caffe的Matlab接口的使用方法

转自：http://blog.csdn.net/ws_20100/article/details/50525879

编译MatCaffe

使用如下命令编译MatCaffe

<code class="language-Shell hljs glsl has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">make <span class="hljs-built_in" style="margin: 0px; padding: 0px; color: rgb(102, 0, 102); box-sizing: border-box;">all</span> matcaffe</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li></ul>

之后，你可以用以下命令测试MatCaffe：

<code class="language-Shell hljs go has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;"><span class="hljs-built_in" style="margin: 0px; padding: 0px; color: rgb(102, 0, 102); box-sizing: border-box;">make</span> mattest</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li></ul>

如果你在运行上面命令时，遇到如下错误：libstdc++.so.6 version ‘GLIBCXX_3.4.15’ not found，说明你的Matlab库不匹配。你需要在启动Matlab之前运行如下命令：

<code class="language-Shell hljs javascript has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">export LD_LIBRARY_PATH=<span class="hljs-regexp" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">/opt/i</span>ntel/mkl/lib/intel64:<span class="hljs-regexp" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">/usr/</span>local/cuda/lib64
export LD_PRELOAD=<span class="hljs-regexp" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">/usr/</span>lib/x86_64-linux-gnu/libstdc++.so<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">.6</span></code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">2</li></ul>

在Caffe根目录启动Matlab之后需要增加路径：

<code class="language-Matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">addpath ./matlab</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li></ul>

你可以使用

savepath

来保存Matlab搜索路径，这样下次就不用再添加路径了。

使用MatCaffe

MatCaffe 和 PyCaffe 的使用方法很相似。

下面将用一个例子来解释MatCaffe的具体使用细节，假设你已经下载了BVLC CaffeNet，并且在caffe根目录启动matlab。

<code class="language-Matlab hljs bash has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">model = <span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'./models/bvlc_reference_caffenet/deploy.prototxt'</span>;
weights = <span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'./models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel'</span>;
</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">2</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">3</li></ul>

1.设置模式和设备

模式和设备的设置必须在创建一个net或solver之前。

使用CPU：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;"><span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">caffe.</span>set_mode_cpu();</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li></ul>

使用GPU并指定gpu_id：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;"><span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">caffe.</span>set_mode_gpu();
<span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">caffe.</span>set_device(gpu_id);</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">2</li></ul>

2.创建一个网络并访问它的layers和blobs

1.创建网络

创建一个网络：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">net = <span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">caffe.</span>Net(model, weights, <span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'test'</span>); <span class="hljs-comment" style="margin: 0px; padding: 0px; color: rgb(136, 0, 0); box-sizing: border-box;">% create net and load weights</span></code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li></ul>

或者

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">net = <span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">caffe.</span>Net(model, <span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'test'</span>); <span class="hljs-comment" style="margin: 0px; padding: 0px; color: rgb(136, 0, 0); box-sizing: border-box;">% create net but not load weights</span>
<span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>copy_from(weights); <span class="hljs-comment" style="margin: 0px; padding: 0px; color: rgb(136, 0, 0); box-sizing: border-box;">% load weights</span></code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">2</li></ul>

