Created
April 30, 2018 16:56
-
-
Save wandering007/d7f3db11285f382fefa39a9cda99ad0d to your computer and use it in GitHub Desktop.
benchmark memory-efficient densenet
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
import numpy as np | |
import sys | |
import time | |
import torch | |
from torch.autograd import Variable | |
import torchvision.models as models | |
import torch.backends.cudnn as cudnn | |
from models import DenseNet | |
import torch.nn as nn | |
def measure(model, x, y): | |
# synchronize gpu time and measure fp | |
torch.cuda.synchronize() | |
t0 = time.time() | |
y_pred = model(x) | |
torch.cuda.synchronize() | |
elapsed_fp = time.time() - t0 | |
# zero gradients, synchronize time and measure | |
model.zero_grad() | |
t0 = time.time() | |
y_pred.backward(y) | |
torch.cuda.synchronize() | |
elapsed_bp = time.time() - t0 | |
return elapsed_fp, elapsed_bp | |
def benchmark(model, x, y): | |
# transfer the model on GPU | |
model.cuda() | |
# DRY RUNS | |
for i in range(5): | |
_, _ = measure(model, x, y) | |
print('DONE WITH DRY RUNS, NOW BENCHMARKING') | |
# START BENCHMARKING | |
t_forward = [] | |
t_backward = [] | |
for i in range(10): | |
t_fp, t_bp = measure(model, x, y) | |
t_forward.append(t_fp) | |
t_backward.append(t_bp) | |
# free memory | |
del model | |
return t_forward, t_backward | |
multigpus = True | |
# set cudnn backend to benchmark config | |
cudnn.benchmark = True | |
# instantiate the models | |
densnet_effi = DenseNet(efficient=True) | |
if multigpus: | |
densnet_effi = nn.DataParallel(densnet_effi, device_ids=[0, 1]) | |
# build dummy variables to input and output | |
x = torch.randn(128, 3, 32, 32).cuda() | |
y = torch.randn(128, 10).cuda() | |
# loop over architectures and measure them | |
t_fp, t_bp = benchmark(densnet_effi, x, y) | |
# print results | |
print('FORWARD PASS: ', np.mean(np.asarray(t_fp) * 1e3), '+/-', np.std(np.asarray(t_fp) * 1e3)) | |
print('BACKWARD PASS: ', np.mean(np.asarray(t_bp) * 1e3), '+/-', np.std(np.asarray(t_bp) * 1e3)) | |
print('RATIO BP/FP:', np.mean(np.asarray(t_bp)) / np.mean(np.asarray(t_fp))) |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment