fully convolutional detection encoder #718
Conversation
Codecov Report
@@ Coverage Diff @@
## master #718 +/- ##
==========================================
+ Coverage 83.82% 83.84% +0.02%
==========================================
Files 27 27
Lines 3839 3838 -1
==========================================
Hits 3218 3218
+ Misses 621 620 -1
Flags with carried forward coverage won't be shown. Click here to find out more.
📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more |
There was a problem hiding this comment.
Thanks @jeff-regier for this PR. In terms of changes, there is only very minor ones in one of the config files. The other comments are just questions which you can feel free to delay until the meeting today.
There was a problem hiding this comment.
@jeff-regier I think you can merge now. Feel free to delete the checkerboard later or not depending on your experiments.
|
The checkerboard wasn't helpful/necessary after all. Deleted. I fixed the config files too. |
|
I improved the learning rate schedule for the new architecture. (Dropped one of the milestones.) Here are results for the old (grey) and new (red) encoders, both trained for 24 hours. The new encoder finishes 4x as many iterations during that time, but even per iteration it appears to have an edge.
|
This PR adds a case study that demonstrates a new fully convolutional encoder architecture. Because it doesn't form padded tiles, this encoder is 4 times than our current detection encoder (>6 training iterations per second, whereas the old detection encoder performed about 1.5 iterations / second). The graphs below suggest that it is also more accurate. The old encoder, which was trained for 24 hours, appears in grey. The new encoder, which was trained for 6 hours, appears in red.