CIFAR-10のトレインニング・その一

データセットの紹介

CIFAR-10は１０類の物のデータセットである。
CIFARの完全の名前は知らないけど「なぜあいつのホムページにはこういうのは全然見つかなかったの？」、とりあえず、CIFARはカナダのリサーチセンターです。
CIFARのホムページはこっち：CIFARの公式サイト
CIFAR-10またCIFAR-100は全てこいつ作り出した公開データセットですし、CIFAR-100の難しさが全く違いのレベルだ。
そして、初心者によって、CIFAR-10はCIFAR-100より適当な入門挑戦だと思います。

ニューラルネットワークの建造

基本タイプはCNNを選んでください、続いてのは、具体的なパラメーター、例はこちらです。

def Build_IINN(n_class): dim_x = [1, None, None, 3] dim_y = [1, n_class]

</span><span class="sc1"># configure the convolution layers</span><span class="sc0">
</span><span class="sc11">n_conv</span><span class="sc0"> </span><span class="sc10">=</span><span class="sc0"> </span><span class="sc2">8</span><span class="sc0">
</span><span class="sc11">conv_config</span><span class="sc0"> </span><span class="sc10">=</span><span class="sc0"> </span><span class="sc10">[</span><span class="sc5">None</span><span class="sc10">]</span><span class="sc0"> </span><span class="sc10">*</span><span class="sc0"> </span><span class="sc11">n_conv</span><span class="sc0">
</span><span class="sc5">for</span><span class="sc0"> </span><span class="sc11">i</span><span class="sc0"> </span><span class="sc5">in</span><span class="sc0"> </span><span class="sc11">range</span><span class="sc10">(</span><span class="sc11">n_conv</span><span class="sc10">):</span><span class="sc0">
    </span><span class="sc11">conv_config</span><span class="sc10">[</span><span class="sc11">i</span><span class="sc10">]</span><span class="sc0"> </span><span class="sc10">=</span><span class="sc0"> </span><span class="sc11">new_conv_config</span><span class="sc10">(</span><span class="sc2">3</span><span class="sc10">,</span><span class="sc0"> </span><span class="sc2">3</span><span class="sc10">,</span><span class="sc0"> </span><span class="sc11">i</span><span class="sc10">%</span><span class="sc2">2</span><span class="sc10">+</span><span class="sc2">1</span><span class="sc10">,</span><span class="sc0"> </span><span class="sc11">i</span><span class="sc10">%</span><span class="sc2">2</span><span class="sc10">+</span><span class="sc2">1</span><span class="sc10">,</span><span class="sc0"> </span><span class="sc2">8</span><span class="sc10">&lt;&lt;(</span><span class="sc11">i</span><span class="sc10">//</span><span class="sc2">2</span><span class="sc10">))</span><span class="sc0">

</span><span class="sc1"># configure the fully connectied layers</span><span class="sc0">
</span><span class="sc11">n_fc</span><span class="sc0"> </span><span class="sc10">=</span><span class="sc0"> </span><span class="sc2">3</span><span class="sc0">
</span><span class="sc11">fc_config</span><span class="sc0"> </span><span class="sc10">=</span><span class="sc0"> </span><span class="sc10">[</span><span class="sc5">None</span><span class="sc10">]</span><span class="sc0"> </span><span class="sc10">*</span><span class="sc0"> </span><span class="sc11">n_fc</span><span class="sc0">
</span><span class="sc5">for</span><span class="sc0"> </span><span class="sc11">i</span><span class="sc0"> </span><span class="sc5">in</span><span class="sc0"> </span><span class="sc11">range</span><span class="sc10">(</span><span class="sc11">n_fc</span><span class="sc10">):</span><span class="sc0">
    </span><span class="sc11">fc_config</span><span class="sc10">[</span><span class="sc11">i</span><span class="sc10">]</span><span class="sc0"> </span><span class="sc10">=</span><span class="sc0"> </span><span class="sc11">new_fc_config</span><span class="sc10">(</span><span class="sc2">16</span><span class="sc0"> </span><span class="sc10">&lt;&lt;</span><span class="sc0"> </span><span class="sc11">i</span><span class="sc10">)</span><span class="sc0">

</span><span class="sc1"># configure the special module : feedback attention</span><span class="sc0">
</span><span class="sc11">n_att</span><span class="sc0"> </span><span class="sc10">=</span><span class="sc0"> </span><span class="sc2">3</span><span class="sc0">
</span><span class="sc11">att_config</span><span class="sc0"> </span><span class="sc10">=</span><span class="sc0"> </span><span class="sc10">[</span><span class="sc5">None</span><span class="sc10">]</span><span class="sc0"> </span><span class="sc10">*</span><span class="sc0"> </span><span class="sc11">n_att</span><span class="sc0">
</span><span class="sc5">for</span><span class="sc0"> </span><span class="sc11">i</span><span class="sc0"> </span><span class="sc5">in</span><span class="sc0"> </span><span class="sc11">range</span><span class="sc10">(</span><span class="sc11">n_att</span><span class="sc10">):</span><span class="sc0">
    </span><span class="sc11">att_config</span><span class="sc10">[</span><span class="sc11">i</span><span class="sc10">]</span><span class="sc0"> </span><span class="sc10">=</span><span class="sc0"> </span><span class="sc11">new_fc_config</span><span class="sc10">(</span><span class="sc2">64</span><span class="sc0"> </span><span class="sc10">&gt;&gt;</span><span class="sc0"> </span><span class="sc11">i</span><span class="sc10">)</span><span class="sc0">

</span><span class="sc5">return</span><span class="sc0"> </span><span class="sc11">IINN</span><span class="sc10">(</span><span class="sc11">dim_x</span><span class="sc10">,</span><span class="sc0"> </span><span class="sc11">dim_y</span><span class="sc10">,</span><span class="sc0">
            </span><span class="sc11">conv_config</span><span class="sc10">,</span><span class="sc0">
            </span><span class="sc11">fc_config</span><span class="sc10">,</span><span class="sc0">
            </span><span class="sc11">att_config</span><span class="sc10">)</span></div></body>

研究のため、一応batch normを利用しない。
トレインニングは３００エポックでした。結果は悪いです。

TRAINING STAGE#1:
TRAIN = 1.00000 0.99968 0.99954 0.99922 0.99824 0.99698 0.98404
TEST = 0.66570 0.66420 0.66380 0.66330 0.66300 0.65990 0.65810