[Python] Tensorflow - validation metric does not show up, it gets the same name as the train...

I am using the model.compile(metrics=[MyMetric])

I was wondering why I only see loss and val_loss, but only se my_metric and not val_my_metric after the evaluation at end of each epoch completes.

I debugged therefore tensorflow code I see in trainer.py:fit() it is ok:



but in CallbackList.on_epoch_end the dict is flattened using logs = python_utils.pythonify_logs(logs), therefore the keys from compile_metrics and val_compile_metrics are saved under the same key:

Note that I did not change anything in the Trainer logic. I just subclassed the train and test_step, as documented on the web:

def update_metrics(self, loss, f1, f2):
# From Cholet's Keras docs:
for metric in self.metrics:
if metric.name == "loss":
metric.update_state(loss)
else:
metric.update_state(f1, f2)

def test_step(self, data):
"""

Args:
data (tf.Tensor): Input batch of shape [B, 2, H, W, C]

Returns:
dict: Dictionary containing the loss
"""
f1, f2, p1, p2, z1, z2 = self(data, training=False)
loss = self.compute_loss(p1, p2, z1, z2)

self.update_metrics(loss, f1, f2)

# Return metrics
return {m.name: m.result() for m in self.metrics}

Continue reading...