module 'sonnet' has no attribute 'AbstractModule'

[poweryin]

Hi,I've installed sonnet by using pip install dm-sonnet
However when I run "python evaluate_sample.py"
I got AttributeError: module 'sonnet' has no attribute 'AbstractModule'
Thanks!

[malcolmreynolds]

@poweryin where is this evaluate_sample.py from? Is it from the Kinetics repository?

Also, could you paste the output of running the following in your interpreter:

import sonnet as snt
print(dir(snt))
print(snt.__version__)

[poweryin]

thanks,it is from kenetics-I3D repository.I figure it out .it is due to tensorflow's version is too low.thanks

[szq261299]

when i try to install the graph-nets , i meet the same problem.......

[tomhennigan]

@szq261299 would you mind printing the output of this in your Python interpreter:

import sonnet as snt
print(dir(snt))
print(snt.__version__)

[szq261299]

@tomhennigan
hello, my python version is 3.6.4, I followed your instructions and the results are as follows：

import sonnet as snt
print(dir(snt))
['__doc__', '__loader__', '__name__', '__package__', '__path__', '__spec__']

print(snt.__version__)
AttributeError                            Traceback (most recent call last)
<ipython-input-10-980fa547e124> in <module>
      1 import sonnet as snt
      2 print(dir(snt))
----> 3 print(snt.__version__)

AttributeError: module 'sonnet' has no attribute '__version__'

In order to find the version number of sonnet, I type the following on the command line.

pip list 
...
dm-sonnet                          1.33
...

so what should I do ?

[tomhennigan]

I can't reproduce this in a clean colab. Perhaps you have a file in your current directory called sonnet.py and you're importing that instead of the library? Can you try printing the output of print(snt.__file__)?

[szq261299]

print(snt.__file__)

AttributeError                            Traceback (most recent call last)
<ipython-input-5-86bcd137c072> in <module>
----> 1 print(snt.__file__)

AttributeError: module 'sonnet' has no attribute '__file__'

[malcolmreynolds]

@szq261299 what about print(snt.__path__)?

[szq261299]

When I uninstalled and reinstalled , it works now . But I can't reproduce this problem. It is so strange....

[szq261299]

Thanks a lot~

[ghost]

When I uninstalled and reinstalled , it works now . But I can't reproduce this problem. It is so strange....

Uninstalling + reinstalling seems to be a solution that worked for me = =.

[ArnaudC]

Uninstalling and reinstalling is not working for me. (tf-gpu)

pip3 uninstall "dm-sonnet>=2.0.0b0"
pip3 uninstall "tensorflow-probability>=0.8.0rc0"

Then

pip3 install --upgrade tensorflow-gpu
pip3 install "tensorflow-probability>=0.8.0rc0" --pre
pip3 install "dm-sonnet>=2.0.0b0" --pre

My tf and Sonnet version :

print("TensorFlow version {}".format(tf.__version__))
print("Sonnet version {}".format(snt.__version__))

TensorFlow version 2.0.0
Sonnet version 2.0.0b0
Python 3.6.8
library cudart64_100.dll

<module 'sonnet' from 'F:\Python\Python 3.6.8\lib\site-packages\sonnet\init.py'>
module 'sonnet' has no attribute 'AbstractModule'
File "C:\Users\a\ImageGeneration\VAEv2.py", line 100, in
class Encoder(snt.AbstractModule):

[tomhennigan]

Hey @ArnaudC, if the code you're using wants snt.AbstractModule then it's likely that it's been designed for Sonnet 1 + TensorFlow 1 (in Sonnet 2 we just have snt.Module, docs here).

I'll see if we can give a better error message in this case, but for now you'll want to use the 1.X series of both TensorFlow and Sonnet.

Also (to save you some time) the current stable version of tensorflow-probability does not support the latest TF1 release so we should use the older version for now:

$ pip install "tensorflow-gpu<2" "dm-sonnet<2" "tensorflow-probability==0.7.0"

If you're developing new code then I would suggest using TensorFlow 2 and Sonnet 2 but if you have existing code that you don't want to rewrite then downgrading to TF1/Sonnet 1 should do the trick.

