This commit is contained in:
Dev Singh 2020-03-06 11:50:07 -06:00
parent eb8914aa26
commit d38744438b
3 changed files with 51 additions and 46 deletions

View File

@ -7,10 +7,12 @@
# current benchmark of optimization: 1.33 times faster
# setup:
__version__ = "1.1.13.001"
__version__ = "1.1.13.002"
# changelog should be viewed using print(analysis.__changelog__)
__changelog__ = """changelog:
1.1.13.002:
- removed torch requirement, and moved Regression back to regression.py
1.1.13.001:
- bug fix with linear regression not returning a proper value
- cleaned up regression

View File

@ -1,27 +1,28 @@
# Titan Robotics Team 2022: CUDA-based Regressions Module
# Written by Arthur Lu & Jacob Levine
# Notes:
# this should be imported as a python module using 'import regression'
# this should be included in the local directory or environment variable
# this module is cuda-optimized and vectorized (except for one small part)
# this module has been automatically inegrated into analysis.py, and should be callable as a class from the package
# this module is cuda-optimized and vectorized (except for one small part)
# setup:
__version__ = "1.0.0.002"
__version__ = "1.0.0.003"
# changelog should be viewed using print(regression.__changelog__)
# changelog should be viewed using print(analysis.regression.__changelog__)
__changelog__ = """
1.0.0.002:
-Added more parameters to log, exponential, polynomial
-Added SigmoidalRegKernelArthur, because Arthur apparently needs
to train the scaling and shifting of sigmoids
1.0.0.001:
-initial release, with linear, log, exponential, polynomial, and sigmoid kernels
-already vectorized (except for polynomial generation) and CUDA-optimized
1.0.0.003:
- bug fixes
1.0.0.002:
-Added more parameters to log, exponential, polynomial
-Added SigmoidalRegKernelArthur, because Arthur apparently needs
to train the scaling and shifting of sigmoids
1.0.0.001:
-initial release, with linear, log, exponential, polynomial, and sigmoid kernels
-already vectorized (except for polynomial generation) and CUDA-optimized
"""
__author__ = (
"Jacob Levine <jlevine@imsa.edu>",
"Arthur Lu <learthurgo@gmail.com>"
)
__all__ = [
@ -39,35 +40,13 @@ __all__ = [
'CustomTrain'
]
# imports (just one for now):
import torch
global device
device = "cuda:0" if torch.torch.cuda.is_available() else "cpu"
#todo: document completely
def factorial(n):
if n==0:
return 1
else:
return n*factorial(n-1)
def num_poly_terms(num_vars, power):
if power == 0:
return 0
return int(factorial(num_vars+power-1) / factorial(power) / factorial(num_vars-1)) + num_poly_terms(num_vars, power-1)
def take_all_pwrs(vec,pwr):
#todo: vectorize (kinda)
combins=torch.combinations(vec, r=pwr, with_replacement=True)
out=torch.ones(combins.size()[0])
for i in torch.t(combins):
out *= i
return torch.cat(out,take_all_pwrs(vec, pwr-1))
def set_device(new_device):
global device
def set_device(self, new_device):
device=new_device
class LinearRegKernel():
@ -154,20 +133,39 @@ class PolyRegKernel():
power=None
def __init__(self, num_vars, power):
self.power=power
num_terms=num_poly_terms(num_vars, power)
num_terms=self.num_poly_terms(num_vars, power)
self.weights=torch.rand(num_terms, requires_grad=True, device=device)
self.bias=torch.rand(1, requires_grad=True, device=device)
self.parameters=[self.weights,self.bias]
def num_poly_terms(self,num_vars, power):
if power == 0:
return 0
return int(self.factorial(num_vars+power-1) / self.factorial(power) / self.factorial(num_vars-1)) + self.num_poly_terms(num_vars, power-1)
def factorial(self,n):
if n==0:
return 1
else:
return n*self.factorial(n-1)
def take_all_pwrs(self, vec, pwr):
#todo: vectorize (kinda)
combins=torch.combinations(vec, r=pwr, with_replacement=True)
out=torch.ones(combins.size()[0]).to(device).to(torch.float)
for i in torch.t(combins).to(device).to(torch.float):
out *= i
if pwr == 1:
return out
else:
return torch.cat((out,self.take_all_pwrs(vec, pwr-1)))
def forward(self,mtx):
#TODO: Vectorize the last part
cols=[]
for i in torch.t(mtx):
cols.append(take_all_pwrs(i,self.power))
cols.append(self.take_all_pwrs(i,self.power))
new_mtx=torch.t(torch.stack(cols))
long_bias=self.bias.repeat([1,mtx.size()[1]])
return torch.matmul(self.weights,new_mtx)+long_bias
def SGDTrain(kernel, data, ground, loss=torch.nn.MSELoss(), iterations=1000, learning_rate=.1, return_losses=False):
def SGDTrain(self, kernel, data, ground, loss=torch.nn.MSELoss(), iterations=1000, learning_rate=.1, return_losses=False):
optim=torch.optim.SGD(kernel.parameters, lr=learning_rate)
data_cuda=data.to(device)
ground_cuda=ground.to(device)
@ -192,7 +190,7 @@ def SGDTrain(kernel, data, ground, loss=torch.nn.MSELoss(), iterations=1000, lea
optim.step()
return kernel
def CustomTrain(kernel, optim, data, ground, loss=torch.nn.MSELoss(), iterations=1000, return_losses=False):
def CustomTrain(self, kernel, optim, data, ground, loss=torch.nn.MSELoss(), iterations=1000, return_losses=False):
data_cuda=data.to(device)
ground_cuda=ground.to(device)
if (return_losses):
@ -214,4 +212,4 @@ def CustomTrain(kernel, optim, data, ground, loss=torch.nn.MSELoss(), iterations
ls=loss(pred,ground_cuda)
ls.backward()
optim.step()
return kernel
return kernel

View File

@ -3,11 +3,16 @@
# Notes:
# setup:
__version__ = "0.0.3.000"
__version__ = "0.0.4.000"
# changelog should be viewed using print(analysis.__changelog__)
__changelog__ = """changelog:
0.0.3.00:
0.0.4.000:
- fixed spelling issue in __changelog__
- addressed nan bug in regression
- fixed errors on line 335 with metrics calling incorrect key "glicko2"
- fixed errors in metrics computing
0.0.3.000:
- added analysis to pit data
0.0.2.001:
- minor stability patches
@ -124,7 +129,7 @@ def main():
d.set_analysis_flags(apikey, "latest_update", {"latest_update":current_time})
print(" pushing to database")
push_to_database(apikey, competition, results, metrics, pit)
push_to_database(apikey, competition, results, pit)
print(" pushed to database")
def load_config(file):