Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -1,85 +1,102 @@
|
|
1 |
import numpy as np
|
2 |
import torch
|
3 |
import shap
|
4 |
-
from transformers import
|
5 |
-
|
|
|
|
|
|
|
|
|
6 |
import gradio as gr
|
7 |
|
8 |
# 1) Device setup
|
9 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
10 |
|
11 |
-
# 2) Load ADR classifier
|
12 |
model_name = "paragon-analytics/ADRv1"
|
13 |
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
14 |
model = AutoModelForSequenceClassification.from_pretrained(model_name).to(device)
|
15 |
|
16 |
-
# 3)
|
17 |
pred_pipeline = pipeline(
|
18 |
"text-classification",
|
19 |
model=model,
|
20 |
tokenizer=tokenizer,
|
21 |
return_all_scores=True,
|
22 |
-
device=0 if device == "cuda" else -1
|
23 |
)
|
24 |
|
25 |
-
# 4)
|
26 |
def predict_proba(texts):
|
27 |
if isinstance(texts, str):
|
28 |
texts = [texts]
|
29 |
results = pred_pipeline(texts)
|
30 |
-
# results
|
31 |
probs = np.array([[d["score"] for d in sample] for sample in results])
|
32 |
return probs
|
33 |
|
34 |
-
# 5)
|
35 |
-
|
36 |
-
#
|
37 |
-
|
38 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
39 |
explainer = shap.Explainer(
|
40 |
-
|
41 |
masker=masker,
|
42 |
output_names=class_labels
|
43 |
)
|
44 |
|
45 |
-
#
|
46 |
-
|
47 |
-
|
|
|
48 |
ner_pipe = pipeline(
|
49 |
"ner",
|
50 |
model=ner_model,
|
51 |
tokenizer=ner_tokenizer,
|
52 |
aggregation_strategy="simple",
|
53 |
-
device=0 if device == "cuda" else -1
|
54 |
)
|
55 |
|
56 |
-
#
|
57 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
# a) Predict probabilities
|
59 |
-
probs = predict_proba(text)[0]
|
60 |
prob_dict = {label: float(probs[i]) for i, label in enumerate(class_labels)}
|
61 |
|
62 |
-
# b) SHAP explanation
|
63 |
shap_values = explainer([text])
|
64 |
fig = shap.plots.text(shap_values[0], display=False)
|
65 |
|
66 |
# c) NER highlighting
|
67 |
-
|
68 |
-
colors = {
|
69 |
-
"Severity": "red",
|
70 |
-
"Sign_symptom": "green",
|
71 |
-
"Medication": "lightblue",
|
72 |
-
"Age": "yellow",
|
73 |
-
"Sex": "yellow",
|
74 |
-
"Diagnostic_procedure": "gray",
|
75 |
-
"Biological_structure": "silver"
|
76 |
-
}
|
77 |
highlighted = ""
|
78 |
last_idx = 0
|
79 |
-
for ent in
|
80 |
start, end = ent["start"], ent["end"]
|
81 |
word = ent["word"].replace("##", "")
|
82 |
-
color =
|
83 |
highlighted += (
|
84 |
text[last_idx:start]
|
85 |
+ f"<mark style='background-color:{color};'>{word}</mark>"
|
@@ -89,23 +106,31 @@ def adr_predict(text):
|
|
89 |
|
90 |
return prob_dict, fig, highlighted
|
91 |
|
92 |
-
#
|
93 |
with gr.Blocks() as demo:
|
94 |
gr.Markdown("## Welcome to **ADR Detector** 🪐")
|
95 |
gr.Markdown(
|
96 |
-
"Predicts the likelihood your text describes a severe vs. non-severe adverse reaction.
|
97 |
-
"_(Not for medical
|
98 |
)
|
99 |
|
100 |
-
txt = gr.Textbox(
|
|
|
|
|
|
|
|
|
101 |
btn = gr.Button("Analyze")
|
102 |
|
103 |
with gr.Row():
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
|
108 |
-
btn.click(
|
|
|
|
|
|
|
|
|
109 |
|
110 |
gr.Examples(
|
111 |
examples=[
|
@@ -113,9 +138,10 @@ with gr.Blocks() as demo:
|
|
113 |
"A 35-year-old female had minor abdominal pain after Acetaminophen."
|
114 |
],
|
115 |
inputs=txt,
|
116 |
-
outputs=[
|
117 |
fn=adr_predict,
|
118 |
cache_examples=True
|
119 |
)
|
120 |
|
121 |
-
|
|
|
|
1 |
import numpy as np
|
2 |
import torch
|
3 |
import shap
|
4 |
+
from transformers import (
|
5 |
+
pipeline,
|
6 |
+
AutoTokenizer,
|
7 |
+
AutoModelForSequenceClassification,
|
8 |
+
AutoModelForTokenClassification
|
9 |
+
)
|
10 |
import gradio as gr
|
11 |
|
12 |
# 1) Device setup
|
13 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
14 |
|
15 |
+
# 2) Load ADR classifier model & tokenizer
|
16 |
model_name = "paragon-analytics/ADRv1"
|
17 |
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
18 |
model = AutoModelForSequenceClassification.from_pretrained(model_name).to(device)
|
19 |
|
20 |
+
# 3) Build HF text-classification pipeline
|
21 |
pred_pipeline = pipeline(
|
22 |
"text-classification",
|
23 |
model=model,
|
24 |
tokenizer=tokenizer,
|
25 |
return_all_scores=True,
|
26 |
+
device=0 if device.type == "cuda" else -1
|
27 |
)
|
28 |
|
29 |
+
# 4) Base predict_proba: List[str] → np.ndarray of shape (n_samples, n_classes)
|
30 |
def predict_proba(texts):
|
31 |
if isinstance(texts, str):
|
32 |
texts = [texts]
|
33 |
results = pred_pipeline(texts)
|
34 |
+
# results: List[List[{"label":…, "score":…}]]
|
35 |
probs = np.array([[d["score"] for d in sample] for sample in results])
|
36 |
return probs
|
37 |
|
38 |
+
# 5) SHAP-compatible wrapper: joins token lists back into strings
|
39 |
+
def predict_proba_shap(inputs):
|
40 |
+
# inputs: List[str] or List[List[str]]
|
41 |
+
texts = [
|
42 |
+
" ".join(x) if isinstance(x, list) else x
|
43 |
+
for x in inputs
|
44 |
+
]
|
45 |
+
return predict_proba(texts)
|
46 |
+
|
47 |
+
# 6) Instantiate SHAP explainer with a Text masker
|
48 |
+
masker = shap.maskers.Text(tokenizer)
|
49 |
+
# Grab output class labels from a dummy sample
|
50 |
+
_example = pred_pipeline(["test"])[0]
|
51 |
+
class_labels = [d["label"] for d in _example]
|
52 |
+
|
53 |
explainer = shap.Explainer(
|
54 |
+
predict_proba_shap,
|
55 |
masker=masker,
|
56 |
output_names=class_labels
|
57 |
)
|
58 |
|
59 |
+
# 7) Load biomedical NER model & pipeline
|
60 |
+
ner_model_name = "d4data/biomedical-ner-all"
|
61 |
+
ner_tokenizer = AutoTokenizer.from_pretrained(ner_model_name)
|
62 |
+
ner_model = AutoModelForTokenClassification.from_pretrained(ner_model_name).to(device)
|
63 |
ner_pipe = pipeline(
|
64 |
"ner",
|
65 |
model=ner_model,
|
66 |
tokenizer=ner_tokenizer,
|
67 |
aggregation_strategy="simple",
|
68 |
+
device=0 if device.type == "cuda" else -1
|
69 |
)
|
70 |
|
71 |
+
# 8) Mapping for entity highlight colors
|
72 |
+
ENTITY_COLORS = {
|
73 |
+
"Severity": "red",
|
74 |
+
"Sign_symptom": "green",
|
75 |
+
"Medication": "lightblue",
|
76 |
+
"Age": "yellow",
|
77 |
+
"Sex": "yellow",
|
78 |
+
"Diagnostic_procedure": "gray",
|
79 |
+
"Biological_structure": "silver"
|
80 |
+
}
|
81 |
+
|
82 |
+
# 9) Full predict + explain + NER function
|
83 |
+
def adr_predict(text: str):
|
84 |
# a) Predict probabilities
|
85 |
+
probs = predict_proba([text])[0]
|
86 |
prob_dict = {label: float(probs[i]) for i, label in enumerate(class_labels)}
|
87 |
|
88 |
+
# b) SHAP explanation → Matplotlib figure
|
89 |
shap_values = explainer([text])
|
90 |
fig = shap.plots.text(shap_values[0], display=False)
|
91 |
|
92 |
# c) NER highlighting
|
93 |
+
ents = ner_pipe(text)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
94 |
highlighted = ""
|
95 |
last_idx = 0
|
96 |
+
for ent in ents:
|
97 |
start, end = ent["start"], ent["end"]
|
98 |
word = ent["word"].replace("##", "")
|
99 |
+
color = ENTITY_COLORS.get(ent["entity_group"], "lightgray")
|
100 |
highlighted += (
|
101 |
text[last_idx:start]
|
102 |
+ f"<mark style='background-color:{color};'>{word}</mark>"
|
|
|
106 |
|
107 |
return prob_dict, fig, highlighted
|
108 |
|
109 |
+
# 10) Build Gradio UI
|
110 |
with gr.Blocks() as demo:
|
111 |
gr.Markdown("## Welcome to **ADR Detector** 🪐")
|
112 |
gr.Markdown(
|
113 |
+
"Predicts the likelihood your text describes a **severe** vs. **non-severe** adverse reaction. \n"
|
114 |
+
"_(Not for medical or diagnostic use.)_"
|
115 |
)
|
116 |
|
117 |
+
txt = gr.Textbox(
|
118 |
+
label="Enter Your Text Here:",
|
119 |
+
lines=3,
|
120 |
+
placeholder="Type a sentence about an adverse reaction…"
|
121 |
+
)
|
122 |
btn = gr.Button("Analyze")
|
123 |
|
124 |
with gr.Row():
|
125 |
+
label_out = gr.Label(label="Predicted Probabilities")
|
126 |
+
shap_out = gr.Plot(label="SHAP Explanation")
|
127 |
+
ner_out = gr.HTML(label="Biomedical Entities Highlighted")
|
128 |
|
129 |
+
btn.click(
|
130 |
+
fn=adr_predict,
|
131 |
+
inputs=txt,
|
132 |
+
outputs=[label_out, shap_out, ner_out]
|
133 |
+
)
|
134 |
|
135 |
gr.Examples(
|
136 |
examples=[
|
|
|
138 |
"A 35-year-old female had minor abdominal pain after Acetaminophen."
|
139 |
],
|
140 |
inputs=txt,
|
141 |
+
outputs=[label_out, shap_out, ner_out],
|
142 |
fn=adr_predict,
|
143 |
cache_examples=True
|
144 |
)
|
145 |
|
146 |
+
if __name__ == "__main__":
|
147 |
+
demo.launch()
|