From a4e58e22bc59c51fa60efeb5355cdfecb921459a Mon Sep 17 00:00:00 2001 From: Bert Peters Date: Tue, 20 Mar 2018 12:11:31 +0100 Subject: [PATCH] Add limitations to README. --- README.md | 31 ++++++++++++++++++++++++++++++- 1 file changed, 30 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index 3ba1607..f438ccc 100644 --- a/README.md +++ b/README.md @@ -79,7 +79,7 @@ After that, you can, for example, run the program as follows -l ../data/ilsvrc12/synset_words.txt \ ../data/samples/*.jpg -This will run the network on the deduplicated caffenet (see: limitations) +This will run the network on the deduplicated caffenet (see: [limitations](#lmitations)) with the correct weights file and the labels file supplied, on all jpegs located in the samples folder. More advanced usage can be discovered using the `-h` option of the executable. @@ -88,3 +88,32 @@ the `-h` option of the executable. You can move around with the WASD keys, and look around using the mouse. Arrow keys change the currently loaded input. + +## Limitations + +The following documents the limitations currently present in the program, +including possible workarounds. Note that these may change at any time. + +### No in-place layers + +Due to the way Caffe works, this program cannot properly extract state +data from networks that use in-place computation, since it cannot observe +the input- and output state of each layer. + +To work around this, a simple script `deinplace.py` is included in the tools +folder. To use it, first run make (requires `protoc` to be on your PATH) and +then run it on your program. To see what options it supports, use its `-h` +option. + +### Single input/output + +The visualisation structuree only supports linear networks. That is: network +where each layer only ever reads from one input layer and only produces one +output. This is not a fundamental limitation, and could be reworked in future +versions. + +### Data input layer only + +The program only supports reading input from image files, rather than using +LLDB or LevelDB sources. Render your inputs to an image format of choice to +use the program.