Deployment bug-fix

COPYING

@@ -1,14 +1,14 @@
-Copyright (c) 2014 The MatConvNet team.
+Copyright (c) 2014-16 The MatConvNet Team.
 All rights reserved.
 
 Redistribution and use in source and binary forms are permitted
 provided that the above copyright notice and this paragraph are
-duplicated in all such forms and that any documentation,
-advertising materials, and other materials related to such
-distribution and use acknowledge that the software was developed
-by the <organization>. The name of the
-<organization> may not be used to endorse or promote products derived
-from this software without specific prior written permission.
-THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT ANY EXPRESS OR
-IMPLIED WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE IMPLIED
-WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE.
+duplicated in all such forms and that any documentation, advertising
+materials, and other materials related to such distribution and use
+acknowledge that the software was developed by the MatConvNet
+Team. The name of the MatConvNet Team may not be used to endorse or
+promote products derived from this software without specific prior
+written permission.  THIS SOFTWARE IS PROVIDED ``AS IS'' AND WITHOUT
+ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, WITHOUT LIMITATION, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.

Makefile

@@ -20,10 +20,10 @@ DEBUG ?=
 ARCH ?= maci64
 
 # Configure MATLAB
-MATLABROOT ?= /Applications/MATLAB_R2015a.app
+MATLABROOT ?= /Applications/MATLAB_R2017a.app
 
 # Configure CUDA and CuDNN. CUDAMETHOD can be either 'nvcc' or 'mex'.
-CUDAROOT ?= /Developer/NVIDIA/CUDA-6.5
+CUDAROOT ?= /Developer/NVIDIA/CUDA-8.0
 CUDNNROOT ?= $(CURDIR)/local/
 CUDAMETHOD ?= $(if $(ENABLE_CUDNN),nvcc,mex)
 
@@ -38,7 +38,7 @@ CUDAMETHOD ?= $(if $(ENABLE_CUDNN),nvcc,mex)
 
 # Maintenance
 NAME = matconvnet
-VER = 1.0-beta22
+VER = 1.0-beta24
 DIST = $(NAME)-$(VER)
 LATEST = $(NAME)-latest
 RSYNC = rsync
@@ -82,7 +82,7 @@ LDFLAGS =
 LDOPTIMFLAGS =
 LINKLIBS = -lmwblas
 
-NVCCFLAGS_PASS = -gencode=arch=compute_30,code=\"sm_30,compute_30\"
+NVCCFLAGS_PASS = -D_FORCE_INLINES -gencode=arch=compute_30,code=\"sm_30,compute_30\"
 NVCCVER = $(shell $(NVCC) --version | \
 sed -n 's/.*V\([0-9]*\).\([0-9]*\).\([0-9]*\).*/\1 \2 \3/p' | \
 xargs printf '%02d%02d%02d')
@@ -159,12 +159,14 @@ cpp_src+=matlab/src/bits/nnpooling.$(ext)
 cpp_src+=matlab/src/bits/nnnormalize.$(ext)
 cpp_src+=matlab/src/bits/nnbnorm.$(ext)
 cpp_src+=matlab/src/bits/nnbilinearsampler.$(ext)
+cpp_src+=matlab/src/bits/nnroipooling.$(ext)
 mex_src+=matlab/src/vl_nnconv.$(ext)
 mex_src+=matlab/src/vl_nnconvt.$(ext)
 mex_src+=matlab/src/vl_nnpool.$(ext)
 mex_src+=matlab/src/vl_nnnormalize.$(ext)
 mex_src+=matlab/src/vl_nnbnorm.$(ext)
 mex_src+=matlab/src/vl_nnbilinearsampler.$(ext)
+mex_src+=matlab/src/vl_nnroipool.$(ext)
 mex_src+=matlab/src/vl_taccummex.$(ext)
 mex_src+=matlab/src/vl_tmove.$(ext)
 ifdef ENABLE_IMREADJPEG
@@ -180,6 +182,7 @@ cpp_src+=matlab/src/bits/impl/pooling_cpu.cpp
 cpp_src+=matlab/src/bits/impl/normalize_cpu.cpp
 cpp_src+=matlab/src/bits/impl/bnorm_cpu.cpp
 cpp_src+=matlab/src/bits/impl/bilinearsampler_cpu.cpp
+cpp_src+=matlab/src/bits/impl/roipooling_cpu.cpp
 cpp_src+=matlab/src/bits/impl/tinythread.cpp
 ifdef ENABLE_IMREADJPEG
 cpp_src+=matlab/src/bits/impl/imread_$(IMAGELIB).cpp
@@ -195,6 +198,7 @@ cpp_src+=matlab/src/bits/impl/pooling_gpu.cu
 cpp_src+=matlab/src/bits/impl/normalize_gpu.cu
 cpp_src+=matlab/src/bits/impl/bnorm_gpu.cu
 cpp_src+=matlab/src/bits/impl/bilinearsampler_gpu.cu
+cpp_src+=matlab/src/bits/impl/roipooling_gpu.cu
 cpp_src+=matlab/src/bits/datacu.cu
 mex_src+=matlab/src/vl_cudatool.cu
 ifdef ENABLE_CUDNN
@@ -255,7 +259,7 @@ CXXOPTIMFLAGS='$$CXXOPTIMFLAGS $(call nvcc-quote,$(CXXOPTIMFLAGS))'
 MEXFLAGS_LD := $(MEXFLAGS) \
 LDFLAGS='$$LDFLAGS $(LDFLAGS)' \
 LDOPTIMFLAGS='$$LDOPTIMFLAGS $(LDOPTIMFLAGS)' \
-LINKLIBS='$$LINKLIBS $(LINKLIBS)' \
+LINKLIBS='$(LINKLIBS) $$LINKLIBS' \
 
