输入图像以测试张量流模型的正确方法

Question

I have been trying to test the fcn implementation posted here . The only thing I changed is the way for setting up the input images for test against the model. My modification is marked with red curve in the following figure.

However, running the program caused the following error message TypeError: The value of a feed cannot be a tf.Tensor object. Acceptable feed values include Python scalars, strings, lists, or numpy ndarrays. TypeError: The value of a feed cannot be a tf.Tensor object. Acceptable feed values include Python scalars, strings, lists, or numpy ndarrays. , happened at down, up = sess.run(tensors, feed_dict=feed_dict) . I am curious to know what are something wrong in my implementation, and how to modify it. In the original post, the author uses img1 = skimage.io.imread("./test_data/tabby_cat.png") to input the image.

If I change batch_images=tf.expand_dims(images,0) to batch_images=tf.expand_dims(img1,0) the program will output the following error messages.

Answer 1

As the error indicates, the types of values that you can use as feeds are Python scalars, strings, lists or numpy arrays. What you are trying to use as feed is img1 , the output of tf.image.decode_png , which is of type tf.Tensor . Hence the error. You have two options :

1) Convert img1 to a numpy array before feeding it. You can do that by simply evaluating img1 as follows :

feed_dict = {images:img1.eval()}

2) Use img1 itself as the input to the rest of the model. You can do that as follows :

batch_images = tf.expand_dims(img1, 0)