Research parameters:

[GreNait]

• [ ] What different activation functions are available and usable for the model
• [x] What loss functions/optimizers are available
• [x] What amount of layers with how many neurons make sense? <- Practical test
• [ ] Comparison between those different models
• [ ] addin Bounding box polygons to the detected figure - Labelling data https://github.com/tzutalin/labelImg
• [ ] best practice with ML
• [ ] logits of a model (model prediction) https://developers.google.com/machine-learning/glossary#logits
• [ ] what is SparseCategoricalCrossentropy and what are others? https://www.tensorflow.org/tutorials/keras/classification
• [ ] what is leaky relu ? https://www.tensorflow.org/api_docs/python/tf/nn/leaky_relu?hl=lt

[GreNait]

Flatten() needs to be done before have fully conected Layers but after a convolution. Flatten() makes a threediemsnional array flat for a vecctorisation. A model does not need the 3dimensional structure, convolutions do. See: https://stackoverflow.com/questions/44572141/why-do-we-flatten-the-data-before-we-feed-it-into-tensorflow

[GreNait]

A guide to activation functions https://missinglink.ai/guides/neural-network-concepts/7-types-neural-network-activation-functions-right/

[GreNait]

It appears, that it is very important to not have an activation function at the last layer. I used for the small model test an softmax at the end. The accuracy was than stuck to 33%.