Question

我正在尝试创建一个CNN，将图像分为3类。 -我使用以下版本：

Keras 2.2.4
Python 3.6.8
tensorflow 1.13.1

训练时，我的val_acc高于90％。但是当我开始预测时在其他图像上（我什至只使用过验证图像->结果仍然是随机的），这些图像的结果acc是随机的。

加载模型后，我只是打印了图层，输入，输出，model.summary（），

我希望我的预测结果会比这个更好...

这是我的代码：

import cv2
import pandas
from keras.models import Sequential
from keras.models import load_model
from keras.layers import Flatten
from keras.layers import Dense
from keras.layers import Dropout
from keras.layers import Activation
from keras.layers.convolutional import Conv2D
from keras.layers.convolutional import MaxPooling2D
from keras import optimizers
from keras.preprocessing.image import ImageDataGenerator
from keras import callbacks
from sklearn.model_selection import train_test_split
from time import gmtime, strftime

CURRENT_TIME = strftime("%Y-%m-%d_%H-%M-%S", gmtime())
NAME = "CNN-" + CURRENT_TIME
SIZE = 280

NUM_CLASSES = 3
epochs = 25
batch_size = 64

df = pandas.read_csv("./csv/sample_tile.csv", ";")
df.columns = ['filenames', 'category']
df['category'] = df['category'].astype(str)
df = df.sample(frac=1).reset_index(drop=True)

train_df, validation_df = train_test_split(df, test_size=0.30, random_state=42)

train_df = train_df.reset_index(drop=True)
validation_df = validate_df.reset_index(drop=True)


total_train = train_df.shape[0]
total_validate = validate_df.shape[0]


train_datagen=ImageDataGenerator(rescale=1./255.)
train_generator = train_datagen.flow_from_dataframe(
    train_df, 
    "data/", 
    x_col='filenames',
    y_col='category',
    target_size=(SIZE, SIZE),
    class_mode='categorical',
    batch_size=batch_size,
)

# -> Found 6358 images belonging to 3 classes.

validation_datagen = ImageDataGenerator(rescale=1./255)
validation_generator = validation_datagen.flow_from_dataframe(
    validation_df, 
    "data/", 
    x_col='filenames',
    y_col='category',
    target_size=(SIZE, SIZE),
    class_mode='categorical',
    batch_size=batch_size//2,
)

# -> Found 2697 images belonging to 3 classes.

earlystop = callbacks.EarlyStopping(patience=3, monitor='val_acc')

learning_rate_reduction = callbacks.ReduceLROnPlateau(monitor='val_acc', 
                                            patience=2, 
                                            verbose=1, 
                                            factor=0.5, 
                                            min_lr=0.00001)

tensorboard = callbacks.TensorBoard(log_dir='logs/{}'.format(NAME))

model_checkpoint = callbacks.ModelCheckpoint(
    filepath='model/checkpoint/' + NAME +'.h5',
    monitor='val_acc',
    save_best_only=True
)

callbacks = [earlystop, learning_rate_reduction, model_checkpoint, tensorboard]

def mode_architecture():
    model = Sequential()

    input_shape = (SIZE, SIZE, 3)
    model.add(Conv2D(32, (3, 3), 
                     input_shape=input_shape, 
                     padding='same',
                     name='layer_conv1'))
    model.add(Activation('relu'))
    model.add(MaxPooling2D(pool_size=(2, 2),
                           name='maxPool1'))

    model.add(Conv2D(64, (3, 3), 
                     padding='same', 
                     name='layer_conv2'))
    model.add(Activation('relu'))
    model.add(MaxPooling2D(pool_size=(2, 2),
                           name='maxPool2'))

    model.add(Conv2D(128, (3, 3), 
                     padding='same', 
                     name='layer_conv3'))
    model.add(Activation('relu'))
    model.add(MaxPooling2D(pool_size=(2, 2),
                           name='maxPool3'))
    model.add(Flatten())
    model.add(Dense(128,
                    activation='relu',
                    name='fc0'))
    model.add(Dropout(0.2))
    model.add(Dense(64,
                    activation='relu',
                    name='fc1'))
    model.add(Dropout(0.2))
    model.add(Dense(32,
                    activation='relu',
                    name='fc2'))

    model.add(Dropout(0.5))
    model.add(Dense(NUM_CLASSES))
    model.add(Activation('softmax'))

    sgd = optimizers.SGD(lr=0.001, momentum=0.5, decay=0.0, nesterov=False)

    model.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])
    return model

model = mode_architecture()
model.summary()

