s 1 98 Sum{acc_dtype=float64}(Elemwise{Composite{(i0 * log((i1 / i2)))}}[(0, 0)].0)
1.1% 92.7% 0.112s 1.12e-01s 1 58 Sum{axis=[1], acc_dtype=float64}(AdvancedSubtensor.0)
1.0% 93.7% 0.102s 1.02e-01s 1 104 GpuDnnSoftmaxGrad{tensor_format='bc01', mode='channel', algo='accurate'}(GpuContiguous.0, GpuContiguous.0)
... (remaining 120 Apply instances account for 6.27%(0.64s) of the runtime)
Function profiling
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:401
Time in 0 calls to Function.__call__: 0.000000e+00s
Total compile time: 1.072570e+00s
Number of Apply nodes: 153
Theano Optimizer time: 9.243569e-01s
Theano validate time: 6.185031e-02s
Theano Linker time (includes C, CUDA code generation/compiling): 1.079371e-01s
Import time 1.020193e-03s
Time in all call to theano.grad() 5.796099e-02s
Time since theano import 22.320s
Function profiling
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:403
Time in 0 calls to Function.__call__: 0.000000e+00s
Total compile time: 3.260720e-01s
Number of Apply nodes: 60
Theano Optimizer time: 2.644930e-01s
Theano validate time: 7.456541e-03s
Theano Linker time (includes C, CUDA code generation/compiling): 4.158711e-02s
Import time 9.508133e-04s
Time in all call to theano.grad() 5.796099e-02s
Time since theano import 22.320s
Function profiling
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:405
Time in 0 calls to Function.__call__: 0.000000e+00s
Total compile time: 5.981460e-01s
Number of Apply nodes: 75
Theano Optimizer time: 3.249922e-01s
Theano validate time: 9.876013e-03s
Theano Linker time (includes C, CUDA code generation/compiling): 4.987216e-02s
Import time 2.160072e-04s
Time in all call to theano.grad() 5.796099e-02s
Time since theano import 22.320s
Function profiling
==================
Message: build/bdist.linux-x86_64/egg/keras/models.py:407
Time in 0 calls to Function.__call__: 0.000000e+00s
Total compile time: 4.345279e-01s
Number of Apply nodes: 88
Theano Optimizer time: 3.557758e-01s
Theano validate time: 1.038814e-02s
Theano Linker time (includes C, CUDA code generation/compiling): 5.503392e-02s
Import time 0.000000e+00s
Time in all call to theano.grad() 5.796099e-02s
Time since theano import 22.320s
Function profiling
==================
Message: Sum of all(5) printed profiles at exit excluding Scan op profile.
Time in 1 calls to Function.__call__: 1.027354e+01s
Time in Function.fn.__call__: 1.026697e+01s (99.936%)
Time in thunks: 1.017123e+01s (99.004%)
Total compile time: 3.853042e+00s
Number of Apply nodes: 140
Theano Optimizer time: 3.093115e+00s
Theano validate time: 1.528854e-01s
Theano Linker time (includes C, CUDA code generation/compiling): 4.017043e-01s
Import time 4.436684e-02s
Time in all call to theano.grad() 5.796099e-02s
Time since theano import 22.320s
Class
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Class name>
27.1% 27.1% 2.752s 9.17e-01s C 3 3 theano.sandbox.cuda.dnn.GpuDnnConvGradW
25.5% 52.6% 2.595s 2.36e-01s C 11 11 theano.tensor.elemwise.Elemwise
16.1% 68.6% 1.635s 1.63e+00s Py 1 1 theano.tensor.subtensor.AdvancedIncSubtensor
5.9% 74.5% 0.600s 2.00e-01s Py 3 3 theano.tensor.subtensor.AdvancedSubtensor
5.8% 80.3% 0.585s 2.93e-01s C 2 2 theano.sandbox.cuda.dnn.GpuDnnConvGradI
3.8% 84.1% 0.384s 9.59e-02s C 4 4 theano.tensor.elemwise.Sum
3.6% 87.6% 0.362s 3.62e-01s C 1 1 theano.sandbox.cuda.basic_ops.HostFromGpu
2.8% 90.4% 0.287s 1.06e-02s C 27 27 theano.sandbox.cuda.basic_ops.GpuElemwise
2.3% 92.7% 0.232s 7.74e-02s C 3 3 theano.sandbox.cuda.dnn.GpuDnnConv
2.0% 94.7% 0.202s 2.02e-01s Py 1 1 theano.tensor.basic.Nonzero
1.4% 96.1% 0.145s 1.45e-01s C 1 1 theano.sandbox.cuda.dnn.GpuDnnSoftmax
1.2% 97.3% 0.118s 9.86e-03s C 12 12 theano.sandbox.cuda.basic_ops.GpuContiguous
1.0% 98.3% 0.102s 1.02e-01s C 1 1 theano.sandbox.cuda.dnn.GpuDnnSoftmaxGrad
0.8% 99.1% 0.081s 4.06e-02s C 2 2 theano.sandbox.cuda.basic_ops.GpuFromHost
0.5% 99.6% 0.054s 5.45e-02s C 1 1 theano.tensor.basic.Alloc
0.3% 100.0% 0.033s 1.12e-02s C 3 3 theano.sandbox.cuda.basic_ops.GpuCAReduce
0.0% 100.0% 0.003s 3.82e-04s C 8 8 theano.sandbox.cuda.basic_ops.GpuAllocEmpty
0.0% 100.0% 0.000s 3.38e-06s C 12 12 theano.sandbox.cuda.basic_ops.GpuDimShuffle
0.0% 100.0% 0.000s 1.55e-06s C 22 22 theano.compile.ops.Shape_i
0.0% 100.0% 0.000s 2.57e-06s C 9 9 theano.tensor.opt.MakeVector
... (remaining 4 Classes account for 0.00%(0.00s) of the runtime)
Ops
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Op name>
27.1% 27.1% 2.752s 9.17e-01s C 3 3 GpuDnnConvGradW{algo='none', inplace=True}
17.1% 44.1% 1.736s 1.74e+00s C 1 1 Elemwise{Composite{(i0 * log((i1 / i2)))}}[(0, 0)]
16.1% 60.2% 1.635s 1.63e+00s Py 1 1 AdvancedIncSubtensor{inplace=False, set_instead_of_inc=False}
5.9% 66.1% 0.600s 2.00e-01s Py 3 3 AdvancedSubtensor
5.8% 71.9% 0.585s 2.93e-01s C 2 2 GpuDnnConvGradI{algo='none', inplace=True}
4.4% 76.3% 0.451s 4.51e-01s C 1 1 Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((-i3) / i4) + i5))}}[(0, 0)]
3.8% 80.1% 0.390s 3.90e-01s C 1 1 Elemwise{clip,no_inplace}
3.6% 83.7% 0.362s 3.62e-01s C 1 1 HostFromGpu
2.6% 86.3% 0.265s 1.33e-01s C 2 2 Sum{axis=[1], acc_dtype=float64}
2.3% 88.6% 0.232s 7.74e-02s C 3 3 GpuDnnConv{algo='small', inplace=True}
2.0% 90.6% 0.202s 2.02e-01s Py 1 1 Nonzero
1.4% 92.0% 0.145s 1.45e-01s C 1 1 GpuDnnSoftmax{tensor_format='bc01', mode='channel', algo='accurate'}
1.3% 93.3% 0.132s 4.40e-02s C 3 3 GpuElemwise{Composite{((i0 * i1) + (i0 * i1 * sgn(i2)))}}[(0, 1)]
1.2% 94.5% 0.118s 9.86e-03s C 12 12 GpuContiguous
1.2% 95.6% 0.118s 5.92e-02s C 2 2 Sum{acc_dtype=float64}
1.0% 96.6% 0.102s 1.02e-01s C 1 1 GpuDnnSoftmaxGrad{tensor_format='bc01', mode='channel', algo='accurate'}
0.9% 97.5% 0.093s 1.04e-02s C 9 9 GpuElemwise{Add}[(0, 0)]
0.8% 98.3% 0.081s 4.06e-02s C 2 2 GpuFromHost
0.6% 98.9% 0.060s 2.00e-02s C 3 3 GpuElemwise{Composite{(i0 * (i1 + Abs(i1)))},no_inplace}
0.5% 99.5% 0.054s 5.45e-02s C 1 1 Alloc
... (remaining 26 Ops account for 0.54%(0.05s) of the runtime)
Apply
------
<% time> <sum %> <apply time> <time per call> <#call> <id> <Apply name>
19.9% 19.9% 2.023s 2.02e+00s 1 124 GpuDnnConvGradW{algo='none', inplace=True}(GpuContiguous.0, GpuContiguous.0, GpuAllocEmpty.0, GpuDnnConvDesc{border_mode=(1, 1), subsample=(1, 1), conv_mode='conv'}.0, Constant{1.0}, Constant{0.0})
17.1% 37.0% 1.736s 1.74e+00s 1 96 Elemwise{Composite{(i0 * log((i1 / i2)))}}[(0, 0)](AdvancedSubtensor.0, Elemwise{clip,no_inplace}.0, InplaceDimShuffle{0,x}.0)
16.1% 53.0% 1.635s 1.63e+00s 1 97 AdvancedIncSubtensor{inplace=False, set_instead_of_inc=False}(Alloc.0, Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((-i3) / i4) + i5))}}[(0, 0)].0, Subtensor{int64}.0, Subtensor{int64}.0, Subtensor{int64}.0)
4.4% 57.5% 0.451s 4.51e-01s 1 95 Elemwise{Composite{(AND(GE(i0, i1), LE(i0, i2)) * (((-i3) / i4) + i5))}}[(0, 0)](AdvancedSubtensor.0, TensorConstant{(1, 1) of 1e-07}, TensorConstant{(1, 1) of 1.0}, AdvancedSubtensor.0, Elemwise{clip,no_inplace}.0, Elemwise{TrueDiv}[(0, 0)].0)
4.1% 61.6% 0.420s 4.20e-01s 1 113 GpuDnnConvGradW{algo='none', inplace=True}(GpuContiguous.0, GpuContiguous.0, GpuAllocEmpty.0, GpuDnnConvDesc{border_mode=(1, 1), subsample=(1, 1), conv_mode='conv'}.0, Constant{1.0}, Constant{0.0})
3.8% 65.4% 0.390s 3.90e-01s 1 91 Elemwise{clip,no_inplace}(AdvancedSubtensor.0, TensorConstant{(1, 1) of 1e-07}, TensorConstant{(1, 1) of 1.0})
3.6% 69.0% 0.362s 3.62e-01s 1 84 HostFromGpu(GpuReshape{4}.0)
3.5% 72.5% 0.357s 3.57e-01s 1 89 AdvancedSubtensor(HostFromGpu.0, Subtensor{int64}.0, Subtensor{int64}.0, Subtensor{int64}.0)
3.5% 76.0% 0.353s 3.53e-01s 1 123 GpuDnnConvGradI{algo='none', inplace=True}(GpuContiguous.0, GpuContiguous.0, GpuAllocEmpty.0, GpuDnnConvDesc{border_mode=(1, 1), subsample=(1, 1), conv_mode='conv'}.0, Constant{1.0}, Constant{0.0})
3.0% 79.0% 0.309s 3.09e-01s 1 134 GpuDnnConvGradW{algo='none', inplace=True}(GpuContiguous.0, GpuContiguous.0, GpuAllocEmpty.0, GpuDnnConvDesc{border_mode=(1, 1), subsample=(1, 1), conv_mode='conv'}.0, Constant{1.0}, Constant{0.0})
2.3% 81.3% 0.232s 2.32e-01s 1 112 GpuDnnConvGradI{algo='none', inplace=True}(GpuContiguous.0, GpuContiguous.0, GpuAllocEmpty.0, GpuDnnConvDesc{border_mode=(1, 1), subsample=(1, 1), conv_mode='conv'}.0, Constant{1.0}, Constant{0.0})
2.0% 83.3% 0.202s 2.02e-01s 1 23 Nonzero(<TensorType(float32, 4D)>)
1.7% 85.0% 0.173s 1.73e-01s 1 49 AdvancedSubtensor(y, Subtensor{int64}.0, Subtensor{int64}.0, Subtensor{int64}.0)
1.5% 86.5% 0.153s 1.53e-01s 1 92 Sum{axis=[1], acc_dtype=float64}(Elemwise{clip,no_inplace}.0)
1.4% 87.9% 0.145s 1.45e-01s 1 79 GpuDnnSoftmax{tensor_format='bc01', mode='channel', algo='accurate'}(GpuContiguous.0)
1.4% 89.3% 0.145s 1.45e-01s 1 65 GpuDnnConv{algo='small', inplace=True}(GpuContiguous.0, GpuContiguous.0, GpuAllocEmpty.0, GpuDnnConvDesc{border_mode=(1, 1), subsample=(1, 1), conv_mode='conv'}.0, Constant{1.0}, Constant{0.0})
1.2% 90.5% 0.118s 1.18e-01s 1 110 GpuContiguous(GpuElemwise{Composite{((i0 * i1) + (i0 * i1 * sgn(i2)))}}[(0, 1)].0)
1.1% 91.6% 0.114s 1.14e-01s 1 98 Sum{acc_dtype=float64}(Elemwise{Composite{(i0 * log((i1 / i2)))}}[(0, 0)].0)
1.1% 92.7% 0.112s 1.12e-01s 1 58 Sum{axis=[1], acc_dtype=float64}(AdvancedSubtensor.0)
1.0% 93.7% 0.102s 1.02e-01s 1 104 GpuDnnSoftmaxGrad{tensor_format='bc01', mode='channel', algo='accurate'}(GpuContiguous.0, GpuContiguous.0)
... (remaining 120 Apply instances account for 6.27%(0.64s) of the runtime)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment