Tensorflow Autoencoders different results between local (M2 Pro Max) and colab / kaggle

Question

nivmorabin OP

Created Nov ’23

Replies 2

Boosts 1

Participants 2

Hi,

I've been going over this tutorial of autoencoders https://www.tensorflow.org/tutorials/generative/autoencoder#third_example_anomaly_detection

Notebook link https://colab.research.google.com/github/tensorflow/docs/blob/master/site/en/tutorials/generative/autoencoder.ipynb

And when I downloaded and ran the notebook locally on my M2 Pro Max - the results were dramatically different and the plots were way off.

This is the plot in the working notebook: Screenshot 2023-11-08 at 0.30.11.png

This is the local plot: Screenshot 2023-11-08 at 0.27.23.png

I checked every moving piece and the difference seems to be in the output of the autoencoder, these lines:

encoded_data = autoencoder.encoder(normal_test_data).numpy() decoded_data = autoencoder.decoder(encoded_data).numpy()

The working notebook output is: Screenshot 2023-11-08 at 0.34.47.png

The local output: Screenshot 2023-11-08 at 0.34.37.png

And the overall result is notebook:

Accuracy = 0.944

Precision = 0.9941176470588236

Recall = 0.9053571428571429

local:

Accuracy = 0.44

Precision = 0.0

Recall = 0.0

I'm using Mac M2 Pro Max Python 3.10.12 Tensorflow 2.14.0

Can anyone help?

Thanks a lot in advance.

Boost

Answer 1

nivmorabin OP

Nov ’23

Apparently when using Mac M2 you need to only install the tensorflow-macos and not tensorflow alongside with it.

Installing tensorflow-macos alone solved the problem for me :)

Thread can be closed, thanks!

0

Answer 2

Kira1108 OP

Nov ’23

I can produce the same error when following the tutorial: https://www.tensorflow.org/tutorials/generative/autoencoder?hl=zh-cn

TensorFlow packages: tensorflow 2.14.0 tensorflow-estimator 2.14.0 tensorflow-io-gcs-filesystem 0.34.0 tensorflow-macos 2.14.0 tensorflow-metal 1.1.0

When training on linux produce (Google Colab TensorFlow = 2.14.0) Epoch 1/10 1875/1875 [==============================] - 12s 3ms/step - loss: 0.0236 - val_loss: 0.0130 Epoch 2/10 1875/1875 [==============================] - 7s 4ms/step - loss: 0.0115 - val_loss: 0.0105 Epoch 3/10 1875/1875 [==============================] - 6s 3ms/step - loss: 0.0101 - val_loss: 0.0098 Epoch 4/10 1875/1875 [==============================] - 7s 4ms/step - loss: 0.0095 - val_loss: 0.0094 Epoch 5/10 1875/1875 [==============================] - 6s 3ms/step - loss: 0.0092 - val_loss: 0.0092 Epoch 6/10 1875/1875 [==============================] - 7s 4ms/step - loss: 0.0091 - val_loss: 0.0091 Epoch 7/10 1875/1875 [==============================] - 7s 4ms/step - loss: 0.0090 - val_loss: 0.0091 Epoch 8/10 1875/1875 [==============================] - 6s 3ms/step - loss: 0.0089 - val_loss: 0.0089 Epoch 9/10 1875/1875 [==============================] - 7s 3ms/step - loss: 0.0088 - val_loss: 0.0089 Epoch 10/10 1875/1875 [==============================] - 6s 3ms/step - loss: 0.0088 - val_loss: 0.0089 <keras.src.callbacks.History at 0x7fe4fd3afbe0>

Training on Mac gives (Exactly the same code)

Epoch 1/10 1875/1875 [==============================] - 12s 3ms/step - loss: 0.0236 - val_loss: 0.0130 Epoch 2/10 1875/1875 [==============================] - 7s 4ms/step - loss: 0.0115 - val_loss: 0.0105 Epoch 3/10 1875/1875 [==============================] - 6s 3ms/step - loss: 0.0101 - val_loss: 0.0098 Epoch 4/10 1875/1875 [==============================] - 7s 4ms/step - loss: 0.0095 - val_loss: 0.0094 Epoch 5/10 1875/1875 [==============================] - 6s 3ms/step - loss: 0.0092 - val_loss: 0.0092 Epoch 6/10 1875/1875 [==============================] - 7s 4ms/step - loss: 0.0091 - val_loss: 0.0091 Epoch 7/10 1875/1875 [==============================] - 7s 4ms/step - loss: 0.0090 - val_loss: 0.0091 Epoch 8/10 1875/1875 [==============================] - 6s 3ms/step - loss: 0.0089 - val_loss: 0.0089 Epoch 9/10 1875/1875 [==============================] - 7s 3ms/step - loss: 0.0088 - val_loss: 0.0089 Epoch 10/10 1875/1875 [==============================] - 6s 3ms/step - loss: 0.0088 - val_loss: 0.0089 <keras.src.callbacks.History at 0x7fe4fd3afbe0>

0