NeAt (Neural Attention) Vision, is a visualization tool for the attention mechanisms of deep-learning models for Natural Language Processing (NLP) tasks.
- Visualize the attention scores, with lots of options.
- Export the visualization to SVG format. This is very convenient if you want to use the visualization in an academic paper. However, you may have to convert the SVG to PDF.
- Visualize the models predictions. Show the posterior distribution over the classes, the error in regression tasks and more. Useful for debugging your models and inspecting their behavior.
- Support for classification, multilabel classification and regression.
neat-vision is made for visualizing the weights of attention mechanisms for Natural Language Processing (Tasks) tasks. At this moment, neat-vision only supports the visualization of self-attention mechanisms, operating on the sentence-level and for the following tasks:
- Regression: predict a single continuous value.
- Multi-class Classification: a classification task with more than two classes.
Each sample belongs to one of
N
classes. - Multi-label Classification: we have
N
classes and each sample may belong to more than one classes. Essentially, it is a binary classification task for each class.
However in the future there are plans for supporting document-level models (hierarchical) and seq2seq models, such as in Neural Machine Translation (NMT).
Website (live): https://cbaziotis.github.io/neat-vision/
neat-vision takes as input 2 kinds of json
files:
- Data file. This file contains (1) the text (tokenized), (2) the attention scores and (3) the models predictions.
- Label file (optional). This is needed only in classifications tasks and if provided, it is used for mapping each class label to a user-defined description.
Here you will find a detailed overview of how to properly format the output files, for each task. Besides the necessary data needed for visualizing the attention weights, in neat-vision you can also visualise the predictions of the model and gain insights in its behavior. However, it is not required that you provide such data (e.g. posterior probabilities).
In any case, in \samples
you will find some examples,
containing the predictions of our team (NTUA-SLP) in Semeval 2018.
You can use them to test neat-vision and to check the format of the data files.
Notes
- the posteriors don't have to be normalized, which means you can simply use the logits (before the softmax). neat-vision will normalize the logits for you. This is convenient for PyTorch users.
- its ok to include the zero padded attention weights. It simply matches each token with the corresponding attention weight, so the zero padded timesteps in the attention weigths don't matter.
The structure of the data file for a classification task is the following:
{
"text": [], \\ list of strings - the tokens (words, chars) in the text. (required)
"label": 0, \\ float - the actual value. (required)
"prediction": 0, \\ float - the predicted value. (required)
"attention": [], \\ list of floats - the attention weigths. (required)
"id": "sample_11" \\ string - a unique id assigned to each sample. (required)
}
Here is an example of a sample in a data file:
{
"text": [
"i",
"have",
"never",
"been",
"so",
"excited",
"to",
"start",
"a",
"semester",
"!"
],
"label": 0.969,
"prediction": 0.8037105202674866,
"attention": [
0.030253062024712563,
0.04317276179790497,
0.12440750747919083,
0.018600208684802055,
0.023923002183437347,
0.1299467384815216,
0.1300467699766159,
0.13003277778625488,
0.1280088871717453,
0.1151493638753891,
0.12645892798900604,
0.0,
0.0,
...
0.0,
0.0
],
"id": "sample_11"
}
The structure of the data file for a classification task is the following:
{
"text": [], \\ list of strings - the tokens (words, chars) in the text. (required)
"label": 0, \\ integer - the class label. (required)
"prediction": 0, \\ integer - the predicted label. (required)
"posterior": [], \\ list of floats - the posterior probabilities. (optional)
"attention": [], \\ list of floats - the attention weigths. (required)
"id": "sample_99" \\ string - a unique id assigned to each sample. (required)
}
Here is an example of a sample in a data file:
{
"text": [
"20",
"episodes",
"left",
"i",
"am",
"dying",
"over",
"here"
],
"label": 0,
"prediction": 0,
"posterior": [
1.6511023044586182,
0.6472567319869995,
0.10215002298355103,
-1.8493231534957886
],
"attention": [
0.026811618357896805,
0.03429250791668892,
0.16327856481075287,
0.1225932389497757,
0.14799638092517853,
0.17938685417175293,
0.15541180968284607,
0.1702289879322052,
0.0,
0.0,
...
0.0,
0.0
],
"id": "sample_99"
}
The structure of the data file for a classification task is the following:
{
"text": [], \\ list of strings - the tokens (words, chars) in the text. (required)
"label": 0, \\ list of ints - the class labels - binary vector. (required)
"prediction": 0, \\ list of ints - the predicted labels - binary vector. (required)
"posterior": [], \\ list of floats - the posterior probabilities. (optional)
"attention": [], \\ list of floats - the attention weigths. (required)
"id": "sample_55" \\ string - a unique id assigned to each sample. (required)
}
Here is an example of a sample in a data file:
{
"text": [
"<allcaps>",
"fall",
"season",
"starts",
"today",
"</allcaps>",
"!",
"<repeated>"
],
"label": [
0,
1,
0,
0,
1,
0,
1,
0,
0,
0,
1
],
"prediction": [
0,
1,
0,
0,
1,
0,
0,
0,
0,
0,
0
],
"posterior": [
-2.388745069503784,
0.4522533118724823,
-3.0336408615112305,
-2.2636921405792236,
1.1948155164718628,
-2.710108995437622,
-0.09358435124158859,
-3.7859573364257812,
-3.229905605316162,
-2.832045078277588,
-2.1722922325134277
],
"attention": [
0.12348131835460663,
0.12422706931829453,
0.12277955561876297,
0.14215923845767975,
0.12141828238964081,
0.12250666320323944,
0.12207339704036713,
0.12135452032089233,
0.0,
0.0,
...
0.0,
0.0
],
"id": "sample_55"
}
In classification tasks, you can optionally provide a mapping of each class label to a name and description. Here is such an example:
{
"0": {
"name": "β€",
"desc": "_red_heart_"
},
"1": {
"name": "π",
"desc": "_smiling_face_with_hearteyes_"
},
"2": {
"name": "π",
"desc": "_face_with_tears_of_joy_"
},
"3": {
"name": "π",
"desc": "_two_hearts_"
},
"4": {
"name": "π₯",
"desc": "_fire_"
},
"5": {
"name": "π",
"desc": "_smiling_face_with_smiling_eyes_"
},
"6": {
"name": "π",
"desc": "_smiling_face_with_sunglasses_"
},
"7": {
"name": "β¨",
"desc": "_sparkles_"
},
"8": {
"name": "π",
"desc": "_blue_heart_"
},
"9": {
"name": "π",
"desc": "_face_blowing_a_kiss_"
},
"10": {
"name": "π·",
"desc": "_camera_"
},
"11": {
"name": "πΊπΈ",
"desc": "_United_States_"
},
"12": {
"name": "β",
"desc": "_sun_"
},
"13": {
"name": "π",
"desc": "_purple_heart_"
},
"14": {
"name": "π",
"desc": "_winking_face_"
},
"15": {
"name": "π―",
"desc": "_hundred_points_"
},
"16": {
"name": "π",
"desc": "_beaming_face_with_smiling_eyes_"
},
"17": {
"name": "π",
"desc": "_Christmas_tree_"
},
"18": {
"name": "πΈ",
"desc": "_camera_with_flash_"
},
"19": {
"name": "π",
"desc": "_winking_face_with_tongue_"
}
}
# install dependencies
npm install
# serve with hot reload at localhost:8080
npm run dev
# build for production with minification
npm run build
# build for production and view the bundle analyzer report
npm run build --report
For a detailed explanation on how things work, check out the guide and docs for vue-loader.