I'm new to tensorflow, according to my experience, you really don't need to feed a tf.tensor to feed_dict. Once you have a tensor with value, you can get this value with your_tensor.eval() or sess.run(your_tensor) and then you can feed the output to your feed_dict

I have also searched about how to get the value of a tensor to solve this problem and seen the solutions like above. However when I try to get image_batch.eval() or even image_buffer.eval(), the program just keep working without stopping or giving any output. I have tried to get a constant value from a tensor and succeeded. But this is just useless with tensor gotten from tf.read_file() and etc.

If you have your data in tensors anyway, you don't need to use placeholders and feed_dict. You can just directly use image_batch / label_batch instead of images_placeholder / label_placeholder when setting up your model / loss.

Yes, I can see from the examples of training procedures that I can just use tensors as input of any ops in tensorflow. However in this case I am using a predefined network so I am confused whether I can change the network definition. That is to say, I have no idea how to change the input of loss function(the definition of node "images" and "prob" is images = tf.placeholder("float32", [None, 224, 224, 3], name="images") logits = resnet.inference(images, is_training=False, num_blocks=num_blocks, preprocess=True, bottleneck=True) prob = tf.nn.softmax(logits, name='prob') in the predefined model).

StackOverflow is a better venue. One key thing to note is that Tensor is simply a symbolic object. The values of your feed_dict are the actual values, e.g. a Numpy ndarry.

I have the same questions with @wishforgood .
"I have no idea how to change the input of loss function"...

How?

Why is this closed?

Author raises legitimate concerns.

1. Why data already on GPU, cannot be fed into GPU model.

2. Why redefine model to use data already on GPU. This makes code more complicated.

3. What if I have 5 variables with data, do I need to define now 5 models with shared trainable variables?

Yes it is confusing, make you write complicated (or slow) code -- I think you guys should reopen and address this.

I have the same opinion as yours. These days, i use the cnn to do some text work. For the input's dimension is too large, i have to change it to other type in case of the memory error. But , what confused me is the same as yous. Why tensor cannot be a feed_dict value!!!

Looks like you could use a session_handle to do this:

Simple case:

p = tf.placeholder(tf.float32, shape=()) 
c = tf.constant(1.0, dtype=tf.float32)
h = tf.get_session_handle(c)
h = sess.run(h)        # gives you a handle
sess.run(p, feed_dict={p:h})

The below is an example from the get_session_handle doc:

c = tf.multiply(v1, v2)
h = tf.get_session_handle(c)
h = sess.run(h)        # gives you a handle

p, a = tf.get_session_tensor(h.handle, tf.float32)
b = tf.multiply(a, 10)
c = sess.run(b, feed_dict={p: h.handle})

This works for a scalar value but when you try to get handle for a sequence of inputs, it is failing.
Basically, i am also getting the same error "TypeError: The value of a feed cannot be a tf.Tensor object. Acceptable feed values include Python scalars, strings, lists, numpy ndarrays, or TensorHandles."

operation I performed is feeding the sequence of number from -10 to 10 for a function.

part of code:
f = tf.square(x) + 2 * x + 5

val_x = tf.range(-10, 10, 0.1)

val_f = sess.run(f, feed_dict = {x: val_x})

But if i replace val_x = tf.range(-10, 10, 0.1) --> val_x = np.arange(-10, 10, 0.1), it works fine without any errors.

why is sequence generated using tf.range() is not treated as tensor object?

Thanks in Advance!

@Anilkumares For the feed_dict can't accept the tensor object. As for the f = tf.square(x) + 2 * x + 5, it can be fed with scalars generated by numpy

You can do something like this in a loop to make batches without using any TensorFlow backend. This will ensure that the output of it remains non-tensor.

with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
            for index, offset in enumerate(range(0, train, batch_size)):
                x_epoch, y_epoch = np.array(train_x[offset: offset + batch_size,:]), np.array(train_y[offset: offset +  batch_size])

The short answer to the question is because of Tensorflow makeы simple things complicated

I may have a different question but similar with yours: if my input data is already in tensor type, how can i feed them into the network? this is very simple and straightforward in pytorch, but as we know, tensorflow uses static computational graph. we have to pre-define the graph and then change the input in a cycle. if the inputs of network and loss function are defined when creating the graph, how can we change it during training?

One workaround is to construct a graph def and load it in with input_mapping, through this the tf.data.Iterator can be hooked up to any arbitrary graph which previously used placeholders for input. This is especially beneficial for inference, since weights can be frozen.

For anyone who lands on this question, Lakrish provided a very usable solution (only a few lines of code). I have posted an example using this suggestion at this related SO question: https://stackoverflow.com/questions/38618960/tensorflow-how-to-insert-custom-input-to-existing-graph/57015133#57015133