What are the bottlenecks of the code? Where are things slow? What can we optimiz

Resolved by PR <a class="issue-link js-issue-link" data-error-text="Failed to load tit

cProfiling about better-sampling HOT 4 CLOSED

zafarali commented on June 1, 2024

cProfiling

from better-sampling.

Comments (4)

arjunkaruvally commented on June 1, 2024

This is for RVISampler that is taking the most time in all the other samplers

ncalls | tottime | percall | cumtime | percall | filename:lineno(function)
1000 | 5.279 | 0.005279 | 5.279 | 0.005279 | ~:0(<method 'run_backward' of 'torch._C._EngineBase' objects>)
1 | 2.654 | 2.654 | 28.69 | 28.69 | RVISampler.py:53(solve)
49000 | 1.168 | 2.383e-05 | 1.73 | 3.531e-05 | baselines.py:40(__call__)
147000 | 1.111 | 7.555e-06 | 1.111 | 7.555e-06 | ~:0(<built-in method torch._C.addmm>)
49000 | 0.7808 | 1.593e-05 | 0.7808 | 1.593e-05 | ~:0(<built-in method torch._C._nn.softmax>)
294000/49000 | 0.7505 | 1.532e-05 | 7.852 | 0.0001602 | module.py:351(__call__)
99000 | 0.7432 | 7.507e-06 | 1.436 | 1.451e-05 | ~:0(<built-in method apply>)
172522 | 0.7212 | 4.18e-06 | 0.7212 | 4.18e-06 | ~:0(<built-in method numpy.core.multiarray.array>)
49000 | 0.6603 | 1.348e-05 | 1.536 | 3.134e-05 | distributions.py:147(log_prob)
49000 | 0.6353 | 1.297e-05 | 1.945 | 3.97e-05 | random_walk.py:86(step)
49000 | 0.5935 | 1.211e-05 | 4.842 | 9.882e-05 | base.py:92(step)
249000 | 0.521 | 2.092e-06 | 0.521 | 2.092e-06 | ~:0(<method 'copy_' of 'torch._C.FloatTensorBase' objects>)
394951 | 0.5169 | 1.309e-06 | 1.429 | 3.618e-06 | _utils.py:6(_type)
6500 | 0.4307 | 6.627e-05 | 0.4953 | 7.62e-05 | analytic_posterior.py:48(pdf)
690000 | 0.3995 | 5.79e-07 | 0.548 | 7.943e-07 | variable.py:64(__getattr__)
55020 | 0.3922 | 7.129e-06 | 0.3922 | 7.129e-06 | ~:0(<method 'reduce' of 'numpy.ufunc' objects>)
147000 | 0.3843 | 2.614e-06 | 0.3843 | 2.614e-06 | ~:0(<method 't' of 'torch._C._VariableBase' objects>)
1234461 | 0.3797 | 3.076e-07 | 0.3797 | 3.076e-07 | ~:0(<built-in method builtins.getattr>)
147951 | 0.3473 | 2.348e-06 | 0.3473 | 2.348e-06 | ~:0(<built-in method torch._C.from_numpy>)
147000 | 0.3407 | 2.318e-06 | 2.497 | 1.699e-05 | linear.py:54(forward)
247000 | 0.3313 | 1.341e-06 | 0.6787 | 2.748e-06 | _utils.py:83(_import_dotted_name)
147000 | 0.3151 | 2.144e-06 | 2.013 | 1.369e-05 | functional.py:822(linear)
99000 | 0.3033 | 3.063e-06 | 0.3033 | 3.063e-06 | ~:0(<built-in method torch._C.ones>)
49000 | 0.2929 | 5.978e-06 | 7.607 | 0.0001553 | policies.py:49(forward)
149000 | 0.2749 | 1.845e-06 | 0.2749 | 1.845e-06 | ~:0(<method 'view' of 'torch._C.FloatTensorBase' objects>)
49000 | 0.2474 | 5.048e-06 | 3.471 | 7.083e-05 | container.py:65(forward)```

from better-sampling.

zafarali commented on June 1, 2024

Ok so as expected backward() takes the most time. Let's leave that out for now.
Doesn't doesn't look like we can make that faster though. torch.ones() call is necessary to make the output the same size.

Places to look:
random_walk.step, analytic_posterior.pdf Like you identified this line keeps creating new lists everytime it is called

better-sampling/rvi_sampling/distributions/analytic_posterior.py

Lines 57 to 60 in b40cd7d

    
           if not (x in range(-c, c+1, 1) and term in range(0, T+1)): 
        
               return 0 
        
           else: 
        
               return (1/(2*c))*comb(T, term)*(p**term)*((1-p)**(T-term))

A quick way to make this faster is to use self.support which already has created range(c, c+1)
The logic for the conditional term in range(0, T+1)) is that i need it to be <T+1 and >0 but also an integer. Maybe a way to check this is to check if it's an integer then we do not need to create the list.

I'm not sure where random_walk.step can be improved but given that it called analytic_posterior.pdf improving that should make it faster as well.

from better-sampling.

arjunkaruvally commented on June 1, 2024

My thoughts exactly...

from better-sampling.

arjunkaruvally commented on June 1, 2024

Resolved by PR #22

Profilers are attached and some performance issues are fixed

from better-sampling.

cProfiling about better-sampling HOT 4 CLOSED

Comments (4)

Related Issues (14)

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent

	if not (x in range(-c, c+1, 1) and term in range(0, T+1)):
	return 0
	else:
	return (1/(2c))comb(T, term)(pterm)((1-p)**(T-term))