Profiling¶
Sometimes computing the likelihood is not as fast as we would like. Theano provides handy profiling tools which are wrapped in PyMC3 by model.profile. This function returns a ProfileStats object conveying information about the underlying Theano operations. Here we’ll profile the likelihood and gradient for the stochastic volatility example.
First we build the model.
Running on PyMC3 v3.9.3
RANDOM_SEED = 8927
np.random.seed(RANDOM_SEED)
# Load the data
returns = pd.read_csv(pm.get_data("SP500.csv"), index_col=0, parse_dates=True)
# Stochastic volatility example
with pm.Model() as model:
sigma = pm.Exponential("sigma", 1.0 / 0.02, testval=0.1)
nu = pm.Exponential("nu", 1.0 / 10)
s = pm.GaussianRandomWalk("s", sigma ** -2, shape=returns.shape[0])
r = pm.StudentT("r", nu, lam=np.exp(-2 * s), observed=returns["change"])
Then we call the profile function and summarize its return values.
# Profiling of the logp call
model.profile(model.logpt).summary()
Function profiling
==================
Message: /home/amit/Work/OpenSourceDev/pymc3/pymc3/model.py:1192
Time in 1000 calls to Function.__call__: 2.042763e-01s
Time in Function.fn.__call__: 1.832466e-01s (89.705%)
Time in thunks: 1.743519e-01s (85.351%)
Total compile time: 7.987223e-01s
Number of Apply nodes: 22
Theano Optimizer time: 7.478609e-01s
Theano validate time: 2.320766e-03s
Theano Linker time (includes C, CUDA code generation/compiling): 3.071356e-02s
Import time 1.030779e-02s
Node make_thunk time 2.952695e-02s
Node Elemwise{Composite{Switch(Cast{int8}((GT(Composite{exp((i0 * i1))}(i0, i1), i2) * i3 * GT(inv(sqrt(Composite{exp((i0 * i1))}(i0, i1))), i2))), (((i4 + (i5 * log(((i6 * Composite{exp((i0 * i1))}(i0, i1)) / i7)))) - i8) - (i5 * i9 * log1p(((Composite{exp((i0 * i1))}(i0, i1) * i10) / i7)))), i11)}}(TensorConstant{(1,) of -2.0}, s, TensorConstant{(1,) of 0}, Elemwise{gt,no_inplace}.0, Elemwise{Composite{scalar_gammaln((i0 * i1))}}.0, TensorConstant{(1,) of 0.5}, TensorConstant{(1,) of 0...8861837907}, InplaceDimShuffle{x}.0, Elemwise{Composite{scalar_gammaln((i0 * i1))}}.0, Elemwise{add,no_inplace}.0, TensorConstant{[1.0435322..54666e-07]}, TensorConstant{(1,) of -inf}) time 2.954960e-03s
Node Elemwise{Composite{(Switch(Cast{int8}(GE(i0, i1)), (i2 - (i3 * i0)), i4) + i5)}}[(0, 0)](nu, TensorConstant{0}, TensorConstant{-2.3025850929940455}, TensorConstant{0.1}, TensorConstant{-inf}, nu_log__) time 2.277613e-03s
