Add application file

app.py

@@ -1,19 +1,18 @@
-from fastapi import FastAPI, Query
-from transformers import pipeline, AutoTokenizer
 import onnxruntime as ort
+from transformers import AutoTokenizer
 import numpy as np
 import time
+from fastapi import FastAPI
 
-app = FastAPI()
-
-# Initialize the Hugging Face pipeline for sentiment analysis
-pipe = pipeline("text-classification", model="cardiffnlp/twitter-roberta-base-sentiment")
-
-# Load the ONNX model and tokenizer
+# Load the ONNX model
 onnx_model_path = "sentiment_model.onnx"
 session = ort.InferenceSession(onnx_model_path)
+
+# Load the tokenizer
 tokenizer = AutoTokenizer.from_pretrained("cardiffnlp/twitter-roberta-base-sentiment")
 
+app = FastAPI()
+
 def preprocess_tweet(tweet: str) -> str:
     tweet_words = []
     for word in tweet.split(' '):
@@ -24,7 +23,7 @@ def preprocess_tweet(tweet: str) -> str:
         tweet_words.append(word)
     return " ".join(tweet_words)
 
-def run_inference_onnx(tweet: str):
+def run_inference(tweet: str):
     inputs = tokenizer(tweet, return_tensors="np", padding=True, truncation=True)
     ort_inputs = {k: v for k, v in inputs.items()}
     ort_outs = session.run(None, ort_inputs)
@@ -35,42 +34,28 @@ def home():
     return {"message": "Welcome to the sentiment analysis API"}
 
 @app.get("/analyze")
-def analyze_sentiment(tweet: str, method: str = Query("pipeline", enum=["pipeline", "onnx"])):
+def analyze_sentiment(tweet: str):
     # Preprocess the tweet
     tweet_proc = preprocess_tweet(tweet)
 
     # Measure the time taken for the inference
     start_time = time.time()
     
-    if method == "pipeline":
-        # Use the Hugging Face pipeline to get the sentiment analysis result
-        results = pipe(tweet_proc, return_all_scores=True)
-
-        # Find the label with the highest score
-        highest_score_result = max(results[0], key=lambda x: x['score'])
-        label_map = {
-            "LABEL_0": "Negative",
-            "LABEL_1": "Neutral",
-            "LABEL_2": "Positive"
-        }
-        highest_label = label_map[highest_score_result['label']]
-        highest_score = round(highest_score_result['score'], 4)
-
-    elif method == "onnx":
-        # Run inference using the ONNX model
-        logits = run_inference_onnx(tweet_proc)
-        label_map = ["Negative", "Neutral", "Positive"]
-        highest_label_idx = np.argmax(logits)
-        highest_label = label_map[highest_label_idx]
-        highest_score = round(float(np.max(logits)), 4)
+    # Run inference using ONNX model
+    logits = run_inference(tweet_proc)
 
     # Calculate the inference time
     inference_time = time.time() - start_time
 
+    # Find the label with the highest score
+    label_map = ["Negative", "Neutral", "Positive"]
+    highest_label_idx = np.argmax(logits)
+    highest_score = np.max(logits)
+
     # Return the original tweet, the label with the highest score, and the inference time
     return {
         "text": tweet,
-        "label": highest_label,
-        "score": highest_score,
+        "label": label_map[highest_label_idx],
+        "score": round(float(highest_score), 4),
         "inference_time": round(inference_time, 4)  # In seconds
     }

export_to_onnx.py

@@ -0,0 +1,18 @@
+from transformers import AutoModelForSequenceClassification, AutoTokenizer
+import torch
+from transformers.onnx import export
+from transformers.onnx.features import FeaturesManager
+
+# Load the model and tokenizer
+model_name = "cardiffnlp/twitter-roberta-base-sentiment"
+model = AutoModelForSequenceClassification.from_pretrained(model_name)
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+
+# Determine the correct export feature
+feature = FeaturesManager.get_supported_features_for_model_type("text-classification")
+
+# Define the path where the ONNX model will be saved
+onnx_model_path = "https://huggingface.co/spaces/Abineshkumar/demodeploy/sentiment_model.onnx"
+
+# Export the model to ONNX
+export(tokenizer, model, feature, onnx_model_path, opset=11, framework="pt")

requirements.txt

@@ -5,5 +5,8 @@ torch
 scipy
 onnx
 onnxruntime
+onnxruntime-gpu
+numpy
+time