TF placeholder access from within TD block about fold HOT 3 CLOSED

On Fri, Mar 24, 2017 at 11:10 AM, Alex Kuhnle ***@***.***> wrote: Hello, I am trying to build a network with some static inputs and a dynamic part using these values. Here an example: import numpy as npimport tensorflow as tfimport tensorflow_fold as td y = tf.placeholder(dtype=tf.float32, shape=(None,)) def fixed_plus(x): return tf.add(x=x, y=y) compiler = td.Compiler.create(td.Tensor(()) >> td.Function(tf_fn=fixed_plus)) with tf.Session() as session: feed_dict = compiler.build_feed_dict(np.ones(shape=(10,))) feed_dict[y] = np.arange(10) print(session.run(fetches=compiler.output_tensors, feed_dict=feed_dict)) I get the following exception message in the line that creates the compiler: InvalidArgumentError (see above for traceback): Incompatible shapes: [0] vs. [10] [[Node: while/Function_1/Add = Add[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/cpu:0"](while/Function_1/arg_0/Gather, while/Function_1/Add/Enter)]] Is there a way to do this, or is it necessary to feed all input used within a block in compiler.build_feed_dict? I have seen example code with a dropout value given as *external* placeholder, so I assume there should be a way? Many thanks in advance! — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#39>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AGGbTZNGTHhOc0KW--ZLkaSLgWngBqhiks5rpAcDgaJpZM4MomSg> .

On Sat, Mar 25, 2017 at 11:22 AM, Alex Kuhnle ***@***.***> wrote: Thanks for the quick response! Regarding the scalar input, I think this is fine. Since TD is batched, I assume it takes the (10,) array as a batch of 10 0-dim scalars - and in fact, the exception only occurs in the tf.add operation. Moreover, in line feed_dict[y] = np.arange(10) I provide a value for y. After your comment on size mismatch, I instead fed a single-value vector to y: feed_dict[y] = np.arange(1). Then it behaves as expected and adds this y value (0 in this case) to the batch x. In further investigation I added the line x = tf.Print(x, (x, y)) to the fixed_plus function, and got the unexpected output [1 1 1...][0] [][0] [array([ 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.], dtype=float32)] and when feeding the previous 10 values for y I got before the exception occurs [1 1 1...][0 1 2...] [][0 1 2...] Traceback (most recent call last):... rest of the exception ... I am surprised by that, because it seems that there is a second call with an empty tensor for x, besides the ones-vector I give as argument. In particular, the addition works as expected in the first run (when adding it to the printed values). Is that expected behavior? Please let me know if any of the steps I explained is unclear. Regarding the application I have in mind, to not be confusing: I would ideally like to have a situation where each instance in a batch has some inputs which are *static*, and some which are *dynamic*, i.e. they are somehow processed by TD blocks where the *static* values of the respective instance are potentially referred to. Obviously, one can just feed all inputs to the TD blocks and forward the *static* values (in addition to using them), but that would not be very elegant. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#39 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AGGbTdpVnhfmgF66wG7HLLLVmeEHLFg7ks5rpVtKgaJpZM4MomSg> .