Граф коммитов

56 Коммитов

Автор SHA1 Сообщение Дата
Evan Shelhamer 0816a94803 fix wrapper example paths 2014-03-13 20:15:27 -07:00
Jonathan L Long 9c1ceb6bd5 pycaffe: expose ForwardPrefilled
This allows, for example, a net taking input from a LevelDB to be run
through a Python script.
2014-03-10 22:38:14 -07:00
Evan Shelhamer fffae6cd69 fix include order for pycaffe on osx, override lint 2014-02-27 15:53:32 -08:00
Jeff Donahue 52a539dcd1 fix python/matlab wrapper bugs introduced by lint; change linter->lint
in Makefile
2014-02-26 16:54:25 -08:00
Jeff Donahue 4b8fe9bf58 fix matcaffe and pycaffe linter errors 2014-02-26 15:42:38 -08:00
Evan Shelhamer f3edc7c504 name blobs and params for their layers in python wrapper
note that pairs of params with the same layer name are the
params & bias
2014-02-26 15:42:37 -08:00
Evan Shelhamer 6cc0c47641 move imagenet splits + synsets to data 2014-02-26 12:37:43 -08:00
Evan Shelhamer 2092fbc5b0 Properly index windows in detector list mode
Fix iteration over single windows: whether one or more windows are
listed for an image, generate windows as a list.
2014-02-23 15:09:38 -08:00
Evan Shelhamer 5bd5ebfcd0 fix detector's coordinate mapping for images smaller than IMAGE_DIM 2014-02-22 15:42:58 -08:00
Evan Shelhamer 5aa2be8001 python requirements: don't fear the future 2014-02-14 18:14:56 -08:00
Evan Shelhamer 4f49004c28 add shebang to python scripts
for executing in the shell without call to python. Follows the example
of @kloudkl .
2014-02-14 15:40:38 -08:00
Evan Shelhamer 73c10a1ae1 add macro for numpy < 1.7
as suggested by @longjon
https://github.com/BVLC/caffe/issues/44#issuecomment-32832810
2014-02-14 15:33:00 -08:00
Sergey Karayev 779c571705 detector.py refactored with argparse (was gflags) 2014-02-13 01:21:43 -08:00
Sergey Karayev 99ecba3a82 Updated detection demo notebook.
- Works with new detector.py
- Everything is run automatically from cells.
2014-02-13 01:03:35 -08:00
Sergey Karayev 45f213032b moving selective_search_demo notebook to examples/ 2014-02-12 23:58:21 -08:00
Evan Shelhamer e89a4b5da6 include pip requirements.txt for python deps 2014-02-11 00:41:01 -08:00
Evan Shelhamer 1c5df72b03 read single input, load/save csv, and record windows
Load `--input_file` as a list of image filenames if .txt OR read as CSV
with fields filename,ymin,xmin,ymax,xmax (with labeled header) if .csv.

Save `--output_file` as HDF5 if .h5 or CSV if .csv. For CSV, enumerate
the class probabilities as numbered class fields.

Record crop windows in original image coordinates if `center_only` or
`corners` crop mode selected. Previously, these modes didn't report
locations.
2014-01-31 01:23:29 -08:00
Evan Shelhamer 4db88bbe2f remove outdated reference to input size, debugging print 2014-01-31 00:15:07 -08:00
Evan Shelhamer 72758c7dde align detector comments to reality 2014-01-28 18:09:13 -08:00
Evan Shelhamer 2f71deb751 detect by window list
Do window detection on a list of predefined windows by adding a
crop_mode `list` that accepts a path to a text file of format:

