高精度训练但低精度测试/预测

Question

I am using CNN to classify apple type. I achieved high accuracy on train data but really low accuracy on test data. Data is split into 80:20. I am not sure if my data is overfitting or not.

I have 2 folders that contain TraningData and TestData , and each folder has 4 subfolders braeburn, red_apple, red_delicious, rotten (containing corresponding pictures).

TRAIN_DIR = 'apple_fruit'
TEST_DIR = 'apple_fruit'
classes = ['braeburn','red_apples','red_delicious','rotten'] train_datagen = ImageDataGenerator(rescale = 1./255, shear_range=0.2, zoom_range=0.2, horizontal_flip=True, fill_mode='nearest')
 
test_datagen = ImageDataGenerator(rescale = 1./255) 

training_set = train_datagen.flow_from_directory(TRAIN_DIR,
shuffle=True,
target_size = (100,100),
batch_size = 25,
classes =['braeburn','red_apples','red_delicious','rotten'])

test_set= test_datagen.flow_from_directory(TEST_DIR,
target_size = (100, 100),
shuffle=True,
 batch_size = 25,classes = classes)

model =Sequential()
model.add(Conv2D(filters=128, kernel_size=(3,3),input_shape=(100,100,3), activation='relu', padding
= 'same'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(filters=16, kernel_size=(3,3), activation='relu', padding = 'same'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Flatten())
model.add(Dense(256))
model.add(Activation('relu'))
model.add(Dropout(0.6))
model.add(Dense(4,activation='softmax'))
model.compile(optimizer ='adam', loss = 'categorical_crossentropy', metrics = ['accuracy'])

history = model.fit(x=training_set,#y=training_set.labels,
steps_per_epoch=len(training_set),
epochs =10)

model.save('Ripe2_model6.h5')  # creates a HDF5 file 'my_model.h5'

model_path = "Ripe2_model6.h5"
loaded_model = keras.models.load_model(model_path)
classes = ['braeburn','red_apples','red_delicious','rotten']
predictions = model.predict(x=test_set, steps=len(test_set), verbose=True)
pred = np.round(predictions)

y_true=test_set.classes
y_pred=np.argmax(pred, axis=-1)
    > cm = confusion_matrix(y_true=test_set.classes, y_pred=np.argmax(pred, axis=-1))
test_set.classes
np.argmax(pred, axis=-1)
def plot_confusion_matrix(cm, classes,
normalize=False,
title='Confusion matrix',
cmap=plt.cm.Blues):

accuracy = np.trace(cm) / float(np.sum(cm))
misclass = 1 - accuracy

  """
This function prints and plots the confusion matrix.
Normalization can be applied by setting `normalize=True`.
    """
plt.imshow(cm, interpolation='nearest', cmap=cmap)
plt.title(title,color = 'white')
plt.colorbar()
tick_marks = np.arange(len(classes))
plt.xticks(tick_marks, classes, rotation=45,color = 'white')
plt.yticks(tick_marks, classes,color = 'white')
target_names = ['braeburn','red_apples','red_delicious','rotten']

if target_names is not None:
 tick_marks = np.arange(len(target_names))
 plt.xticks(tick_marks, target_names, rotation=45)
 plt.yticks(tick_marks, target_names)

if normalize:
 cm = cm.astype('float') / cm.sum(axis=1)[:, np.newaxis]
 thresh = cm.max() / 1.5 if normalize else cm.max() / 2
 for i, j in itertools.product(range(cm.shape[0]), 
  range(cm.shape[1])):
  if normalize:
   plt.text(j, i, "{:0.4f}".format(cm[i, j]),
   horizontalalignment="center",
   color="white" if cm[i, j] > thresh else "black")
  else:
   plt.text(j, i, "{:,}".format(cm[i, j]),
   horizontalalignment="center",
   color="white" if cm[i, j] > thresh else "black")

plt.tight_layout()
plt.ylabel('True label',color = 'white')
plt.xlabel('Predicted label',color = 'white')

cm_plot_labels = ['braeburn','red_apples','red_delicious','rotten']
plot_confusion_matrix(cm=cm, classes=cm_plot_labels, title='Confusion Matrix')

print(accuracy_score(y_true, y_pred))
print(recall_score(y_true, y_pred, average=None))
print(precision_score(y_true, y_pred, average=None))

The confusion matrix:

• accuracy - 0.2909090909090909
• recall - [0.23484848 0.32319392 0.15151515 0.36213992]
• precision - [0.23308271 0.32319392 0.15151515 0.36363636]

I have tried changing many features but still no progress.

Answer 1

It indicates that the data in the test set is quite different from what the model learned. To understand if it is overfitting or a single unfortunate split:

1. Check if your results are dependent on the initial Train/Test split . To achieve that you can:

• [optional] Merge all the pictures into the whole dataset (train+test) folder.
• Split images into train/test randomly (rather than using initial splitting)
• Implement cross-validation (eg K-Fold)

2. Do you have a sufficient number of samples ? Try to add more samples and check how it affects performance. You can also apply data augmentation technics.

Answer 2

If the accuracy of train data is good, but the test data accuracy is low, then the model tends to overfit. The reason could be a simple dataset where the model tries to catch all data points including noise. In the above case, try to tune the parameters and set a higher batch, perform cross-validation to understand the performance and perform data augmentation.