Metadata-Version: 2.1
Name: KENN2
Version: 0.1
Summary: Knowledge Enhanced Neural Networks
Home-page: https://github.com/DanieleAlessandro/KENN2
Author: Alessandro Daniele, Riccardo Mazzieri, Luciano Serafini
Author-email: daniele@fbk.eu, serafini@fbk.eu
License: UNKNOWN
Project-URL: Bug Tracker, https://github.com/DanieleAlessandro/KENN2/issues
Keywords: kenn,Statistical Relational Learning,Neural Symbolic Learning,Learning with logic
Platform: UNKNOWN
Classifier: Programming Language :: Python :: 3
Classifier: License :: OSI Approved :: BSD License
Classifier: Operating System :: OS Independent
Requires-Python: >=3.6
Description-Content-Type: text/markdown
Requires-Dist: tensorflow
Requires-Dist: numpy

# KENN: Knowledge Enhanced Neural Networks

KENN (Knowledge Enhanced Neural Networks) is a library for python 2.7 built on top of TensorFlow that permit to enhance neural networks models with logical constraints (clauses). It does so by adding a new final layer, called **Knowledge Enhancer (KE)**, to the existing neural network. The KE change the orginal predictions of the standard neural network enforcing the satisfaction of the constraints. Additionally, it contains **clause weights**, learnable parameters which represent the strength of the constraints.

This is an implementation of the model presented in our paper:
[Knowledge Enhanced Neural Networks](https://link.springer.com/chapter/10.1007/978-3-030-29908-8_43).

If you use this software for academic research, please, cite our work using the following BibTeX:

```
@InProceedings{10.1007/978-3-030-29908-8_43,
author="Daniele, Alessandro
and Serafini, Luciano",
editor="Nayak, Abhaya C.
and Sharma, Alok",
title="Knowledge Enhanced Neural Networks",
booktitle="PRICAI 2019: Trends in Artificial Intelligence",
year="2019",
publisher="Springer International Publishing",
address="Cham",
pages="542--554",
isbn="978-3-030-29908-8"
}
```

**NB:** a [new version of KENN](https://github.com/DanieleAlessandro/KENN2) based on TensorFlow 2 has been also released and it implements additional features to work with relational domains.

## Installation

KENN can be installed using pip:

```
pip install kenn
```

## Getting started

To use KENN, you can start from existing TensorFlow code. Few changes must be done to add the logical clauses.
Here a simple example, where we higlighted the 5 changes that are typically needed (for simplicity, we omit most of the standard TensorFlow code):

```python

import tensorflow as tf

# *** 1 ***
from kenn import Knowledge_base as kb

# *** 2 ***
clauses = kb.read_knowledge_base(kb_file_name)

# Standard TensorFlow code
# ....

# Calculate preactivations
z = tf.matmul(h, w) + bias

# Standard code:
# y_hat = tf.nn.sigmoid(z)

# *** 3 ***
_, y_hat = kb.knowledge_enhancer(z, clauses)

# Loss definition
loss = tf.losses.mean_squared_error(y_hat, y)

# optimizer
train_step = tf.train.RMSPropOptimizer(learning_rate).minimize(loss)

# *** 4 ***
clauses_clip_ops = kb.clip_weigths(clauses)

with tf.Session(config=config) as sess:
    sess.run(tf.global_variables_initializer())

    for _ in range(number_of_steps):
        sess.run(train_step)

        # *** 5 ***
        sess.run(clauses_clip_ops)

    # Evaluations and standard operations ...

    # *** 6 (optional) ***
    learned_kb = kb.kb_to_string(sess, clauses)

    with open(kb_output_file, 'w') as kb_file:
        kb_file.writelines(learned_kb)
```

### Example explained

In the previous example, we applies 5 changes to the standard TensorFlow code. Following, the details.

