Preparing submission

Dockerfile

@@ -11,8 +11,8 @@ WORKDIR /app
 
 COPY --chown=user ./requirements.txt requirements.txt
 
-# Needed for dependency errors of opencv
-RUN pip install ultralytics
+# Needed here instead of requirements.txt, because of dependency errors of opencv
+RUN pip install ultralytics==8.3.69
 RUN pip install opencv-python-headless
 
 RUN pip install --no-cache-dir --upgrade -r requirements.txt

README.md

@@ -8,45 +8,39 @@ pinned: false
 ---
 
 
-# Random Baseline Model for Climate Disinformation Classification
+# Object Detector for forest fire smoke
 
 ## Model Description
 
-This is a random baseline model for the Frugal AI Challenge 2024, specifically for the text classification task of identifying climate disinformation. The model serves as a performance floor, randomly assigning labels to text inputs without any learning.
+This is a frugal object detector use to detect fire smoke, as part of the Frugal AI Challenge 2024. It is based of the yolo model series
 
 ### Intended Use
 
-- **Primary intended uses**: Baseline comparison for climate disinformation classification models
+- **Primary intended uses**: Detect fire smoke on photos of forests, in different natural settings
 - **Primary intended users**: Researchers and developers participating in the Frugal AI Challenge
-- **Out-of-scope use cases**: Not intended for production use or real-world classification tasks
 
 ## Training Data
 
-The model uses the QuotaClimat/frugalaichallenge-text-train dataset:
-- Size: ~6000 examples
-- Split: 80% train, 20% test
-- 8 categories of climate disinformation claims
+The model uses the pyronear/pyro-sdis dataset:
+- Size: ~33 600 examples
+- Split: 88% train, 12% test
+- Images with smoke or no smoke
 
 ### Labels
-0. No relevant claim detected
-1. Global warming is not happening
-2. Not caused by humans
-3. Not bad or beneficial
-4. Solutions harmful/unnecessary
-5. Science is unreliable
-6. Proponents are biased
-7. Fossil fuels are needed
+Smoke
 
 ## Performance
 
 ### Metrics
-- **Accuracy**: ~12.5% (random chance with 8 classes)
+- **Accuracy**: ~ 90% 
 - **Environmental Impact**:
   - Emissions tracked in gCO2eq
   - Energy consumption tracked in Wh
 
 ### Model Architecture
-The model implements a random choice between the 8 possible labels, serving as the simplest possible baseline.
+Based of YOLOv11, see https://arxiv.org/abs/2410.17725, fine tuned on the pyronear dataset. The network is pruned and quantized to be as compressed as possible. 
+
+Inference should ideally performed on GPU - the speed bump is drastic, it is more energy efficient than CPU inference which takes much longer.
 
 ## Environmental Impact
 
@@ -57,15 +51,10 @@ Environmental impact is tracked using CodeCarbon, measuring:
 This tracking helps establish a baseline for the environmental impact of model deployment and inference.
 
 ## Limitations
-- Makes completely random predictions
-- No learning or pattern recognition
-- No consideration of input text
-- Serves only as a baseline reference
-- Not suitable for any real-world applications
+- Quantization was performed to FP16 - INT8 could compress even more but the accuracy drop was too big. Finding a way to smartly quantize and calibrate to INT8 could be interesting
+- To maximize inference speed even more, the model can be converted to TensorRT - it is note done in this repository, as the same type of GPU needs to be used both for exporting to TensorRT and inferencing with TensorRT
 
 ## Ethical Considerations
 
-- Dataset contains sensitive topics related to climate disinformation
-- Model makes random predictions and should not be used for actual classification
 - Environmental impact is tracked to promote awareness of AI's carbon footprint
 ```

requirements.txt

@@ -12,3 +12,6 @@ requests>=2.31.0
 librosa==0.10.2.post1
 torch==2.5.1
 torchvision==0.20.1
+onnx==1.17.0
+onnxslim==0.1.48
+onnxruntime==1.20.1

tasks/image.py

@@ -97,16 +97,31 @@ async def evaluate_image(request: ImageEvaluationRequest):
     # YOUR MODEL INFERENCE CODE HERE
     # Update the code below to replace the random baseline with your model inference
     #--------------------------------------------------------------------------------------------  
+    # Import strict minimum
     from pathlib import Path
     from ultralytics import YOLO
-    import torch
+    from torch import device
+    from torch.cuda import is_available
+
+    THRESHOLD = 0.18
 
     # Load model
     model_path = Path("tasks", "models")
-    # model_name = "best.pt" # nano 20e, unpruned
-    model_name = "pruned.pt" # nano 20e, 20% pruned
-    model = YOLO(Path(model_path, model_name))
-    threshold = 0.14
+    # If CUDA is available, load FP16 pytorch
+    # if is_available():
+    # print("CUDA available, loading FP16 pytorch model")
+    model_name = "pruned_fp16.pt"
+    model = YOLO(Path(model_path, model_name), task="detect")
+    # device = device("cuda")
+    device_name = device("cuda" if is_available() else "cpu")
+    IMGSIZE = 1280
+    # # If not, load FP16 ONNX model
+    # else:
+    #     print("CUDA not, available, loading ONNX model")
+    #     model_name = "640_fp16_cpu.onnx"
+    #     model = YOLO(Path(model_path, model_name), task="detect")
+    #     device = device("cpu")
+    #     IMGSIZE = 640 # required to make CPU inference a bit fast 
 
     predictions = []
     true_labels = []
@@ -120,13 +135,8 @@ async def evaluate_image(request: ImageEvaluationRequest):
         true_labels.append(int(has_smoke))
 
         # Make prediction
-        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-        results = model.predict(example["image"], device=device, conf=threshold, verbose=False)[0] # index 0 since we predict on one image at a time
-        if results.boxes.cls.numel()!=0:
-            # This means a fire was detected, hence we append 1
-            pred_has_smoke = 1
-        else:
-            pred_has_smoke = 0
+        results = model.predict(example["image"], device=device_name, conf=THRESHOLD, verbose=False, imgsz=IMGSIZE)[0]
+        pred_has_smoke = len(results) > 0
         predictions.append(int(pred_has_smoke))
         
         # If there's a true box, parse it and add box prediction

tasks/models/pruned.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:0e5e9ef2d0bbe8e8984d6739ccc2d21045844c2be98425b271090de621042ce8
-size 5470665

tasks/models/{best.pt → pruned_fp16.pt}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:08ca51a239f739eab4f3653956abcf303f836e8ea3b9a1c225c85f0cc1d086fa
-size 5443539
+oid sha256:b39f8abf26409f62ce689af44095cfd8debd183eae3a83b18729d6e826fce51a
+size 5558000