 NVCCFLAGS = $(CXXFLAGS) $(NVCCFLAGS_PASS) \
 -I"$(MATLABROOT)/extern/include" \

doc/Makefile

@@ -28,6 +28,7 @@ vl_nnnormalizelp.m \
 vl_nnpdist.m \
 vl_nnpool.m \
 vl_nnrelu.m \
+vl_nnroipool.m \
 vl_nnsigmoid.m \
 vl_nnsoftmax.m \
 vl_nnsoftmaxloss.m \
@@ -38,6 +39,7 @@ vl_imreadjpeg.m \
 vl_imreadjpeg.m \
 vl_taccum.m \
 vl_tmove.m \
+vl_tshow.m \
 simplenn/vl_simplenn.m \
 simplenn/vl_simplenn_diagnose.m \
 simplenn/vl_simplenn_tidy.m \

doc/blocks.tex

@@ -214,6 +214,46 @@ \section{Spatial bilinear resampling}\label{s:spatial-sampler}
 
 See \cref{s:impl-sampler} for implementation details.
 
+% ------------------------------------------------------------------
+\section{Region of interest pooling}\label{s:roi-pooling}
+% ------------------------------------------------------------------
+
+The \emph{region of interest (ROI) pooling} block applies max or average pooling to specified subwindows of a tensor. A region is a rectangular region $R = (u_-,v_-,u_+,v_+)$. The region itself is partitioned into $(H',W')$ tiles along the vertical and horizontal directions. The edges of the tiles have coordinates
+\begin{align*}
+   v_{i'} &= v_- + (v_+ - v_- + 1) (i' - 1), \quad i' = 1,\dots,H',\\
+   u_{j'} &= u_- + (u_+ - u_- + 1) (j' - 1), \quad j' = 1,\dots,W'.
+\end{align*}
+Following the implementation of~\cite{girshick15fast}, the $H'\times W'$ pooling tiles are given by
+\[
+   \Omega_{i'j'} =
+   \{\lfloor v_{i'} \rfloor + 1, \dots, \lceil v_{i'+1} \rceil\}
+   \times
+   \{\lfloor u_{i'} \rfloor + 1, \dots, \lceil u_{i'+1} \rceil\}.
+\]
+Then the input and output tensors are as follows:
+\[
+  \bx \in \mathbb{R}^{H \times W \times C},
+  \qquad
+  \by \in \mathbb{R}^{H' \times W' \times C},
+\]
+where
+\[
+   y_{i'j'c} = \operatornamewithlimits{max}_{(i,j) \in \Omega_{i'j'}} x_{ijc}.
+\]
+Alternatively, $\max$ can be replaced by the averaging operator.
+
+The extent of each region is defined by four coordinates as specified above; however, differently from tensor indexes, these use $(0,0)$ as the coordinate of the top-left pixel. In fact, if there is a single tile ($H'=W'=1$), then the region $(0,0,H-1,W-1)$ covers the whole input image:
+\[
+   \Omega_{11} =
+   \{1, \dots, W\}
+   \times
+   \{1, \dots, H\}.
+\]
+
+In more details, the input of the block is a sequence of $K$ regions. Each region pools one of the $T$ images in the batch stored in $\bx \in \mathbb{R}^{H\times W\times C\times T}$. Regions are therefore specified as a tensor $R \in \mathbb{R}^{5 \times K}$, where the first coordinate is the index of the pooled image in the batch. The output is a $\by \in \mathbb{R}^{H' \times W' \times C \times K}$ tensor.
+
+For compatibility with~\cite{girshick15fast}, furthermore, the region coordinates are rounded to the nearest integer before the definitions above are used. Note also that, due to the discretization details, 1) tiles always contain at least one pixel, 2) there can be a pixel of overlap between them and 3) the discretization has a slight bias towards left-top pixels.
+
 % ------------------------------------------------------------------
 \section{Normalization}\label{s:normalization}
 % ------------------------------------------------------------------