    image_filename xmin ymin xmax ymax
2014-01-27 15:28:18 -08:00
Evan Shelhamer 47f8f53606 replace magic numbers with variable names in comments 2014-01-27 15:17:45 -08:00
Evan Shelhamer ec010d549e automagically set detection batch size from network 2014-01-25 20:53:00 -08:00
Evan Shelhamer 6408ac1098 document power_wrapper -> detector in detection notebook 2014-01-25 20:03:00 -08:00
Evan Shelhamer 3ff61773f0 promote power_wrapper to 'detection' submodule 2014-01-24 21:34:47 -08:00
Evan Shelhamer 7b562d573d groom power_wrapper flags (cont'd) 2014-01-24 21:34:44 -08:00
Evan Shelhamer c3c635d2ab note power_wrapper TODOs 2014-01-24 20:52:39 -08:00
Evan Shelhamer 69671dd30b generalize power_wrapper to different networks and inputs
generalize from the imagenet reference to other detection networks:
- refactor to configure for a given net and data source
- determine dimensions of net input and output automatically from blobs
2014-01-24 20:52:38 -08:00
Evan Shelhamer 2f7fbfa9e4 groom power_wrapper flags 2014-01-24 20:52:38 -08:00
Evan Shelhamer e7fc77b389 python lint 2014-01-24 20:52:38 -08:00
Evan Shelhamer eafd6950bc give batch size efficiency advice 2014-01-23 22:31:29 -08:00
Evan Shelhamer 96da93c8a6 default power_wrapper batch size to 10 (aeca741a69 cont'd)
default command line arg and function arg to 10
remove global BATCH_SIZE in favor of arg
2014-01-23 22:31:29 -08:00
Evan Shelhamer f8039bc22b Merge pull request #11 from longjon/master
Python interface to blobs and blob data through boost::python
2014-01-19 21:17:43 -08:00
forresti aeca741a69 setting default power_wrapper batch size to match the imagenet_deploy prototxt 2014-01-15 12:02:12 -08:00
Jonathan L Long ae72b9e870 Expose params in Python interface 2014-01-14 15:46:52 -08:00
Jonathan L Long 6e18b00492 Add Python interface to layer blobs 2014-01-14 15:46:51 -08:00
Sergey Karayev 4eca379c2f updated detection demo notebook with picture of two cats 2013-12-06 11:29:02 -08:00
Sergey Karayev 837633b0d8 minor edit 2013-12-05 19:35:32 -08:00
Sergey Karayev 5c0693e30a selective search notebook and renaming to power_wrapper 2013-12-05 19:33:13 -08:00
Sergey Karayev c850cc06c7 fixed bug that renormalized window crops on second resize 2013-12-05 19:12:50 -08:00
Sergey Karayev b1aff6dfc3 center_only and corners modes work correctly 2013-12-05 17:21:47 -08:00
Sergey Karayev 5e2823f3f7 processing images in batch, with option to use selective search window proposals 2013-12-05 16:38:36 -08:00
Ross Girshick 41617345db Merge remote-tracking branch 'origin/master' 2013-11-22 14:03:55 -08:00
Yangqing Jia 2b981cf3dc fix Makefile problem 2013-11-21 17:00:57 -08:00
Ross Girshick bcba5961a1 fixed minor typo in imagenet wrapper 2013-11-21 15:45:04 -08:00
Yangqing Jia e08801e728 imagenet_solver.prototxt: added back the missing test line, not sure when it slipped off the repo 2013-11-19 08:51:43 -08:00
Yangqing Jia e24b1dd22b Made a major change: when initializing a network, the input size are no longer provided by an additional vector of blobs, but should be specified in the netparameters proto by the field "input_dim". This avoid the often awkward code of creating a dummy input vector just for the sake of initializing the network. 2013-11-15 16:58:53 -08:00
Yangqing Jia f54e9779ef wrapper update 2013-11-15 08:02:48 -08:00
Yangqing Jia fa7b43062b pycaffe: added a temporary numpy 1.6 compile solution 2013-11-14 14:51:27 -08:00
Yangqing Jia a5f2611988 misc update 2013-11-14 14:33:24 -08:00
Yangqing Jia ff9c69c77c pycaffe update and imagenet wrapper 2013-11-13 16:31:40 -08:00