update space

README.md

@@ -1,13 +1,95 @@
 ---
 title: Frugal AI Agent
-emoji: 🌖
-colorFrom: pink
+emoji: 🌍
+colorFrom: green
 colorTo: blue
 sdk: gradio
 sdk_version: 5.33.0
 app_file: app.py
 pinned: false
 short_description: Estimate carbon footprint and get frugal AI code suggestions
+tags:
+- agent-demo-track
+- mcp-server-track
+thumbnail: >-
+  https://cdn-uploads.huggingface.co/production/uploads/675973b3138cfdc3f3f4f85b/7Iw5hINn5IJRWMOEo3yY_.jpeg
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Frugalize it
+
+This AI agent takes a Python code snippet, uses [CodeCarbon](https://mlco2.github.io/codecarbon/) to estimate its CO2 emissions, and then suggests modifications using frugal AI techniques such as pruning. It also provides general advice on how to make your code more energy-efficient.
+
+Learn more about the principles behind this project by visiting the [Frugal AI Challenge](https://frugalaichallenge.org/about/).
+
+**Special thanks to Anthropic for the free credits, which helped me run more tests and improve my agent!**
+
+## Getting Started
+
+1. **Add an Anthropic API key** in the repository's secrets variables to enable the agent.
+
+2. **Use the example prompts provided:**
+
+   Both examples use a context prompt beforehand to explain to the agent, Fruggy, what is expected from it. This context prompt is followed by a sample user query:
+   - The first example can be used directly: `"What are you capable of?"` - [Video demo agent-demo-track 1](https://youtu.be/3ybSoaxf_6g)
+     It runs a sample code (from the Frugal AI Challenge), estimates its energy consumption, and suggests optimizations using pruning and quantization.
+   - The second example: `"Here is my code, {code}, please give me frugal alternatives"` [Video demo agent-demo-track 2](https://youtu.be/8W7S4x2vcrE)
+     This cannot be used directly. Replace `{code}` with your own Python code. The agent will then return frugal alternatives and optimization suggestions.
+
+   You can also use the MCP client to call the agent: [Video demo mcp-server-track](https://youtu.be/yaRfHPyBdPA)
+
+> ⚠️ Note: This agent is a proof of concept. There are many improvements and features that could be added. Besides, the example codes are quite basic and simple, both to ensure low CPU usage for generating the results and because they run directly in the cloud rather than locally in these examples.
+
+## Contribute
+
+All types of contributions are, of course, welcome! Feel free to submit pull requests or contact me with any questions or feedback via the Hugging Face Discord (@cindydelage_51846).
+
+## What's next ?
+
+As mentioned, there's still a lot to do, but the first improvements could include:
+
+- Adding the context prompt used in the examples as a system prompt for the agent
+- Enabling the Gradio interface to import a code file
+
+## CodeCarbon citation:
+
+@software{benoit_courty_2024_11171501,
+  author       = {Benoit Courty and
+                  Victor Schmidt and
+                  Sasha Luccioni and
+                  Goyal-Kamal and
+                  MarionCoutarel and
+                  Boris Feld and
+                  Jérémy Lecourt and
+                  LiamConnell and
+                  Amine Saboni and
+                  Inimaz and
+                  supatomic and
+                  Mathilde Léval and
+                  Luis Blanche and
+                  Alexis Cruveiller and
+                  ouminasara and
+                  Franklin Zhao and
+                  Aditya Joshi and
+                  Alexis Bogroff and
+                  Hugues de Lavoreille and
+                  Niko Laskaris and
+                  Edoardo Abati and
+                  Douglas Blank and
+                  Ziyao Wang and
+                  Armin Catovic and
+                  Marc Alencon and
+                  Michał Stęchły and
+                  Christian Bauer and
+                  Lucas Otávio N. de Araújo and
+                  JPW and
+                  MinervaBooks},
+  title        = {mlco2/codecarbon: v2.4.1},
+  month        = may,
+  year         = 2024,
+  publisher    = {Zenodo},
+  version      = {v2.4.1},
+    doi          = {10.5281/zenodo.11171501},
+  url          = {https://doi.org/10.5281/zenodo.11171501}
+}
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -0,0 +1,74 @@
+import gradio as gr
+import random
+from smolagents import GradioUI, CodeAgent, InferenceClientModel
+from smolagents import LiteLLMModel
+from smolagents import ToolCallingAgent, PythonInterpreterTool, PromptTemplates
+import os
+from PIL import Image
+#from smolagents.src.smolagents.prompts 
+
+# Import our custom tools from their modules
+from tools import evaluate_consumption, evaluate_consumption_example
+from retriever import FrugalAI_methods
+
+#HF_TOKEN = os.environ.get("HF_TOKEN") 
+try:
+    API_KEY = os.environ.get("API_KEY") 
+except:
+    print("Please provide the Anthropic API_KEY in the secret variables")
+# Initialize the Hugging Face model
+model = LiteLLMModel(
+    model_id="anthropic/claude-3-5-sonnet-latest",
+    temperature=0.2,
+    api_key=API_KEY
+)
+
+evaluate_consumption = evaluate_consumption()
+evaluate_consumption_example = evaluate_consumption_example()
+FrugalAI_methods=FrugalAI_methods()
+
+# Create agent with all the tools
+fruggy = CodeAgent(
+    tools=[evaluate_consumption, evaluate_consumption_example, FrugalAI_methods], 
+    model=model,
+    additional_authorized_imports=['os'],
+    add_base_tools=True,  # Add any additional base tools
+    planning_interval=10   
+)
+
+with gr.Blocks(theme=gr.themes.Ocean()) as demo:
+    with gr.Row():
+        with gr.Column(scale=1):
+            image = gr.Image(value=Image.open("frugal.jpg"), label="Frugal Image")
+            hello = gr.Interface(
+                fn=fruggy,
+                inputs="text",
+                outputs="text",
+                title="Frugalize it!",
+                examples = [
+                "You are an AI agent that receives Python code from a manager. "
+                "Your task is to test its resource consumption using a custom tool. If no code is provided, use evaluate_consumption_example. If a code is provided by the manager, use evaluate_consumption."
+                "Then, propose frugal alternatives, such as pruning, using your dedicated tool — always use it. "
+                "Additionally, feel free to search the internet for frugal methods and present them to the manager. "
+                "Examples include Knowledge Distillation, Transfer Learning, and others. "
+                "Show me what you can do using your custom tool evaluate_consumption_example, and provide examples based on the code in that tool to optimize it for frugality. "
+                "Use both your tools and web searches if needed. If implementation guidance is given by your custom tools, always give it in your final answer to the manager.",
+            
+                "You are an AI agent that receives Python code from a manager. "
+                "Your task is to test its resource consumption using a custom tool. If no code is provided, use evaluate_consumption_example. If a code is provided by the manager, use evaluate_consumption."
+                "Then, propose frugal alternatives, such as pruning, using your dedicated tool — always use it. "
+                "Additionally, feel free to search the internet for frugal methods and present them to the manager. "
+                "Examples include Knowledge Distillation, Transfer Learning, and others. If implementation guidance is given by your custom tools, always give it in your final answer to the manager. "
+                "Here is my code: "
+                "import anthropic "
+                "API_KEY = os.environ.get('API_KEY') "
+                "client = anthropic.Anthropic(api_key=API_KEY) "
+                "message = client.messages.create(model='claude-sonnet-4-20250514',max_tokens=1024,messages=[{'role': 'user', 'content': 'Hello, Claude'}]) "
+                "print(message.content)     "
+                "Please give me frugal alternatives."
+            ],
+                description="Share your Python code with this AI agent! It will track its CO2 emissions using CodeCarbon and recommend greener, frugal AI alternatives."
+            )
+    
+if __name__ == "__main__":
+    demo.launch(mcp_server=True, share=True)    

requirements.txt

@@ -0,0 +1,18 @@
+transformers
+torch
+smolagents
+pillow
+fastapi
+datasets
+scikit-learn
+numpy
+pydantic
+psutil
+librosa
+soundfile
+codecarbon
+smolagents[litellm]
+smolagents[mcp]
+duckduckgo_search
+sentencepiece
+anthropic

retriever.py

@@ -0,0 +1,39 @@
+from smolagents import Tool
+
+class FrugalAI_methods(Tool):
+    name = "FrugalAI_methods"
+    description = "Retrieves methods for model frugalization. It will return ideas to frugalize a code, please use it."
+    inputs = {
+        "method": {
+            "type": "string",
+            "description": "Always set it to 'Frugal' or it will raise an error."
+        }
+    }
+    output_type = "string"
+    
+    def forward(self, method):
+        ideas=[]
+        ideas.append(pruning())
+        ideas.append(quantization())
+        return ideas
+        
+def pruning():
+    """
+    Optimizes models by removing unnecessary components, such as certain weights in a neural network. 
+    This function demonstrates how to apply pruning.
+    """
+    code = "model = apply_pruning(model, amount=0.3)"
+    return (
+            f"To apply pruning to a model, use the following code snippet: {code}. "
+            f"You should adapt it to your actual implementation. In particular, the 'amount' parameter "
+            f"can be increased or decreased depending on the initial number of weights and the complexity of your use case (minimu value: 0, maximum value: 1). You should give these information to the manager, you can not prune the model yourself."
+        )
+
+def quantization():
+    """
+    Converts high-precision weights into lower-precision one to reduce cost. 
+    """
+    code = "model = torch.quantization.quantize_dynamic(model, dtype=torch.qint8)"
+    return (
+            f"To apply quantization to a model, use the following code snippet: {code}. You should give these information to the manager, you can not quantize the model yourself."
+        )

tools.py

@@ -0,0 +1,143 @@
+from fastapi import APIRouter
+from datetime import datetime
+from datasets import load_dataset  
+from sklearn.metrics import accuracy_score  
+import numpy as np
+import os
+import torch
+import gc
+import psutil
+from transformers import Wav2Vec2ForSequenceClassification, AutoFeatureExtractor, pipeline
+from utils.evaluation import AudioEvaluationRequest
+from utils.emissions import tracker, clean_emissions_data, get_space_info
+from dotenv import load_dotenv
+import logging
+import csv
+import torch.nn.utils.prune as prune
+from typing import Optional
+from pydantic import BaseModel, Field
+from smolagents import Tool
+
+# Configurer le logging
+logging.basicConfig(level=logging.INFO)
+logging.info("Début du fichier python")
+load_dotenv()
+
+router = APIRouter()
+
+DESCRIPTION = "Random Baseline"
+ROUTE = "/audio"
+
+device = 0 if torch.cuda.is_available() else -1  
+
+def preprocess_function(example, feature_extractor):
+    return feature_extractor(
+        [x["array"] for x in example["audio"]], 
+        sampling_rate=feature_extractor.sampling_rate, padding="longest", max_length=16000, truncation=True, return_tensors="pt"
+    )
+
+def apply_pruning(model, amount=0.3):
+    for name, module in model.named_modules():
+        if isinstance(module, torch.nn.Linear):
+            prune.l1_unstructured(module, name="weight", amount=amount)
+            prune.remove(module, "weight")
+    return model
+
+class BaseEvaluationRequest(BaseModel):
+    test_size: float = Field(0.2, ge=0.0, le=1.0, description="Size of the test split (between 0 and 1)")
+    test_seed: int = Field(42, ge=0, description="Random seed for reproducibility")
+    
+class AudioEvaluationRequest(BaseEvaluationRequest):
+    dataset_name: str = Field("rfcx/frugalai", 
+    description="The name of the dataset on HuggingFace Hub") 
+    
+class evaluate_consumption_example(Tool):
+    name = "evaluate_consumption_example"
+    description = "This is only an example. If a manager wants to know what you are capable of, use it : it will use code carbon to evaluate the CO2 emissions from an example Python code"
+    inputs = {
+        "code": {
+            "type": "string",
+            "description": "The code to evaluate. Here, it is an example, so just set it to 'None'."
+        }
+    }
+    output_type = "string"
+    
+    def forward(self, code :  str):
+        request = AudioEvaluationRequest() 
+        logging.info("Chargement des données")
+        dataset = load_dataset(request.dataset_name, streaming=True, token=os.getenv("HF_TOKEN"))
+        logging.info("Données chargées")
+        
+        test_dataset = dataset["test"]
+        del dataset
+        
+        # Start tracking emissions
+        tracker.start()
+        tracker.start_task("inference")
+    
+        feature_extractor = AutoFeatureExtractor.from_pretrained("facebook/wav2vec2-base")
+    
+        test_dataset = test_dataset.map(preprocess_function, fn_kwargs={"feature_extractor": feature_extractor}, remove_columns="audio", batched=True, batch_size=32)  
+        
+        gc.collect()
+    
+        model_name = "CindyDelage/Challenge_HuggingFace_DFG_FrugalAI"
+        model = Wav2Vec2ForSequenceClassification.from_pretrained(model_name)
+        
+        # Appliquer la quantification dynamique et le pruning
+        model.eval()
+        #model = torch.quantization.quantize_dynamic(model, dtype=torch.qint8)
+        #model = apply_pruning(model, amount=0.3)  # Prune 30% des poids linéaires
+        
+        classifier = pipeline("audio-classification", model=model, feature_extractor=feature_extractor, device=device)
+        predictions = []  
+        logging.info("Début des prédictions par batch")
+        i=0
+        for data in iter(test_dataset):
+            print(i)
+            if (i<=5):
+                with torch.no_grad():
+                    result = classifier(np.asarray(data["input_values"]), batch_size=64)
+                predicted_label = result[0]['label']
+                label = 1 if predicted_label == 'environment' else 0
+                predictions.append(label)  
+                
+                # Nettoyer la mémoire après chaque itération
+                del result
+                del label 
+                torch.cuda.empty_cache()
+                gc.collect()
+                i=i+1
+            if(i>5):
+                break
+        logging.info("Fin des prédictions")
+        del classifier 
+        del feature_extractor 
+        
+        gc.collect()
+        # Stop tracking emissions
+        emissions_data = tracker.stop_task()
+        
+        return emissions_data
+
+class evaluate_consumption(Tool):
+    name = "evaluate_consumption"
+    description = "If the manager gave you its Python code, this function uses code carbon to evaluate the CO2 emissions from the given Python code"
+    inputs = {
+        "code": {
+            "type": "string",
+            "description": "The code to evaluate."
+        }
+    }
+    output_type = "string"
+    
+    def forward(self, code : str):
+        
+        # Start tracking emissions
+        tracker.start()
+        tracker.start_task("inference")
+        exec(code)
+        # Stop tracking emissions
+        emissions_data = tracker.stop_task()
+        
+        return emissions_data

utils/emissions.py

@@ -0,0 +1,28 @@
+from codecarbon import EmissionsTracker
+import os
+
+# Initialize tracker
+tracker = EmissionsTracker(allow_multiple_runs=True)
+
+class EmissionsData:
+    def __init__(self, energy_consumed: float, emissions: float):
+        self.energy_consumed = energy_consumed
+        self.emissions = emissions
+
+def clean_emissions_data(emissions_data):
+    """Remove unwanted fields from emissions data"""
+    data_dict = emissions_data.__dict__
+    fields_to_remove = ['timestamp', 'project_name', 'experiment_id', 'latitude', 'longitude']
+    return {k: v for k, v in data_dict.items() if k not in fields_to_remove}
+
+def get_space_info():
+    """Get the space username and URL from environment variables"""
+    space_name = os.getenv("SPACE_ID", "")
+    if space_name:
+        try:
+            username = space_name.split("/")[0]
+            space_url = f"https://huggingface.co/spaces/{space_name}"
+            return username, space_url
+        except Exception as e:
+            print(f"Error getting space info: {e}")
+    return "local-user", "local-development" 

utils/evaluation.py

@@ -0,0 +1,18 @@
+from typing import Optional
+from pydantic import BaseModel, Field
+
+class BaseEvaluationRequest(BaseModel):
+    test_size: float = Field(0.2, ge=0.0, le=1.0, description="Size of the test split (between 0 and 1)")
+    test_seed: int = Field(42, ge=0, description="Random seed for reproducibility")
+
+class TextEvaluationRequest(BaseEvaluationRequest):
+    dataset_name: str = Field("QuotaClimat/frugalaichallenge-text-train", 
+                            description="The name of the dataset on HuggingFace Hub")
+
+class ImageEvaluationRequest(BaseEvaluationRequest):
+    dataset_name: str = Field("pyronear/pyro-sdis", 
+                            description="The name of the dataset on HuggingFace Hub")
+
+class AudioEvaluationRequest(BaseEvaluationRequest):
+    dataset_name: str = Field("rfcx/frugalai", 
+                            description="The name of the dataset on HuggingFace Hub")