Question

我使用经过预先培训的VGG16网络跟踪Building powerful image classification models using very little data的教程。

我能够获得相当准确的数据集:)但是，我如何重用经过重新训练的网络？

model = Sequential()
model.add(Flatten(input_shape=train_data.shape[1:]))
model.add(Dense(256, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(nb_class, activation='softmax'))

model.compile(optimizer='rmsprop',
      loss='categorical_crossentropy', metrics=['accuracy'])
model.fit(train_data, train_labels,
      epochs=epochs,
      batch_size=batch_size,
      validation_data=(validation_data, validation_labels))

# After training, I save my model
model_json = model.to_json()
with open(top_model_json_path, "w") as json_file:
    json_file.write(model_json)
model.save_weights(top_model_weights_path)

我希望将dog.1234.jpg重新用于def main(_): # vgg16 vgg_model = applications.VGG16(include_top=False, weights='imagenet') # load top model and weights json_file = open('retrain_VGG16_dogcat2000.json', 'r') loaded_model_json = json_file.read() json_file.close() top_model = model_from_json(loaded_model_json) top_model.load_weights("retrain_VGG16_dogcat2000.h5") # concatenate vgg16 and top classifier model = Model(inputs=vgg_model.input, outputs=top_model(vgg_model.output)) # error occurs # test image = cv2.imread(FLAGS.img_dir) image = cv2.resize(image, (224, 224)).astype(np.float32) / 255. image = image.reshape(1 , image.shape[0], image.shape[1], image.shape[2]) model.predict(image)

The shape of the input to "Flatten" is not fully defined (got (None, None, 512). Make sure to pass a complete "input_shape" or "batch_input_shape" argument to the first layer in your model.

发生错误：{{1}}

有什么想法吗？

Answer 1

您的VGG模型没有明确定义的输入形状。它接受任何大小的图像。由于它是纯粹的卷积模型，因此输出形状也是可变的。这就是形状显示为Kernel.php的原因。

但是(None, None, 512)图层（存在于您自己的Flatten中）不支持变量输入形状。它必须使用定义明确的一个。

您有两种可能的解决方案：

对于固定图像尺寸：使用top_model定义VGG模型的输入形状（这不需要您重新训练此模型，权重不依赖于纯卷积模型中图像的大小）
对于可变图像尺寸：在顶部模型的开头添加全局池图层以消除空间变量尺寸。使用input_shape=(224,224,3)。（这可能会降低你的模型的能力，而且肯定需要重新训练一切）

Answer 2

我目前能够以这种方式重复使用我的再训练模型。但是，如何将重新训练的模型连接到一个机器学习模型中呢？

classes = ['cat', 'dog']
# load VGG model
vgg_model = applications.VGG16(include_top=False, weights='imagenet')

# load model and weights
json_file = open('retrain_VGG16_dogcat2000.json', 'r')
loaded_model_json = json_file.read()
json_file.close()
top_model = model_from_json(loaded_model_json)
top_model.load_weights("retrain_VGG16_dogcat2000.h5")

# test image
image = cv2.imread(FLAGS.img_dir)
image = cv2.resize(image, (224, 224)).astype(np.float32) / 255.
image = image.reshape(1 , image.shape[0], image.shape[1], image.shape[2])

# predict the image throughout vgg net
vgg_out = vgg_model.predict(image)

# predict the top_model with input from vgg's output
prediction = top_model.predict(vgg_out)

print(prediction, classes[np.argmax(prediction)])

python evaluate.py --img_dir ~/datasets/dog_cat/dog.4457.jpg

[[5.281641e-32 1.000000e + 00]]狗

python evaluate.py --img_dir ~/datasets/dog_cat/cat.2345.jpg

[[1.000000e + 00 9.587834e-36]] cat

如何重用再培训模型

2 个答案: