一 項目展示

下面可以看到驗證集可以到了0.9823了，實際上，在下面的另外一個訓練，可以得到0.9839，我保守的寫了0.982

二 項目參數展示

我們先來看看LeNet 5 的結構與參數，參數有61，706個。

這個是我用keras寫的，可以看到參數只有886個。

項目代碼

我們來看一下模型的代碼，模型由傳統卷積改成了可分離卷積；
這里padding 主要是為了后面降維方便而設置
這里設置了5層卷積，是考慮到如果層數比較少的話，感受野比較小，增加卷積層，從而增加感受野（我覺得感受野，這時候比模型的復雜度重要）
因為mnist的字體大概占圖像面積70-85之間

代碼的注釋也寫得比較多，可以自行研究。

# 我用的是GPU訓練，如果訓練報錯的話，可以把SperableConv2D改成Conv2D # 先訓練1個epoch，再改成如下的sperableconv2d def my_network(input_shape=(32, 32, 1), classes=10):X_input = Input(input_shape)X = X_input# 這里padding 主要是為了后面降維方便而設置# 這里設置了5層卷積，是考慮到如果層數比較少的話，感受野比較小，增加卷積層，從而增加感受野# 因為mnist的字體大概占圖像面積70-85之間X = ZeroPadding2D((2, 2))(X_input)X = Conv2D(4, (3, 3), strides=(1, 1), padding='valid', activation='elu', name='conv1')(X)X = Conv2D(6, (3, 3), strides=(2, 2), padding='valid', activation='elu', name='conv2')(X)X = SeparableConv2D(8, (3, 3), strides=(2, 2), padding='valid', activation='elu', name='conv3')(X)X = SeparableConv2D(10, (3, 3), strides=(2, 2), padding='same', activation='elu', name='conv4')(X)X = SeparableConv2D(12, (3, 3), strides=(2, 2), padding='valid', activation='elu', name='conv5')(X)# 利用浮點卷積做為輸出，注意激活函數是softmaxX = Conv2D(classes, (1, 1), strides=(1, 1), padding='same', activation='softmax')(X)X = keras.layers.Reshape((classes,))(X)model = Model(inputs=X_input, outputs=X, name='my_network')return modelmodel_con = my_network(input_shape=(28, 28, 1), classes=10)# 這里的lr，如果用了學習衰減的話，就可以不設置了 optimizer = keras.optimizers.Adam(lr=0.001) model_con.compile(optimizer=optimizer, loss='categorical_crossentropy', metrics=['accuracy']) model_con.summary()

訓練過程

訓練的過程，用的是Adam優化器，利用了學習率衰減的策略，先由大的學習，加快訓練，再調整比較小的學習率來訓練，慢慢訓練得出的結果比較穩定。
下面是100個epochs的全部，識別的輸出，其它lr初始值為0.01，而不是0.001。

Epoch 1/100 lr change to 0.01 235/235 [==============================] - 2s 7ms/step - loss: 0.1065 - acc: 0.9673 - val_loss: 0.0792 - val_acc: 0.9755 Epoch 2/100 lr change to 0.01 235/235 [==============================] - 1s 6ms/step - loss: 0.0886 - acc: 0.9726 - val_loss: 0.1013 - val_acc: 0.9686 Epoch 3/100 lr change to 0.01 235/235 [==============================] - 1s 6ms/step - loss: 0.0884 - acc: 0.9732 - val_loss: 0.0744 - val_acc: 0.9768 Epoch 4/100 lr change to 0.01 235/235 [==============================] - 2s 6ms/step - loss: 0.0873 - acc: 0.9730 - val_loss: 0.0815 - val_acc: 0.9740 Epoch 5/100 lr change to 0.01 235/235 [==============================] - 1s 6ms/step - loss: 0.0851 - acc: 0.9733 - val_loss: 0.0724 - val_acc: 0.9768 Epoch 6/100 lr change to 0.01 235/235 [==============================] - 2s 6ms/step - loss: 0.0817 - acc: 0.9752 - val_loss: 0.0826 - val_acc: 0.9745 Epoch 7/100 lr change to 0.01 235/235 [==============================] - 1s 6ms/step - loss: 0.0855 - acc: 0.9741 - val_loss: 0.0830 - val_acc: 0.9740 Epoch 8/100 lr change to 0.01 235/235 [==============================] - 2s 6ms/step - loss: 0.0870 - acc: 0.9728 - val_loss: 0.0847 - val_acc: 0.9729 Epoch 9/100 lr change to 0.01 235/235 [==============================] - 2s 7ms/step - loss: 0.0813 - acc: 0.9752 - val_loss: 0.0701 - val_acc: 0.9787 Epoch 10/100 lr change to 0.01 235/235 [==============================] - 2s 7ms/step - loss: 0.0815 - acc: 0.9748 - val_loss: 0.0742 - val_acc: 0.9765 Epoch 11/100 lr change to 0.01 235/235 [==============================] - 2s 7ms/step - loss: 0.0809 - acc: 0.9749 - val_loss: 0.0721 - val_acc: 0.9770 Epoch 12/100 lr change to 0.005 235/235 [==============================] - 1s 6ms/step - loss: 0.0672 - acc: 0.9796 - val_loss: 0.0635 - val_acc: 0.9788 Epoch 13/100 lr change to 0.005 235/235 [==============================] - 2s 6ms/step - loss: 0.0662 - acc: 0.9800 - val_loss: 0.0575 - val_acc: 0.9808 Epoch 14/100 lr change to 0.005 235/235 [==============================] - 2s 6ms/step - loss: 0.0659 - acc: 0.9796 - val_loss: 0.0624 - val_acc: 0.9803 Epoch 15/100 lr change to 0.005 235/235 [==============================] - 1s 6ms/step - loss: 0.0662 - acc: 0.9793 - val_loss: 0.0631 - val_acc: 0.9791 Epoch 16/100 lr change to 0.005 235/235 [==============================] - 2s 6ms/step - loss: 0.0655 - acc: 0.9802 - val_loss: 0.0561 - val_acc: 0.9816 Epoch 17/100 lr change to 0.005 235/235 [==============================] - 2s 7ms/step - loss: 0.0659 - acc: 0.9799 - val_loss: 0.0653 - val_acc: 0.9808 Epoch 18/100 lr change to 0.005 235/235 [==============================] - 2s 6ms/step - loss: 0.0653 - acc: 0.9798 - val_loss: 0.0765 - val_acc: 0.9748 Epoch 19/100 lr change to 0.005 235/235 [==============================] - 2s 7ms/step - loss: 0.0656 - acc: 0.9792 - val_loss: 0.0565 - val_acc: 0.9811 Epoch 20/100 lr change to 0.005 235/235 [==============================] - 2s 7ms/step - loss: 0.0645 - acc: 0.9800 - val_loss: 0.0569 - val_acc: 0.9814 Epoch 21/100 lr change to 0.005 235/235 [==============================] - 2s 7ms/step - loss: 0.0639 - acc: 0.9804 - val_loss: 0.0566 - val_acc: 0.9814 Epoch 22/100 lr change to 0.005 235/235 [==============================] - 2s 7ms/step - loss: 0.0637 - acc: 0.9801 - val_loss: 0.0672 - val_acc: 0.9778 Epoch 23/100 lr change to 0.005 235/235 [==============================] - 2s 7ms/step - loss: 0.0631 - acc: 0.9807 - val_loss: 0.0607 - val_acc: 0.9804 Epoch 24/100 lr change to 0.0025 235/235 [==============================] - 2s 6ms/step - loss: 0.0562 - acc: 0.9825 - val_loss: 0.0587 - val_acc: 0.9804 Epoch 25/100 lr change to 0.0025 235/235 [==============================] - 2s 6ms/step - loss: 0.0550 - acc: 0.9831 - val_loss: 0.0542 - val_acc: 0.9819 Epoch 26/100 lr change to 0.0025 235/235 [==============================] - 2s 7ms/step - loss: 0.0541 - acc: 0.9833 - val_loss: 0.0541 - val_acc: 0.9818 Epoch 27/100 lr change to 0.0025 235/235 [==============================] - 2s 7ms/step - loss: 0.0544 - acc: 0.9829 - val_loss: 0.0545 - val_acc: 0.9825 Epoch 28/100 lr change to 0.0025 235/235 [==============================] - 2s 7ms/step - loss: 0.0548 - acc: 0.9829 - val_loss: 0.0514 - val_acc: 0.9827 Epoch 29/100 lr change to 0.0025 235/235 [==============================] - 2s 7ms/step - loss: 0.0536 - acc: 0.9833 - val_loss: 0.0587 - val_acc: 0.9805 Epoch 30/100 lr change to 0.0025 235/235 [==============================] - 2s 7ms/step - loss: 0.0534 - acc: 0.9831 - val_loss: 0.0683 - val_acc: 0.9783 Epoch 31/100 lr change to 0.0025 235/235 [==============================] - 2s 7ms/step - loss: 0.0535 - acc: 0.9839 - val_loss: 0.0517 - val_acc: 0.9831 Epoch 32/100 lr change to 0.0025 235/235 [==============================] - 2s 7ms/step - loss: 0.0537 - acc: 0.9831 - val_loss: 0.0524 - val_acc: 0.9821 Epoch 33/100 lr change to 0.0025 235/235 [==============================] - 2s 8ms/step - loss: 0.0533 - acc: 0.9833 - val_loss: 0.0543 - val_acc: 0.9820 Epoch 34/100 lr change to 0.0025 235/235 [==============================] - 2s 8ms/step - loss: 0.0527 - acc: 0.9836 - val_loss: 0.0529 - val_acc: 0.9814 Epoch 35/100 lr change to 0.0025 235/235 [==============================] - 2s 7ms/step - loss: 0.0534 - acc: 0.9835 - val_loss: 0.0558 - val_acc: 0.9814 Epoch 36/100 lr change to 0.00125 235/235 [==============================] - 2s 7ms/step - loss: 0.0487 - acc: 0.9852 - val_loss: 0.0523 - val_acc: 0.9823 Epoch 37/100 lr change to 0.00125 235/235 [==============================] - 2s 7ms/step - loss: 0.0478 - acc: 0.9854 - val_loss: 0.0500 - val_acc: 0.9837 Epoch 38/100 lr change to 0.00125 235/235 [==============================] - 2s 7ms/step - loss: 0.0482 - acc: 0.9851 - val_loss: 0.0532 - val_acc: 0.9823 Epoch 39/100 lr change to 0.00125 235/235 [==============================] - 2s 7ms/step - loss: 0.0480 - acc: 0.9849 - val_loss: 0.0509 - val_acc: 0.9832 Epoch 40/100 lr change to 0.00125 235/235 [==============================] - 2s 7ms/step - loss: 0.0477 - acc: 0.9850 - val_loss: 0.0509 - val_acc: 0.9824 Epoch 41/100 lr change to 0.00125 235/235 [==============================] - 2s 7ms/step - loss: 0.0481 - acc: 0.9850 - val_loss: 0.0509 - val_acc: 0.9831 Epoch 42/100 lr change to 0.00125 235/235 [==============================] - 2s 7ms/step - loss: 0.0482 - acc: 0.9847 - val_loss: 0.0511 - val_acc: 0.9827 Epoch 43/100 lr change to 0.00125 235/235 [==============================] - 2s 7ms/step - loss: 0.0480 - acc: 0.9851 - val_loss: 0.0513 - val_acc: 0.9838 Epoch 44/100 lr change to 0.00125 235/235 [==============================] - 2s 7ms/step - loss: 0.0474 - acc: 0.9850 - val_loss: 0.0530 - val_acc: 0.9831 Epoch 45/100 lr change to 0.00125 235/235 [==============================] - 2s 7ms/step - loss: 0.0476 - acc: 0.9852 - val_loss: 0.0513 - val_acc: 0.9831 Epoch 46/100 lr change to 0.00125 235/235 [==============================] - 2s 7ms/step - loss: 0.0470 - acc: 0.9855 - val_loss: 0.0554 - val_acc: 0.9817 Epoch 47/100 lr change to 0.00125 235/235 [==============================] - 2s 7ms/step - loss: 0.0471 - acc: 0.9850 - val_loss: 0.0527 - val_acc: 0.9817 Epoch 48/100 lr change to 0.000625 235/235 [==============================] - 2s 7ms/step - loss: 0.0451 - acc: 0.9857 - val_loss: 0.0506 - val_acc: 0.9828 Epoch 49/100 lr change to 0.000625 235/235 [==============================] - 2s 7ms/step - loss: 0.0448 - acc: 0.9861 - val_loss: 0.0501 - val_acc: 0.9838 Epoch 50/100 lr change to 0.000625 235/235 [==============================] - 2s 7ms/step - loss: 0.0448 - acc: 0.9859 - val_loss: 0.0490 - val_acc: 0.9829 Epoch 51/100 lr change to 0.000625 235/235 [==============================] - 2s 7ms/step - loss: 0.0447 - acc: 0.9865 - val_loss: 0.0501 - val_acc: 0.9834 Epoch 52/100 lr change to 0.000625 235/235 [==============================] - 2s 7ms/step - loss: 0.0444 - acc: 0.9864 - val_loss: 0.0488 - val_acc: 0.9826 Epoch 53/100 lr change to 0.000625 235/235 [==============================] - 2s 7ms/step - loss: 0.0446 - acc: 0.9860 - val_loss: 0.0507 - val_acc: 0.9826 Epoch 54/100 lr change to 0.000625 235/235 [==============================] - 2s 7ms/step - loss: 0.0446 - acc: 0.9856 - val_loss: 0.0511 - val_acc: 0.9827 Epoch 55/100 lr change to 0.000625 235/235 [==============================] - 2s 7ms/step - loss: 0.0446 - acc: 0.9862 - val_loss: 0.0506 - val_acc: 0.9830 Epoch 56/100 lr change to 0.000625 235/235 [==============================] - 2s 7ms/step - loss: 0.0446 - acc: 0.9858 - val_loss: 0.0500 - val_acc: 0.9830 Epoch 57/100 lr change to 0.000625 235/235 [==============================] - 2s 7ms/step - loss: 0.0443 - acc: 0.9861 - val_loss: 0.0517 - val_acc: 0.9831 Epoch 58/100 lr change to 0.000625 235/235 [==============================] - 2s 7ms/step - loss: 0.0442 - acc: 0.9860 - val_loss: 0.0507 - val_acc: 0.9828 Epoch 59/100 lr change to 0.000625 235/235 [==============================] - 2s 7ms/step - loss: 0.0445 - acc: 0.9864 - val_loss: 0.0504 - val_acc: 0.9833 Epoch 60/100 lr change to 0.0003125 235/235 [==============================] - 2s 7ms/step - loss: 0.0430 - acc: 0.9868 - val_loss: 0.0499 - val_acc: 0.9825 Epoch 61/100 lr change to 0.0003125 235/235 [==============================] - 2s 7ms/step - loss: 0.0429 - acc: 0.9865 - val_loss: 0.0507 - val_acc: 0.9834 Epoch 62/100 lr change to 0.0003125 235/235 [==============================] - 2s 7ms/step - loss: 0.0430 - acc: 0.9866 - val_loss: 0.0494 - val_acc: 0.9830 Epoch 63/100 lr change to 0.0003125 235/235 [==============================] - 2s 7ms/step - loss: 0.0432 - acc: 0.9865 - val_loss: 0.0501 - val_acc: 0.9836 Epoch 64/100 lr change to 0.0003125 235/235 [==============================] - 2s 7ms/step - loss: 0.0426 - acc: 0.9866 - val_loss: 0.0518 - val_acc: 0.9830 Epoch 65/100 lr change to 0.0003125 235/235 [==============================] - 2s 7ms/step - loss: 0.0428 - acc: 0.9865 - val_loss: 0.0500 - val_acc: 0.9824 Epoch 66/100 lr change to 0.0003125 235/235 [==============================] - 2s 7ms/step - loss: 0.0427 - acc: 0.9866 - val_loss: 0.0506 - val_acc: 0.9824 Epoch 67/100 lr change to 0.0003125 235/235 [==============================] - 2s 7ms/step - loss: 0.0430 - acc: 0.9866 - val_loss: 0.0504 - val_acc: 0.9826 Epoch 68/100 lr change to 0.0003125 235/235 [==============================] - 2s 7ms/step - loss: 0.0427 - acc: 0.9868 - val_loss: 0.0501 - val_acc: 0.9832 Epoch 69/100 lr change to 0.0003125 235/235 [==============================] - 2s 7ms/step - loss: 0.0425 - acc: 0.9866 - val_loss: 0.0505 - val_acc: 0.9823 Epoch 70/100 lr change to 0.0003125 235/235 [==============================] - 2s 7ms/step - loss: 0.0427 - acc: 0.9869 - val_loss: 0.0505 - val_acc: 0.9839 Epoch 71/100 lr change to 0.0003125 235/235 [==============================] - 2s 7ms/step - loss: 0.0432 - acc: 0.9868 - val_loss: 0.0494 - val_acc: 0.9838 Epoch 72/100 lr change to 0.00015625 235/235 [==============================] - 2s 7ms/step - loss: 0.0419 - acc: 0.9870 - val_loss: 0.0493 - val_acc: 0.9833 Epoch 73/100 lr change to 0.00015625 235/235 [==============================] - 2s 7ms/step - loss: 0.0419 - acc: 0.9871 - val_loss: 0.0498 - val_acc: 0.9835 Epoch 74/100 lr change to 0.00015625 235/235 [==============================] - 2s 7ms/step - loss: 0.0419 - acc: 0.9871 - val_loss: 0.0496 - val_acc: 0.9831 Epoch 75/100 lr change to 0.00015625 235/235 [==============================] - 2s 7ms/step - loss: 0.0419 - acc: 0.9868 - val_loss: 0.0499 - val_acc: 0.9834 Epoch 76/100 lr change to 0.00015625 235/235 [==============================] - 2s 7ms/step - loss: 0.0419 - acc: 0.9868 - val_loss: 0.0496 - val_acc: 0.9835 Epoch 77/100 lr change to 0.00015625 235/235 [==============================] - 2s 7ms/step - loss: 0.0420 - acc: 0.9870 - val_loss: 0.0500 - val_acc: 0.9834 Epoch 78/100 lr change to 0.00015625 235/235 [==============================] - 2s 7ms/step - loss: 0.0419 - acc: 0.9871 - val_loss: 0.0496 - val_acc: 0.9833 Epoch 79/100 lr change to 0.00015625 235/235 [==============================] - 2s 7ms/step - loss: 0.0419 - acc: 0.9868 - val_loss: 0.0497 - val_acc: 0.9835 Epoch 80/100 lr change to 0.00015625 235/235 [==============================] - 2s 7ms/step - loss: 0.0419 - acc: 0.9869 - val_loss: 0.0499 - val_acc: 0.9835 Epoch 81/100 lr change to 0.00015625 235/235 [==============================] - 2s 7ms/step - loss: 0.0420 - acc: 0.9866 - val_loss: 0.0499 - val_acc: 0.9830 Epoch 82/100 lr change to 0.00015625 235/235 [==============================] - 2s 7ms/step - loss: 0.0418 - acc: 0.9870 - val_loss: 0.0503 - val_acc: 0.9839 Epoch 83/100 lr change to 0.00015625 235/235 [==============================] - 2s 7ms/step - loss: 0.0420 - acc: 0.9868 - val_loss: 0.0494 - val_acc: 0.9832 Epoch 84/100 lr change to 7.8125e-05 235/235 [==============================] - 2s 7ms/step - loss: 0.0414 - acc: 0.9873 - val_loss: 0.0499 - val_acc: 0.9837 Epoch 85/100 lr change to 7.8125e-05 235/235 [==============================] - 2s 7ms/step - loss: 0.0416 - acc: 0.9871 - val_loss: 0.0496 - val_acc: 0.9833 Epoch 86/100 lr change to 7.8125e-05 235/235 [==============================] - 2s 7ms/step - loss: 0.0414 - acc: 0.9872 - val_loss: 0.0501 - val_acc: 0.9837 Epoch 87/100 lr change to 7.8125e-05 235/235 [==============================] - 2s 7ms/step - loss: 0.0415 - acc: 0.9871 - val_loss: 0.0502 - val_acc: 0.9836 Epoch 88/100 lr change to 7.8125e-05 235/235 [==============================] - 2s 7ms/step - loss: 0.0415 - acc: 0.9871 - val_loss: 0.0501 - val_acc: 0.9834 Epoch 89/100 lr change to 7.8125e-05 235/235 [==============================] - 2s 7ms/step - loss: 0.0414 - acc: 0.9870 - val_loss: 0.0496 - val_acc: 0.9835 Epoch 90/100 lr change to 7.8125e-05 235/235 [==============================] - 2s 7ms/step - loss: 0.0414 - acc: 0.9872 - val_loss: 0.0498 - val_acc: 0.9837 Epoch 91/100 lr change to 7.8125e-05 235/235 [==============================] - 2s 7ms/step - loss: 0.0416 - acc: 0.9870 - val_loss: 0.0502 - val_acc: 0.9839 Epoch 92/100 lr change to 7.8125e-05 235/235 [==============================] - 2s 7ms/step - loss: 0.0414 - acc: 0.9872 - val_loss: 0.0503 - val_acc: 0.9834 Epoch 93/100 lr change to 7.8125e-05 235/235 [==============================] - 2s 7ms/step - loss: 0.0415 - acc: 0.9870 - val_loss: 0.0500 - val_acc: 0.9836 Epoch 94/100 lr change to 7.8125e-05 235/235 [==============================] - 2s 7ms/step - loss: 0.0415 - acc: 0.9869 - val_loss: 0.0499 - val_acc: 0.9832 Epoch 95/100 lr change to 7.8125e-05 235/235 [==============================] - 2s 7ms/step - loss: 0.0415 - acc: 0.9871 - val_loss: 0.0502 - val_acc: 0.9836 Epoch 96/100 lr change to 6.25e-05 235/235 [==============================] - 2s 7ms/step - loss: 0.0413 - acc: 0.9871 - val_loss: 0.0501 - val_acc: 0.9838 Epoch 97/100 lr change to 6.25e-05 235/235 [==============================] - 2s 7ms/step - loss: 0.0413 - acc: 0.9872 - val_loss: 0.0501 - val_acc: 0.9837 Epoch 98/100 lr change to 6.25e-05 235/235 [==============================] - 2s 7ms/step - loss: 0.0413 - acc: 0.9872 - val_loss: 0.0503 - val_acc: 0.9834 Epoch 99/100 lr change to 6.25e-05 235/235 [==============================] - 2s 7ms/step - loss: 0.0413 - acc: 0.9870 - val_loss: 0.0499 - val_acc: 0.9833 Epoch 100/100 lr change to 6.25e-05 235/235 [==============================] - 2s 7ms/step - loss: 0.0417 - acc: 0.9870 - val_loss: 0.0500 - val_acc: 0.9834

總結

通過此次，認識到了感受野有時候比模型的復雜度更重要，還有學習率衰減策略也對模型有比較大的影響。
就是你用CPU訓練，這么少的參數，也會很快的。

下面附上本項目的代碼的Jupyter notebook。

鏈接：https://pan.baidu.com/s/1EEzfRDD_PAgSeS999Eq9-Q 提取碼：3131

如果你覺得好，請給我個star，如有問題，也可以與我聯系。

總結

以上是生活随笔為你收集整理的MNIST 手写数字识别，我是如何做到886个可训练参数，识别率达到98.2%? (参数、模型压缩）， Keras实现，模型优化的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。