Huggingface Flask FAQ & Answers

Install HuggingFace Transformers using pip in a virtual environment. First create and activate a virtual environment, then install with PyTorch support using: pip install transformers[torch]. For the basic installation: pip install transformers. To install from source for the latest features: git clone https://github.com/huggingface/transformers.git && cd transformers && pip install .[torch]. Verify installation with: python -c "from transformers import pipeline; print(pipeline('sentiment-analysis')('test'))"

CRITICAL: Use --extra-index-url (not --index-url) to keep PyPI access for other packages. Correct command: pip install torch transformers flask --extra-index-url https://download.pytorch.org/whl/cpu. This installs CPU-only PyTorch (~185MB vs ~2GB CUDA) while still accessing transformers from PyPI. Alternative two-step method: pip install torch --index-url https://download.pytorch.org/whl/cpu && pip install transformers flask. The --index-url REPLACES PyPI (breaking transformers install), while --extra-index-url ADDS to PyPI.

Use from_pretrained() to download a model, then save_pretrained() to save it locally. Example: from transformers import AutoModelForSequenceClassification, AutoTokenizer; model = AutoModelForSequenceClassification.from_pretrained('distilbert-base-uncased-finetuned-sst-2-english'); tokenizer = AutoTokenizer.from_pretrained('distilbert-base-uncased-finetuned-sst-2-english'); model.save_pretrained('/path/to/local/directory'); tokenizer.save_pretrained('/path/to/local/directory'). This saves the model weights and config.json to the specified directory.

Use from_pretrained() with the local directory path. Example: from transformers import AutoModelForSequenceClassification, AutoTokenizer; model = AutoModelForSequenceClassification.from_pretrained('/path/to/local/directory'); tokenizer = AutoTokenizer.from_pretrained('/path/to/local/directory'). The directory must contain the model weights and config.json file saved by save_pretrained(). For offline use, add local_files_only=True parameter.

Use Flask's request.json or request.get_json() to access JSON data from POST requests. Example: from flask import Flask, request, jsonify; app = Flask(name); @app.route('/endpoint', methods=['POST']); def handle_post(): data = request.json; text = data.get('text'); return jsonify({'result': text}). The client must send Content-Type: application/json header. Use request.is_json to check if the request contains JSON. Return responses using jsonify() for proper JSON formatting.

Use nohup to run Flask in the background. IMPORTANT: After starting, ALWAYS sleep to let server initialize before testing. Full sequence: nohup python /app/app.py > /app/app.log 2>&1 & followed by sleep 5 (or longer if loading ML models). The sleep is CRITICAL because Flask needs time to start, especially when loading large models. Without sleep, subsequent requests will get 'Connection refused'. Check server is running with: curl http://localhost:5000/health or check app.log for errors.

For sentiment analysis, use AutoModelForSequenceClassification with a pre-trained sentiment model. Example: from transformers import AutoTokenizer, AutoModelForSequenceClassification; import torch; tokenizer = AutoTokenizer.from_pretrained('distilbert-base-uncased-finetuned-sst-2-english'); model = AutoModelForSequenceClassification.from_pretrained('distilbert-base-uncased-finetuned-sst-2-english'); inputs = tokenizer(text, return_tensors='pt', padding=True, truncation=True); with torch.no_grad(): outputs = model(**inputs); probs = torch.nn.functional.softmax(outputs.logits, dim=-1). The model returns logits which you convert to probabilities using softmax. Index 0 is negative, index 1 is positive for SST-2 models.

Use app.run() with host and port parameters. Example: if name == 'main': app.run(host='0.0.0.0', port=5000). Setting host='0.0.0.0' makes the server accessible from any IP address on the network, not just localhost. Default is host='127.0.0.1' (localhost only) and port=5000. For CLI: flask run --host=0.0.0.0 --port=5000. This is for development only - use Gunicorn or uWSGI in production.

Use the requests library with the json parameter which automatically sets Content-Type to application/json. Example: import requests; url = 'http://localhost:5000/sentiment'; response = requests.post(url, json={'text': 'sample text'}); print(f'Status: {response.status_code}'); print(f'Response: {response.json()}'). For multiple test cases: test_inputs = ['I love this!', 'This is terrible.', 'It was okay.']; for text in test_inputs: resp = requests.post(url, json={'text': text}); print(f'Input: {text}'); print(f'Result: {resp.json()}\n'). Install with: pip install requests. The json= parameter handles serialization and headers automatically. Use response.status_code (200=success), response.json() to parse response, response.raise_for_status() to raise exception on HTTP errors.