resnet katex fix

This commit is contained in:
Varuna Jayasiri
2022-04-10 13:41:33 +05:30
parent 1d2ce11ce4
commit 42257b7130
2 changed files with 40 additions and 40 deletions

View File

@ -73,16 +73,16 @@
<p>ResNets train layers as residual functions to overcome the <em>degradation problem</em>. The degradation problem is the accuracy of deep neural networks degrading when the number of layers becomes very high. The accuracy increases as the number of layers increase, then saturates, and then starts to degrade.</p>
<p>The paper argues that deeper models should perform at least as well as shallower models because the extra layers can just learn to perform an identity mapping.</p>
<h2>Residual Learning</h2>
<p>If <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqh" style=""><span class="mord mathcal" style="margin-right:0.00965em">H</span><span class="mopen" style="">(</span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span><span class="mclose" style="">)</span></span></span></span></span> is the mapping that needs to be learned by a few layers, they train the residual function</p>
<p><span class="katex-display"><span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqg" style=""><span class="mord coloredeq eqj" style=""><span class="mord coloredeq eqt" style=""><span class="mord mathcal" style="margin-right:0.09931em">F</span></span></span><span class="mopen" style="">(</span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span><span class="mclose" style="">)</span></span><span class="mspace" style="margin-right:0.2777777777777778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2777777777777778em;"></span></span><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqh" style=""><span class="mord mathcal" style="margin-right:0.00965em">H</span><span class="mopen" style="">(</span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span><span class="mclose" style="">)</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin"></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.43056em;vertical-align:0em;"></span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span></span></p>
<p>instead. And the original function becomes <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqg" style=""><span class="mord coloredeq eqj" style=""><span class="mord coloredeq eqt" style=""><span class="mord mathcal" style="margin-right:0.09931em">F</span></span></span><span class="mopen" style="">(</span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span><span class="mclose" style="">)</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">+</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.43056em;vertical-align:0em;"></span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span>.</p>
<p>In this case, learning identity mapping for <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqh" style=""><span class="mord mathcal" style="margin-right:0.00965em">H</span><span class="mopen" style="">(</span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span><span class="mclose" style="">)</span></span></span></span></span> is equivalent to learning <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqg" style=""><span class="mord coloredeq eqj" style=""><span class="mord coloredeq eqt" style=""><span class="mord mathcal" style="margin-right:0.09931em">F</span></span></span><span class="mopen" style="">(</span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span><span class="mclose" style="">)</span></span></span></span></span> to be <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord">0</span></span></span></span>, which is easier to learn.</p>
<p>If <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqh" style=""><span class="mord mathcal" style="margin-right:0.00965em">H</span><span class="mopen" style="">(</span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span><span class="mclose" style="">)</span></span></span></span></span> is the mapping that needs to be learned by a few layers, they train the residual function</p>
<p><span class="katex-display"><span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqg" style=""><span class="mord coloredeq eqj" style=""><span class="mord coloredeq eqr" style=""><span class="mord mathcal" style="margin-right:0.09931em">F</span></span></span><span class="mopen" style="">(</span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span><span class="mclose" style="">)</span></span><span class="mspace" style="margin-right:0.2777777777777778em;"></span><span class="mrel">=</span><span class="mspace" style="margin-right:0.2777777777777778em;"></span></span><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqh" style=""><span class="mord mathcal" style="margin-right:0.00965em">H</span><span class="mopen" style="">(</span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span><span class="mclose" style="">)</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin"></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.43056em;vertical-align:0em;"></span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span></span></p>
<p>instead. And the original function becomes <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqg" style=""><span class="mord coloredeq eqj" style=""><span class="mord coloredeq eqr" style=""><span class="mord mathcal" style="margin-right:0.09931em">F</span></span></span><span class="mopen" style="">(</span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span><span class="mclose" style="">)</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">+</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.43056em;vertical-align:0em;"></span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span>.</p>
