使用Python，Keras和TensorFlow训练第一个CNN
Posted 程序媛一枚~
tags:
篇首语：本文由小常识网(cha138.com)小编为大家整理，主要介绍了使用Python，Keras和TensorFlow训练第一个CNN相关的知识，希望对你有一定的参考价值。
使用Python，Keras和TensorFlow训练第一个CNN
这篇博客将介绍如何使用Python和Keras训练第一个卷积神经网络架构——ShallowNet，并在动物和CIFAR-10数据集上对其进行了训练。ShallowNet对动物的分类准确率为71%，比以前使用简单前馈神经网络的最佳分类准确率提高了12%。当应用于CIFAR-10时，ShallowNet达到了60%的精度，比以前使用简单多层神经网络的57%的最佳精度提高了（并且没有显著的过拟合）。
- ShallowNet是一种非常简单的CNN，只使用一个CONV层-通过使用多组CONV=>RELU=>POOL 操作训练更深层次的网络，可以获得更高的精度。
- ShallowNet架构只包含几个层-整个网络架构可以概括为：INPUT => CONV => RELU => FC。这种简单的网络架构将允许通过使用Keras库实现卷积神经网络来达到目的。
- 它是一个非常浅的CNN，然而ShallowNet能够在CIFAR-10和动物数据集上获得比许多其他方法更高的分类精度。
- ShallowNet CNN能够显著优于许多其他图像分类方法。
1. 效果图
python shallownet_animals.py --dataset datasets/animals
[INFO] loading images...
[INFO] processed500/36
[INFO] processed1000/36
[INFO] processed1500/36
[INFO] processed2000/36
[INFO] processed2500/38
[INFO] processed3000/38
[INFO] compiling model...
2022-07-03 12:28:08.856627: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2
[INFO] training network...
Train on 4500 samples, validate on 1500 samples
Epoch 1/100
4500/4500 [==============================] - 6s 1ms/sample - loss: 0.9715 - accuracy: 0.4960 - val_loss: 0.9313 - val_accuracy: 0.5147
Epoch 2/100
4500/4500 [==============================] - 2s 536us/sample - loss: 0.8726 - accuracy: 0.5662 - val_loss: 0.8771 - val_accuracy: 0.5387
Epoch 3/100
4500/4500 [==============================] - 2s 509us/sample - loss: 0.8299 - accuracy: 0.5856 - val_loss: 0.8338 - val_accuracy: 0.5580
Epoch 4/100
4500/4500 [==============================] - 2s 504us/sample - loss: 0.8025 - accuracy: 0.6100 - val_loss: 0.8457 - val_accuracy: 0.5607
Epoch 5/100
4500/4500 [==============================] - 2s 501us/sample - loss: 0.7837 - accuracy: 0.6169 - val_loss: 0.8012 - val_accuracy: 0.6313
Epoch 6/100
4500/4500 [==============================] - 2s 516us/sample - loss: 0.7635 - accuracy: 0.6413 - val_loss: 0.7617 - val_accuracy: 0.6513
Epoch 7/100
4500/4500 [==============================] - 2s 541us/sample - loss: 0.7469 - accuracy: 0.6456 - val_loss: 0.7499 - val_accuracy: 0.6380
Epoch 8/100
4500/4500 [==============================] - 2s 534us/sample - loss: 0.7319 - accuracy: 0.6618 - val_loss: 0.7531 - val_accuracy: 0.6407
Epoch 9/100
4500/4500 [==============================] - 2s 525us/sample - loss: 0.7202 - accuracy: 0.6642 - val_loss: 0.7483 - val_accuracy: 0.6200
Epoch 10/100
4500/4500 [==============================] - 2s 549us/sample - loss: 0.7030 - accuracy: 0.6880 - val_loss: 0.7450 - val_accuracy: 0.6507
Epoch 11/100
4500/4500 [==============================] - 2s 526us/sample - loss: 0.6838 - accuracy: 0.6960 - val_loss: 0.7061 - val_accuracy: 0.6753
Epoch 12/100
4500/4500 [==============================] - 2s 517us/sample - loss: 0.6748 - accuracy: 0.6962 - val_loss: 0.7228 - val_accuracy: 0.6593
Epoch 13/100
4500/4500 [==============================] - 2s 528us/sample - loss: 0.6592 - accuracy: 0.7076 - val_loss: 0.6786 - val_accuracy: 0.6947
Epoch 14/100
4500/4500 [==============================] - 2s 523us/sample - loss: 0.6414 - accuracy: 0.7187 - val_loss: 0.6656 - val_accuracy: 0.7053
Epoch 15/100
4500/4500 [==============================] - 2s 513us/sample - loss: 0.6278 - accuracy: 0.7327 - val_loss: 0.6977 - val_accuracy: 0.6553
Epoch 16/100
4500/4500 [==============================] - 2s 531us/sample - loss: 0.6140 - accuracy: 0.7373 - val_loss: 0.7598 - val_accuracy: 0.6173
Epoch 17/100
4500/4500 [==============================] - 2s 509us/sample - loss: 0.5979 - accuracy: 0.7493 - val_loss: 0.6814 - val_accuracy: 0.6500
Epoch 18/100
4500/4500 [==============================] - 2s 512us/sample - loss: 0.5892 - accuracy: 0.7442 - val_loss: 0.6723 - val_accuracy: 0.6567
Epoch 19/100
4500/4500 [==============================] - 2s 503us/sample - loss: 0.5743 - accuracy: 0.7524 - val_loss: 0.6594 - val_accuracy: 0.6620
Epoch 20/100
4500/4500 [==============================] - 2s 534us/sample - loss: 0.5661 - accuracy: 0.7653 - val_loss: 0.6620 - val_accuracy: 0.6753
Epoch 21/100
4500/4500 [==============================] - 2s 524us/sample - loss: 0.5478 - accuracy: 0.7787 - val_loss: 0.6299 - val_accuracy: 0.6893
Epoch 22/100
4500/4500 [==============================] - 2s 515us/sample - loss: 0.5390 - accuracy: 0.7742 - val_loss: 0.5977 - val_accuracy: 0.7460
Epoch 23/100
4500/4500 [==============================] - 2s 532us/sample - loss: 0.5294 - accuracy: 0.7818 - val_loss: 0.6104 - val_accuracy: 0.7407
Epoch 24/100
4500/4500 [==============================] - 2s 518us/sample - loss: 0.5167 - accuracy: 0.7889 - val_loss: 0.5828 - val_accuracy: 0.7407
Epoch 25/100
4500/4500 [==============================] - 3s 561us/sample - loss: 0.5027 - accuracy: 0.7960 - val_loss: 0.6251 - val_accuracy: 0.7053
Epoch 26/100
4500/4500 [==============================] - 3s 588us/sample - loss: 0.4924 - accuracy: 0.8029 - val_loss: 0.6016 - val_accuracy: 0.7093
Epoch 27/100
4500/4500 [==============================] - 2s 547us/sample - loss: 0.4837 - accuracy: 0.8064 - val_loss: 0.5647 - val_accuracy: 0.7507
Epoch 28/100
4500/4500 [==============================] - 2s 513us/sample - loss: 0.4808 - accuracy: 0.8058 - val_loss: 0.5967 - val_accuracy: 0.7087
Epoch 29/100
4500/4500 [==============================] - 2s 517us/sample - loss: 0.4622 - accuracy: 0.8238 - val_loss: 0.5568 - val_accuracy: 0.7513
Epoch 30/100
4500/4500 [==============================] - 2s 524us/sample - loss: 0.4536 - accuracy: 0.8238 - val_loss: 0.5760 - val_accuracy: 0.7247
Epoch 31/100
4500/4500 [==============================] - 2s 537us/sample - loss: 0.4477 - accuracy: 0.8282 - val_loss: 0.5729 - val_accuracy: 0.7427
Epoch 32/100
4500/4500 [==============================] - 3s 565us/sample - loss: 0.4406 - accuracy: 0.8300 - val_loss: 0.5676 - val_accuracy: 0.7333
Epoch 33/100
4500/4500 [==============================] - 2s 539us/sample - loss: 0.4270 - accuracy: 0.8371 - val_loss: 0.5434 - val_accuracy: 0.7640
Epoch 34/100
4500/4500 [==============================] - 2s 530us/sample - loss: 0.4210 - accuracy: 0.8418 - val_loss: 0.5660 - val_accuracy: 0.7507
Epoch 35/100
4500/4500 [==============================] - 2s 531us/sample - loss: 0.4111 - accuracy: 0.8451 - val_loss: 0.5258 - val_accuracy: 0.7773
Epoch 36/100
4500/4500 [==============================] - 2s 511us/sample - loss: 0.4043 - accuracy: 0.8524 - val_loss: 0.5369 - val_accuracy: 0.7527
Epoch 37/100
4500/4500 [==============================] - 3s 574us/sample - loss: 0.3980 - accuracy: 0.8518 - val_loss: 0.5137 - val_accuracy: 0.7840
Epoch 38/100
4500/4500 [==============================] - 2s 537us/sample - loss: 0.3853 - accuracy: 0.8598 - val_loss: 0.5773 - val_accuracy: 0.7107
Epoch 39/100
4500/4500 [==============================] - 2s 509us/sample - loss: 0.3818 - accuracy: 0.8578 - val_loss: 0.5110 - val_accuracy: 0.7753
Epoch 40/100
4500/4500 [==============================] - 2s 509us/sample - loss: 0.3731 - accuracy: 0.8669 - val_loss: 0.5063 - val_accuracy: 0.7773
Epoch 41/100
4500/4500 [==============================] - 2s 527us/sample - loss: 0.3639 - accuracy: 0.8707 - val_loss: 0.5468 - val_accuracy: 0.7720
Epoch 42/100
4500/4500 [==============================] - 2s 512us/sample - loss: 0.3588 - accuracy: 0.8764 - val_loss: 0.5168 - val_accuracy: 0.7607
Epoch 43/100
4500/4500 [==============================] - 3s 582us/sample - loss: 0.3509 - accuracy: 0.8749 - val_loss: 0.4909 - val_accuracy: 0.8113
Epoch 44/100
4500/4500 [==============================] - 3s 612us/sample - loss: 0.3460 - accuracy: 0.8813 - val_loss: 0.4830 - val_accuracy: 0.8087
Epoch 45/100
4500/4500 [==============================] - 3s 604us/sample - loss: 0.3385 - accuracy: 0.8824 - val_loss: 0.4841 - val_accuracy: 0.8080
Epoch 46/100
4500/4500 [==============================] - 3s 574us/sample - loss: 0.3321 - accuracy: 0.8867 - val_loss: 0.4977 - val_accuracy: 0.7747
Epoch 47/100
4500/4500 [==============================] - 3s 581us/sample - loss: 0.3237 - accuracy: 0.8940 - val_loss: 0.4790 - val_accuracy: 0.8100
Epoch 48/100
4500/4500 [==============================] - 2s 524us/sample - loss: 0.3195 - accuracy: 0.8909 - val_loss: 0.4732 - val_accuracy: 0.8073
Epoch 49/100
4500/4500 [==============================] - 2s 535us/sample - loss: 0.3139 - accuracy: 0.8964 - val_loss: 0.5134 - val_accuracy: 0.7687
Epoch 50/100
4500/4500 [==============================] - 2s 519us/sample - loss: 0.3089 - accuracy: 0.8949 - val_loss: 0.4775 - val_accuracy: 0.7960
Epoch 51/100
4500/4500 [==============================] - 3s 558us/sample - loss: 0.2988 - accuracy: 0.9076 - val_loss: 0.4618 - val_accuracy: 0.8160
Epoch 52/100
4500/4500 [==============================] - 2s 538us/sample - loss: 0.2974 - accuracy: 0.9049 - val_loss: 0.4629 - val_accuracy: 0.8147
Epoch 53/100
4500/4500 [==============================] - 2s 542us/sample - loss: 0.2949 - accuracy: 0.9047 - val_loss: 0.4793 - val_accuracy: 0.7953
Epoch 54/100
4500/4500 [==============================] - 2s 534us/sample - loss: 0.2883 - accuracy: 0.9096 - val_loss: 0.4598 - val_accuracy: 0.8047
Epoch 55/100
4500/4500 [==============================] - 2s 535us/sample - loss: 0.2810 - accuracy: 0.9122 - val_loss: 0.4782 - val_accuracy: 0.7920
Epoch 56/100
4500/4500 [==============================] - 2s 519us/sample - loss: 0.2800 - accuracy: 0.9131 - val_loss: 0.4675 - val_accuracy: 0.8120
Epoch 57/100
4500/4500 [==============================] - 2s 544us/sample - loss: 0.2707 - accuracy: 0.9180 - val_loss: 0.4547 - val_accuracy: 0.8153
Epoch 58/100
4500/4500 [==============Keras + Tensorflow 和 Python 中的多处理
tensorflow和theano的python3 keras导入错误
使用Python，Keras和TensorFlow训练第一个CNN
使用Python，Keras和TensorFlow训练第一个CNN