Enhance model loading and caching: add support for .pkl files, set cache directories, and improve error handling

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+models/*.pkl filter=lfs diff=lfs merge=lfs -text

Dockerfile

@@ -4,6 +4,14 @@ FROM python:3.9-slim
 # Set the working directory in the container
 WORKDIR /app
 
+# Create cache directories with proper permissions
+RUN mkdir -p /tmp/transformers_cache /tmp/hf_home && \
+    chmod 777 /tmp/transformers_cache /tmp/hf_home
+
+# Set environment variables for transformers cache
+ENV TRANSFORMERS_CACHE=/tmp/transformers_cache
+ENV HF_HOME=/tmp/hf_home
+
 # Copy the requirements file and install dependencies first to leverage caching
 COPY requirements.txt .
 RUN pip install --no-cache-dir -r requirements.txt
@@ -16,5 +24,5 @@ COPY . .
 EXPOSE 7860
 
 # --- CORRECTED COMMAND ---
-# This now correctly points to the 'app' variable inside the 'api/index.py' file
+# This correctly points to the 'app' variable inside the 'app/main.py' file
 CMD ["uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "7860"]

api/index.py

@@ -0,0 +1,4 @@
+from app.main import app
+
+# This file is required for Vercel deployment
+# It imports the FastAPI app from app.main

app/core/models.py

@@ -4,20 +4,30 @@ import os
 
 MODEL_DIR = "models"
 
+# Initialize variables
+tfidf_vectorizer = None
+le = None
+rf_model = None
+sentence_model = None
+
 try:
     tfidf_vectorizer = joblib.load(os.path.join(MODEL_DIR, "tfidf_vectorizer.pkl"))
     le = joblib.load(os.path.join(MODEL_DIR, "label_encoder.pkl"))
     rf_model = joblib.load(os.path.join(MODEL_DIR, "random_forest_model.pkl"))
     print("Classification models loaded.")
 
-    sentence_model = SentenceTransformer('all-MiniLM-L6-v2')
+    # Set cache directory to a writable location for Hugging Face Spaces
+    os.environ['TRANSFORMERS_CACHE'] = '/tmp/transformers_cache'
+    os.environ['HF_HOME'] = '/tmp/hf_home'
+
+    sentence_model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
     print("SentenceTransformer model loaded.")
 
 except FileNotFoundError as e:
     print(f"MODEL LOADING ERROR: {e}")
     print("Make sure the .pkl files are in the 'models' directory.")
-    # In a real production app, you might want the app to exit or handle this more gracefully.
-    tfidf_vectorizer, le, rf_model, sentence_model = None, None, None, None
+    raise e  # Re-raise to prevent the application from starting with None models
 
 except Exception as e:
     print(f"An unexpected error occurred during model loading: {e}")
+    raise e  # Re-raise to prevent the application from starting with None models

app/main.py

@@ -1,14 +1,14 @@
-from fastapi import FastAPI
-from .api.routes import recommendations
-
-app = FastAPI(
-    title="Student Recommendation API",
-    description="An API that uses machine learning to predict job categories and recommend internships.",
-    version="1.0.0"
-)
-
-app.include_router(recommendations.router, prefix="/api/v1")
-
-@app.get("/", tags=["Root"])
-def read_root():
-    return {"message": "Welcome to the Student Recommendation API"}
+from fastapi import FastAPI
+from .api.routes import recommendations
+
+app = FastAPI(
+    title="Student Recommendation API",
+    description="An API that uses machine learning to predict job categories and recommend internships.",
+    version="1.0.0"
+)
+
+app.include_router(recommendations.router, prefix="/api/v1")
+
+@app.get("/", tags=["Root"])
+def read_root():
+    return {"message": "Welcome to the Student Recommendation API"}

app/services/ranking.py

@@ -55,6 +55,9 @@ def geo_coords(city_name: str) -> tuple | None:
 
 def get_category_prediction(profile_text: str) -> str:
     """Processes text and predicts the job category."""
+    if not all([tfidf_vectorizer, le, rf_model]):
+        raise RuntimeError("Classification models are not properly loaded")
+
     cleaned_text = profile_text.lower()
     vectorized_text = tfidf_vectorizer.transform([cleaned_text])
     prediction_encoded = rf_model.predict(vectorized_text)[0]
@@ -63,6 +66,8 @@ def get_category_prediction(profile_text: str) -> str:
 
 def  get_ranked_internships(request: RecommendationRequest) -> list[int]:
     """Performs two-stage ranking with dynamic geocoding."""
+    if not sentence_model:
+        raise RuntimeError("SentenceTransformer model is not properly loaded")
 
     profile_text_to_encode = request.profile_text
 

requirements.txt

@@ -2,8 +2,8 @@
 fastapi
 uvicorn
 
-# Machine Learning & Data
-scikit-learn
+# Machine Learning & Data - Pin scikit-learn version to match your models
+scikit-learn==1.5.1
 joblib
 sentence-transformers
 torch