Wrong Results using tpr, tnr, fpr, fnr

[thogaertner]

Hi,

I using the metrics (fpr, fnr, tpr, tnr) for my model in Keras (Tensorflow.python keras v2.1.2) as you can see:

model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['binary_accuracy', tpr, tnr, fpr, fnr])

But running this with history = model.fit_generator(..) am getting the accuracy on tpr and fpr; and 1-acc for tnr and fnr. For debugging I returned also the outputs of the contingency_table and tp=tn = all true predicitons and fp=fn = all false predicitons.

Do you have any suggestions how to fix this issue and get the right values? Thanks in advance!

[Avsecz]

The problem with Keras metrics is that they are computed for each batch and averaged on the fly. Consider a metric computed as A/B. If the denominator B is the same for all the batches (which is the case for accuracy), then the average across batches is the same as the metric computed across the whole dataset. Hower, if B varies across batches, then the average of batch metrics is not the same as the global metric.

Can you compute your metrics using

from concise.eval_metrics import *
y_pred = model.predict(x)
tpr(y_true, y_pred)
....
fpr(y_true, y_pred)

and print those together with the metrics you obtain from fig_generator?

[thogaertner]

Thanks a lot for the fast reply.

"The problem with Keras metrics is that they are computed for each batch and averaged on the fly. Consider a metric computed as A/B. If the denominator B is the same for all the batches (which is the case for accuracy), then the average across batches is the same as the metric computed across the whole dataset. Hower, if B varies across batches, then the average of batch metrics is not the same as the global metric." - total agree. But my goal is to have a metrics with the contingency table.

I could not compute the metrics in model.compile() using the code from concise.eval_metrics.

[Avsecz]

For model.compile you should use the metrics from concise.metrics and for the evaluation of the prediction you should use metrics from concise.eval_metrics. The reason is that the former have to be implemented with Keras functions.

[thogaertner]

Thanks a lot. I used these functions in 'model.compile'. But the issue is, that the tpr and tnr are similar as you can see below. For testing I used just a small data set, but the values should distinguish more.

Epoch 1/3
62/62 [==============================] - 108s 2s/step - loss: 0.7520 - binary_accuracy: 0.6280 - tpr: 0.6280 - tnr: 0.6280 - val_loss: 2.5995 - val_binary_accuracy: 0.4271 - val_tpr: 0.4271 - val_tnr: 0.4271
Epoch 2/3
62/62 [==============================] - 28s 459ms/step - loss: 0.6160 - binary_accuracy: 0.6845 - tpr: 0.6845 - tnr: 0.6845 - val_loss: 0.9102 - val_binary_accuracy: 0.4870 - val_tpr: 0.4870 - val_tnr: 0.4870
Epoch 3/3
62/62 [==============================] - 27s 439ms/step - loss: 0.5802 - binary_accuracy: 0.7016 - tpr: 0.7016 - tnr: 0.7016 - val_loss: 1.6536 - val_binary_accuracy: 0.4818 - val_tpr: 0.4818 - val_tnr: 0.4818

[thogaertner]

Solved the issue:

I set my data set to a categorical class mode with to classes, so the values were one hot encoded. Applying K.argmax() y and y_pred is solving the issue.

[Kuselokusi]

Good-day, Hope you are well, @tommysft I would like to ask how you accessed the tpr and tnr in the model.compile() line of code. I also have a binary classification problem, however I am having a hard time getting the tpr and tnr.

I used the code model.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ["accuracy", "tpr", "tnr"])

Please indicate what I am doing wrong.