<p>In this case, learning identity mapping for <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqh" style=""><span class="mord mathcal" style="margin-right:0.00965em">H</span><span class="mopen" style="">(</span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span><span class="mclose" style="">)</span></span></span></span></span> is equivalent to learning <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqg" style=""><span class="mord coloredeq eqj" style=""><span class="mord coloredeq eqr" style=""><span class="mord mathcal" style="margin-right:0.09931em">F</span></span></span><span class="mopen" style="">(</span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span><span class="mclose" style="">)</span></span></span></span></span> to be <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord">0</span></span></span></span>, which is easier to learn.</p>
<p>In the parameterized form this can be written as,</p>
<p><span class="katex-display"><span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqd" style=""><span class="mord coloredeq eqj" style=""><span class="mord coloredeq eqt" style=""><span class="mord mathcal" style="margin-right:0.09931em">F</span></span></span><span class="mopen" style="">(</span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span><span class="mpunct" style="">,</span><span class="mspace" style="margin-right:0.16666666666666666em;"></span><span class="mopen" style="">{</span><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.31166399999999994em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mclose" style="">})</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">+</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.43056em;vertical-align:0em;"></span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span></span></p>
<p>and when the feature map sizes of <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqj" style=""><span class="mord coloredeq eqt" style=""><span class="mord mathcal" style="margin-right:0.09931em">F</span></span></span><span class="mopen">(</span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.16666666666666666em;"></span><span class="mord"><span class="mord"><span class="mord mathnormal" style="margin-right:0.13889em;">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.31166399999999994em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span><span class="mclose">)</span></span></span></span> and <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.43056em;vertical-align:0em;"></span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span> are different the paper suggests doing a linear projection, with learned weights <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.83333em;vertical-align:-0.15em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.151392em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">s</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span></span></span></span>.</p>
<p><span class="katex-display"><span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqd" style=""><span class="mord coloredeq eqj" style=""><span class="mord coloredeq eqt" style=""><span class="mord mathcal" style="margin-right:0.09931em">F</span></span></span><span class="mopen" style="">(</span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span><span class="mpunct" style="">,</span><span class="mspace" style="margin-right:0.16666666666666666em;"></span><span class="mopen" style="">{</span><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.31166399999999994em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mclose" style="">})</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">+</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.83333em;vertical-align:-0.15em;"></span><span class="mord coloredeq eqm" style=""><span class="mord coloredeq eqp" style=""><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.151392em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">s</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span></span></span></p>
<p><span class="katex-display"><span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqd" style=""><span class="mord coloredeq eqj" style=""><span class="mord coloredeq eqr" style=""><span class="mord mathcal" style="margin-right:0.09931em">F</span></span></span><span class="mopen" style="">(</span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span><span class="mpunct" style="">,</span><span class="mspace" style="margin-right:0.16666666666666666em;"></span><span class="mopen" style="">{</span><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.31166399999999994em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mclose" style="">})</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">+</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.43056em;vertical-align:0em;"></span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span></span></p>
<p>and when the feature map sizes of <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqj" style=""><span class="mord coloredeq eqr" style=""><span class="mord mathcal" style="margin-right:0.09931em">F</span></span></span><span class="mopen">(</span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span><span class="mpunct">,</span><span class="mspace" style="margin-right:0.16666666666666666em;"></span><span class="mord"><span class="mord"><span class="mord mathnormal" style="margin-right:0.13889em;">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.31166399999999994em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span><span class="mclose">)</span></span></span></span> and <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.43056em;vertical-align:0em;"></span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span> are different the paper suggests doing a linear projection, with learned weights <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.83333em;vertical-align:-0.15em;"></span><span class="mord coloredeq eqn" style=""><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.151392em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">s</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span></span></span></span>.</p>
