Vfrdtyky

I trained a classifier for the Iris dataset using TF Estimators, but each prediction call I'm getting different results. I wonder if I'm making something wrong in the training or if it is an issue in the prediction.

I'm loading an already trained model and just making the .predict call. This is my input function for prediction.

def get_predict_fn(features,batch_size):

    def predict_input_fn():

        dataset = tf.data.Dataset.from_tensor_slices(dict(features))

        dataset = dataset.batch(batch_size)

        return dataset.make_one_shot_iterator().get_next()



    return predict_input_fn

This is the result of one call

[{'logits': array([-3.5082035, -1.074667 , -3.8533034],
dtype=float32), 'probabilities': array([0.07629351, 0.8696793 ,
0.05402722], dtype=float32), 'class_ids': array([1]), 'classes': array([b'Iris-versicolor'], dtype=object)}]

This is another call

[{'logits': array([ 3.0530725, -1.0889677, 2.3922846],
dtype=float32), 'probabilities': array([0.6525989 , 0.01037006,
0.337031 ], dtype=float32), 'class_ids': array([0]), 'classes': array([b'Iris-setosa'], dtype=object)}]

Both are making a call to the same model, sending the same example DataFrame.

sepal_length sepal_width petal_length petal_width

5.7 2.5 5.0 2.0