How To Use tf.estimator.DNNClassifier (Scikit Flow?)

Question

Could someone point me to a basic working example for tf.estimator.DNNClassifier (originally skflow)?

Since I'm familiar with Sklearn, I was excited to read about Scikit Flow on this blog. Especially the api looked pretty much the same as SK-Learn.

However, I was having a problem getting the code from the blog to work.

Then I read from Scikit Flow Github that it moved to tensorflow/tensorflow/contrib/learn/python/learn.

Upon further investigation, I found tf.contrib.learn.DNNClassifier moved to tf.estimator.DNNClassifier.

However, now api for estimator seems pretty different than sklearn classifier.

I would appreciate if someone could point me to a basic working example.

Here's the code from the blog above.

import tensorflow.contrib.learn as skflow
from sklearn import datasets, metrics

iris = datasets.load_iris()
classifier = skflow.TensorFlowDNNClassifier(hidden_units=[10, 20, 10], n_classes=3)
classifier.fit(iris.data, iris.target)
score = metrics.accuracy_score(iris.target, classifier.predict(iris.data))
print("Accuracy: %f" % score)

score 1 · Accepted Answer · answered Apr 06 '19 at 16:31

The API was changed very much, so now you can do something like this (an official example is available here):

import tensorflow as tf
from sklearn import datasets, metrics

def train_input_fn(features, labels, batch_size):
    dataset = tf.data.Dataset.from_tensor_slices((dict(features), labels))
    return dataset.shuffle(1000).repeat().batch(batch_size)

iris = datasets.load_iris()
train_x = {
    '0': iris.data[:, 0],
    '1': iris.data[:, 1],
    '2': iris.data[:, 2],
    '3': iris.data[:, 3],
}

my_feature_columns = []
for key in train_x.keys():
    my_feature_columns.append(tf.feature_column.numeric_column(key=key))

clf = tf.estimator.DNNClassifier(hidden_units=[10, 20, 10], feature_columns=my_feature_columns, n_classes=3)
clf.train(input_fn=lambda: train_input_fn(train_x, iris.target, 32), steps=10000)

preds = list()
for idx, p in enumerate(classifier.predict(input_fn=lambda: train_input_fn(train_x, iris.target, 32))):
    preds.append(p['class_ids'][0])
    if idx == 99:
        break

print(metrics.accuracy_score(iris.target[:100], preds))

But nowadays it is better to use TF Keras API like this:

import tensorflow as tf
from sklearn import datasets, metrics

def train_input_fn(features, labels, batch_size):
    dataset = tf.data.Dataset.from_tensor_slices((dict(features), labels))
    return dataset.shuffle(1000).repeat().batch(batch_size)

iris = datasets.load_iris()

clf = tf.keras.models.Sequential([
    tf.keras.layers.Dense(10, activation='sigmoid'),
    tf.keras.layers.Dense(20, activation='sigmoid'),
    tf.keras.layers.Dense(10, activation='sigmoid'),
    tf.keras.layers.Dense(3, activation='softmax'),
])
clf.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
clf.fit(iris.data, iris.target, batch_size=32)

Awesome!! Thank you so much! Since api changed quite a bit, I assume I can't use cross_validate from sklearn.model_selection to get different scores like f1_macro? — jl303, Apr 07 '19 at 13:10
It looks like this, so you can implement your own cross validation approach with `KFold` from `sklearn` as shown [here](https://stackoverflow.com/a/46589591/6453992) if you want. — Petr Vytovtov, Apr 07 '19 at 18:05

How To Use tf.estimator.DNNClassifier (Scikit Flow?)

1 Answers1