doc/impl.tex

@@ -167,7 +167,7 @@ \section{Spatial pooling}\label{s:impl-pooling}
 \frac{d z}{d (\vv \by)^\top}
 S(\bx),
 $
-for all but a null set of points, where the operator is not differentiable (this usually does not pose problems in optimization by stochastic gradient). For max-pooling, similar relations exists with two differences: $S$ does not depend on the input $\bx$ and it is not binary, in order to account for the normalization factors. In summary, we have the expressions:
+for all but a null set of points, where the operator is not differentiable (this usually does not pose problems in optimization by stochastic gradient). For average pooling, similar relations exists with two differences: $S$ does not depend on the input $\bx$ and it is not binary, in order to account for the normalization factors. In summary, we have the expressions:
 \begin{equation}\label{e:max-mat}
 \boxed{
 \vv\by = S(\bx) \vv \bx,
@@ -429,12 +429,12 @@ \subsection{Spatial normalization}\label{s:impl-spnorm}
 The derivative of spatial normalization can be obtained as follows:
 \begin{align*}
 \frac{dz}{dx_{ijd}} 
-&= \sum_{i''j''d}
+&= \sum_{i''j''}
 \frac{dz}{d y_{i''j''d}} 
 \frac{d y_{i''j''d}}{d x_{ijd}}
 \\
 &=
-\sum_{i''j''d}
+\sum_{i''j''}
 \frac{dz}{d y_{i''j''d}} 
 (1 + \alpha n_{i''j''d}^2)^{-\beta}
 \frac{dx_{i''j''d}}{d x_{ijd}} 
@@ -450,7 +450,7 @@ \subsection{Spatial normalization}\label{s:impl-spnorm}
 (1 + \alpha n_{ijd}^2)^{-\beta}
 -2\alpha\beta x_{ijd}
 \left[
-\sum_{i''j''d}
+\sum_{i''j''}
 \frac{dz}{d y_{i''j''d}} 
 (1 + \alpha n_{i''j''d}^2)^{-\beta-1}
 x_{i''j''d}
@@ -462,7 +462,7 @@ \subsection{Spatial normalization}\label{s:impl-spnorm}
 (1 + \alpha n_{ijd}^2)^{-\beta}
 -2\alpha\beta x_{ijd}
 \left[
-\sum_{i''j''d}
+\sum_{i''j''}
 \eta_{i''j''d}
 \frac{dn_{i''j''d}^2}{d (x_{ijd}^2)}
 \right],

doc/intro.tex

@@ -25,8 +25,8 @@ \section{Getting started}\label{s:getting-statrted}
 \begin{lstlisting}[escapechar=!]
 % install and compile MatConvNet (run once)
 untar(['http://www.vlfeat.org/matconvnet/download/' ...
-   'matconvnet-1.0-beta12.tar.gz']) ;
-cd matconvnet-1.0-beta12
+   'matconvnet-1.0-beta24.tar.gz']) ;
+cd matconvnet-1.0-beta24
 run matlab/vl_compilenn
 
