07_food_vision_milestone_project_1 can't work in Tensorflow 2.5.0

[PhyoLinMg]

I am now currently following 07_food_vision_milestone_project_1_template and I can't run the feature extraction model.

from tensorflow.keras import layers
from tensorflow.keras.layers.experimental import preprocessing

# Create base model
input_shape = (224, 224, 3)
base_model = tf.keras.applications.EfficientNetB0(include_top=False)
base_model.trainable = False # freeze base model layers

# Create Functional model 
inputs = layers.Input(shape=input_shape, name="input_layer")
# Note: EfficientNetBX models have rescaling built-in but if your model didn't you could have a layer like below
# x = preprocessing.Rescaling(1./255)(x)
x = base_model(inputs, training=False) # set base_model to inference mode only
x = layers.GlobalAveragePooling2D(name="pooling_layer")(x)
x = layers.Dense(len(class_names))(x) # want one output neuron per class 
# Separate activation of output layer so we can output float32 activations
outputs = layers.Activation("softmax", dtype=tf.float32, name="softmax_float32")(x) 
model = tf.keras.Model(inputs, outputs)

# Compile the model
model.compile(loss="sparse_categorical_crossentropy", # Use sparse_categorical_crossentropy when labels are *not* one-hot
              optimizer=tf.keras.optimizers.Adam(),
              metrics=["accuracy"])

and the output error is

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
/usr/local/lib/python3.7/dist-packages/tensorflow/python/ops/math_ops.py in binary_op_wrapper(x, y)
   1244             r_op = getattr(y, "__r%s__" % op_name)
-> 1245             out = r_op(x)
   1246             if out is NotImplemented:

17 frames
ValueError: Tensor conversion requested dtype float16 for Tensor with dtype float32: <tf.Tensor 'normalization/Cast:0' shape=(None, None, None, 3) dtype=float32>

During handling of the above exception, another exception occurred:

TypeError                                 Traceback (most recent call last)
/usr/local/lib/python3.7/dist-packages/tensorflow/python/framework/op_def_library.py in _apply_op_helper(op_type_name, name, **keywords)
    556                 "%s type %s of argument '%s'." %
    557                 (prefix, dtypes.as_dtype(attrs[input_arg.type_attr]).name,
--> 558                  inferred_from[input_arg.type_attr]))
    559 
    560         types = [values.dtype]

TypeError: Input 'y' of 'Sub' Op has type float16 that does not match type float32 of argument 'x'.

You can take a look at my current code here

I have tried to fix by typecasting it but it didn't work.
Hope someone can help me.

[gauravreddy08]

Hey @PhyoLinMg,

It's a bug with the tensorflow 2.5.0 and EfficientNetBX I downgraded the version to tensorflow 2.4.1 and it works fine now.

Weird that TensorFlow actually didn't look over this.

Here a link to another discussion which reports the same issue : link

To know further about the issue click here

[Parvez13]

I have downgraded to TensorFlow 2.4.x, Now the code is working. Thank you, @gauravreddy08

[uzorchukwu]

I have to say this tutorials has been awesome. It has been smooth sailing to this point. Please @gaurxvreddy I encountered this issue and downgraded as instructed but it doesn't seem to solve the problem as of today because it no longer downgrades.
I tried downgrading from both the jupyter notebook

...and terminal

Please is there a more recent solution to the problem? I would really appreciate your reply.
Thanks.

[gauravreddy08]

Hey, you dont have to downgrade to TF 2.4.0 anymore, Tensorflow have fixed this issue. So train your models normally, without any changes in the version, it will work (hopefully).

[uzorchukwu]

Can you elaborate on what you mean by "train your models normally".
This is because I encountered the same errors that @PhyoLinMg encountered on the 07_food_vision_milestone_project_1_template and can't seem to run the feature extraction model.

I am just looking for a way to move forward (by moving forward, I mean to be able to run the extraction model.) This is important to me because I am coding everything by hand on my system and would really love to have a solution to this.

I hope I was able to make you understand my current situation.

[uzorchukwu]

After waiting for hours with no feedback from you and hours of trying numerous options, luckily I finally found a way around it.
This will be of benefit to anyone trying to run this on your system and not on google collab.
Simply update your TensorFlow to the latest version.
Unfortunately, this issue seems to be a bug with TF 2.5.0 and you cannot downgrade to TF 2.4.1 currently (evidence above). Also, TF 2.6.0 might not be of help to all EfficientNetBX models.
I hope this was helpful.

NB: TF 2.8.0 is the latest at the time of this post.