Merge pull request #6 from atrifat/feat-support-custom-and-hybrid-model

Support custom and hybrid model (Fast Inference for CPU-only device)

Author: atrifat

.env.example

@@ -13,4 +13,19 @@ ENABLE_CACHE=false
 CACHE_DURATION_SECONDS=60
 
 # (Optional. Default: auto. Options: auto,cpu,cuda) Set torch default device for detoxify library. Automatically detect if cuda/gpu device is available
-TORCH_DEVICE=auto
+TORCH_DEVICE=auto
+
+# (Required. Default: detoxify. Options: detoxify, hybrid, custom). Set classification model to be used in prediction. "hybrid" and "custom" can be used as fast model if application run on machine without GPU.
+HATE_SPEECH_MODEL=detoxify
+
+# (Optional. Default: "". Options: "any potential words") Some potential toxic words can be included to assist hybrid model detection. Hybrid approach uses both custom and detoxify model based on probability thresold. 
+# POTENTIAL_TOXIC_WORDS="f**k,n***a,ni**er"
+
+# (Optional. Default: 0.5 . Options: float value between 0.0 and 1.0) Probability thresold when using "hybrid" model. The thresold will determine whether to continue classify using detoxify model after using custom model
+HYBRID_THRESOLD_CHECK=0.5
+
+# (Optional. Default: "./experiments/model_voting_partial_best.pkl") Custom model path. Custom Pretrained model (pickle) of scikit-learn which implement predict_proba
+# CUSTOM_MODEL_PATH="./experiments/model_voting_partial_best.pkl"
+
+# (Optional. Default: "./experiments/vectorizer_count_no_stop_words.pkl") Custom vectorizer path. Custom Vectorizer model (pickle) of scikit-learn which implement vector transform for text
+# CUSTOM_VECTORIZER_PATH="./experiments/vectorizer_count_no_stop_words.pkl"

Dockerfile

@@ -26,7 +26,10 @@ ENV HOME=/home/user
 
 COPY --from=builder --chown=user:user /builder/venv /app/venv
 
-COPY --chown=user:user app.py app.py 
+COPY --chown=user:user app.py app.py
+
+RUN mkdir -p /app/experiments
+COPY --chown=user:user experiments/*.pkl experiments/ 
 
 RUN chown -R user:user /app && chown -R user:user /home/user
 

README.md

@@ -1,10 +1,12 @@
 # hate-speech-detector-api
 
-A Simple PoC (Proof of Concept) of Hate-speech (Toxic content) Detector API Server using model from [detoxify](https://github.com/unitaryai/detoxify). Detoxify (unbiased model) achieves score of 93.74% compared to top leaderboard score with 94.73% in [Jigsaw Unintended Bias in Toxicity Classification](https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification).
+A Simple PoC (Proof of Concept) of Hate-speech (Toxic content) Detector API Server using model from [detoxify](https://github.com/unitaryai/detoxify) and/or [custom traditional machine learning](experiments/model_voting_partial_best.pkl) model. Detoxify (unbiased model) achieves AUC score of 93.74% compared to top leaderboard score with AUC 94.73% in [Jigsaw Unintended Bias in Toxicity Classification](https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification). To those who are interested in training custom machine learning model based on [Jigsaw Unintended Bias in Toxicity Classification](https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification) can take a look at our [Jupyter Notebook](experiments/hate-speech-classification.ipynb).
+
+hate-speech-detector-api is a core dependency of [nostr-filter-relay](https://github.com/atrifat/nostr-filter-relay).
 
 ## Demo
 
-A demo instance is available on [HuggingFace Spaces - https://atrifat-hate-speech-detector-api-demo.hf.space](https://atrifat-hate-speech-detector-api-demo.hf.space). There is no guarantee for the uptime, but feel free to test.
+A demo gradio showcase is available on [HuggingFace Spaces - https://huggingface.co/spaces/rifatramadhani/hate-speech-detector](https://huggingface.co/spaces/rifatramadhani/hate-speech-detector). There is no guarantee for the uptime, but feel free to test.
 
 ## Requirements
 

app.py

@@ -4,10 +4,11 @@
 from flask import Flask, request, jsonify
 import functools
 import datetime
-import torch
 from detoxify import Detoxify
 import logging
+import torch
 from flask_caching import Cache
+import pickle
 
 load_dotenv()
 
@@ -16,11 +17,27 @@
 APP_ENV = os.getenv("APP_ENV", "production")
 LISTEN_HOST = os.getenv("LISTEN_HOST", "0.0.0.0")
 LISTEN_PORT = os.getenv("LISTEN_PORT", "7860")
+
+CUSTOM_MODEL_PATH = os.getenv(
+    "CUSTOM_MODEL_PATH",
+    os.path.dirname(os.path.abspath(__file__))
+    + "/experiments/model_voting_partial_best.pkl",
+)
+CUSTOM_VECTORIZER_PATH = os.getenv(
+    "CUSTOM_VECTORIZER_PATH",
+    os.path.dirname(os.path.abspath(__file__))
+    + "/experiments/vectorizer_count_no_stop_words.pkl",
+)
+
 DETOXIFY_MODEL = os.getenv("DETOXIFY_MODEL", "unbiased-small")
+HATE_SPEECH_MODEL = os.getenv("HATE_SPEECH_MODEL", "detoxify")
 CACHE_DURATION_SECONDS = int(os.getenv("CACHE_DURATION_SECONDS", 60))
 ENABLE_CACHE = os.getenv("ENABLE_CACHE", "false") == "true"
+POTENTIAL_TOXIC_WORDS = list(
+    filter(None, os.getenv("POTENTIAL_TOXIC_WORDS", "").split(","))
+)
+HYBRID_THRESOLD_CHECK = float(os.getenv("HYBRID_THRESOLD_CHECK", 0.5))
 TORCH_DEVICE = os.getenv("TORCH_DEVICE", "auto")
-
 APP_VERSION = "0.2.0"
 
 # Setup logging configuration
@@ -47,7 +64,22 @@
 else:
     torch_device = TORCH_DEVICE
 
-model = Detoxify(DETOXIFY_MODEL, device=torch_device)
+if HATE_SPEECH_MODEL in ["hybrid", "detoxify"]:
+    model = Detoxify(DETOXIFY_MODEL, device=torch_device)
+
+if HATE_SPEECH_MODEL in ["hybrid", "custom"]:
+    try:
+        with open(CUSTOM_VECTORIZER_PATH, "rb") as f:
+            vectorizer = pickle.load(f)
+    except Exception as e:
+        vectorizer = None
+
+    try:
+        with open(CUSTOM_MODEL_PATH, "rb") as f:
+            model_custom = pickle.load(f)
+    except Exception as e:
+        raise e
+
 
 app = Flask(__name__)
 
@@ -107,6 +139,55 @@ def perform_hate_speech_analysis(query):
     return result
 
 
+def perform_hate_speech_analysis_custom(query):
+    query_vector = vectorizer.transform([query]) if vectorizer != None else ["query"]
+
+    result = {
+        "identity_attack": 0.0,
+        "insult": 0.0,
+        "obscene": 0.0,
+        "severe_toxicity": 0.0,
+        "sexual_explicit": 0.0,
+        "threat": 0.0,
+        "toxicity": 0.0,
+    }
+
+    temp_result = model_custom.predict_proba(query_vector)
+    result["toxicity"] = temp_result[0][1].round(3).astype("float")
+
+    return result
+
+
+def perform_hate_speech_analysis_hybrid(
+    query, thresold_check=0.5, potential_toxic_words=[]
+):
+    result = {
+        "identity_attack": 0.0,
+        "insult": 0.0,
+        "obscene": 0.0,
+        "severe_toxicity": 0.0,
+        "sexual_explicit": 0.0,
+        "threat": 0.0,
+        "toxicity": 0.0,
+    }
+    temp_result_custom = perform_hate_speech_analysis_custom(query)
+
+    has_potential_toxic_word = False
+    for word in potential_toxic_words:
+        if word in query:
+            has_potential_toxic_word = True
+            break
+
+    if temp_result_custom["toxicity"] > thresold_check or has_potential_toxic_word:
+        temp_result_detoxify = perform_hate_speech_analysis(query)
+        if temp_result_detoxify["toxicity"] > thresold_check:
+            result = temp_result_detoxify
+    else:
+        result = temp_result_custom
+
+    return result
+
+
 @app.errorhandler(Exception)
 def handle_exception(error):
     res = {"error": str(error)}
@@ -120,7 +201,16 @@ def predict():
     data = request.json
     q = data["q"]
     start_time = datetime.datetime.now()
-    result = perform_hate_speech_analysis(q)
+
+    if HATE_SPEECH_MODEL == "custom":
+        result = perform_hate_speech_analysis_custom(q)
+    elif HATE_SPEECH_MODEL == "hybrid":
+        result = perform_hate_speech_analysis_hybrid(
+            q, HYBRID_THRESOLD_CHECK, POTENTIAL_TOXIC_WORDS
+        )
+    else:
+        result = perform_hate_speech_analysis(q)
+
     end_time = datetime.datetime.now()
     elapsed_time = end_time - start_time
     logging.debug("elapsed predict time: %s", str(elapsed_time))

requirements.txt

@@ -2,5 +2,6 @@ detoxify==0.5.1
 Flask==3.0.0
 Flask-Caching==2.3.0
 gunicorn==21.2.0
-pandas==2.1.1
-python-dotenv==1.0.0
+pandas==2.2.2
+python-dotenv==1.0.0
+scikit-learn==1.5.0