Using Lambda layers to take different slices of a prevous layer's output causes earlier Lambda layers to be overwritten about tf-keras

[qlzh727]

@hosford42 who is the original reporter of the issue here.

[qlzh727]

I think the root cause here is the mix usage of lambda and for loop here, which is particularly errorprone.

The for loop cause the lambda function that captured by the lambda layer to be always the last loop step. You can actually verify this by visiting the labmda_layer.function.

If I change your model building function as below (without the for loop), it will actually run properly.

def slice_func_0_2(x):
  return x[..., 0:2]

def slice_func_2_5(x):
  return x[..., 2:5]


i = tf.keras.layers.Input((channels,), name='i')
o1 = tf.keras.layers.Lambda(slice_func_0_2, name="slice_0_2")(i)
o2 = tf.keras.layers.Lambda(slice_func_2_5, name="slice_2_5")(i)

o = [o1, o2]
m = tf.keras.Model(i, o, name='m')

[qlzh727]

I am closing this issue since this is a user code error. Feel free to reopen this if there is anything else we need to address.

In general, please take a look for https://gist.github.com/gisbi-kim/2e5648225cc118fc72ac933ef63c2d64 for pitfall for using lambda in a loop.

[hosford42]

I don't understand. Are we not allowed to use loops when building models? This code is meant to build a model from a spec, which is not known in advance, so the loop is necessary. It makes no sense that simply using a for loop to produce the layers would behave any differently than unrolling the loop.

…

On Thu, Oct 12, 2023, 1:42 PM Qianli Scott Zhu ***@***.***> wrote: Closed #671 <#671> as completed. — Reply to this email directly, view it on GitHub <#671 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABC2LM635ZV4FDGS3XNWK23X7A2ZVAVCNFSM6AAAAAA5572FOKVHI2DSMVQWIX3LMV45UABCJFZXG5LFIV3GK3TUJZXXI2LGNFRWC5DJN5XDWMJQGYZTKNJXG44TOMY> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[hosford42]

Ah, never mind. I see now. The lambda needs default parameters for start & end due to scoping. Thanks for pointing out my mistake.

…

On Fri, Oct 13, 2023, 1:28 AM Aaron Hosford ***@***.***> wrote: I don't understand. Are we not allowed to use loops when building models? This code is meant to build a model from a spec, which is not known in advance, so the loop is necessary. It makes no sense that simply using a for loop to produce the layers would behave any differently than unrolling the loop. On Thu, Oct 12, 2023, 1:42 PM Qianli Scott Zhu ***@***.***> wrote: > Closed #671 <#671> as > completed. > > — > Reply to this email directly, view it on GitHub > <#671 (comment)>, > or unsubscribe > <https://github.com/notifications/unsubscribe-auth/ABC2LM635ZV4FDGS3XNWK23X7A2ZVAVCNFSM6AAAAAA5572FOKVHI2DSMVQWIX3LMV45UABCJFZXG5LFIV3GK3TUJZXXI2LGNFRWC5DJN5XDWMJQGYZTKNJXG44TOMY> > . > You are receiving this because you were mentioned.Message ID: > ***@***.***> >