<p><span class="katex-display"><span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqd" style=""><span class="mord coloredeq eqj" style=""><span class="mord coloredeq eqr" style=""><span class="mord mathcal" style="margin-right:0.09931em">F</span></span></span><span class="mopen" style="">(</span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span><span class="mpunct" style="">,</span><span class="mspace" style="margin-right:0.16666666666666666em;"></span><span class="mopen" style="">{</span><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.31166399999999994em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mclose" style="">})</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">+</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.83333em;vertical-align:-0.15em;"></span><span class="mord coloredeq eqm" style=""><span class="mord coloredeq eqn" style=""><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.151392em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">s</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span></span></span></p>
<p>Paper experimented with zero padding instead of linear projections and found linear projections to work better. Also when the feature map sizes match they found identity mapping to be better than linear projections.</p>
<p><span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.68333em;vertical-align:0em;"></span><span class="mord coloredeq eqj" style=""><span class="mord coloredeq eqt" style=""><span class="mord mathcal" style="margin-right:0.09931em">F</span></span></span></span></span></span> should have more than one layer, otherwise the sum <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqd" style=""><span class="mord coloredeq eqj" style=""><span class="mord coloredeq eqt" style=""><span class="mord mathcal" style="margin-right:0.09931em">F</span></span></span><span class="mopen" style="">(</span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span><span class="mpunct" style="">,</span><span class="mspace" style="margin-right:0.16666666666666666em;"></span><span class="mopen" style="">{</span><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.31166399999999994em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mclose" style="">})</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">+</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.83333em;vertical-align:-0.15em;"></span><span class="mord coloredeq eqm" style=""><span class="mord coloredeq eqp" style=""><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.151392em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">s</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span></span> also won&#x27;t have non-linearities and will be like a linear layer.</p>
<p><span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.68333em;vertical-align:0em;"></span><span class="mord coloredeq eqj" style=""><span class="mord coloredeq eqr" style=""><span class="mord mathcal" style="margin-right:0.09931em">F</span></span></span></span></span></span> should have more than one layer, otherwise the sum <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqd" style=""><span class="mord coloredeq eqj" style=""><span class="mord coloredeq eqr" style=""><span class="mord mathcal" style="margin-right:0.09931em">F</span></span></span><span class="mopen" style="">(</span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span><span class="mpunct" style="">,</span><span class="mspace" style="margin-right:0.16666666666666666em;"></span><span class="mopen" style="">{</span><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.31166399999999994em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mclose" style="">})</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">+</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.83333em;vertical-align:-0.15em;"></span><span class="mord coloredeq eqm" style=""><span class="mord coloredeq eqn" style=""><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.151392em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">s</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span></span> also won&#x27;t have non-linearities and will be like a linear layer.</p>
<p>Here is <a href="experiment.html">the training code</a> for training a ResNet on CIFAR-10.</p>
<p><a href="https://app.labml.ai/run/fc5ad600e4af11ebbafd23b8665193c1"><img alt="View Run" src="https://img.shields.io/badge/labml-experiment-brightgreen"></a></p>
@ -102,7 +102,7 @@
<a href='#section-1'>#</a>
</div>
<h2>Linear projections for shortcut connection</h2>
<p>This does the <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.83333em;vertical-align:-0.15em;"></span><span class="mord coloredeq eqm" style=""><span class="mord coloredeq eqp" style=""><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.151392em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">s</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span></span> projection described above.</p>
<p>This does the <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.83333em;vertical-align:-0.15em;"></span><span class="mord coloredeq eqm" style=""><span class="mord coloredeq eqn" style=""><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.151392em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">s</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span></span> projection described above.</p>
</div>
<div class='code'>
@ -115,11 +115,11 @@
<a href='#section-2'>#</a>
</div>
<ul><li><code class="highlight"><span></span><span class="n">in_channels</span></code>
is the number of channels in <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.43056em;vertical-align:0em;"></span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span> </li>