Node Elemwise{Composite{Switch(i0, (i1 * ((-(i2 * sqr((i3 - i4)))) + i5)), i6)}}(Elemwise{Composite{Cast{int8}(GT(i0, i1))}}.0, TensorConstant{(1,) of 0.5}, Elemwise{Composite{inv(sqr(i0))}}[(0, 0)].0, Subtensor{int64::}.0, Subtensor{:int64:}.0, Elemwise{Composite{log((i0 * i1))}}.0, TensorConstant{(1,) of -inf}) time 2.170801e-03s
Node Elemwise{Composite{(Switch(Cast{int8}(GE(i0, i1)), (i2 - (i3 * i0)), i4) + i5)}}(sigma, TensorConstant{0}, TensorConstant{3.912023005428146}, TensorConstant{50.0}, TensorConstant{-inf}, sigma_log__) time 2.163649e-03s
Node Elemwise{Composite{scalar_gammaln((i0 * i1))}}(TensorConstant{(1,) of 0.5}, InplaceDimShuffle{x}.0) time 1.938581e-03s
Time in all call to theano.grad() 0.000000e+00s
Time since theano import 5.477s
Class
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Class name>
93.1% 93.1% 0.162s 1.16e-05s C 14000 14 theano.tensor.elemwise.Elemwise
5.8% 99.0% 0.010s 3.39e-06s C 3000 3 theano.tensor.elemwise.Sum
0.5% 99.5% 0.001s 4.64e-07s C 2000 2 theano.tensor.elemwise.DimShuffle
0.3% 99.8% 0.001s 2.97e-07s C 2000 2 theano.tensor.subtensor.Subtensor
0.2% 100.0% 0.000s 3.08e-07s C 1000 1 theano.tensor.opt.MakeVector
... (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>
87.9% 87.9% 0.153s 1.53e-04s C 1000 1 Elemwise{Composite{Switch(Cast{int8}((GT(Composite{exp((i0 * i1))}(i0, i1), i2) * i3 * GT(inv(sqrt(Composite{exp((i0 * i1))}(i0, i1))), i2))), (((i4 + (i5 * log(((i6 * Composite{exp((i0 * i1))}(i0, i1)) / i7)))) - i8) - (i5 * i9 * log1p(((Composite{exp((i0 * i1))}(i0, i1) * i10) / i7)))), i11)}}
5.8% 93.7% 0.010s 3.39e-06s C 3000 3 Sum{acc_dtype=float64}
3.4% 97.1% 0.006s 5.88e-06s C 1000 1 Elemwise{Composite{Switch(i0, (i1 * ((-(i2 * sqr((i3 - i4)))) + i5)), i6)}}
0.5% 97.6% 0.001s 4.64e-07s C 2000 2 InplaceDimShuffle{x}
0.4% 98.0% 0.001s 3.19e-07s C 2000 2 Elemwise{Composite{scalar_gammaln((i0 * i1))}}
0.3% 98.3% 0.000s 2.49e-07s C 2000 2 Elemwise{exp,no_inplace}
0.2% 98.5% 0.000s 3.47e-07s C 1000 1 Elemwise{Composite{(Switch(Cast{int8}(GE(i0, i1)), (i2 - (i3 * i0)), i4) + i5)}}[(0, 0)]
0.2% 98.7% 0.000s 3.37e-07s C 1000 1 Elemwise{Composite{(Switch(Cast{int8}(GE(i0, i1)), (i2 - (i3 * i0)), i4) + i5)}}
0.2% 98.9% 0.000s 3.35e-07s C 1000 1 Subtensor{int64::}
0.2% 99.1% 0.000s 3.30e-07s C 1000 1 Elemwise{Composite{log((i0 * i1))}}
0.2% 99.2% 0.000s 3.08e-07s C 1000 1 MakeVector{dtype='float64'}
0.2% 99.4% 0.000s 2.86e-07s C 1000 1 Elemwise{add,no_inplace}
0.2% 99.6% 0.000s 2.83e-07s C 1000 1 Elemwise{Composite{Cast{int8}(GT(i0, i1))}}
0.1% 99.7% 0.000s 2.58e-07s C 1000 1 Subtensor{:int64:}
0.1% 99.8% 0.000s 2.28e-07s C 1000 1 Elemwise{gt,no_inplace}
0.1% 99.9% 0.000s 1.50e-07s C 1000 1 Elemwise{Composite{inv(sqrt(inv(sqr(i0))))}}[(0, 0)]
0.1% 100.0% 0.000s 1.30e-07s C 1000 1 Elemwise{Composite{inv(sqr(i0))}}[(0, 0)]
... (remaining 0 Ops account for 0.00%(0.00s) of the runtime)
Apply
------
<% time> <sum %> <apply time> <time per call> <#call> <id> <Apply name>
87.9% 87.9% 0.153s 1.53e-04s 1000 14 Elemwise{Composite{Switch(Cast{int8}((GT(Composite{exp((i0 * i1))}(i0, i1), i2) * i3 * GT(inv(sqrt(Composite{exp((i0 * i1))}(i0, i1))), i2))), (((i4 + (i5 * log(((i6 * Composite{exp((i0 * i1))}(i0, i1)) / i7)))) - i8) - (i5 * i9 * log1p(((Composite{exp((i0 * i1))}(i0, i1) * i10) / i7)))), i11)}}(TensorConstant{(1,) of -2.0}, s, TensorConstant{(1,) of 0}, Elemwise{gt,no_inplace}.0, Elemwise{Composite{scalar_gammaln((i0 * i1))}}.0, TensorConstant{(1,)
3.4% 91.3% 0.006s 5.88e-06s 1000 18 Elemwise{Composite{Switch(i0, (i1 * ((-(i2 * sqr((i3 - i4)))) + i5)), i6)}}(Elemwise{Composite{Cast{int8}(GT(i0, i1))}}.0, TensorConstant{(1,) of 0.5}, Elemwise{Composite{inv(sqr(i0))}}[(0, 0)].0, Subtensor{int64::}.0, Subtensor{:int64:}.0, Elemwise{Composite{log((i0 * i1))}}.0, TensorConstant{(1,) of -inf})
2.9% 94.2% 0.005s 5.02e-06s 1000 19 Sum{acc_dtype=float64}(Elemwise{Composite{Switch(i0, (i1 * ((-(i2 * sqr((i3 - i4)))) + i5)), i6)}}.0)
2.9% 97.0% 0.005s 4.99e-06s 1000 17 Sum{acc_dtype=float64}(Elemwise{Composite{Switch(Cast{int8}((GT(Composite{exp((i0 * i1))}(i0, i1), i2) * i3 * GT(inv(sqrt(Composite{exp((i0 * i1))}(i0, i1))), i2))), (((i4 + (i5 * log(((i6 * Composite{exp((i0 * i1))}(i0, i1)) / i7)))) - i8) - (i5 * i9 * log1p(((Composite{exp((i0 * i1))}(i0, i1) * i10) / i7)))), i11)}}.0)
0.3% 97.3% 0.001s 5.51e-07s 1000 4 InplaceDimShuffle{x}(nu)
0.2% 97.6% 0.000s 4.05e-07s 1000 8 Elemwise{Composite{scalar_gammaln((i0 * i1))}}(TensorConstant{(1,) of 0.5}, InplaceDimShuffle{x}.0)
0.2% 97.8% 0.000s 3.76e-07s 1000 6 InplaceDimShuffle{x}(sigma)
0.2% 98.0% 0.000s 3.47e-07s 1000 15 Elemwise{Composite{(Switch(Cast{int8}(GE(i0, i1)), (i2 - (i3 * i0)), i4) + i5)}}[(0, 0)](nu, TensorConstant{0}, TensorConstant{-2.3025850929940455}, TensorConstant{0.1}, TensorConstant{-inf}, nu_log__)
0.2% 98.2% 0.000s 3.39e-07s 1000 1 Elemwise{exp,no_inplace}(sigma_log__)
0.2% 98.4% 0.000s 3.37e-07s 1000 5 Elemwise{Composite{(Switch(Cast{int8}(GE(i0, i1)), (i2 - (i3 * i0)), i4) + i5)}}(sigma, TensorConstant{0}, TensorConstant{3.912023005428146}, TensorConstant{50.0}, TensorConstant{-inf}, sigma_log__)
0.2% 98.6% 0.000s 3.35e-07s 1000 3 Subtensor{int64::}(s, Constant{1})
0.2% 98.7% 0.000s 3.30e-07s 1000 16 Elemwise{Composite{log((i0 * i1))}}(TensorConstant{(1,) of 0...4309189535}, Elemwise{Composite{inv(sqr(i0))}}[(0, 0)].0)
