我正在研究叶子疾病，并尝试在SpMohanty's PlantVillage-dataset上实施CNN。
它有38 classes，每个类别中的图像可变，范围为1500至3000图像/类。 Total images = 54303
这是我的Colab notebook。问题是，在预测图像时，它总是抛出模型训练的第一类。我不明白为什么。
这是模型摘要：
Model: "sequential_1"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_1 (Conv2D) (None, 64, 64, 32) 896
_________________________________________________________________
activation_1 (Activation) (None, 64, 64, 32) 0
_________________________________________________________________
batch_normalization_1 (Batch (None, 64, 64, 32) 128
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 21, 21, 32) 0
_________________________________________________________________
dropout_1 (Dropout) (None, 21, 21, 32) 0
_________________________________________________________________
conv2d_2 (Conv2D) (None, 21, 21, 64) 18496
_________________________________________________________________
activation_2 (Activation) (None, 21, 21, 64) 0
_________________________________________________________________
batch_normalization_2 (Batch (None, 21, 21, 64) 256
_________________________________________________________________
conv2d_3 (Conv2D) (None, 21, 21, 64) 36928
_________________________________________________________________
activation_3 (Activation) (None, 21, 21, 64) 0
_________________________________________________________________
batch_normalization_3 (Batch (None, 21, 21, 64) 256
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 10, 10, 64) 0
_________________________________________________________________
dropout_2 (Dropout) (None, 10, 10, 64) 0
_________________________________________________________________
conv2d_4 (Conv2D) (None, 10, 10, 128) 73856
_________________________________________________________________
activation_4 (Activation) (None, 10, 10, 128) 0
_________________________________________________________________
batch_normalization_4 (Batch (None, 10, 10, 128) 512
_________________________________________________________________
conv2d_5 (Conv2D) (None, 10, 10, 128) 147584
_________________________________________________________________
activation_5 (Activation) (None, 10, 10, 128) 0
_________________________________________________________________
batch_normalization_5 (Batch (None, 10, 10, 128) 512
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 5, 5, 128) 0
_________________________________________________________________
dropout_3 (Dropout) (None, 5, 5, 128) 0
_________________________________________________________________
flatten_1 (Flatten) (None, 3200) 0
_________________________________________________________________
dense_1 (Dense) (None, 1024) 3277824
_________________________________________________________________
activation_6 (Activation) (None, 1024) 0
_________________________________________________________________
batch_normalization_6 (Batch (None, 1024) 4096
_________________________________________________________________
dropout_4 (Dropout) (None, 1024) 0
_________________________________________________________________
dense_2 (Dense) (None, 38) 38950
_________________________________________________________________
activation_7 (Activation) (None, 38) 0
=================================================================
Total params: 3,600,294
Trainable params: 3,597,414
Non-trainable params: 2,880
_________________________________________________________________
尽管该模型达到了98.94％的验证准确度和99.7％的火车准确度。
型号历史如下：
Epoch 1/10
678/678 [==============================] - 691s 1s/step - loss: 0.0527 - accuracy: 0.9836 - val_loss: 0.0521 - val_accuracy: 0.9840
Epoch 2/10
678/678 [==============================] - 705s 1s/step - loss: 0.0283 - accuracy: 0.9900 - val_loss: 0.0926 - val_accuracy: 0.9787
Epoch 3/10
678/678 [==============================] - 680s 1s/step - loss: 0.0205 - accuracy: 0.9925 - val_loss: 0.0228 - val_accuracy: 0.9924
Epoch 4/10
678/678 [==============================] - 692s 1s/step - loss: 0.0170 - accuracy: 0.9938 - val_loss: 0.0741 - val_accuracy: 0.9828
Epoch 5/10
678/678 [==============================] - 679s 1s/step - loss: 0.0148 - accuracy: 0.9946 - val_loss: 0.0503 - val_accuracy: 0.9860
Epoch 6/10
678/678 [==============================] - 682s 1s/step - loss: 0.0129 - accuracy: 0.9953 - val_loss: 0.0323 - val_accuracy: 0.9918
Epoch 7/10
678/678 [==============================] - 691s 1s/step - loss: 0.0110 - accuracy: 0.9960 - val_loss: 0.0393 - val_accuracy: 0.9890
Epoch 8/10
678/678 [==============================] - 701s 1s/step - loss: 0.0098 - accuracy: 0.9965 - val_loss: 0.0420 - val_accuracy: 0.9875
Epoch 9/10
678/678 [==============================] - 692s 1s/step - loss: 0.0090 - accuracy: 0.9967 - val_loss: 0.0687 - val_accuracy: 0.9855
Epoch 10/10
678/678 [==============================] - 690s 1s/step - loss: 0.0082 - accuracy: 0.9971 - val_loss: 0.0414 - val_accuracy: 0.9894
这看起来像是模型过度拟合的情况，但是我对各种train_test_splits做了一个混淆矩阵，发现其中80％的训练是最好的。因此，我认为该模型不适合。它还总是预测第一个类，因此我尝试将类的名称更改为[eg：Apple_scab更改为Z_Apple_Scab]，然后开始将输出打印为：Apple_Blackrot（Apple_scab之后按字母顺序的下一个标签，现在该标签在训练中成为第一个标签。）
T.I.A。
[更新]：
这是我用来预测新图像的预测函数...
imAr = cvtim(impath)
savedclfmodel = pickle.load(open(model_file,'rb'))
pred = savedclfmodel.predict(imAr)
lb = pickle.load(open(lb_file,'rb'))
ret_data = lb.inverse_transform(pred)[0]
print("Predicted: ",ret_data)
其中cvtim（image_path）使用keras.preprocessing.image.img_to_array将图像转换为数组
然后savedclfmodel使用pickle.load()打开先前保存的模型权重文件并存储在lb中，然后使用该文件来查找inverse_transform以获得预测的标签
答案 0 :(得分：0)
您的训练数据看起来很好，并且训练和验证准确性很高。我看不出过度拟合的迹象。验证损失在最低点附近小幅波动是正常的。您没有显示足够的代码来找出问题所在。我假设您在测试集上使用model.predict。像
predictions=model.predict(data,batch_size=batch_size, steps=steps, verbose=0 )
# data is your test data provided as an array or from a generator
# then use the code below to find the predicted class
for i in range (0,len(predictions):
predicted_class_number=predictions[i].argmax()
# this will be an integer denoting the class