 % download a pre-trained CNN from the web (run once)

doc/matdocparser.py

@@ -51,7 +51,7 @@
 import re
 
 __mpname__           = 'MatDocParser'
-__version__          = '1.0-beta15'
+__version__          = '1.0-beta24'
 __date__             = '2015-09-20'
 __description__      = 'MatDoc MATLAB inline function description interpreter.'
 __long_description__ = __doc__

doc/site/docs/about.md

@@ -23,6 +23,28 @@ here.
 <a name='changes'></a>
 # Changes
 
+-   1.0-beta24 (March 2017).
+
+    **New features**
+
+    * New toy example `cnn_toy_data.m` demonstrating using a
+      customized `imdb`.
+    * `vl_argparse.m` now supports dot paths and ignoring missing
+      defaults.
+    * Support for different example solvers (AdaGrad, Adam, AdaDelta,
+      RMSProp) and ability to add new ones.
+    * A new function `vl_tshow.m` to glance at tensors.
+    * Bugfixes.
+
+-   1.0-beta23 (September 2016).
+
+    **New features**
+
+    * A new function `vl_nnroipool.m` for region of interest pooling,
+      supporting networks such as Fast-RCNN.
+    * Imported Fast-RCNN models from Caffe.
+    * An example Fast-RCNN implementation, training and testing.
+
 -   1.0-beta22 (Spetember 2016).
 
     * Bugfixes.

doc/site/docs/css/fixes.css

@@ -66,7 +66,7 @@ a  { color: #00438E ; }
 
 #Functions .dropdown-menu {
   color: #000;
-  max-height: 400px;
+  max-height: 800px;
   width: 342px;
 }
 

doc/site/docs/functions.md

@@ -33,6 +33,7 @@ showing how to train CNNs.
 - [`vl_nnpdist`](mfiles/vl_nnpdist.md) Pairwise distances.
 - [`vl_nnpool`](mfiles/vl_nnpool.md) Max and sum pooling.
 - [`vl_nnrelu`](mfiles/vl_nnrelu.md) Rectified Linear Unit.
+- [`vl_nnroipool`](mfiles/vl_nnroipool.md) Reegion of interest pooling.
 - [`vl_nnsigmoid`](mfiles/vl_nnsigmoid.md) Sigmoid.
 - [`vl_nnsoftmax`](mfiles/vl_nnsoftmax.md) Channel soft-max.
 - [`vl_nnsoftmaxloss`](mfiles/vl_nnsoftmaxloss.md) *Deprecated*
@@ -70,3 +71,4 @@ showing how to train CNNs.
 - [`vl_imreadjpeg`](mfiles/vl_imreadjpeg.md) Quickly load a batch of JPEG images.
 - [`vl_taccum`](mfiles/vl_taccum.md) Accumulate tensors operating in-place when possible.
 - [`vl_tmove`](mfiles/vl_tmove.md) Exchange tensors between MATLAB processes and GPUs.
+- [`vl_tshow`](mfiles/vl_tshow.md) Show a tensor on screen.

doc/site/docs/index.md

@@ -2,7 +2,7 @@
 
 <div class="row" style="white-space: nowrap;">
 <div class="col-sm-3">
-<a href="download/matconvnet-1.0-beta21.tar.gz">
+<a href="download/matconvnet-1.0-beta24.tar.gz">
 <div class="menuicon"><span class="fa fa-download fa-2x"></span></div>
 Download</a>
 </div>
@@ -31,6 +31,12 @@ efficient, and can run and learn state-of-the-art CNNs. Many
 pre-trained CNNs for image classification, segmentation, face
 recognition, and text detection are available.
 
+> **New:** [1.0-beta24](about.md#changes) released with bugfixes, new
+> examples, and utility functions.
+>
+> **New:** [1.0-beta23](about.md#changes) released with
+> [`vl_nnroipool`](mfiles/vl_nnroipool) and a Fast-RCNN demo.
+>
 > **New:** [1.0-beta22](about.md#changes) released with a few bugfixes.
 >
 > **New:** [1.0-beta21](about.md#changes) provides two new tools,
@@ -52,7 +58,7 @@ recognition, and text detection are available.
 > numerous other improvements and bugfixes.
 
 ## Obtaining MatConvNet
-- <span class="fa fa-file-archive-o"></span>&nbsp;Tarball for [version 1.0-beta22](download/matconvnet-1.0-beta22.tar.gz); [older versions](download/) (<span class="fa fa-apple"/> <span class="fa fa-windows"/> <span class="fa fa-linux"/>)
+- <span class="fa fa-file-archive-o"></span>&nbsp;Tarball for [version 1.0-beta24](download/matconvnet-1.0-beta24.tar.gz); [older versions](download/) (<span class="fa fa-apple"/> <span class="fa fa-windows"/> <span class="fa fa-linux"/>)
 - <span class="fa fa-github"></span>&nbsp;[GIT repository](http://www.github.com/vlfeat/matconvnet.git)
 - <span class="fa fa-pencil-square-o"></span>&nbsp;<a href="javascript:void(0);"
   onclick="toggle_visibility('citation');">Citation</a>