0.2% 98.9% 0.000s 3.08e-07s 1000 20 MakeVector{dtype='float64'}(__logp_sigma_log__, __logp_nu_log__, __logp_s, __logp_r)
0.2% 99.1% 0.000s 2.86e-07s 1000 7 Elemwise{add,no_inplace}(TensorConstant{(1,) of 1.0}, InplaceDimShuffle{x}.0)
0.2% 99.2% 0.000s 2.83e-07s 1000 12 Elemwise{Composite{Cast{int8}(GT(i0, i1))}}(Elemwise{Composite{inv(sqrt(inv(sqr(i0))))}}[(0, 0)].0, TensorConstant{(1,) of 0})
0.1% 99.4% 0.000s 2.58e-07s 1000 2 Subtensor{:int64:}(s, Constant{-1})
0.1% 99.5% 0.000s 2.33e-07s 1000 11 Elemwise{Composite{scalar_gammaln((i0 * i1))}}(TensorConstant{(1,) of 0.5}, Elemwise{add,no_inplace}.0)
0.1% 99.7% 0.000s 2.28e-07s 1000 9 Elemwise{gt,no_inplace}(InplaceDimShuffle{x}.0, TensorConstant{(1,) of 0})
0.1% 99.8% 0.000s 1.58e-07s 1000 0 Elemwise{exp,no_inplace}(nu_log__)
0.1% 99.8% 0.000s 1.50e-07s 1000 10 Elemwise{Composite{inv(sqrt(inv(sqr(i0))))}}[(0, 0)](InplaceDimShuffle{x}.0)
... (remaining 2 Apply instances account for 0.16%(0.00s) of the runtime)
Here are tips to potentially make your code run faster
(if you think of new ones, suggest them on the mailing list).
Test them first, as they are not guaranteed to always provide a speedup.
- Try the Theano flag floatX=float32
We don't know if amdlibm will accelerate this scalar op. scalar_gammaln
We don't know if amdlibm will accelerate this scalar op. scalar_gammaln
- Try installing amdlibm and set the Theano flag lib.amdlibm=True. This speeds up only some Elemwise operation.
# Profiling of the gradient call dlogp/dx
model.profile(pm.gradient(model.logpt, model.vars)).summary()
Function profiling
==================
Message: /home/amit/Work/OpenSourceDev/pymc3/pymc3/model.py:1192
Time in 1000 calls to Function.__call__: 2.495768e-01s
Time in Function.fn.__call__: 2.238679e-01s (89.699%)
Time in thunks: 2.018890e-01s (80.893%)
Total compile time: 1.827419e+00s
Number of Apply nodes: 47
Theano Optimizer time: 1.549069e+00s
Theano validate time: 8.029699e-03s
Theano Linker time (includes C, CUDA code generation/compiling): 2.464304e-01s
Import time 1.863251e-01s
Node make_thunk time 2.439311e-01s
Node Elemwise{Composite{Switch(i0, ((i1 * i2 * i3 * i4) / i5), i6)}}(Elemwise{Composite{Cast{int8}((GT(i0, i1) * i2 * GT(inv(sqrt(i0)), i1)))}}.0, TensorConstant{(1,) of 0.5}, Elemwise{add,no_inplace}.0, Elemwise{Composite{exp((i0 * i1))}}.0, TensorConstant{[1.0435322..54666e-07]}, Elemwise{Add}[(0, 1)].0, TensorConstant{(1,) of 0}) time 3.346968e-02s
Node Elemwise{switch,no_inplace}(Elemwise{Composite{Cast{int8}(GT(i0, i1))}}.0, TensorConstant{(1,) of 1.0}, TensorConstant{(1,) of 0.0}) time 3.334451e-02s