它可以创建一个如下的net对象：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">  Net with <span class="hljs-keyword" style="margin: 0px; padding: 0px; color: rgb(0, 0, 136); box-sizing: border-box;">properties</span>:
layer_vec: <span class="hljs-matrix" style="margin: 0px; padding: 0px; box-sizing: border-box;">[<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">1</span>x23 caffe.Layer]</span>
blob_vec: <span class="hljs-matrix" style="margin: 0px; padding: 0px; box-sizing: border-box;">[<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">1</span>x15 caffe.Blob]</span>
inputs: <span class="hljs-cell" style="margin: 0px; padding: 0px; box-sizing: border-box;">{<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'data'</span>}</span>
outputs: <span class="hljs-cell" style="margin: 0px; padding: 0px; box-sizing: border-box;">{<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'prob'</span>}</span>
name2layer_index: <span class="hljs-matrix" style="margin: 0px; padding: 0px; box-sizing: border-box;">[<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">23</span>x1 containers.Map]</span>
name2blob_index: <span class="hljs-matrix" style="margin: 0px; padding: 0px; box-sizing: border-box;">[<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">15</span>x1 containers.Map]</span>
layer_names: <span class="hljs-cell" style="margin: 0px; padding: 0px; box-sizing: border-box;">{<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">23</span>x1 cell}</span>
blob_names: <span class="hljs-cell" style="margin: 0px; padding: 0px; box-sizing: border-box;">{<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">15</span>x1 cell}</span></code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">2</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">3</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">4</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">5</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">6</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">7</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">8</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">9</li></ul>

两个

containers.Map

对象可以通过layer或者blob的名称找到对应的索引。

2.访问blob

你可以访问网络中的每一个blob，将

data

的blob填充为全一：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;"><span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>blobs(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'data'</span>).set_data(<span class="hljs-built_in" style="margin: 0px; padding: 0px; color: rgb(102, 0, 102); box-sizing: border-box;">ones</span>(<span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>blobs(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'data'</span>).shape));</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li></ul>

将

data

的blob中数值全部乘以10：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;"><span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>blobs(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'data'</span>).set_data(<span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>blobs(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'data'</span>).get_data() * <span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">10</span>);</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li></ul>

注意：因为Matlab是以1作为起始单元，且以列为主，在Matlab中使用四维blob为[width, height, channels, num]，且width是最快的维度，而且要在BGR通道。而且Caffe使用单精度浮点型数据。如果你的数据不是浮点型的，

set_data

将会自动转换为single。

3.访问layer

你也可以访问网络的每一层，以便你作一些网络调整。例如把

conv1

参数乘以10：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;"><span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>params(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'conv1'</span>, <span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">1</span>).set_data(<span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>params(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'conv1'</span>, <span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">1</span>).get_data() * <span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">10</span>); <span class="hljs-comment" style="margin: 0px; padding: 0px; color: rgb(136, 0, 0); box-sizing: border-box;">% set weights</span>
<span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>params(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'conv1'</span>, <span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">2</span>).set_data(<span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>params(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'conv1'</span>, <span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">2</span>).get_data() * <span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">10</span>); <span class="hljs-comment" style="margin: 0px; padding: 0px; color: rgb(136, 0, 0); box-sizing: border-box;">% set bias</span></code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">2</li></ul>

你也可以如下代码：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;"><span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>layers(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'conv1'</span>).params(<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">1</span>).set_data(<span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>layers(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'conv1'</span>).params(<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">1</span>).get_data() * <span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">10</span>);
<span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>layers(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'conv1'</span>).params(<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">2</span>).set_data(<span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>layers(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'conv1'</span>).params(<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">2</span>).get_data() * <span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">10</span>);</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">2</li></ul>

4.保存网络

你仅仅需要如下代码保存网络：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;"><span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>save(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'my_net.caffemodel'</span>);</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li></ul>

5.获得一层的类型(string)

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">layer_type = <span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>layers(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'conv1'</span>).type;</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li></ul>

3.前向和后向计算

前向和后向计算可以使用

net.forward

或者

net.forward_prefilled

实现。函数

net.forward

将一个包含输入blob(s)的cell数组作为输入，并输出一个包含输出blob(s)的cell数组。函数

net.forward_prefilled

将使用输入blob(s)中的已有数据进行计算，没有输入数据，没有输出数据。

在通过一些方法（如：

data = rand(net.blobs('data').shape);

）产生输入数据后，你可以运行：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">res = <span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>forward(<span class="hljs-cell" style="margin: 0px; padding: 0px; box-sizing: border-box;">{data}</span>);
prob = res<span class="hljs-cell" style="margin: 0px; padding: 0px; box-sizing: border-box;">{<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">1</span>}</span>;</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">2</li></ul>