is the number of channels in <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.43056em;vertical-align:0em;"></span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span> </li>
<li><code class="highlight"><span></span><span class="n">out_channels</span></code>
is the number of channels in <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqd" style=""><span class="mord coloredeq eqj" style=""><span class="mord coloredeq eqt" style=""><span class="mord mathcal" style="margin-right:0.09931em">F</span></span></span><span class="mopen" style="">(</span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span><span class="mpunct" style="">,</span><span class="mspace" style="margin-right:0.16666666666666666em;"></span><span class="mopen" style="">{</span><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.31166399999999994em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mclose" style="">})</span></span></span></span></span> </li>
is the number of channels in <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:1em;vertical-align:-0.25em;"></span><span class="mord coloredeq eqd" style=""><span class="mord coloredeq eqj" style=""><span class="mord coloredeq eqr" style=""><span class="mord mathcal" style="margin-right:0.09931em">F</span></span></span><span class="mopen" style="">(</span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span><span class="mpunct" style="">,</span><span class="mspace" style="margin-right:0.16666666666666666em;"></span><span class="mopen" style="">{</span><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.31166399999999994em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">i</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mclose" style="">})</span></span></span></span></span> </li>
<li><code class="highlight"><span></span><span class="n">stride</span></code>
is the stride length in the convolution operation for <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.68333em;vertical-align:0em;"></span><span class="mord coloredeq eqt" style=""><span class="mord mathnormal" style="margin-right:0.13889em">F</span></span></span></span></span>. We do the same stride on the shortcut connection, to match the feature-map size.</li></ul>
is the stride length in the convolution operation for <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.68333em;vertical-align:0em;"></span><span class="mord coloredeq eqr" style=""><span class="mord mathnormal" style="margin-right:0.13889em">F</span></span></span></span></span>. We do the same stride on the shortcut connection, to match the feature-map size.</li></ul>
</div>
<div class='code'>
@ -142,7 +142,7 @@
<div class='section-link'>
<a href='#section-4'>#</a>
</div>
<p>Convolution layer for linear projection <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.83333em;vertical-align:-0.15em;"></span><span class="mord coloredeq eqm" style=""><span class="mord coloredeq eqp" style=""><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.151392em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">s</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span></span> </p>
<p>Convolution layer for linear projection <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.83333em;vertical-align:-0.15em;"></span><span class="mord coloredeq eqm" style=""><span class="mord coloredeq eqn" style=""><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.151392em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">s</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span></span> </p>
</div>
<div class='code'>
@ -197,7 +197,7 @@
to <code class="highlight"><span></span><span class="n">out_channels</span></code>
, where the <code class="highlight"><span></span><span class="n">out_channels</span></code>
is higher than <code class="highlight"><span></span><span class="n">in_channels</span></code>
when we reduce the feature map size with a stride length greater than <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqr" style=""><span class="mord" style="">1</span></span></span></span></span>.</p>
when we reduce the feature map size with a stride length greater than <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style="">1</span></span></span></span></span>.</p>
<p>The second convolution layer maps from <code class="highlight"><span></span><span class="n">out_channels</span></code>
to <code class="highlight"><span></span><span class="n">out_channels</span></code>
and always has a stride length of 1.</p>
@ -214,7 +214,7 @@
<a href='#section-9'>#</a>
</div>
<ul><li><code class="highlight"><span></span><span class="n">in_channels</span></code>
is the number of channels in <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.43056em;vertical-align:0em;"></span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span> </li>
is the number of channels in <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.43056em;vertical-align:0em;"></span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span> </li>
<li><code class="highlight"><span></span><span class="n">out_channels</span></code>
is the number of output channels </li>
<li><code class="highlight"><span></span><span class="n">stride</span></code>
@ -302,7 +302,7 @@
<div class='section-link'>
<a href='#section-16'>#</a>
</div>
<p>Shortcut connection should be a projection if the stride length is not <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqr" style=""><span class="mord" style="">1</span></span></span></span></span> of if the number of channels change </p>
<p>Shortcut connection should be a projection if the stride length is not <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style="">1</span></span></span></span></span> of if the number of channels change </p>
</div>
<div class='code'>
@ -314,7 +314,7 @@
<div class='section-link'>
<a href='#section-17'>#</a>
</div>
