Sanity check not working
In "Exercises - Training and Optimizing Neural Networks-FS24.ipynb" there is a sanity check in section "6.2 Overfit a tiny subset of data".
We should see zero cost (loss roughly zero). However, loss stays at initial value (roughly 2.3) after 100 epochs which means the network is randomly guessing.
Epoch 98/100
1/1 [==============================] - 0s 10ms/step - loss: 2.2498 - accuracy: 0.2000
Epoch 99/100
1/1 [==============================] - 0s 10ms/step - loss: 2.2493 - accuracy: 0.2000
Epoch 100/100
1/1 [==============================] - 0s 12ms/step - loss: 2.2488 - accuracy: 0.2000
I'm running the notebook on google colab.