microsoft
/
caffe
зеркало из https://github.com/microsoft/caffe.git
1
0
Форкнуть 0
Код Задачи Пакеты Проекты Релизы Вики Активность
1 764 коммитов 3 Ветки 11 Теги 51 MiB
Граф коммитов

141 Коммитов

Автор SHA1 Сообщение Дата
Evan Shelhamer f8b612d8cd default raw_scale in python scripts to ImageNet model value 2014-08-09 20:43:40 -07:00
Evan Shelhamer b97b88f0c2 LICENSE governs the whole project so strip file headers 2014-08-06 18:35:29 -07:00
Evan Shelhamer 0db94786a7 drop np.asarray() in favor of declaration (~1.75x speedup) 2014-08-05 23:17:59 -07:00
Evan Shelhamer 4f7726916c fix pycaffe context cropping with or without mean 2014-08-05 23:17:59 -07:00
Evan Shelhamer e1b3413708 take array in pycaffe `Net.set_mean()` instead of file path 2014-08-05 23:17:59 -07:00
Evan Shelhamer d5c3cef471 fix pycaffe input processing 
- load an image as [0,1] single / np.float32 according to Python convention
- fix input scaling during preprocessing:
  - scale input for preprocessing by `raw_scale` e.g. to map an image
    to [0, 255] for the CaffeNet and AlexNet ImageNet models
  - scale feature space by `input_scale` after mean subtraction
  - switch examples to raw scale for ImageNet models
  - fix #525
- preserve type after resizing.
- resize 1, 3, or K channel images with special casing between
  skimage.transform (1 and 3) and scipy.ndimage (K) for speed
 2014-08-05 23:17:59 -07:00
Evan Shelhamer 99f27fc082 define caffe.Net input preprocessing members by boost::python 
define `Net.{mean, input_scale, channel_swap}` on the boost::python side
so that the members always exist. drop ugly initialization logic.
 2014-08-05 15:55:45 -07:00
Evan Shelhamer d842f4a24f Merge pull request #733 from longjon/pycaffe-tweaks 
pycaffe fixes
 2014-07-28 14:14:03 -07:00
Jonathan L Long fb2f7c1c27 pycaffe: test channel_order and input_scale against None 
An input_scale of zero should produce warnings/errors/infinities rather
than being the same as an input_scale of one. A channel_order of () or 0
should not be equivalent to the default ordering.
 2014-07-18 17:08:49 -07:00
Jonathan L Long 32fefa96f9 pycaffe: reorder exceptions 
It doesn't make sense to complain about input not being batch-sized if
it isn't even 4-d in the first place.
 2014-07-18 17:08:49 -07:00
Jonathan L Long fc48d5de5c pycaffe: allow unspecified mean. Fixes #671. 2014-07-18 17:08:49 -07:00
Jonathan L Long 3ef43e41e7 pycaffe: expose Forward/Backward From/To as kwargs start and end 2014-07-18 15:39:55 -07:00
Evan Shelhamer 506e476047 Merge pull request #455 from shelhamer/pycaffe-save 
Save from python for net surgery
 2014-06-12 16:00:49 -07:00
Evan Shelhamer 6a5b2c19c7 save from python for net surgery 
0. Scheme desired parameters.
1. Do surgery on the net through `net.params['name'][idx].data[...] = `.
2. Save post-operation net params by `net.save('fname')`.

Handwoven deep nets, anyone?
 2014-06-12 14:27:36 -07:00
Evan Shelhamer a7e397abbd Merge pull request #482 from shelhamer/rcnn-detector-example 
Make R-CNN the Caffe detection example
 2014-06-11 15:22:04 -07:00
Evan Shelhamer 8dc270e691 pycaffe: leave grayscale images gray according to arg 2014-06-11 10:17:17 -07:00
Evan Shelhamer 206bf86872 make selective search proposals with R-CNN configuration 2014-06-09 21:06:18 -07:00
Evan Shelhamer 4992abecec pycaffe Detector crops with surrounding context 
- caffe.Detector learned how to crop windows with context in the R-CNN
  style s.t. the bordero of the network input is a given amount of
  context.
- add --context_pad arg to detect.py for amount of context. Default is
  16, as in R-CNN.
 2014-06-09 18:14:25 -07:00
Evan Shelhamer a57b8d5b40 fix old detect.py default 2014-06-09 18:14:25 -07:00
Evan Shelhamer 8c609da736 caffe.Net preprocessing members belong to object, not class 2014-05-26 21:50:39 -07:00
Evan Shelhamer 211c3c29bb follow-up on #443 to invert k channels (instead of 3) 2014-05-23 09:21:17 -07:00
James Thewlis e5fcf030fb Correctly invert the swapping of colour channels 
In the 'deprocess' method, get back the image with the original channel order
by inverting the original transform, rather than reversing the tuple which is
incorrect.
 2014-05-23 09:21:17 -07:00
Evan Shelhamer 1cb6feaf5a fix draw_net python script 
include caffe.draw for drawing functions.
 2014-05-22 01:02:56 -07:00
Evan Shelhamer 8830dc5b81 preprocess single inputs instead of lists 
For compositionality and expectations.
 2014-05-20 11:42:41 -07:00
Evan Shelhamer 02ecf1dacf windowed detection in python 2014-05-20 11:42:41 -07:00
Evan Shelhamer 111df0ecfc squash infuriating loop assignment bug in batching 2014-05-19 23:56:48 -07:00
Evan Shelhamer 2fc32d558a image classification in python 2014-05-19 23:56:48 -07:00
Evan Shelhamer bf4d7262bd fix padding for the last batch 2014-05-19 23:56:16 -07:00
Evan Shelhamer 6b85fd006d split drawnet into module code and script 
Don't run scripts in the module dir to avoid import collisions between
io and caffe.io.
 2014-05-19 23:55:22 -07:00
Evan Shelhamer 50d0b6d9c6 add caffe.io submodule for conversions, image loading and resizing 2014-05-19 23:55:21 -07:00
Evan Shelhamer 738c8758e3 fix python mean subtraction 2014-05-19 23:54:25 -07:00
Evan Shelhamer 37123a5126 drop cute names in favor of Net.{pre,de}process() for input formatting 
...and refer to inputs as inputs and not images since general vectors
and matrices are perfectly fine.
 2014-05-16 16:10:25 -07:00
Evan Shelhamer 51024131cc Net.caffeinate() and Net.decaffeinate() format/unformat lists 2014-05-16 16:10:25 -07:00
Evan Shelhamer 5d584c27f0 take blob args as ndarrays and assign on the python side 
Take blob args and give blob returns as single ndarrays instead of lists
of arrays.

Assign the net blobs and diffs as needed on the python side, which
reduces copies and simplifies the C++ side of the wrapper.

Thanks @longjon for the suggestion.
 2014-05-16 16:10:25 -07:00
Evan Shelhamer 025c64e71d resize to input dimensions when formatting in python 2014-05-15 13:52:53 -07:00
Evan Shelhamer 459c8c1dac replace iterator with indices for consistency 2014-05-15 13:06:17 -07:00
Evan Shelhamer a7f6750cc3 python style 2014-05-15 12:35:32 -07:00
Evan Shelhamer a3b307a385 fix accidental revert of Init() from f5c28581 2014-05-15 10:20:43 -07:00
Evan Shelhamer 1b23680244 batch inputs in python by forward_all() and forward_backward_all() 2014-05-14 23:48:01 -07:00
Evan Shelhamer 8af33e8ca2 don't squeeze blob arrays for python 
Preserve the non-batch dimensions of blob arrays, even for singletons.

The forward() and backward() helpers take lists of ndarrays instead of a
single ndarray per blob, and lists of ndarrays are likewise returned.

Note that for output the blob array could actually be returned as a
single ndarray instead of a list.
 2014-05-14 20:17:16 -07:00
Evan Shelhamer af0b857d54 python forward() and backward() extract any blobs and diffs 2014-05-14 20:14:55 -07:00
Evan Shelhamer ac5e6fa1ce python Net.backward() helper and Net.BackwardPrefilled() 2014-05-14 14:37:33 -07:00
Evan Shelhamer 9d4324e5e7 bad forward/backward inputs throw exceptions instead of crashing python 2014-05-14 14:02:54 -07:00
Evan Shelhamer 0e5a5cf50e pycaffe Net.forward() helper 
Do forward pass by prefilled or packaging input + output blobs and
returning a {output blob name: output list} dict.
 2014-05-14 13:44:02 -07:00
Evan Shelhamer 96cd02dd53 set input preprocessing per blob in python 2014-05-14 13:44:02 -07:00
Evan Shelhamer 56ca978c4e expose input and output blob names to python as lists 2014-05-14 13:44:02 -07:00
Evan Shelhamer 872ddf3f81 pycaffe comments, lint 2014-05-13 18:10:42 -07:00
Evan Shelhamer 8da2a3209c add python io getters, mean helper, and image caffeinator/decaffeinator 2014-05-13 18:10:41 -07:00
Evan Shelhamer 47ec9ace41 make python wrapper mean match binaryproto dimensions 
ilsvrc_2012_mean.npy has dims K x H x W.

Code written for the old D x D x K mean needs to be rewritten!
 2014-05-13 18:10:41 -07:00
Evan Shelhamer 51f276e6b2 match existing python formatting 2014-05-13 18:10:41 -07:00