<p>Projection <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.83333em;vertical-align:-0.15em;"></span><span class="mord coloredeq eqm" style=""><span class="mord coloredeq eqp" style=""><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.151392em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">s</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span></span> </p>
<p>Projection <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.83333em;vertical-align:-0.15em;"></span><span class="mord coloredeq eqm" style=""><span class="mord coloredeq eqn" style=""><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.151392em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">s</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span></span> </p>
</div>
<div class='code'>
@ -327,7 +327,7 @@
<div class='section-link'>
<a href='#section-18'>#</a>
</div>
<p>Identity <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.43056em;vertical-align:0em;"></span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span> </p>
<p>Identity <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.43056em;vertical-align:0em;"></span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span> </p>
</div>
<div class='code'>
@ -415,24 +415,24 @@
</div>
<p> <a id="bottleneck_residual_block"></a></p>
<h2>Bottleneck Residual Block</h2>
<p>This implements the bottleneck block described in the paper. It has <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.72777em;vertical-align:-0.08333em;"></span><span class="mord coloredeq eqr" style=""><span class="mord" style="">1</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqr" style=""><span class="mord" style="">1</span></span></span></span></span>, <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.72777em;vertical-align:-0.08333em;"></span><span class="mord">3</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord">3</span></span></span></span>, and <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.72777em;vertical-align:-0.08333em;"></span><span class="mord coloredeq eqr" style=""><span class="mord" style="">1</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqr" style=""><span class="mord" style="">1</span></span></span></span></span> convolution layers.</p>
<p>This implements the bottleneck block described in the paper. It has <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.72777em;vertical-align:-0.08333em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style="">1</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style="">1</span></span></span></span></span>, <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.72777em;vertical-align:-0.08333em;"></span><span class="mord">3</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord">3</span></span></span></span>, and <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.72777em;vertical-align:-0.08333em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style="">1</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style="">1</span></span></span></span></span> convolution layers.</p>
<p><img alt="Bottlenext Block" src="bottleneck_block.svg"></p>
<p>The first convolution layer maps from <code class="highlight"><span></span><span class="n">in_channels</span></code>
to <code class="highlight"><span></span><span class="n">bottleneck_channels</span></code>
with a <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqr" style=""><span class="mord" style="">1</span></span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span><span class="mord coloredeq eqr" style=""><span class="mord" style="">1</span></span></span></span></span> convolution, where the <code class="highlight"><span></span><span class="n">bottleneck_channels</span></code>
with a <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.72777em;vertical-align:-0.08333em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style="">1</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style="">1</span></span></span></span></span> convolution, where the <code class="highlight"><span></span><span class="n">bottleneck_channels</span></code>
is lower than <code class="highlight"><span></span><span class="n">in_channels</span></code>
.</p>
<p>The second <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord">3</span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span><span class="mord">3</span></span></span></span> convolution layer maps from <code class="highlight"><span></span><span class="n">bottleneck_channels</span></code>
<p>The second <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.72777em;vertical-align:-0.08333em;"></span><span class="mord">3</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord">3</span></span></span></span> convolution layer maps from <code class="highlight"><span></span><span class="n">bottleneck_channels</span></code>
to <code class="highlight"><span></span><span class="n">bottleneck_channels</span></code>
. This can have a stride length greater than <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqr" style=""><span class="mord" style="">1</span></span></span></span></span> when we want to compress the feature map size.</p>
<p>The third, final <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqr" style=""><span class="mord" style="">1</span></span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span><span class="mord coloredeq eqr" style=""><span class="mord" style="">1</span></span></span></span></span> convolution layer maps to <code class="highlight"><span></span><span class="n">out_channels</span></code>
. This can have a stride length greater than <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style="">1</span></span></span></span></span> when we want to compress the feature map size.</p>
<p>The third, final <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.72777em;vertical-align:-0.08333em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style="">1</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style="">1</span></span></span></span></span> convolution layer maps to <code class="highlight"><span></span><span class="n">out_channels</span></code>
