Update app.py

app.py

@@ -1,6 +1,8 @@
 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
@@ -12,11 +14,21 @@ 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)
@@ -50,13 +62,20 @@ def adr_predict(x):
     
     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)
+    # 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
+
+    return {"Severe Reaction": float(scores.numpy()[1]), "Non-severe Reaction": float(scores.numpy()[0])}, word_attributions,plt
 
 def main(text):
     text = str(text).lower()
     obj = adr_predict(text)
-    return obj[0],obj[1]
+    return obj[0],obj[1],obj[2]
 
 title = "Welcome to **ADR Detector** 🪐"
 description1 = """
@@ -82,14 +101,16 @@ with gr.Blocks(title=title) as demo:
                                                 "--": "darkblue",
                                                 "-": "blue", "NA":"white"})
 
+        shap = gr.HighlightedText(label="SHAP Scores",combine_adjacent=False)
+
 
     submit_btn.click(
         main,
         [text],
-        [label,intp], api_name="adr"
+        [label,intp,shap], 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], main, cache_examples=True)
+    gr.Examples([["I have minor pain."],["I have severe pain."]], [text], [label,intp,shap], main, cache_examples=True)
     
 demo.launch()