app.py

from functools import lru_cache
from fastai.text.all import *
from fastcore.all import *
import matplotlib.cm as cm
import html
import gradio as gr

learn_inf = load_learner("20210928-model.pkl")


def _value2rgba(x, cmap=cm.RdYlGn, alpha_mult=1.0):
    "Convert a value `x` from 0 to 1 (inclusive) to an RGBA tuple according to `cmap` times transparency `alpha_mult`."
    c = cmap(x)
    rgb = (np.array(c[:-1]) * 255).astype(int)
    a = c[-1] * alpha_mult
    return tuple(rgb.tolist() + [a])


def _eval_dropouts(mod):
    module_name = mod.__class__.__name__
    if "Dropout" in module_name or "BatchNorm" in module_name:
        mod.training = False
    for module in mod.children():
        _eval_dropouts(module)


def _piece_attn_html(pieces, attns, sep=" ", **kwargs):
    html_code, spans = ['<span style="font-family: monospace;">'], []
    for p, a in zip(pieces, attns):
        p = html.escape(p)
        c = str(_value2rgba(a, alpha_mult=0.5, **kwargs))
        spans.append(
            f'<span title="{a:.3f}" style="background-color: rgba{c};">{p}</span>'
        )
    html_code.append(sep.join(spans))
    html_code.append("</span>")
    return "".join(html_code)


def _show_piece_attn(*args, **kwargs):
    from IPython.display import display, HTML

    display(HTML(_piece_attn_html(*args, **kwargs)))


@lru_cache(maxsize=1024 * 2)
def _intrinsic_attention(learn, text, class_id=None):
    "Calculate the intrinsic attention of the input w.r.t to an output `class_id`, or the classification given by the model if `None`."
    learn.model.train()
    _eval_dropouts(learn.model)
    learn.model.zero_grad()
    learn.model.reset()
    dl = learn.dls.test_dl([text])
    batch = next(iter(dl))[0]
    emb = learn.model[0].module.encoder(batch).detach().requires_grad_(True)
    emb.retain_grad()
    lstm = learn.model[0].module(emb, True)
    learn.model.eval()
    cl = learn.model[1]((lstm, torch.zeros_like(batch).bool(),))[
        0
    ].softmax(dim=-1)
    if class_id is None:
        class_id = cl.argmax()
    cl[0][class_id].backward()
    attn = emb.grad.squeeze().abs().sum(dim=-1)
    attn /= attn.max()
    tok, _ = learn.dls.decode_batch((*tuplify(batch), *tuplify(cl)))[0]
    return tok, attn


@patch
def intrinsic_attention(x: TextLearner, text: str, class_id: int = None, **kwargs):
    "Shows the `intrinsic attention for `text`, optional `class_id`"
    if isinstance(x, LMLearner):
        raise Exception("Language models are not supported")
    text, attn = _intrinsic_attention(x, text, class_id)
    return _piece_attn_html(text.split(), to_np(attn), **kwargs)


labels = learn_inf.dls.vocab[1]


@lru_cache(maxsize=1024 * 2)
def predict_label(title):
    *_, probs = learn_inf.predict(title)
    return probs


def predict(title):
    # *_, probs = learn_inf.predict(title)

    probs = predict_label(title)
    return learn_inf.intrinsic_attention(title), {
        labels[i]: float(probs[i]) for i in range(len(labels))
    }


sample_text = [
    [
        "Poems on various subjects. Whereto is prefixed a short essay on the structure of English verse"
    ],
    [
        "Journal of a Residence in China and the neighbouring countries from 1830 to 1833. With an introductory essay by the Hon. and Rev. Baptist Wriothesley Noel. [With a map.]"
    ],
    ["The Adventures of Oliver Twist. [With plates.]"],
    ["['The Adventures of Sherlock Holmes', 'Single Works']"],
    [
        "['Coal, Iron, and Oil; or, the Practical American miner. A plain and popular work on our mines and mineral resources ... With numerous maps and engravings, etc']"
    ],
    [
        "Summer Travelling in Iceland; being the narrative of two journeys across the island ... With a chapter on Askja by E. Delmar Morgan ... Containing also a literal translation of three sagas. Maps, etc'"
    ],
    [
        "Histoire de France au moyen aÃÇge, depuis Philippe-Auguste jusqu'aÃÄ la fin du reÃÄgne de Louis XI. 1223-1483. Troisieme eÃÅdition"
    ],
    [
        "Two Centuries of Soho: its institutions, firms, and amusements. By the Clergy of St. Anne's, Soho, J. H. Cardwell ... H. B. Freeman ... G. C. Wilton ... assisted by other contributors, etc"
    ],
    ["""A Christmas Carol"""],
]

description = """
British Library Books genre detection model
"""

article = """
[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.5245175.svg)](https://doi.org/10.5281/zenodo.5245175)

# British Library Books genre detection demo

This demo alows you to play with a 'genre' detection model which has been trained to predict, from the title of a book, whether it is 'fiction' or 'non-fiction'.
The demo also shows you which parts of the input the model is using most to make its prediction. You can hover over the words to see the attenton score assigned to that word. This gives you some sense of which words are important to the model in making a prediction.

## Background 

This model was developed as part of work by the [Living with Machines](https://livingwithmachines.ac.uk/). The process of training the model and working with the data is documented in a tutorial which will be released soon. 

## Model description

This model is intended to predict, from the title of a book, whether it is 'fiction' or 'non-fiction'. This model was trained on data created from the [Digitised printed books (18th-19th Century)](https://www.bl.uk/collection-guides/digitised-printed-books) book collection. 
This dataset is dominated by English language books though it includes books in several other languages in much smaller numbers. This model was originally developed for use as part of the Living with Machines project to be able to 'segment' this large dataset of books into different categories based on a 'crude' classification of genre i.e. whether the title was `fiction` or `non-fiction`.


## Training data

The model is trained on a particular collection of books digitised by the British Library. As a result the model may do less well on titles that look different to this data.
In particular the training data, was mostly English, and mostly from the 19th Century. You can find more information about the model [here]((https://doi.org/10.5281/zenodo.5245175))

## Model performance

The models performance on a held-out test set is as follows:


```
             precision    recall  f1-score   support

     Fiction       0.91      0.88      0.90       296
 Non-fiction       0.94      0.95      0.95       554

    accuracy                           0.93       850
   macro avg       0.93      0.92      0.92       850
weighted avg       0.93      0.93      0.93       850
```

> Credit: This work was partly supported by [Living with Machines](https://livingwithmachines.ac.uk/). This project, funded by the UK Research and Innovation (UKRI) Strategic Priority Fund, is a multidisciplinary collaboration delivered by the Arts and Humanities Research Council (AHRC), with The Alan Turing Institute, the British Library and the Universities of Cambridge, East Anglia, Exeter, and Queen Mary University of London. 

"""

gr_interface = gr.Interface(
    fn=predict,
    inputs=gr.inputs.Textbox(),
    outputs=[
        gr.outputs.HTML("Intrinsic attention"),
        gr.outputs.Label(num_top_classes=len(labels), label="Confidence"),
    ],
    title="British Library 19th Century Books Genre Classifier",
    description=description,
    article=article,
    examples=sample_text,
    allow_screenshot=True,
)
gr_interface.launch(inline=False, share=False)