. <code class="highlight"><span></span><span class="n">out_channels</span></code>
is higher than <code class="highlight"><span></span><span class="n">in_channels</span></code>
if the stride length is greater than <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqr" style=""><span class="mord" style="">1</span></span></span></span></span>; otherwise, <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.84444em;vertical-align:-0.15em;"></span><span class="mord mathnormal">o</span><span class="mord mathnormal">u</span><span class="mord"><span class="mord mathnormal">t</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.151392em;"><span style="top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight">c</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mord mathnormal">hann</span><span class="mord mathnormal">e</span><span class="mord mathnormal" style="margin-right:0.01968em;">l</span><span class="mord mathnormal">s</span></span></span></span> is equal to <code class="highlight"><span></span><span class="n">in_channels</span></code>
if the stride length is greater than <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style="">1</span></span></span></span></span>; otherwise, <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.84444em;vertical-align:-0.15em;"></span><span class="mord mathnormal">o</span><span class="mord mathnormal">u</span><span class="mord"><span class="mord mathnormal">t</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.151392em;"><span style="top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight"><span class="mord mathnormal mtight">c</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span><span class="mord mathnormal">hann</span><span class="mord mathnormal">e</span><span class="mord mathnormal" style="margin-right:0.01968em;">l</span><span class="mord mathnormal">s</span></span></span></span> is equal to <code class="highlight"><span></span><span class="n">in_channels</span></code>
.</p>
<p><code class="highlight"><span></span><span class="n">bottleneck_channels</span></code>
is less than <code class="highlight"><span></span><span class="n">in_channels</span></code>
and the <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord">3</span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span><span class="mord">3</span></span></span></span> convolution is performed on this shrunk space (hence the bottleneck). The two <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqr" style=""><span class="mord" style="">1</span></span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span><span class="mord coloredeq eqr" style=""><span class="mord" style="">1</span></span></span></span></span> convolution decreases and increases the number of channels.</p>
and the <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.72777em;vertical-align:-0.08333em;"></span><span class="mord">3</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord">3</span></span></span></span> convolution is performed on this shrunk space (hence the bottleneck). The two <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.72777em;vertical-align:-0.08333em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style="">1</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style="">1</span></span></span></span></span> convolution decreases and increases the number of channels.</p>
</div>
<div class='code'>
@ -445,13 +445,13 @@
<a href='#section-26'>#</a>
</div>
<ul><li><code class="highlight"><span></span><span class="n">in_channels</span></code>
is the number of channels in <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.43056em;vertical-align:0em;"></span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span> </li>
is the number of channels in <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.43056em;vertical-align:0em;"></span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span> </li>
<li><code class="highlight"><span></span><span class="n">bottleneck_channels</span></code>
is the number of channels for the <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord">3</span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span><span class="mord">3</span></span></span></span> convlution </li>
is the number of channels for the <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.72777em;vertical-align:-0.08333em;"></span><span class="mord">3</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord">3</span></span></span></span> convlution </li>
<li><code class="highlight"><span></span><span class="n">out_channels</span></code>
is the number of output channels </li>
<li><code class="highlight"><span></span><span class="n">stride</span></code>
is the stride length in the <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord">3</span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span><span class="mord">3</span></span></span></span> convolution operation.</li></ul>
is the stride length in the <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.72777em;vertical-align:-0.08333em;"></span><span class="mord">3</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord">3</span></span></span></span> convolution operation.</li></ul>
</div>
<div class='code'>
@ -474,7 +474,7 @@
<div class='section-link'>
<a href='#section-28'>#</a>
</div>
<p>First <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.72777em;vertical-align:-0.08333em;"></span><span class="mord coloredeq eqr" style=""><span class="mord" style="">1</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqr" style=""><span class="mord" style="">1</span></span></span></span></span> convolution layer, this maps to <code class="highlight"><span></span><span class="n">bottleneck_channels</span></code>
