app.py

import numpy as np
import pickle
import urllib
from transformers import pipeline
from transformers import AutoModelForMaskedLM, AutoTokenizer

import gradio as gr
import matplotlib.pyplot as plt

plot_url = "https://huggingface.co/spaces/fvancesco/test_time_1.1/resolve/main/plot_example.p"

dates = []
dates.extend([f"18 {m}" for m in range(1,13)])
dates.extend([f"19 {m}" for m in range(1,13)])
dates.extend([f"20 {m}" for m in range(1,13)])
dates.extend([f"21 {m}" for m in range(1,13)])
months = [x.split(" ")[-1] for x in dates]

model_name = "fvancesco/tmp_date"
model = AutoModelForMaskedLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)
model.eval()
#pipe = pipeline('fill-mask', model=model, tokenizer=tokenizer, device=0)
pipe = pipeline('fill-mask', model=model, tokenizer=tokenizer)


def get_mf_dict(text):

    # predictions
    texts = []
    for d in dates:
        texts.append(f"{d} {text}")
    tmp_preds = pipe(texts, top_k=50265)
    preds = {}
    for i in range(len(tmp_preds)):
        preds[dates[i]] = tmp_preds[i]

    # get preds summary (only top words)
    top_n = 5 # top n for each prediction
    most_freq_tokens = set()
    for d in dates:
        tmp = [t['token_str'] for t in preds[d][:top_n]]
        most_freq_tokens.update(tmp)

    token_prob = {}
    for d in dates:
        token_prob[d] = {p['token_str']:p['score'] for p in preds[d]}

    mf_dict = {p:np.zeros(len(dates)) for p in most_freq_tokens}

    c=0
    for d in dates:
        for t in most_freq_tokens:
            mf_dict[t][c] = token_prob[d][t]
        c+=1

    return mf_dict

def plot_time(text):
    mf_dict = get_mf_dict(text)
    
    #max_tokens = 10

    fig = plt.figure(figsize=(16,9))
    ax = fig.add_subplot(111)
    #fig, ax = plt.subplots(figsize=(16,9))

    x = [i for i in range(len(dates))]

    ax.set_xlabel('Month')
    ax.set_xlim(0)
    ax.set_xticks(x)
    ax.set_xticklabels(months)
    # ax.set_yticks([-1,0,1])

    ax2 = ax.twiny()
    ax2.set_xlabel('Year')
    ax2.set_xlim(0)
    ax2.set_xticks([0,12,24,36,47])

    ax2.set_xticklabels('')
    ax2.set_xticks([6,18,30,42,47], minor=True)
    ax2.set_xticklabels(['2018','2019','2020','2021',''], minor=True)

    ax2.grid()

    # plot lines
    for k in mf_dict.keys():
        ax.plot(x, mf_dict[k], label = k)
    # k = list(mf_dict.keys())
    # for i in range(max_tokens):
    #     ax.plot(x, mf_dict[k[i]], label = k[i])

    ax.legend(loc='center left', bbox_to_anchor=(1.0, 0.5))
    
    return fig

def add_mask(text):
    out = ""
    if len(text) == 0 or text[-1] == " ": 
        out = text+"<mask>"
    else: 
        out = text+" <mask>"
    return out

with gr.Blocks() as demo:
    
    text_description="""
    # TimeLMs Demo

    This is a demo for **timeLMs**:
    - [Github](https://github.com/cardiffnlp/timelms)
    - [Paper](https://aclanthology.org/2022.acl-demo.25.pdf)

    Input any text with a *\<mask\>* token as in the example, and (the demo does not
    use GPUs, and it takes about 1 min). In the graph, we show the probability of 
    some token candidates for mask over different months. 
    
    In this demo we run use a roberta-base model trained on tweets, where the first two 
    tokens are the year and the month ("21 1" for January 2021). It was trained 
    for tweets between January 2018 to December 2021).
    """
    
    description = gr.Markdown(text_description)
    
    textbox = gr.Textbox(value="Happy <mask>!", max_lines=1)

    with gr.Row():
        generate_btn = gr.Button("Generate Plot")
        mask_btn = gr.Button("Add <mask>")

    # plot (with starting example already loaded)
    f = urllib.request.urlopen(plot_url)
    plot_example = pickle.load(f)
    plot = gr.Plot(plot_example)
    
    #textbox.change(fn=plot_time, inputs=textbox, outputs=plot)
    generate_btn.click(fn=plot_time, inputs=textbox, outputs=plot)
    mask_btn.click(fn=add_mask, inputs=textbox, outputs=textbox)

demo.launch(debug=True)