Edit a TensorFlow v2 training model for deep learning insights

Before uploading a TensorFlow v2 training model, edit the model to work with the deep learning insights feature in Deep Learning Impact.

1. The model must import the monitor_tf module, a Python module for metric collection.

   import monitor_tf

2. If the TensorFlow model uses Keras API model.fit() for training, pass a TFKerasMonitorCallback to the fit() function. Example:

   from monitor_tf import TFKerasMonitorCallback
   model=...
   model.compile
   monitor=TFKerasMonitorCallback(test_ds)
   model.fit(train_ds, epochs=1, steps_per_epoch=1, callbacks=[monitor])

   Note: Using this method the following insights are unavailable: activation histogram, gradient histogram, gradient ratio or weight ratio.
3. If the TensorFlow model uses a custom loop for training, use TFKerasMonitor. Example:

   code example:
   from monitor_tf import TFKerasMonitor
   model=...
   tfmonitor = TFKerasMonitor(model)
   for epoch in range(100):
     print('Start of epoch %d' % (epoch,))
   
     for step, (x_batch, y_batch) in enumerate(train_ds):
       ...
       tfmonitor.log_train(step, grads, loss_value, accuracy, x_batch)
       ...
    
     for step, (x_batch, y_batch) in enumerate(test_ds):
         ...
         tfmonitor.log_test(step, loss_value, accuracy)
         ...