<p>First <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.72777em;vertical-align:-0.08333em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style="">1</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style="">1</span></span></span></span></span> convolution layer, this maps to <code class="highlight"><span></span><span class="n">bottleneck_channels</span></code>
</p>
</div>
@ -547,7 +547,7 @@
<div class='section-link'>
<a href='#section-34'>#</a>
</div>
<p>Third <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.72777em;vertical-align:-0.08333em;"></span><span class="mord coloredeq eqr" style=""><span class="mord" style="">1</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqr" style=""><span class="mord" style="">1</span></span></span></span></span> convolution layer, this maps to <code class="highlight"><span></span><span class="n">out_channels</span></code>
<p>Third <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.72777em;vertical-align:-0.08333em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style="">1</span></span><span class="mspace" style="margin-right:0.2222222222222222em;"></span><span class="mbin">×</span><span class="mspace" style="margin-right:0.2222222222222222em;"></span></span><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style="">1</span></span></span></span></span> convolution layer, this maps to <code class="highlight"><span></span><span class="n">out_channels</span></code>
. </p>
</div>
@ -572,7 +572,7 @@
<div class='section-link'>
<a href='#section-36'>#</a>
</div>
<p>Shortcut connection should be a projection if the stride length is not <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqr" style=""><span class="mord" style="">1</span></span></span></span></span> of if the number of channels change </p>
<p>Shortcut connection should be a projection if the stride length is not <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.64444em;vertical-align:0em;"></span><span class="mord coloredeq eqp" style=""><span class="mord" style="">1</span></span></span></span></span> of if the number of channels change </p>
</div>
<div class='code'>
@ -584,7 +584,7 @@
<div class='section-link'>
<a href='#section-37'>#</a>
</div>
<p>Projection <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.83333em;vertical-align:-0.15em;"></span><span class="mord coloredeq eqm" style=""><span class="mord coloredeq eqp" style=""><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.151392em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">s</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span></span> </p>
<p>Projection <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.83333em;vertical-align:-0.15em;"></span><span class="mord coloredeq eqm" style=""><span class="mord coloredeq eqn" style=""><span class="mord" style=""><span class="mord mathnormal" style="margin-right:0.13889em">W</span><span class="msupsub"><span class="vlist-t vlist-t2"><span class="vlist-r"><span class="vlist" style="height:0.151392em;"><span style="top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em;"><span class="pstrut" style="height:2.7em;"></span><span class="sizing reset-size6 size3 mtight" style=""><span class="mord mathnormal mtight" style="">s</span></span></span></span><span class="vlist-s"></span></span><span class="vlist-r"><span class="vlist" style="height:0.15em;"><span></span></span></span></span></span></span></span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span></span> </p>
</div>
<div class='code'>
@ -597,7 +597,7 @@
<div class='section-link'>
<a href='#section-38'>#</a>
</div>
<p>Identity <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.43056em;vertical-align:0em;"></span><span class="mord coloredeq equ" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span> </p>
<p>Identity <span class="katex"><span aria-hidden="true" class="katex-html"><span class="base"><span class="strut" style="height:0.43056em;vertical-align:0em;"></span><span class="mord coloredeq eqs" style=""><span class="mord mathnormal" style="">x</span></span></span></span></span> </p>
</div>
<div class='code'>

View File

@ -166,29 +166,29 @@ class BottleneckResidualBlock(Module):
![Bottlenext Block](bottleneck_block.svg)
The first convolution layer maps from `in_channels` to `bottleneck_channels` with a $1x1$
The first convolution layer maps from `in_channels` to `bottleneck_channels` with a $1 \times 1$
convolution,
where the `bottleneck_channels` is lower than `in_channels`.
The second $3x3$ convolution layer maps from `bottleneck_channels` to `bottleneck_channels`.
The second $3 \times 3$ convolution layer maps from `bottleneck_channels` to `bottleneck_channels`.
This can have a stride length greater than $1$ when we want to compress the
feature map size.
The third, final $1x1$ convolution layer maps to `out_channels`.
The third, final $1 \times 1$ convolution layer maps to `out_channels`.
`out_channels` is higher than `in_channels` if the stride length is greater than $1$;
otherwise, $out_channels$ is equal to `in_channels`.
`bottleneck_channels` is less than `in_channels` and the $3x3$ convolution is performed
on this shrunk space (hence the bottleneck). The two $1x1$ convolution decreases and increases
`bottleneck_channels` is less than `in_channels` and the $3 \times 3$ convolution is performed
on this shrunk space (hence the bottleneck). The two $1 \times 1$ convolution decreases and increases
the number of channels.
"""
def __init__(self, in_channels: int, bottleneck_channels: int, out_channels: int, stride: int):
"""
* `in_channels` is the number of channels in $x$
* `bottleneck_channels` is the number of channels for the $3x3$ convlution
* `bottleneck_channels` is the number of channels for the $3 \times 3$ convlution
* `out_channels` is the number of output channels
* `stride` is the stride length in the $3x3$ convolution operation.
* `stride` is the stride length in the $3 \times 3$ convolution operation.
"""
super().__init__()