When training, the model returns Nan rather than a valid loss number. This could be due to the fact that the model's loss is incorrectly setup for comparison against single-channel binary image mask.

Compiled model

model.compile(optimizer='adam',
              loss=tf.keras.losses.CategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])
model_history = model.fit(
    train_gen, 
    epochs=EPOCHS,
    steps_per_epoch=TRAIN_STEPS_PER_EPOCH,
    validation_steps=VAL_STEPS_PER_EPOCH,
    validation_data=val_gen,
    #callbacks=[DisplayCallback()])
)

Corresponding output

Epoch 1/20
 4/26 [===>..........................] - ETA: 9:44 - loss: Nan - accuracy: 0.7641

For the sake of creating a prototype, it's unreasonable to start from scratch with a completely untrained model. The purpose of this issue is to identify what model should be used as a starting point.

When running the model in browser I can only see the black masks.
How I can solve it

One underlying component of the project is that it will require using a server framework for model serving. Short-term, this is just a web page, but it could evolve to something more complex over time (think potentially something like P2P video streams).

The purpose of this issue if two-fold:

• Research server frameworks across a variety of platforms (multiple languages should be considered, although node and Python seem like the most natural fits given their platforms will most likely already be used in other portions of the project)
• Implement the skeleton of the server framework (proper routes and such will be provided in later issues)

Display predictions in notebook after training model