或者

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;"><span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>blobs(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'data'</span>).set_data(data);
<span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>forward_prefilled();
prob = <span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>blobs(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'prob'</span>).get_data();</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">2</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">3</li></ul>

后向计算使用

net.backward

或者

net.backward_prefilled

，并且把

get_data

和

set_data

替换为

get_diff

和

set_diff

。在通过一些方法（例如

prob_diff
= rand(net.blobs('prob').shape);

）产生输出blobs的梯度后，你可以运行：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">res = <span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>backward(<span class="hljs-cell" style="margin: 0px; padding: 0px; box-sizing: border-box;">{prob_diff}</span>);
data_diff = res<span class="hljs-cell" style="margin: 0px; padding: 0px; box-sizing: border-box;">{<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">1</span>}</span>;</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">2</li></ul>

或者

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;"><span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>blobs(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'prob'</span>).set_diff(prob_diff);
<span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>backward_prefilled();
data_diff = <span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>blobs(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'data'</span>).get_diff();</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">2</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">3</li></ul>

然而，如上的后向计算并不能得到正确的结果，因为Caffe默认网络不需要后向计算。为了获取正确的后向计算结果，你需要在你的网络prototxt文件中设置

force_backward: true

在完成前向和后向计算之后，你可以获得中间blobs的data和diff。例如，你可以在前向计算后获取

pool5

的特征。

4.Reshape

假设你想要运行1幅图像，而不是10幅时：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;"><span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span>blobs(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'data'</span>).<span class="hljs-built_in" style="margin: 0px; padding: 0px; color: rgb(102, 0, 102); box-sizing: border-box;">reshape</span>(<span class="hljs-matrix" style="margin: 0px; padding: 0px; box-sizing: border-box;">[<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">227</span> <span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">227</span> <span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">3</span> <span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">1</span>]</span>); <span class="hljs-comment" style="margin: 0px; padding: 0px; color: rgb(136, 0, 0); box-sizing: border-box;">% reshape blob 'data'</span>
<span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">net.</span><span class="hljs-built_in" style="margin: 0px; padding: 0px; color: rgb(102, 0, 102); box-sizing: border-box;">reshape</span>();</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">2</li></ul>

然后，整个网络就reshape了，此时

net.blobs('prob').shape

应该是[1000 1];

5.训练网络

假设你按照ImageNET Tutorial的方法创建了训练lmdb和验证lmdb，产生一个solver并且在ILSVRC 2012 分类数据集上训练：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">solver = <span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">caffe.</span>Solver(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'./models/bvlc_reference_caffenet/solver.prototxt'</span>);</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li></ul>

这样可以创建一个

solver

对象：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">  Solver with <span class="hljs-keyword" style="margin: 0px; padding: 0px; color: rgb(0, 0, 136); box-sizing: border-box;">properties</span>:

net: <span class="hljs-matrix" style="margin: 0px; padding: 0px; box-sizing: border-box;">[<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">1</span>x1 caffe.Net]</span>
test_nets: <span class="hljs-matrix" style="margin: 0px; padding: 0px; box-sizing: border-box;">[<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">1</span>x1 caffe.Net]</span></code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">2</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">3</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">4</li></ul>

训练代码：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;"><span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">solver.</span>solve();</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li></ul>

如果只想训练迭代1000次：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;"><span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">solver.</span>step(<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">1000</span>);</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li></ul>

来获取迭代数量：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">iter = <span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">solver.</span>iter();</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li></ul>

来获取这个网络：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">train_net = <span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">solver.</span>net;
test_net = <span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">solver.</span>test_nets(<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">1</span>);</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">2</li></ul>

假设从一个snapshot中恢复网络训练：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;"><span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">solver.</span>restore(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'your_snapshot.solverstate'</span>);</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li></ul>

