module 'tensorflow.python.keras.optimizers' has no attribute 'SGD'

Question

I am working with python 3.9, tensorflow 2.7.0 with a modified version of Mask RCNN https://github.com/leekunhee/Mask_RCNN/blob/tensorflow2.0/mrcnn). I am working with tensorflow .keras and specifically import

import tensorflow as tf
from tensorflow.python import keras 
from tensorflow.python.keras import backend as K
from tensorflow.python.keras import layers as KL
from tensorflow.python.keras import utils as KU
from tensorflow.python.eager import context
from tensorflow.python.keras import initializers as KI#nuevo
from tensorflow.python.keras import engine as KE
from tensorflow.python.keras import models as KM
from tensorflow.keras.optimizers import SGD

I call this module in this code

`optimizer = SGD(lr=learning_rate, momentum=momentum, clipnorm=self.config.GRADIENT_CLIP_NORM)`

and get this error

~/work/model.py in compile(self, learning_rate, momentum)
   2234         """
   2235 
-> 2236         # Optimizer object
   2237         optimizer = SGD(
   2238             lr=learning_rate, momentum=momentum,
AttributeError: module 'tensorflow.python.keras.optimizers' has no attribute 'SGD' 

I have tried calling tensforflow.keras.optimizers.SGD or tensorflow.python.keras.optimizers.SGD instead but nothing seems to work.

I am working on a jupyter notebook in IBM Watson Studio

Thanks!

Answer 1

I ran into this issue and came across your post. I'm running python v3.9.7 on a MacBook Pro with an M1 processor. (I'm not quite sure how to check my tensor flow version, as I'm still learning).

To better understand what was happening, I ran specifically:

from tensorflow.python.keras.optimizers import SGD

did not work.

However, it did work when I ran:

from tensorflow.keras.optimizers import SGD

from tensorflow.keras.optimizers import SGD

It seems as if the tensorflow.python.keras.optimizers has different optimisers than tensorflow.keras.optimizers. However, I can't find any documentation on this and am surprised as something as fundamental as stochastic gradient descent (SGD) would be missing?