GradientTape "AttributeError: 'KerasTensor' object has no attribute '_id'"

[sirimet]

Brief intro
I apologize if this is not the correct forum for this question, but any help would be greatly appreciated! I already asked here, but it's a very tensorflow specific question.

I'm trying to build a Wasserstein GAN with gradient penalty, following this paper. Essentially, I'm trying to recreate their python code in R. All was going well until I tried to implement their gradient penalty. In the original code, they've defined it as a class, and they use K.gradients to calculate the gradients:

class GradientPenalty(Layer):
    def __init__(self, **kwargs):
        super(GradientPenalty, self).__init__(**kwargs)

    def call(self, inputs):
        target, wrt = inputs
        grad = K.gradients(target, wrt)[0]
        return K.sqrt(K.sum(K.batch_flatten(K.square(grad)), axis=1, keepdims=True))-1

    def compute_output_shape(self, input_shapes):
        return (input_shapes[1][0], 1)

I generally try to avoid classes in R, and write things as functions instead. I think the rest will be better explained with what I'm hoping can work as a minimal reproducible example.

Session info

R version 4.0.2 (2020-06-22)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: Ubuntu 18.04.6 LTS

Matrix products: default
BLAS:   /usr/lib/x86_64-linux-gnu/atlas/libblas.so.3.10.3
LAPACK: /usr/lib/x86_64-linux-gnu/atlas/liblapack.so.3.10.3

locale:
  [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C               LC_TIME=nb_NO.UTF-8        LC_COLLATE=en_US.UTF-8
[5] LC_MONETARY=nb_NO.UTF-8    LC_MESSAGES=en_US.UTF-8    LC_PAPER=nb_NO.UTF-8       LC_NAME=C
[9] LC_ADDRESS=C               LC_TELEPHONE=C             LC_MEASUREMENT=nb_NO.UTF-8 LC_IDENTIFICATION=C

attached base packages:
  [1] stats     graphics  grDevices utils     datasets  methods   base

other attached packages:
  [1] keras_2.9.0      tensorflow_2.9.0

loaded via a namespace (and not attached):
  [1] Rcpp_1.0.8       here_0.1         lattice_0.20-41  png_0.1-7        rprojroot_1.3-2  zeallot_0.1.0    rappdirs_0.3.3
[8] grid_4.0.2       R6_2.5.1         backports_1.1.10 jsonlite_1.8.0   magrittr_2.0.2   rlang_0.4.11     tfruns_1.4
[15] whisker_0.4      Matrix_1.2-18    reticulate_1.25  generics_0.1.2   tools_4.0.2      compiler_4.0.2   base64enc_0.1-3

Reproducible example

library(tensorflow)
library(keras)

tf$executing_eagerly() # should be true, because eager execution is needed in a different part of the code

# these are not initialized here in the real thing, but the resulting tensors are the same:
disc_out_avg <- layer_input(shape = list(1)) 
disc_in_avg  <- layer_input(shape = list(NULL, NULL, 1)) 

# my first attempt at translating gradient penalty:
gradient_penalty <- function(inputs){
  c(target, wrt) %<-% inputs
  grad <- k_gradients(loss = target, variables = wrt)[1]
  return(k_sqrt(k_sum(k_batch_flatten(k_square(grad)), axis = 1, keepdims = TRUE)) -1)
}

disc_gp       <- gradient_penalty(list(disc_out_avg, disc_in_avg))

This produces the error

> Error in py_call_impl(callable, dots$args, dots$keywords) : 
> RuntimeError: tf.gradients is not supported when eager execution is enabled. Use tf.GradientTape instead.

I won't try to hide the fact that I don't quite understand what GradientTape is, which might explain why I can't get the following to work. However, I've been googling it for weeks and I'm not getting any wiser!

library(tensorflow)
library(keras)

tf$executing_eagerly() # should be true, because eager execution is needed in a different part of the code

# these are not initialized here in the real thing, but the resulting tensors are the same:
disc_out_avg <- layer_input(shape = list(1)) 
disc_in_avg  <- layer_input(shape = list(NULL, NULL, 1)) 

# my second attempt at translating gradient penalty:
gradient_penalty <- function(inputs){
  c(target, wrt) %<-% inputs
  with(tf$GradientTape() %as% tape, {
    # tape$watch(wrt)
  })
  grad <- tape$gradient(target, wrt)
  return(k_sqrt(k_sum(k_batch_flatten(k_square(grad)), axis = 1, keepdims = TRUE)) -1)
}

disc_gp       <- gradient_penalty(list(disc_out_avg, disc_in_avg))

Error in py_call_impl(callable, dots$args, dots$keywords) :
AttributeError: 'KerasTensor' object has no attribute '_id'
Called from: py_call_impl(callable, dots$args, dots$keywords)

If I try "watching" one variable or the other, I get:

Error in py_call_impl(callable, dots$args, dots$keywords) :
ValueError: Passed in object of type <class 'keras.engine.keras_tensor.KerasTensor'>, not tf.Tensor
Called from: py_call_impl(callable, dots$args, dots$keywords)

Any clues as to what I might be doing wrong?

[t-kalinowski]

Hi, thanks for posting! You're on the right track. :) k_gradients()/K.gradients() is indeed, effectively if not officially, deprecated. It only works in non-eager mode.

The reason this works in the Python project associated w/ that paper is because they disable eager mode: https://github.com/jleinonen/downscaling-rnn-gan/blob/e136192e4786e7d0a1832f42854e5ef642698570/dsrnngan/train.py#L7

You could do the same in R, if you want. :)

But, there is a way to calculate this gradient penalty with GradientTape() while in eager mode: there is a WGAN.gradient_penalty() example here.

A PR with an adaptation of that example to R would be very welcome :) It could then be hosted here: https://tensorflow.rstudio.com/examples/ (via https://github.com/rstudio/tensorflow.rstudio.com/).

[sirimet]

Thanks, this looks like a huge nudge in the right direction! I'll give it a go :)

[t-kalinowski]

There is also a GAN implementation, and a callback_gan_monitor(), in the 2nd edition of the Deep Learning with R book, which may be helpful or save you some time in adapting the Python example.

[nickschurch]

@sirimet did you solve this? I'm getting the same problem trying to compute the gradients...

[lainconn]

@nickschurch me too

[t-kalinowski]

@lainconn Can you please open a new issue with instructions to reproduce the error?

[lainconn]

@t-kalinowski Sorry for bothering, I fixed my issue

[ZFH-AI]

Can you tell me how you solved it? I had the same problem, thanks

[ZFH-AI]

@t-kalinowski Sorry for bothering, I fixed my issue

Can you tell me how you solved it? I had the same problem, thanks

[t-kalinowski]

@ZFH-AI, can you please open a new issue with a minimal example code snippet that reproduces the error?