atereoyinn
commited on
Commit
•
cd33ba8
1
Parent(s):
08af9e7
first commit
Browse files- app.py +208 -0
- requirements.txt +5 -0
app.py
ADDED
@@ -0,0 +1,208 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import gradio as gr
|
2 |
+
from faster_whisper import WhisperModel
|
3 |
+
from pydantic import BaseModel, Field, AliasChoices, field_validator, ValidationError
|
4 |
+
from typing import List
|
5 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
|
6 |
+
import csv
|
7 |
+
import json
|
8 |
+
import tempfile
|
9 |
+
import torch
|
10 |
+
|
11 |
+
|
12 |
+
# Initiate checkpoints for model loading
|
13 |
+
numind_checkpoint = "numind/NuExtract-tiny"
|
14 |
+
llama_checkpoint = "Atereoyin/Llama3_finetuned_for_medical_entity_extraction"
|
15 |
+
whisper_checkpoint = "large-v3"
|
16 |
+
|
17 |
+
quantization_config = BitsAndBytesConfig(
|
18 |
+
load_in_8bit=True,
|
19 |
+
)
|
20 |
+
|
21 |
+
# Load models with the correct device
|
22 |
+
whisper_model = WhisperModel(whisper_checkpoint, device="cuda")
|
23 |
+
numind_model = AutoModelForCausalLM.from_pretrained(numind_checkpoint, quantization_config=quantization_config, torch_dtype=torch.float16, trust_remote_code=True)
|
24 |
+
numind_tokenizer = AutoTokenizer.from_pretrained(numind_checkpoint)
|
25 |
+
llama_model = AutoModelForCausalLM.from_pretrained(llama_checkpoint, quantization_config=quantization_config, trust_remote_code=True)
|
26 |
+
llama_tokenizer = AutoTokenizer.from_pretrained(llama_checkpoint)
|
27 |
+
|
28 |
+
# Function to transcribe audio
|
29 |
+
def transcribe_audio(audio_file_path):
|
30 |
+
try:
|
31 |
+
segments, info = whisper_model.transcribe(audio_file_path, beam_size=5)
|
32 |
+
text = "".join([segment.text for segment in segments])
|
33 |
+
return text
|
34 |
+
except Exception as e:
|
35 |
+
return str(e)
|
36 |
+
|
37 |
+
# Functions for Person entity extraction
|
38 |
+
def predict_NuExtract(model, tokenizer, text, schema, example=["","",""]):
|
39 |
+
schema = json.dumps(json.loads(schema), indent=4)
|
40 |
+
input_llm = "<|input|>\n### Template:\n" + schema + "\n"
|
41 |
+
for i in example:
|
42 |
+
if i != "":
|
43 |
+
input_llm += "### Example:\n"+ json.dumps(json.loads(i), indent=4)+"\n"
|
44 |
+
|
45 |
+
input_llm += "### Text:\n"+text +"\n<|output|>\n"
|
46 |
+
input_ids = tokenizer(input_llm, return_tensors="pt", truncation=True, max_length=4000).to("cuda")
|
47 |
+
|
48 |
+
output = tokenizer.decode(model.generate(**input_ids)[0], skip_special_tokens=True)
|
49 |
+
return output.split("<|output|>")[1].split("<|end-output|>")[0]
|
50 |
+
|
51 |
+
|
52 |
+
#Function for generating promtps for Llama
|
53 |
+
def prompt_format(text):
|
54 |
+
prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
|
55 |
+
|
56 |
+
### Instruction:
|
57 |
+
{}
|
58 |
+
|
59 |
+
### Input:
|
60 |
+
{}
|
61 |
+
|
62 |
+
### Response:
|
63 |
+
{}"""
|
64 |
+
|
65 |
+
instruction = """Extract the following entities from the medical conversation:
|
66 |
+
* **Symptoms:** List all the symptoms the patient mentions.
|
67 |
+
* **Diagnosis:** List the doctor's diagnosis or potential diagnoses.
|
68 |
+
* **Medical History:** Summarize the patient's relevant medical history.
|
69 |
+
* **Action Plan:** List the recommended actions or treatment plan.
|
70 |
+
|
71 |
+
Provide the result in the following JSON format:
|
72 |
+
{
|
73 |
+
"Symptoms": [...],
|
74 |
+
"Diagnosis": [...],
|
75 |
+
"Medical history": [...],
|
76 |
+
"Action plan": [...]
|
77 |
+
}"""
|
78 |
+
full_prompt = prompt.format(instruction, text, "")
|
79 |
+
return full_prompt
|
80 |
+
|
81 |
+
|
82 |
+
#Pydantic Validator to validate Llama's response
|
83 |
+
def validate_medical_record(response):
|
84 |
+
|
85 |
+
class MedicalRecord(BaseModel):
|
86 |
+
Symptoms: List[str] = Field(default_factory=list)
|
87 |
+
Diagnosis: List[str] = Field(default_factory=list)
|
88 |
+
Medical_history: List[str] = Field(
|
89 |
+
default_factory=list,
|
90 |
+
validation_alias=AliasChoices('Medical history', 'History of Patient')
|
91 |
+
)
|
92 |
+
Action_plan: List[str] = Field(
|
93 |
+
default_factory=list,
|
94 |
+
validation_alias=AliasChoices('Action plan', 'Plan of Action')
|
95 |
+
)
|
96 |
+
|
97 |
+
@field_validator('*', mode='before')
|
98 |
+
def ensure_list(cls, v):
|
99 |
+
if isinstance(v, str):
|
100 |
+
return [item.strip() for item in v.split(',')]
|
101 |
+
return v
|
102 |
+
|
103 |
+
try:
|
104 |
+
validated_data = MedicalRecord(**response)
|
105 |
+
return validated_data.dict()
|
106 |
+
except ValidationError as e:
|
107 |
+
return response
|
108 |
+
|
109 |
+
|
110 |
+
|
111 |
+
# Function to predict medical entities using Llama
|
112 |
+
def predict_Llama(model, tokenizer, text):
|
113 |
+
inputs = tokenizer(prompt_format(text), return_tensors="pt", truncation=True).to("cuda")
|
114 |
+
|
115 |
+
try:
|
116 |
+
outputs = model.generate(**inputs, max_new_tokens=128, temperature=0.2, use_cache=True)
|
117 |
+
extracted_entities = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
118 |
+
|
119 |
+
response = extracted_entities.split("### Response:", 1)[-1].strip()
|
120 |
+
response_dict = {k.strip(): v.strip() for k, v in (line.split(': ', 1) for line in response.splitlines() if ': ' in line)}
|
121 |
+
|
122 |
+
validated_response = validate_medical_record(response_dict)
|
123 |
+
|
124 |
+
return validated_response
|
125 |
+
except Exception as e:
|
126 |
+
print(f"Error during Llama prediction: {str(e)}")
|
127 |
+
return {}
|
128 |
+
|
129 |
+
|
130 |
+
#Control function that cordinates communication of other functions to map entities to form fields
|
131 |
+
def process_audio(audio):
|
132 |
+
if isinstance(audio, str):
|
133 |
+
with open(audio, 'rb') as f:
|
134 |
+
audio_bytes = f.read()
|
135 |
+
else:
|
136 |
+
audio_bytes = audio
|
137 |
+
|
138 |
+
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as temp_audio:
|
139 |
+
temp_audio.write(audio_bytes)
|
140 |
+
temp_audio.flush()
|
141 |
+
audio_path = temp_audio.name
|
142 |
+
|
143 |
+
transcription = transcribe_audio(audio_path)
|
144 |
+
|
145 |
+
person_schema = """{"Name": "","Age": "","Gender": ""}"""
|
146 |
+
person_entities_raw = predict_NuExtract(numind_model, numind_tokenizer, transcription, person_schema)
|
147 |
+
|
148 |
+
try:
|
149 |
+
person_entities = json.loads(person_entities_raw)
|
150 |
+
except json.JSONDecodeError as e:
|
151 |
+
return f"Error in NuExtract response: {str(e)}"
|
152 |
+
|
153 |
+
medical_entities = predict_Llama(llama_model, llama_tokenizer, transcription)
|
154 |
+
|
155 |
+
return (
|
156 |
+
person_entities.get("Name", ""),
|
157 |
+
person_entities.get("Age", ""),
|
158 |
+
person_entities.get("Gender", ""),
|
159 |
+
", ".join(medical_entities.get("Symptoms", [])),
|
160 |
+
", ".join(medical_entities.get("Diagnosis", [])),
|
161 |
+
", ".join(medical_entities.get("Medical_history", [])),
|
162 |
+
", ".join(medical_entities.get("Action_plan", []))
|
163 |
+
)
|
164 |
+
|
165 |
+
|
166 |
+
|
167 |
+
#Function that allows users to download information
|
168 |
+
def download_csv(name, age, gender, symptoms, diagnosis, medical_history, action_plan):
|
169 |
+
csv_file = tempfile.NamedTemporaryFile(delete=False, suffix=".csv")
|
170 |
+
|
171 |
+
with open(csv_file.name, mode='w', newline='') as file:
|
172 |
+
writer = csv.writer(file)
|
173 |
+
writer.writerow(["Name", "Age", "Gender", "Symptoms", "Diagnosis", "Medical History", "Plan of Action"])
|
174 |
+
writer.writerow([name, age, gender, symptoms, diagnosis, medical_history, action_plan])
|
175 |
+
|
176 |
+
return csv_file.name
|
177 |
+
|
178 |
+
|
179 |
+
|
180 |
+
# Gradio interface to create a web-based form for users to input audio and fill the medical diagnostic form
|
181 |
+
demo = gr.Interface(
|
182 |
+
fn=process_audio,
|
183 |
+
inputs=[
|
184 |
+
gr.Audio(type="filepath")
|
185 |
+
],
|
186 |
+
outputs=[
|
187 |
+
gr.Textbox(label="Name"),
|
188 |
+
gr.Textbox(label="Age"),
|
189 |
+
gr.Textbox(label="Gender"),
|
190 |
+
gr.Textbox(label="Symptoms"),
|
191 |
+
gr.Textbox(label="Diagnosis"),
|
192 |
+
gr.Textbox(label="Medical History"),
|
193 |
+
gr.Textbox(label="Plan of Action"),
|
194 |
+
],
|
195 |
+
title="Medical Diagnostic Form Assistant",
|
196 |
+
description="Upload an audio file or record audio to generate a medical diagnostic form."
|
197 |
+
)
|
198 |
+
|
199 |
+
with demo:
|
200 |
+
download_button = gr.Button("Download CSV")
|
201 |
+
download_button.click(
|
202 |
+
fn=lambda name, age, gender, symptoms, diagnosis, medical_history, action_plan: download_csv(name, age, gender, symptoms, diagnosis, medical_history, action_plan),
|
203 |
+
inputs=demo.output_components,
|
204 |
+
outputs=gr.File(label="Download CSV")
|
205 |
+
)
|
206 |
+
|
207 |
+
demo.launch(share=True)
|
208 |
+
|
requirements.txt
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
gradio
|
2 |
+
transformers
|
3 |
+
torch
|
4 |
+
faster-whisper
|
5 |
+
pydantic
|