model training accuracy issue

[ssuhas76]

I am running the CNN - finetuning notebook and while training the model I get the below issue. It seems for alternate epochs, there accuracy is completely 0 which doesn't make any sense. Has anyone face this issue?

I tried this in both colab and kaggle, both give the same output

tf version - 2.17.0

Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/efficientnet_v2/efficientnetv2-b0_notop.h5
24274472/24274472 ━━━━━━━━━━━━━━━━━━━━ 0s 0us/step
Epoch 1/5
24/24 ━━━━━━━━━━━━━━━━━━━━ 49s 955ms/step - accuracy: 0.2408 - loss: 2.1961 - val_accuracy: 0.7089 - val_loss: 1.3862
Epoch 2/5
24/24 ━━━━━━━━━━━━━━━━━━━━ 1s 43ms/step - accuracy: 0.0000e+00 - loss: 0.0000e+00 - val_accuracy: 0.7484 - val_loss: 1.3528
Epoch 3/5
24/24 ━━━━━━━━━━━━━━━━━━━━ 5s 91ms/step - accuracy: 0.7453 - loss: 1.2831 - val_accuracy: 0.8109 - val_loss: 0.9363
Epoch 4/5
24/24 ━━━━━━━━━━━━━━━━━━━━ 1s 42ms/step - accuracy: 0.0000e+00 - loss: 0.0000e+00 - val_accuracy: 0.8191 - val_loss: 0.8962
Epoch 5/5
24/24 ━━━━━━━━━━━━━━━━━━━━ 28s 1s/step - accuracy: 0.8088 - loss: 0.9278 - val_accuracy: 0.8676 - val_loss: 0.7205

[WNastasia]

Hey, I ran into the same problem. I tried a few things and at the end I commented out the steps_per_epoch in the .fit function, then the model runs without a problem. Apparently it has something to do with the version 2,17.0. If you run it with version 2.15 it should also work.