张量流中的不同图像大小与batch_size = 1

Question

I want to achieve a python class, which can load a tensorflow model and implement a inference. However I have no idea about how can I input image with variable image size. :(

class ArtGenerater():
    def __init__(self,model_path):
            self.model_path = model_path
            # vary shape?
            self.x = tf.placeholder(tf.float32,shape=(1,512,512,3))
            self.gen = model.resnet(self.x)
            self.out = tf.saturate_cast(self.gen,tf.uint8)

            self.sess = tf.Session()
            file = tf.train.lastest_checkpoint(self.model_path)
            saver = tf.train.Saver()
            saver.restore(self.sess,file)
    def pic(self,image_path):
            img =np.asarray(Image.open(image_path)).astype(np.float32)
            img = np.expand_dims(img,0)
            output_t = self.sess.run(self.out,feed_dict={self.x:img})
            return output_t

Now I just use tf.placeholder(tf.float32,shape=(1,512,512,3)) , but my image have different sizes(eg. 1000*900).

How can i achieve this function? Thank you.

EDIT:

Thank you everyone.I have solved problem by using:

x = tf.placeholder(tf.string)
img = tf.image.decode_jpeg(x,channels=3)

And this can feed network (my ConvNet include many conv2d & conv2d_tranpose ) with different image size. :)

Answer 1

Basically you can define a various size input using None as follows

self.x = tf.placeholder(tf.float32, [1, None, None, 3]) 

and then you can feed different input

feed_dict={self.x: current_data} etc..

But be careful about your neural net structure. If you flatten your last conv layer as input to the first dense layer then your network only works at that size, and you need to either stretch or crop the image to make it work.

A more flexible approach is to use something like Global Average Pooling or Spatial Pyramid Pooling which both fix this problem.