Keras: Precision, Recall and F1 Metrics Removed

[Lif3line]

It appears Precision, Recall and F1 metrics have been removed from metrics.py as of today but I couldn't find any reference to their removal in the commit logs. Was this intentional?

[soldni]

Yes, it was intentional. See https://github.com/fchollet/keras/wiki/Keras-2.0-release-notes

[Lif3line]

Ah missed that, thank you very much.

[inexxt]

What was the reason behind removing them?

[fchollet]

Basically these are all global metrics that were approximated
batch-wise, which is more misleading than helpful. This was mentioned in
the docs but it's much cleaner to remove them altogether. It was a mistake
to merge them in the first place.

On 19 March 2017 at 11:28, Jacek Karwowski [e-mail] wrote:

What was the reason behind removing them?

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
https://github.com/fchollet/keras/issues/5794#issuecomment-287636509,
or mute the thread
https://github.com/notifications/unsubscribe-auth/AArWb3iZAJPZC-FrViC7eESFBO4ceX2yks5rnXPPgaJpZM4MeOw0
.

[karimpedia]

@fchollet Are there any plans to add another implementation for these metrics (evaluated globally) ?

[iuria21]

@karimpedia What I did was creating a Callback and calculate them on the end of each epoch with the validation data.

class Metrics(keras.callbacks.Callback):
    def on_epoch_end(self, batch, logs={}):
        predict = np.asarray(self.model.predict(self.validation_data[0]))
        targ = self.validation_data[1]
        self.f1s=f1(targ, predict)
        return
metrics = Metrics()
model.fit(X_train, y_train, epochs=epochs, batch_size=batch_size, validation_data=[X_test,y_test], 
       verbose=1, callbacks=[metrics])

And implement a function that calculates the f1 score or instead use Scikit Learn's Fscore function.

[ndor]

@imanoluria, i was using your code (thanks for posting it BTW) with the sklearn's F1 in my model. the model has 3 inputs and one output.
i used two calbacks: callbacks=[checkpointer, metrics].
alas, i get the error:

  File "build/bdist.linux-x86_64/egg/keras/engine/training.py", line 1631, in fit
  File "build/bdist.linux-x86_64/egg/keras/engine/training.py", line 1233, in _fit_loop
  File "build/bdist.linux-x86_64/egg/keras/callbacks.py", line 73, in on_epoch_end
  File "rafael.py", line 29, in on_epoch_end
    predict = np.asarray(self.model.predict(self.validation_data[0]))
  File "build/bdist.linux-x86_64/egg/keras/engine/training.py", line 1730, in predict
  File "build/bdist.linux-x86_64/egg/keras/engine/training.py", line 121, in _standardize_input_data
ValueError: The model expects 3  arrays, but only received one array. Found: array with shape (34374, 15, 9)

the fit portion:

    history = model.fit([T_train, S_train, P_train], y_train,
                        batch_size=size_batch,
                        nb_epoch=epochs,
                        verbose=1,
                        callbacks=[checkpointer, metrics],
                        class_weight=weights_dict,
                        validation_data=[[T_validation, S_validation, P_validation], y_validation],
                        shuffle=True)

my input to the validation is (first dim is #samples):
(34374, 15, 9)-temporal samples
(34374, 3)-1D vector samples
(34374, 7)-1D vector samples

my model's output is a one hot 25 category vector...

can you see what's the problem? thanks a lot!

@fchollet - will there be a future innate API for metrics in check pointer / similar method?

[ShiangYong]

Hi @basque21, I tried to extend your example to work with predict_generator but it did not work. Any ideas?

The error I got was AttributeError: 'Metrics' object has no attribute 'validation_data'.

My code snippet is below:

class Metrics(keras.callbacks.Callback):
    def on_epoch_end(self, batch, logs={}):
        predict = self.model.predict_generator(
            self.validation_data,
            steps=self.validation_steps,
            workers=6
        )
        targ = self.targ
        self.f1s=f1(targ, predict)
        return

[dakl]

@ShiangYong Did you set validation_data when you fit()?

[gabrielam2018]

Em sex, 20 de abr de 2018 às 2:11 AM, Daniel Klevebring <
[email protected]> escreveu:

@ShiangYong https://github.com/ShiangYong Did you set validation_data
when you fit()?

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
https://github.com/keras-team/keras/issues/5794#issuecomment-382978514,
or mute the thread
https://github.com/notifications/unsubscribe-auth/Ai62Hu897l4u2QQZEvkx3QTfGiAu-HhCks5tqW3mgaJpZM4MeOw0
.

no

[ShiangYong]

@dakl No, I need to use fit_generator or predict_generator for my applications, those expect generators and not validation_data

[arcticriki]

If someone needs to implement it I suggest this work-around:

• install keras_metrics package by ybubnov
• call model.fit(nb_epoch=1, ...) inside a for loop taking advantage of the precision/recall metrics outputted after every epoch

Something like this:

```
for mini_batch in range(epochs):
model_hist = model.fit(X_train, Y_train, batch_size=batch_size, epochs=1,
verbose=2, validation_data=(X_val, Y_val))

    precision = model_hist.history['val_precision'][0]
    recall = model_hist.history['val_recall'][0]
    f_score = (2.0 * precision * recall) / (precision + recall) ```