How to feed multiple inputs through feed_dict in tensorflow

Question

I have a network consist of multiple sub-network (multiple convolution net and at last one fully connected + soft max layer). Every ConvNet feed with specific region and size of images. so, to feed my network I write image place holder for every convnet input and one label place holder to feed label of all images in one batch (all of the input images in all convnet inputs have the same labels). Unfortunately I don't have any idea for feed_dict part. for example this code is for only one convnet training:

    images_r, labels_r = sess.run([images, labels])
    train_feed = {images_placeholder: images_r,
              labels_placeholder: labels_r}
    _, loss_value = sess.run([train_op, loss_func], feed_dict=train_feed)

How can I extend above code for feed all conv nets?

I'm not sure I get your question right. You have multiple input points for the network? And if so, at each training step you need to provide an input for all such input points? — GPhilo, Jul 04 '17 at 08:05

Vijay Mariappan · Accepted Answer · 2017-07-05T12:17:00.120

5

So for each of the conv networks, if the placeholders for inputs are: conv_1_input, conv_2_input.... conv_N_input, then you pass the list in the feed_dict like this:

train_feed = {`conv_1_input`: image_1, `conv_2_input`: image_2,.. `conv_N_input`: image_N,
          labels_placeholder: labels_r}
_, loss_value = sess.run([train_op, loss_func], feed_dict=train_feed)

edited Jul 05 '17 at 12:17

answered Jul 04 '17 at 08:32

Vijay Mariappan

16,921
3
40
59

after using this method I've got this error: TypeError: Cannot interpret feed_dict key as Tensor: Can not convert a function into a Tensor. – hajbabaeim Jul 04 '17 at 11:42
The variables `conv_1_input`, `conv_2_input` should be tensors, The error says its not so. – Vijay Mariappan Jul 04 '17 at 11:48

score -2 · Answer 2 · answered Jul 04 '17 at 07:17

-2

You should split/slice your images inside computation graph and use a single input instead.

answered Jul 04 '17 at 07:17

nttstar

341
2
10

Unfortunately all of my patches (from original images) have not the identical feature and I extract them from original images based on some landmarks. (for example consider we have an original image of a human face and my landmarks are 5 points of center of eyes, nose tip, and corner of lip). So, I have to use my extracted patch to feed all related feed_dict. – hajbabaeim Jul 04 '17 at 07:35

How to feed multiple inputs through feed_dict in tensorflow

2 Answers2