File size: 4,245 Bytes
e669abf
 
f387393
1abbe27
 
e669abf
 
0e20e10
e669abf
 
 
 
0a1e37c
e669abf
 
 
1abbe27
 
e669abf
0a1e37c
 
 
 
1abbe27
 
 
 
 
 
 
 
e669abf
 
 
 
 
f387393
0a1e37c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1abbe27
 
 
 
98c11b8
1abbe27
98c11b8
0a1e37c
1abbe27
98c11b8
e669abf
 
82818a6
e669abf
1abbe27
e669abf
 
 
 
 
 
 
 
 
 
d2c6fef
e669abf
 
 
 
 
 
 
 
 
0a1e37c
aac4588
6981fa0
 
0a1e37c
98c11b8
1abbe27
e669abf
 
 
 
98c11b8
e669abf
 
 
98c11b8
e669abf
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
import streamlit as st
import gradio as gr
import shap
import numpy as np
import scipy as sp
import torch
import tensorflow as tf
import transformers
from transformers import RobertaTokenizer, RobertaModel
from transformers import AutoModelForSequenceClassification
from transformers import TFAutoModelForSequenceClassification
from transformers import AutoTokenizer
from transformers_interpret import SequenceClassificationExplainer

tokenizer = AutoTokenizer.from_pretrained("paragon-analytics/ADRv1")  
model = AutoModelForSequenceClassification.from_pretrained("paragon-analytics/ADRv1")
modelc = AutoModelForSequenceClassification.from_pretrained("paragon-analytics/ADRv1").cuda


cls_explainer = SequenceClassificationExplainer(
    model,
    tokenizer)

# define a prediction function
def f(x):
    tv = torch.tensor([tokenizer.encode(v, padding='max_length', max_length=500, truncation=True) for v in x]).cuda()
    outputs = modelc(tv)[0].detach().cpu().numpy()
    scores = (np.exp(outputs).T / np.exp(outputs).sum(-1)).T
    val = sp.special.logit(scores[:,1]) # use one vs rest logit units
    return val

def adr_predict(x):
    encoded_input = tokenizer(x, return_tensors='pt')
    output = model(**encoded_input)
    scores = output[0][0].detach().numpy()
    scores = tf.nn.softmax(scores)
    
    # # build a pipeline object to do predictions
    # pred = transformers.pipeline("text-classification", model=model, 
    #                          tokenizer=tokenizer, device=0, return_all_scores=True)
    # explainer = shap.Explainer(pred)
    # shap_values = explainer([x])
    # shap_plot = shap.plots.text(shap_values)

    word_attributions = cls_explainer(str(x))
    letter = []
    score = []
    for i in word_attributions:
        if i[1]>0.5:
            a = "++"
        elif (i[1]<=0.5) and (i[1]>0.1):
            a = "+"
        elif (i[1]>=-0.5) and (i[1]<-0.1):
            a = "-"
        elif i[1]<-0.5:
            a = "--"
        else: 
            a = "NA"
    
        letter.append(i[0])
        score.append(a)
    
    word_attributions = [(letter[i], score[i]) for i in range(0, len(letter))]

    # SHAP:
    # build an explainer using a token masker
    explainer = shap.Explainer(f, tokenizer)
    shap_values = explainer(str(x), fixed_context=1)
    scores = list(zip(shap_values.data[0], shap_values.values[0, :, 1]))
    # plot the first sentence's explanation
    # plt = shap.plots.text(shap_values[0],display=False)


    return {"Severe Reaction": float(scores.numpy()[1]), "Non-severe Reaction": float(scores.numpy()[0])}, word_attributions,scores

def main(text):
    text = str(text).lower()
    obj = adr_predict(text)
    return obj[0],obj[1],obj[2]

title = "Welcome to **ADR Detector** 🪐"
description1 = """
This app takes text (up to a few sentences) and predicts to what extent the text describes severe (or non-severe) adverse reaction to medicaitons.
"""

with gr.Blocks(title=title) as demo:
    gr.Markdown(f"## {title}")
    gr.Markdown(description1)
    gr.Markdown("""---""")
    text = gr.Textbox(label="Enter Your Text Here:",lines=2, placeholder="Type it here ...")
    submit_btn = gr.Button("Analyze")


    with gr.Column(visible=True) as output_col:
        label = gr.Label(label = "Predicted Label")
        # impplot = gr.HighlightedText(label="Important Words", combine_adjacent=False).style(
        # color_map={"+++": "royalblue","++": "cornflowerblue",
        #  "+": "lightsteelblue", "NA":"white"})
        # NER = gr.HTML(label = 'NER:')
        intp = gr.HighlightedText(label="Word Scores",
        combine_adjacent=False).style(color_map={"++": "darkred","+": "red", 
                                                "--": "darkblue",
                                                "-": "blue", "NA":"white"})

        interpretation = gr.components.Interpretation(text)


    submit_btn.click(
        main,
        [text],
        [label,intp,interpretation], api_name="adr"
    )

    gr.Markdown("### Click on any of the examples below to see to what extent they contain resilience messaging:")
    gr.Examples([["I have minor pain."],["I have severe pain."]], [text], [label,intp,interpretation], main, cache_examples=True)
    
demo.launch()