[lucky096]

Hey @tomhennigan, I need to upgrade to sonnet2/tf2, can you confirm is there any way to upgrade the code written in sonnet1 to sonnet2? For instance, the snt.AbstractModule from sonnet1 is now just snt.Module. I checked and they are not same. It seems backward compatibility is not there. So did you mean that to maintain models designed in sonnet1, we should stick to sonnet1/tf1 and we can't upgrade them to sonnet2/tf2?

[tomhennigan]

Hi @lucky096 there is no automated way to convert Sonnet 1 -> Sonnet 2 and the library is not backwards compatible.

The main change between the Sonnet 1 and 2 is how variables and submodules are created and used. Sonnet 2 doesn't have any magical reuse of modules or variables, instead users are expected to use normal Python code to manage this (e.g. save parameters as self.w = tf.Variable(...)).

We provide @snt.once to make a method run once and only once, this can make data dependent initialization (very common in most neural nets) more convenient.

# TensorFlow 2 / Sonnet 2
class AddNModule(snt.Module):

  @snt.once
  def _initialize(self, x):
    # Make sure to initialize variables and submodules once and only once.
    self.n = tf.Variable(tf.ones(x.shape))
    self.s = SomeSubModule()

  def __call__(self, x):
    self._initialize(x)
    return self.s(self.n + x)

# TensorFlow 1 / Sonnet 1
class AddNModule(snt.AbstractModule):
  def __call__(self, x):
    n = tf.get_variable("n", x.shape, initializer=tf.ones)
    return SomeSubModule()(n + x)

I have ported a lot of code between Sonnet 1 -> Sonnet 2 and for most modules moving initialization into a method annotated with @snt.once is all you need to change (aside from changing to snt.Module name and dealing with things that have been renamed in TensorFlow).

[tfqbasha]

Hey,
I was using deepmind's kinetics-i3d repository and got this AttributeError. Below are the versions of tf and snt I am using.
TensorFlow version 2.1.0
Sonnet version 2.0.0b0

Based on the discussion above, I see that AbstractModule is not used in sonnet2.0. Do I have to go back to sonnet1.0 or is there a kinetics-i3d based on sonnet2.0?

[malcolmreynolds]

@tfqbasha as mentioned above, models written in Sonnet v1 will not work unmodified with V2. I would suggest asking the authors of Kinetics directly whether they are planning on making a Sonnet v2 version available.

[AlexanderMelde]

@tfqbasha you can follow this issue google-deepmind/kinetics-i3d#85 to see if the developers will respond to the update request.

[BigfishWa]

@szq261299你介意在你的 Python 解释器中打印这个输出吗：

import sonnet as snt
print(dir(snt))
print(snt.__version__)

2022-05-20 11:39:34.302485: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cudart64_101.dll
['BaseBatchNorm', 'BatchApply', 'BatchNorm', 'Bias', 'Conv1D', 'Conv1DLSTM', 'Conv1DTranspose', 'Conv2D', 'Conv2DLSTM', 'Conv2DTranspose', 'Conv3D', 'Conv3DLSTM', 'Conv3DTranspose', 'DeepRNN', 'Deferred', 'DepthwiseConv2D', 'Dropout', 'Embed', 'ExponentialMovingAverage', 'Flatten', 'GRU', 'GroupNorm', 'InstanceNorm', 'LSTM', 'LSTMState', 'LayerNorm', 'Linear', 'Mean', 'Metric', 'Module', 'Optimizer', 'RNNCore', 'Reshape', 'Sequential', 'Sum', 'TrainableState', 'UnrolledLSTM', 'UnrolledRNN', 'VanillaRNN', 'all', 'builtins', 'cached', 'doc', 'file', 'loader', 'name', 'package', 'path', 'spec', 'version', 'absolute_import', 'allow_empty_variables', 'build', 'custom_variable_getter', 'deep_rnn_with_residual_connections', 'deep_rnn_with_skip_connections', 'distribute', 'division', 'dynamic_unroll', 'flatten', 'format_variables', 'initializers', 'leaky_clip_by_value', 'log_variables', 'lstm_with_recurrent_dropout', 'merge_leading_dims', 'mixed_precision', 'nets', 'no_name_scope', 'once', 'optimizers', 'pad', 'print_function', 'regularizers', 'reshape', 'scale_gradient', 'split_leading_dim', 'static_unroll']
2.0.0