#### 1. Import knowledge_base module

The first change is trivial, we need to import the library:

```python
from kenn import Knowledge_base as kb
```

#### 2. Read the knowledge base file

```python
clauses = kb.read_knowledge_base(kb_file_name)
```

The `read_knowledge_base` function takes as input the path of the file containing the logical constraints. Following, an example of knowledge base file:

```
Dog,Cat,Animal,Car,Truck,Chair

1.5:nDog,Animal
_:nCat,Animal
2.0:nDog,nCat
_:nCar,Animal
_:nAnimal,Dog,Cat
```

The first row contains a list of predicates separated with a comma with no spaces. Each predicate must start with a capital letter.
The second row must be empty.
Other rows contain the clauses.

Each clause is in a separate row and must be written respecting this properties:

1. logical disjunctions are represented with commas
1. if a literal is negated, it must be precedeed by the lowercase 'n'
1. they must contain only predicates specified in the first row
1. there shouldn't be spaces

Additionally, each clause must be precedeed by a positive weight that represents the strength of the clause. More precisly, the weight could be a numeric value or an underscore: in the first case, the weight is fixed and determined by the specfied value, in the second case the weight is learned during training.

For example, the third line represents the clause <a href="https://www.codecogs.com/eqnedit.php?latex=\lnot&space;Dog&space;\lor&space;Animal" target="_blank"><img src="https://latex.codecogs.com/gif.latex?\lnot&space;Dog&space;\lor&space;Animal" title="\lnot Dog \lor Animal" /></a> and tells us that a dog should also be an animal. In this case, the clause weigth is fixed to the value 1.5. A more interesting clause is the last one, that tells us that in our domain only cats and dogs are animals. Moreover, the corresponding weight is learned and if the constraint is not satisfied in the training set, KENN learn to ignore it.

#### 3. Knowledge Enhancer (KE)

This is the most relevant change. Usually, in the last layer of a neural network, we apply an activation function to the preactivations calculated by previous layer.

```python
y_hat = tf.nn.sigmoid(z)
```

Here, instead, we feed the preactivations to the knowledge enhancer:

```python
_, y_hat = kb.knowledge_enhancer(z, clauses)
```

The `knowledge_enhancer` function takes as input the preactivations of the final layer and the set of clauses previously generated by the `read_knowledge_base` function (see 2.). It defines a new layer that changes the results by enforcing the clauses satisfaction. More precisely, it returns a tuple containing the preactivations (usefull when using the `cross_entropy_with_logits` loss function) and final activations of the KE layer.

At this point, we can use the output of the `knowledge_enhancer` as we would have used the original activations: we can define the loss function directly on such predictions.

**NB:** in order to work properly, the preactivations tensor given as input to `knowledge_enhancer` should be a matrix whose rows represent a possible grounding. The columns represent the predicates of the knowledge base and should be in the same order as specified in the first row of knowledge base file.

#### 4-5. Clauses weights clipping

Another change that is needed is to define the clauses_clip_ops.

```python
clauses_clip_ops = kb.clip_weigths(clauses)
```

Such operation set to zero the clause weights that became negative. It should be called after each train step in order to keep the clauses weights positive:

```python
_ = sess.run(train_step)
sess.run(clauses_clip_ops)
```

#### 6. Saving learned clauses weigths

Optionally, by calling `kb_to_string` function, it is possible to obtain a string containing the clauses with the learned weights and storing it into a file.

```python
# *** 6 (optional) ***
learned_kb = kb.kb_to_string(sess, clauses)

with open(kb_output_file, 'w') as kb_file:
    kb_file.writelines(learned_kb)
```

The content of the string will be something like this:

```
1.5:nDog,Animal
1.301:nCat,Animal
2.0:nDog,nCat
3.02:nCar,Animal
0.0:nAnimal,Dog,Cat
```

As you can see, these are the same clauses given as input through the knowledge base file (see 2.) with the only difference that the underscores are sustituted by actual numbers, that are the learned clause weights. For instance, in this example the method learned that the last constraint shouldn't be taken in consideration by KENN.

**NB:** notice that writing this last piece of code is not mandatory, but reading the learned clause weights could help the user to better understand the importance of each clause in the final predictions made by KENN, increasing in this way the interpretability.

## License

Copyright (c) 2019, Daniele Alessandro, Serafini Luciano
All rights reserved.

Licensed under the BSD 3-Clause License.