Node Shape_i{0}(s) time 2.912211e-02s
Node Elemwise{Composite{inv(sqr(i0))}}(sigma) time 2.812171e-02s
Node Elemwise{Composite{Switch(i0, (i1 * sqr((i2 - i3))), i4)}}(Elemwise{Composite{Cast{int8}(GT(i0, i1))}}.0, TensorConstant{(1,) of 0.5}, Subtensor{int64::}.0, Subtensor{:int64:}.0, TensorConstant{(1,) of 0}) time 1.561642e-02s
Time in all call to theano.grad() 2.058767e+00s
Time since theano import 9.911s
Class
---
<% time> <sum %> <apply time> <time per call> <type> <#call> <#apply> <Class name>
72.2% 72.2% 0.146s 5.83e-06s C 25000 25 theano.tensor.elemwise.Elemwise
17.6% 89.8% 0.036s 5.09e-06s C 7000 7 theano.tensor.elemwise.Sum
5.1% 94.9% 0.010s 5.13e-06s C 2000 2 theano.tensor.subtensor.IncSubtensor
1.7% 96.7% 0.003s 3.50e-06s C 1000 1 theano.tensor.basic.Join
1.4% 98.1% 0.003s 2.85e-06s C 1000 1 theano.tensor.basic.Alloc
0.9% 99.0% 0.002s 4.49e-07s C 4000 4 theano.tensor.elemwise.DimShuffle
0.4% 99.4% 0.001s 4.38e-07s C 2000 2 theano.tensor.subtensor.Subtensor
0.3% 99.7% 0.001s 3.29e-07s C 2000 2 theano.tensor.basic.Reshape
0.2% 99.9% 0.000s 3.31e-07s C 1000 1 theano.compile.ops.Shape_i
0.1% 100.0% 0.000s 1.12e-07s C 2000 2 theano.compile.ops.Rebroadcast
... (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>
20.5% 20.5% 0.041s 4.14e-05s C 1000 1 Elemwise{Composite{Switch(i0, (-log1p((i1 / i2))), i3)}}
17.6% 38.2% 0.036s 5.09e-06s C 7000 7 Sum{acc_dtype=float64}
16.6% 54.8% 0.034s 3.36e-05s C 1000 1 Elemwise{Composite{Cast{int8}((GT(i0, i1) * i2 * GT(inv(sqrt(i0)), i1)))}}
12.9% 67.7% 0.026s 2.59e-05s C 1000 1 Elemwise{Composite{exp((i0 * i1))}}
4.6% 72.2% 0.009s 2.30e-06s C 4000 4 Elemwise{switch,no_inplace}
3.8% 76.1% 0.008s 7.75e-06s C 1000 1 Elemwise{Composite{Switch(i0, (i1 * (i2 + ((i3 * i4 * i5 * i6) / i7))), i8)}}[(0, 6)]
3.3% 79.4% 0.007s 6.68e-06s C 1000 1 Elemwise{Composite{Switch(i0, ((i1 * i2 * i3 * i4) / i5), i6)}}
3.2% 82.5% 0.006s 6.44e-06s C 1000 1 IncSubtensor{InplaceInc;int64::}
2.4% 85.0% 0.005s 4.94e-06s C 1000 1 Elemwise{Composite{Switch(i0, (i1 * i2 * (i3 - i4)), i5)}}
2.3% 87.2% 0.005s 4.55e-06s C 1000 1 Elemwise{Composite{Switch(i0, (i1 * (i2 - i3)), i4)}}
2.2% 89.4% 0.004s 4.42e-06s C 1000 1 Elemwise{Composite{Switch(i0, (i1 * sqr((i2 - i3))), i4)}}
1.9% 91.3% 0.004s 3.83e-06s C 1000 1 IncSubtensor{InplaceInc;:int64:}
1.7% 93.1% 0.003s 3.50e-06s C 1000 1 Join
1.4% 94.5% 0.003s 2.85e-06s C 1000 1 Alloc
1.0% 95.5% 0.002s 2.09e-06s C 1000 1 Elemwise{mul,no_inplace}
0.9% 96.4% 0.002s 4.49e-07s C 4000 4 InplaceDimShuffle{x}
0.4% 96.8% 0.001s 8.92e-07s C 1000 1 Elemwise{Add}[(0, 1)]
