search

libornovax / master_thesis_code

Code for my master thesis: Vehicle Detection and Pose Estimation for Autonomous Driving
MIT License
187 stars 69 forks source link

Stride 2 on x1 to x2 instead of pooling #36

Closed libornovax closed 7 years ago

libornovax commented 7 years ago

I want to do this because it saves a lot of GPU memory and I don't think it should affect the result...

libornovax commented 7 years ago

Memory consumption of net macc_0.3_r2_x2_to_x8_s2 vs macc_0.3_r2_x2_to_x8:

TEST: 9046 MiB vs. 11610 MiB, TRAIN: 8760 MiB vs. 10666 MiB

libornovax commented 7 years ago

I can happily say that it does not affect performance in a negative way. The net even performs slightly better, but that is probably just a coincidence:

With pooling after x1: pr_jura_test_short_fix

With stride=2 of the last conv layer in x1: pr_jura_test_short_fix

This is great news, because the net consumes way less memory and it is also faster to compute!