Keras + tensorflow批处理图像分类

Question

I have service which should make response on classification of 10-20 images per request.

now I process then one by one:

 with model_store.some_graph.as_default():
        with model_store.some_session.as_default():
            for i in images:
                x = image.img_to_array(i)
                x = preprocess_input(x)
                x = np.expand_dims(x, axis=0)
                pred = model_store.top_model.predict(x)[0]

It takes 3-4 seconds( On CPU instance ) Is it possible to process them at one call? And if yes can it give any performance boost?

Answer 1

You need to give a batch of images, something of shape (batch_size, W, H, C) . In your case collect images to a list and then concatenate into single tensor:

def process_img(x):
  x = image.img_to_array(x)
  x = preprocess_input(x)
  x = np.expand_dims(x, axis=0)
  return x
imgs = [process_img(i) for i in images]
imgs = np.concatenate(*imgs, axis=0)
preds = model_store.top_model.predict(imgs, batch_size=32)

Now the model will predict with batches of 32 images at a time. This might speed things up depending on your hardware. As a result, you will get all the predictions in one go as well.