0.4% 97.2% 0.001s 7.88e-07s C 1000 1 Elemwise{Composite{(Switch(Cast{int8}(GE(i0, i1)), (i2 * i0), i1) + i3 + (i4 * (((i5 * i6 * Composite{inv(Composite{(sqr(i0) * i0)}(i0))}(i7)) / i8) - (i9 * Composite{inv(Composite{(sqr(i0) * i0)}(i0))}(i7))) * (i10 ** i11) * inv(Composite{(sqr(i0) * i0)}(i0)) * i0))}}[(0, 0)]
0.4% 97.6% 0.001s 3.89e-07s C 2000 2 Elemwise{exp,no_inplace}
0.4% 98.0% 0.001s 7.70e-07s C 1000 1 Elemwise{Composite{(Switch(Cast{int8}(GE(i0, i1)), (i2 * i0), i1) + i3 + (i4 * i5 * psi((i4 * (i6 + i0))) * i0) + (i7 * i8) + (i4 * i9 * psi((i4 * i0)) * i0) + (i4 * i10 * i0) + i11)}}[(0, 0)]
... (remaining 11 Ops account for 2.00%(0.00s) of the runtime)
Apply
------
<% time> <sum %> <apply time> <time per call> <#call> <id> <Apply name>
20.5% 20.5% 0.041s 4.14e-05s 1000 19 Elemwise{Composite{Switch(i0, (-log1p((i1 / i2))), i3)}}(Elemwise{Composite{Cast{int8}((GT(i0, i1) * i2 * GT(inv(sqrt(i0)), i1)))}}.0, Elemwise{mul,no_inplace}.0, InplaceDimShuffle{x}.0, TensorConstant{(1,) of 0})
16.6% 37.2% 0.034s 3.36e-05s 1000 15 Elemwise{Composite{Cast{int8}((GT(i0, i1) * i2 * GT(inv(sqrt(i0)), i1)))}}(Elemwise{Composite{exp((i0 * i1))}}.0, TensorConstant{(1,) of 0}, Elemwise{gt,no_inplace}.0)
12.9% 50.0% 0.026s 2.59e-05s 1000 5 Elemwise{Composite{exp((i0 * i1))}}(TensorConstant{(1,) of -2.0}, s)
3.8% 53.9% 0.008s 7.75e-06s 1000 35 Elemwise{Composite{Switch(i0, (i1 * (i2 + ((i3 * i4 * i5 * i6) / i7))), i8)}}[(0, 6)](Elemwise{Composite{Cast{int8}((GT(i0, i1) * i2 * GT(inv(sqrt(i0)), i1)))}}.0, TensorConstant{(1,) of -2.0}, TensorConstant{(1,) of 0.5}, TensorConstant{(1,) of -0.5}, Elemwise{add,no_inplace}.0, TensorConstant{[1.0435322..54666e-07]}, Elemwise{Composite{exp((i0 * i1))}}.0, Elemwise{Add}[(0, 1)].0, TensorConstant{(1,) of 0})
3.3% 57.2% 0.007s 6.68e-06s 1000 32 Elemwise{Composite{Switch(i0, ((i1 * i2 * i3 * i4) / i5), i6)}}(Elemwise{Composite{Cast{int8}((GT(i0, i1) * i2 * GT(inv(sqrt(i0)), i1)))}}.0, TensorConstant{(1,) of 0.5}, Elemwise{add,no_inplace}.0, Elemwise{Composite{exp((i0 * i1))}}.0, TensorConstant{[1.0435322..54666e-07]}, Elemwise{Add}[(0, 1)].0, TensorConstant{(1,) of 0})
3.2% 60.4% 0.006s 6.44e-06s 1000 38 IncSubtensor{InplaceInc;int64::}(Elemwise{Composite{Switch(i0, (i1 * (i2 + ((i3 * i4 * i5 * i6) / i7))), i8)}}[(0, 6)].0, Elemwise{Composite{Switch(i0, (i1 * i2 * (i3 - i4)), i5)}}.0, Constant{1})
2.6% 62.9% 0.005s 5.17e-06s 1000 37 Sum{acc_dtype=float64}(Alloc.0)
2.5% 65.5% 0.005s 5.13e-06s 1000 36 Sum{acc_dtype=float64}(Elemwise{Composite{Switch(i0, ((i1 * i2 * i3 * i4) / i5), i6)}}.0)
2.5% 68.0% 0.005s 5.08e-06s 1000 31 Sum{acc_dtype=float64}(Elemwise{Switch}.0)
2.5% 70.5% 0.005s 5.08e-06s 1000 33 Sum{acc_dtype=float64}(Elemwise{Composite{Switch(i0, (i1 * sqr((i2 - i3))), i4)}}.0)
2.5% 73.0% 0.005s 5.08e-06s 1000 28 Sum{acc_dtype=float64}(Elemwise{Composite{Switch(i0, (-log1p((i1 / i2))), i3)}}.0)
2.5% 75.5% 0.005s 5.04e-06s 1000 29 Sum{acc_dtype=float64}(Elemwise{switch,no_inplace}.0)
2.5% 78.0% 0.005s 5.03e-06s 1000 30 Sum{acc_dtype=float64}(Elemwise{switch,no_inplace}.0)
2.4% 80.4% 0.005s 4.94e-06s 1000 25 Elemwise{Composite{Switch(i0, (i1 * i2 * (i3 - i4)), i5)}}(Elemwise{Composite{Cast{int8}(GT(i0, i1))}}.0, TensorConstant{(1,) of -1.0}, InplaceDimShuffle{x}.0, Subtensor{int64::}.0, Subtensor{:int64:}.0, TensorConstant{(1,) of 0})
2.3% 82.7% 0.005s 4.55e-06s 1000 24 Elemwise{Composite{Switch(i0, (i1 * (i2 - i3)), i4)}}(Elemwise{Composite{Cast{int8}(GT(i0, i1))}}.0, InplaceDimShuffle{x}.0, Subtensor{int64::}.0, Subtensor{:int64:}.0, TensorConstant{(1,) of 0})
2.2% 84.9% 0.004s 4.42e-06s 1000 26 Elemwise{Composite{Switch(i0, (i1 * sqr((i2 - i3))), i4)}}(Elemwise{Composite{Cast{int8}(GT(i0, i1))}}.0, TensorConstant{(1,) of 0.5}, Subtensor{int64::}.0, Subtensor{:int64:}.0, TensorConstant{(1,) of 0})
1.9% 86.8% 0.004s 3.83e-06s 1000 41 IncSubtensor{InplaceInc;:int64:}(IncSubtensor{InplaceInc;int64::}.0, Elemwise{Composite{Switch(i0, (i1 * (i2 - i3)), i4)}}.0, Constant{-1})
1.7% 88.5% 0.003s 3.50e-06s 1000 46 Join(TensorConstant{0}, Rebroadcast{1}.0, Rebroadcast{1}.0, (d__logp/ds))
1.6% 90.1% 0.003s 3.17e-06s 1000 22 Elemwise{Switch}(Elemwise{Composite{Cast{int8}((GT(i0, i1) * i2 * GT(inv(sqrt(i0)), i1)))}}.0, TensorConstant{(1,) of 1.0}, TensorConstant{(1,) of 0.0})
1.4% 91.5% 0.003s 2.92e-06s 1000 21 Elemwise{switch,no_inplace}(Elemwise{Composite{Cast{int8}((GT(i0, i1) * i2 * GT(inv(sqrt(i0)), i1)))}}.0, TensorConstant{(1,) of -0..4309189535}, TensorConstant{(1,) of 0})
... (remaining 27 Apply instances account for 8.47%(0.02s) of the runtime)
Here are tips to potentially make your code run faster
(if you think of new ones, suggest them on the mailing list).
Test them first, as they are not guaranteed to always provide a speedup.
- Try the Theano flag floatX=float32
We don't know if amdlibm will accelerate this scalar op. psi
We don't know if amdlibm will accelerate this scalar op. psi
- Try installing amdlibm and set the Theano flag lib.amdlibm=True. This speeds up only some Elemwise operation.
%load_ext watermark
%watermark -n -u -v -iv -w
pymc3 3.9.3
numpy 1.18.5
pandas 1.0.5
last updated: Sun Aug 30 2020
CPython 3.8.3
IPython 7.16.1
watermark 2.0.2