Small difference in metrics in KERAS for the same model

Question

I see that the MSE metric provided by the model.fit (history) is slightly different from the MSE calculated by model.evaluate?

Can anyone help?

# fit model
Hist = model_rna.fit(x_train, y_train,
                  validation_data=(x_val, y_val),
                  callbacks=[early_stopping],
                  verbose=2,
                  epochs=epochs)

# get last trained mse
hist = pd.DataFrame(Hist.history)
mse_train = [i for i in np.array(hist['mse']).tolist()]
print(mse_train[-1])

The result is 0.03789380192756653

# evaluate the trained model
model_rna.evaluate(x_train, y_train)

The result is: 5/5 [==============================] - 0s 4ms/step - loss: 0.0379 - mse: 0.0379 - acc: 0.0000e+00 [0.03786146640777588, 0.03786146640777588, 0.0]

If I do the "manual" calculation:

Sum_of_Squared_Errors= np.sum( (y_train - modelo_rna.predict(x_train))**2 )
print(Sum_of_Squared_Errors/len(y_train))

The result is: 0.03786148292614872

This is exactly what I found via model.evaluate() but slightly different of History of model.fit().

Why am I finding this tiny difference?

My training and validation samples are fixed.

Answer 1

I found explanation here:

https://github.com/tensorflow/tensorflow/issues/29964

https://stackoverflow.com/questions/59118430/keras-model-evaluate-on-training-and-val-set-differ-from-the-acc-and-val-acc

https://stackoverflow.com/questions/44843581/what-is-the-difference-between-model-fit-an-model-evaluate-in-keras Hope this help others.