6.输入和输出

caffe.io

类提供了基本的输入函数

load_image

和

read_mean

。例如，读取ILSVRC
2012 mean文件（假设你已经通过运行

./data/ilsvrc12/get_ilsvrc_aux.sh

下载imagenet例程辅助文件）

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">mean_data = <span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">caffe.</span><span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">io.</span>read_mean(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'./data/ilsvrc12/imagenet_mean.binaryproto'</span>);</code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li></ul>

为了读取Caffe例程图片，并且resize到[width, height]，且假设

width = 256; height = 256;

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">im_data = <span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">caffe.</span><span class="hljs-transposed_variable" style="margin: 0px; padding: 0px; box-sizing: border-box;">io.</span>load_image(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'./examples/images/cat.jpg'</span>);
im_data = imresize(im_data, <span class="hljs-matrix" style="margin: 0px; padding: 0px; box-sizing: border-box;">[width, height]</span>); <span class="hljs-comment" style="margin: 0px; padding: 0px; color: rgb(136, 0, 0); box-sizing: border-box;">% resize using Matlab's imresize</span></code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">2</li></ul>

注意：

width

是最快的维度，通道为BGR，与Matlab存取图片的一般方式不一样。如果你不想要使用

caffe.io.load_image

，且想自己导入一幅图片：

<code class="language-matlab hljs  has-numbering" style="display: block; padding: 0px; color: inherit; box-sizing: border-box; font-family: 'Source Code Pro', monospace;font-size:undefined; white-space: pre; border-radius: 0px; word-wrap: normal; background: transparent;">im_data = imread(<span class="hljs-string" style="margin: 0px; padding: 0px; color: rgb(0, 136, 0); box-sizing: border-box;">'./examples/images/cat.jpg'</span>); <span class="hljs-comment" style="margin: 0px; padding: 0px; color: rgb(136, 0, 0); box-sizing: border-box;">% read image</span>
im_data = im_data(:, :, <span class="hljs-matrix" style="margin: 0px; padding: 0px; box-sizing: border-box;">[<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">3</span>, <span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">2</span>, <span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">1</span>]</span>); <span class="hljs-comment" style="margin: 0px; padding: 0px; color: rgb(136, 0, 0); box-sizing: border-box;">% convert from RGB to BGR</span>
im_data = <span class="hljs-built_in" style="margin: 0px; padding: 0px; color: rgb(102, 0, 102); box-sizing: border-box;">permute</span>(im_data, <span class="hljs-matrix" style="margin: 0px; padding: 0px; box-sizing: border-box;">[<span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">2</span>, <span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">1</span>, <span class="hljs-number" style="margin: 0px; padding: 0px; color: rgb(0, 102, 102); box-sizing: border-box;">3</span>]</span>); <span class="hljs-comment" style="margin: 0px; padding: 0px; color: rgb(136, 0, 0); box-sizing: border-box;">% permute width and height</span>
im_data = single(im_data); <span class="hljs-comment" style="margin: 0px; padding: 0px; color: rgb(136, 0, 0); box-sizing: border-box;">% convert to single precision</span></code><ul class="pre-numbering" style="margin: 0px; padding: 6px 0px 40px; box-sizing: border-box; position: absolute; width: 50px; top: 0px; left: 0px; border-right-width: 1px; border-right-style: solid; border-right-color: rgb(221, 221, 221); list-style: none; text-align: right; background-color: rgb(238, 238, 238);"><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">1</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">2</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">3</li><li style="margin: 0px; padding: 0px 5px; box-sizing: border-box;">4</li></ul>

你也可以看一下

caffe/matlab/demo/classification_demo.m

文件，了解如何将输入图片crop成多个输入块。

你可以查看

caffe/matlab/hdf5creation

，了解如何通过Matlab读和写HDF5数据。但不提供额外的数据输出函数，因为在Matlab本身已经具有了强大的功能。

7.清除nets和solvers

调用

caffe.reset_all()

来清理你所创建的所有的solvers，和stand-alone nets。