-
- move label processing up to setup section - also rerun the first image through the trained net
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-make nets using net spec -solvers created inline as well -two-stage finetuning (first fc8 only, then end-to-end) -show a couple sample results -
-
-
-
Subheadings and minor edits.
-
-
This pass adds subheadings, cleans up the text, adds comments and tries to make the code more understandable, and adds a new section summarizing steps and encouraging the user to try their own image.
-
- include net surgery in sequence - drop detection entirely since it is not standard, and it is better to favor the true R-CNN / fast R-CNN / faster R-CNN - make new and complete editions of the notebooks so that users need to execute the examples to see output
-
Remove precreated solvers
-
Nicely prints gpu names
-
-
Include proto definition in distribute dir
-
use relative paths on making build/tools/ links
-
The old uses `abspath`, which I think is so harmful: * If I `cp -a` the whole project, `build/tools/caffe` still refer to the old file, until `make clean`, making debugging very hard * For `tar` and `scp`, the soft links can not work unless the target project folder has the same path