Layer (type)                 Output Shape              Param #   
_________________________________________________________________
layer_conv1 (Conv2D)         (None, 280, 280, 32)      896       
_________________________________________________________________
activation_1 (Activation)    (None, 280, 280, 32)      0         
_________________________________________________________________
maxPool1 (MaxPooling2D)      (None, 140, 140, 32)      0         
_________________________________________________________________
layer_conv2 (Conv2D)         (None, 140, 140, 64)      18496     
_________________________________________________________________
activation_2 (Activation)    (None, 140, 140, 64)      0         
_________________________________________________________________
maxPool2 (MaxPooling2D)      (None, 70, 70, 64)        0         
_________________________________________________________________
layer_conv3 (Conv2D)         (None, 70, 70, 128)       73856     
_________________________________________________________________
activation_3 (Activation)    (None, 70, 70, 128)       0         
_________________________________________________________________
maxPool3 (MaxPooling2D)      (None, 35, 35, 128)       0         
_________________________________________________________________
flatten_1 (Flatten)          (None, 156800)            0         
_________________________________________________________________
fc0 (Dense)                  (None, 128)               20070528  
_________________________________________________________________
dropout_1 (Dropout)          (None, 128)               0         
_________________________________________________________________
fc1 (Dense)                  (None, 64)                8256      
_________________________________________________________________
dropout_2 (Dropout)          (None, 64)                0         
_________________________________________________________________
fc2 (Dense)                  (None, 32)                2080      
_________________________________________________________________
dropout_3 (Dropout)          (None, 32)                0         
_________________________________________________________________
dense_1 (Dense)              (None, 3)                 99        
_________________________________________________________________
activation_4 (Activation)    (None, 3)                 0         
_________________________________________________________________
Total params: 20,174,211
Trainable params: 20,174,211
Non-trainable params: 0
_________________________________________________________________

history = model.fit_generator(
    train_generator, 
    epochs=epochs,
    validation_data=validation_generator,
    validation_steps=total_validate//batch_size,
    steps_per_epoch=total_train//batch_size,
    verbose=1,
    callbacks=callbacks
)

Epoch 25/25->损失：0.1062-acc：0.9610-val_loss：0.0721-val_acc：0.9780

model.save("model/" + str(batch_size) + "_" + str(epochs) + "_" +  CURRENT_TIME + ".h5")

del model

model = load_model("weights/" + str(batch_size) + "_" + str(epochs) + "_" +  CURRENT_TIME + ".h5")

files_df = pandas.DataFrame(os.listdir('path/to/images/'))
files_df.columns = ['filenames']

#shuffle
files_df = files_df.sample(frac=1).reset_index(drop=True)
files_df = files_df.reset_index(drop=True)

files_datagen=ImageDataGenerator(rescale=1./255.)
files_generator = files_datagen.flow_from_dataframe(
            files_df, 
            path_to_image_slices, 
            x_col='filenames',
            target_size=(SIZE, SIZE),
            class_mode=None,
            batch_size=64,
            )
predictions = model.predict_generator(files_generator,
                                      steps=value//batch_size,
                                      verbose=1)

编辑：

train_df_plot：

train_df['category'].value_counts().plot.bar()

validation_df图：

validation_df['category'].value_counts().plot.bar()

Dataframe

df = pandas.read_csv("./csv/sample_tile.csv", ";")

这就是我如何绘制结果的方式

# select the index with the maximum probability
test_list_df = pandas.DataFrame(prediction_test)
test_list_df.columns = ['class1', 'class2', 'class3']
test_list = test_list_df.values.tolist()
max_value_idx = test_list_df.idxmax(axis=1)

test_filenames = test_generator.filenames
results=pandas.DataFrame({"Filenames":test_filenames,
                      "Predictions":max_value_idx})
length = len(test_list)/2
length = int(length)

# settings
nrows, ncols = 4, 2  # array of sub-plots
figsize = [25, 25]     # figure size, inches

# create figure (fig), and array of axes (ax)
fig, ax = plt.subplots(nrows=nrows, ncols=ncols, figsize=figsize)

# plot simple raster image on each sub-plot
for i, axi in enumerate(ax.flat):
    # i runs from 0 to (nrows*ncols-1)
    # axi is equivalent with ax[rowid][colid]
    img = cv2.imread('data/' + results.iloc[i]['Filenames'])
    #img_s = cv2.resize(img, (SIZE,SIZE))
    axi.imshow(img)
    # get indices of row/column
    rowid = i // ncols
    colid = i % ncols
    # write row/col indices as axes' title for identification
    axi.set_title("Row:"+str(rowid)+", Col:"+str(colid) + results.iloc[i]['Predictions'], size=20)

plt.tight_layout(True)
plt.show()

训练期间Keras acc和val_acc很高，但预测确实很差

Epoch 25/25->损失：0.1062-acc：0.9610-val_loss：0.0721-val_acc：0.9780

这就是我如何绘制结果的方式

0 个答案: