"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"labels_distr = df.split.value_counts()\n",
"\n",
"plt.pie(\n",
" labels_distr,\n",
" labels=labels_distr.index,\n",
" autopct=make_autopct(labels_distr),\n",
")\n",
"plt.show()"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# [Model Training](#toc0_)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## [Load configuration](#toc0_)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# [π§ MODEL CHOICE](#toc0_)\n",
"\n",
"- https://paperswithcode.com/sota/image-classification-on-imagenet\n",
"- https://keras.io/api/applications/efficientnet_v2/#efficientnetv2m-function\n",
"- [ViTForImageClassification](https://huggingface.co/docs/transformers/main/en/model_doc/vit)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# [π¦π¦ CHECKPOINT π¦π¦](#toc0_)"
]
},
{
"cell_type": "code",
"execution_count": 17,
"metadata": {},
"outputs": [
{
"ename": "NameError",
"evalue": "name 'stop' is not defined",
"output_type": "error",
"traceback": [
"\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
"\u001b[0;31mNameError\u001b[0m Traceback (most recent call last)",
"Cell \u001b[0;32mIn[17], line 1\u001b[0m\n\u001b[0;32m----> 1\u001b[0m \u001b[43mstop\u001b[49m\n",
"\u001b[0;31mNameError\u001b[0m: name 'stop' is not defined"
]
}
],
"source": [
"stop"
]
},
{
"cell_type": "code",
"execution_count": 18,
"metadata": {},
"outputs": [],
"source": [
"# Model config\n",
"model_name = \"EfficientNetV2M\"\n",
"input_size = (224, 224)\n",
"batch_size = 32\n",
"optimizer = AdamW(learning_rate=0.0002, weight_decay=0.05)\n",
"# model_name = \"EfficientNetB0\"\n",
"# input_size = (224, 224)\n",
"# batch_size = 48\n",
"# optimizer = AdamW(learning_rate=0.0002, weight_decay=0.05)\n",
"n_epochs = 100\n",
"loss = \"binary_crossentropy\"\n",
"metrics = [\"accuracy\", Precision(), Recall()]\n",
"# metrics = [\"accuracy\", Precision(), Recall(), \"f1_score\"]\n",
"\n",
"# Input paths\n",
"data_dir = os.path.join(OUTPUT_DIR, \"keras\")\n",
"train_dir = os.path.join(data_dir, \"train\")\n",
"val_dir = os.path.join(data_dir, \"val\")\n",
"test_dir = os.path.join(data_dir, \"test\")\n",
"# Output paths\n",
"MODEL_PATH = os.path.join(MODELS_ROOT_DIR, model_name)\n",
"LOG_PATH = os.path.join(MODEL_PATH, LOG_DIR)\n",
"CHKPT_PATH = os.path.join(MODEL_PATH, CHKPT_DIR)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## [Prepare model](#toc0_)"
]
},
{
"cell_type": "code",
"execution_count": 19,
"metadata": {},
"outputs": [
{
"data": {
"text/html": [
"Model: \"functional\"\n",
"\n"
],
"text/plain": [
"\u001b[1mModel: \"functional\"\u001b[0m\n"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"data": {
"text/html": [
"βββββββββββββββββββββββββββββββ³ββββββββββββββββββββββββ³βββββββββββββ³ββββββββ\n",
"β Layer (type) β Output Shape β Param # β Traiβ¦ β\n",
"β‘βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ©\n",
"β input_layer_1 (InputLayer) β (None, 224, 224, 3) β 0 β - β\n",
"βββββββββββββββββββββββββββββββΌββββββββββββββββββββββββΌβββββββββββββΌββββββββ€\n",
"β efficientnetv2-m β (None, 7, 7, 1280) β 53,150,388 β N β\n",
"β (Functional) β β β β\n",
"βββββββββββββββββββββββββββββββΌββββββββββββββββββββββββΌβββββββββββββΌββββββββ€\n",
"β flatten (Flatten) β (None, 62720) β 0 β - β\n",
"βββββββββββββββββββββββββββββββΌββββββββββββββββββββββββΌβββββββββββββΌββββββββ€\n",
"β dense (Dense) β (None, 1) β 62,721 β Y β\n",
"βββββββββββββββββββββββββββββββ΄ββββββββββββββββββββββββ΄βββββββββββββ΄ββββββββ\n",
"\n"
],
"text/plain": [
"βββββββββββββββββββββββββββββββ³ββββββββββββββββββββββββ³βββββββββββββ³ββββββββ\n",
"β\u001b[1m \u001b[0m\u001b[1mLayer (type) \u001b[0m\u001b[1m \u001b[0mβ\u001b[1m \u001b[0m\u001b[1mOutput Shape \u001b[0m\u001b[1m \u001b[0mβ\u001b[1m \u001b[0m\u001b[1m Param #\u001b[0m\u001b[1m \u001b[0mβ\u001b[1m \u001b[0m\u001b[1mTraiβ¦\u001b[0m\u001b[1m \u001b[0mβ\n",
"β‘βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ©\n",
"β input_layer_1 (\u001b[38;5;33mInputLayer\u001b[0m) β (\u001b[38;5;45mNone\u001b[0m, \u001b[38;5;34m224\u001b[0m, \u001b[38;5;34m224\u001b[0m, \u001b[38;5;34m3\u001b[0m) β \u001b[38;5;34m0\u001b[0m β \u001b[1m-\u001b[0m β\n",
"βββββββββββββββββββββββββββββββΌββββββββββββββββββββββββΌβββββββββββββΌββββββββ€\n",
"β efficientnetv2-m β (\u001b[38;5;45mNone\u001b[0m, \u001b[38;5;34m7\u001b[0m, \u001b[38;5;34m7\u001b[0m, \u001b[38;5;34m1280\u001b[0m) β \u001b[38;5;34m53,150,388\u001b[0m β \u001b[1;91mN\u001b[0m β\n",
"β (\u001b[38;5;33mFunctional\u001b[0m) β β β β\n",
"βββββββββββββββββββββββββββββββΌββββββββββββββββββββββββΌβββββββββββββΌββββββββ€\n",
"β flatten (\u001b[38;5;33mFlatten\u001b[0m) β (\u001b[38;5;45mNone\u001b[0m, \u001b[38;5;34m62720\u001b[0m) β \u001b[38;5;34m0\u001b[0m β \u001b[1m-\u001b[0m β\n",
"βββββββββββββββββββββββββββββββΌββββββββββββββββββββββββΌβββββββββββββΌββββββββ€\n",
"β dense (\u001b[38;5;33mDense\u001b[0m) β (\u001b[38;5;45mNone\u001b[0m, \u001b[38;5;34m1\u001b[0m) β \u001b[38;5;34m62,721\u001b[0m β \u001b[1;38;5;34mY\u001b[0m β\n",
"βββββββββββββββββββββββββββββββ΄ββββββββββββββββββββββββ΄βββββββββββββ΄ββββββββ\n"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"data": {
"text/html": [
" Total params: 53,213,109 (202.99 MB)\n",
"\n"
],
"text/plain": [
"\u001b[1m Total params: \u001b[0m\u001b[38;5;34m53,213,109\u001b[0m (202.99 MB)\n"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"data": {
"text/html": [
" Trainable params: 62,721 (245.00 KB)\n",
"\n"
],
"text/plain": [
"\u001b[1m Trainable params: \u001b[0m\u001b[38;5;34m62,721\u001b[0m (245.00 KB)\n"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"data": {
"text/html": [
" Non-trainable params: 53,150,388 (202.75 MB)\n",
"\n"
],
"text/plain": [
"\u001b[1m Non-trainable params: \u001b[0m\u001b[38;5;34m53,150,388\u001b[0m (202.75 MB)\n"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# Load pre-trained model without top layers\n",
"base_model = EfficientNetV2M(\n",
" weights=\"imagenet\", # pre-trained weights\n",
" include_top=False, # no dense layer\n",
" input_shape=(input_size[0], input_size[1], 3), # input shape,\n",
")\n",
"\n",
"# For feature extraction\n",
"base_model.trainable = False\n",
"# # For partial fine-tuning: freeze bottom layers\n",
"# for layer in base_model.layers[:735]:\n",
"# layer.trainable = False\n",
"\n",
"# Create explicit input layer\n",
"inputs = tf.keras.Input(shape=(input_size[0], input_size[1], 3))\n",
"\n",
"x = base_model(inputs)\n",
"\n",
"# # Dropout for regularization\n",
"# x = Dropout(0.2)(x)\n",
"# Flatten output\n",
"x = Flatten()(x)\n",
"# New FC layer for binary classification\n",
"predictions = Dense(1, activation=\"sigmoid\")(x)\n",
"\n",
"# Define new model\n",
"model_ready = Model(inputs=inputs, outputs=predictions)\n",
"\n",
"# Display model summary\n",
"model_ready.summary(show_trainable=True)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## [Prepare data](#toc0_)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Create datasets from local images and labels"
]
},
{
"cell_type": "code",
"execution_count": 20,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Train dataset:\n",
"Found 39654 files belonging to 2 classes.\n",
"\n",
"Val dataset:\n",
"Found 6670 files belonging to 2 classes.\n",
"\n",
"Test dataset:\n",
"Found 5908 files belonging to 2 classes.\n"
]
}
],
"source": [
"# Prepare for outputs\n",
"os.makedirs(MODELS_ROOT_DIR, exist_ok=True)\n",
"y = df[\"label\"]\n",
"X = df[\"uri\"]\n",
"\n",
"# Create datasets\n",
"print(\"Train dataset:\")\n",
"train_ds = image_dataset_from_directory(\n",
" train_dir,\n",
" labels=\"inferred\", # class names upon folders structure\n",
" label_mode=\"int\", # integer encoding\n",
" shuffle=True, # shuffle images\n",
" seed=42, # random seed\n",
" image_size=input_size, # automatic resizing\n",
" batch_size=batch_size, # tensor shape[0]\n",
")\n",
"\n",
"print(\"\\nVal dataset:\")\n",
"val_ds = image_dataset_from_directory(\n",
" val_dir,\n",
" labels=\"inferred\", # class names upon folders structure\n",
" label_mode=\"int\", # integer encoding\n",
" shuffle=True, # shuffle images\n",
" seed=42, # random seed\n",
" image_size=input_size, # automatic resizing\n",
" batch_size=batch_size, # tensor shape[0]\n",
")\n",
"\n",
"print(\"\\nTest dataset:\")\n",
"test_ds = image_dataset_from_directory(\n",
" test_dir,\n",
" labels=\"inferred\", # class names upon folders structure\n",
" label_mode=\"int\", # integer encoding\n",
" shuffle=False, # do not shuffle images\n",
" seed=42, # random seed\n",
" image_size=input_size, # automatic resizing\n",
" batch_size=batch_size, # tensor shape[0]\n",
")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## [Training](#toc0_)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Follow with : `tensorboard --logdir models/EfficientNetB0/runs`"
]
},
{
"cell_type": "code",
"execution_count": 101,
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"INFO:root:βοΈ compiling\n",
"INFO:root:ποΈ declaring callbacks\n",
"INFO:root:πͺ starting training\n"
]
},
{
"name": "stdout",
"output_type": "stream",
"text": [
"Epoch 1/100\n",
"\u001b[1m 7/1240\u001b[0m \u001b[37mββββββββββββββββββββ\u001b[0m \u001b[1m36:58\u001b[0m 2s/step - accuracy: 0.4936 - loss: 0.8204 - precision: 0.5297 - recall: 0.5054"
]
}
],
"source": [
"model_trained, history = eval_pretrained_model(\n",
" model=model_ready,\n",
" train_ds=train_ds,\n",
" val_ds=val_ds,\n",
" test_ds=test_ds,\n",
" LOG_DIR=LOG_PATH,\n",
" CHKPT_DIR=CHKPT_PATH,\n",
" model_name=model_name,\n",
" input_size=input_size,\n",
" batch_size=batch_size,\n",
" n_epochs=n_epochs,\n",
" optimizer=optimizer,\n",
" loss=loss,\n",
" metrics=metrics,\n",
")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [
{
"ename": "NameError",
"evalue": "name 'stop' is not defined",
"output_type": "error",
"traceback": [
"\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
"\u001b[0;31mNameError\u001b[0m Traceback (most recent call last)",
"Cell \u001b[0;32mIn[33], line 1\u001b[0m\n\u001b[0;32m----> 1\u001b[0m \u001b[43mstop\u001b[49m\n",
"\u001b[0;31mNameError\u001b[0m: name 'stop' is not defined"
]
}
],
"source": [
"stop"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"**Training** with Ultralytics YOLO"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# from huggingface_hub import hf_hub_download\n",
"\n",
"# # Correctly set repo_id and repo_type\n",
"# repo_id = \"pyronear/pyro-sdis\"\n",
"# filename = \"data.yaml\"\n",
"\n",
"# # Download data.yaml to the current directory\n",
"# yaml_path = hf_hub_download(repo_id=repo_id, filename=filename, repo_type=\"dataset\", local_dir=\".\")\n",
"# print(f\"data.yaml downloaded to: {yaml_path}\")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Train with Yolo (command line)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# yolo task=detect mode=train data=data.yaml model=yolov8n.pt epochs=50 imgsz=640 single_cls=True"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## [Random Baseline](#toc0_)"
]
},
{
"cell_type": "code",
"execution_count": 8,
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"[codecarbon WARNING @ 17:11:39] Already started tracking\n",
"[codecarbon INFO @ 17:11:39] A task is already under measure\n"
]
}
],
"source": [
"# Start tracking emissions\n",
"tracker.start()\n",
"tracker.start_task(\"inference\")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# π§ INFERENCE"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# import keras\n",
"# from keras.applications.resnet50 import ResNet50\n",
"# from keras.applications.resnet50 import preprocess_input, decode_predictions\n",
"# import numpy as np\n",
"\n",
"# model = ResNet50(weights='imagenet')\n",
"\n",
"# img_path = 'elephant.jpg'\n",
"# img = keras.utils.load_img(img_path, target_size=(224, 224))\n",
"# x = keras.utils.img_to_array(img)\n",
"# x = np.expand_dims(x, axis=0)\n",
"# x = preprocess_input(x)\n",
"\n",
"# preds = model.predict(x)\n",
"# # decode the results into a list of tuples (class, description, probability)\n",
"# # (one such list for each sample in the batch)\n",
"# print('Predicted:', decode_predictions(preds, top=3)[0])\n",
"# # Predicted: [(u'n02504013', u'Indian_elephant', 0.82658225), (u'n01871265', u'tusker', 0.1122357), (u'n02504458', u'African_elephant', 0.061040461)]\n"
]
},
{
"cell_type": "code",
"execution_count": 11,
"metadata": {},
"outputs": [],
"source": [
"# --------------------------------------------------------------------------------------------\n",
"# YOUR MODEL INFERENCE CODE HERE\n",
"# Update the code below to replace the random baseline by your model inference within the inference pass where the energy consumption and emissions are tracked.\n",
"# --------------------------------------------------------------------------------------------\n",
"\n",
"# Make random predictions (placeholder for actual model inference)\n",
"\n",
"predictions = []\n",
"true_labels = []\n",
"pred_boxes = []\n",
"true_boxes_list = [] # List of lists, each inner list contains boxes for one image\n",
"\n",
"for example in test_dataset:\n",
" # Parse true annotation (YOLO format: class_id x_center y_center width height)\n",
" annotation = example.get(\"annotations\", \"\").strip()\n",
" has_smoke = len(annotation) > 0\n",
" true_labels.append(int(has_smoke))\n",
"\n",
" # Make random classification prediction\n",
" pred_has_smoke = random.random() > 0.5\n",
" predictions.append(int(pred_has_smoke))\n",
"\n",
" # If there's a true box, parse it and make random box prediction\n",
" if has_smoke:\n",
" # Parse all true boxes from the annotation\n",
" image_true_boxes = parse_boxes(annotation)\n",
" true_boxes_list.append(image_true_boxes)\n",
"\n",
" # For baseline, make one random box prediction per image\n",
" # In a real model, you might want to predict multiple boxes\n",
" random_box = [\n",
" random.random(), # x_center\n",
" random.random(), # y_center\n",
" random.random() * 0.5, # width (max 0.5)\n",
" random.random() * 0.5, # height (max 0.5)\n",
" ]\n",
" pred_boxes.append(random_box)\n",
"\n",
"\n",
"# --------------------------------------------------------------------------------------------\n",
"# YOUR MODEL INFERENCE STOPS HERE\n",
"# --------------------------------------------------------------------------------------------"
]
},
{
"cell_type": "code",
"execution_count": 9,
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"[codecarbon WARNING @ 17:12:24] Background scheduler didn't run for a long period (1885s), results might be inaccurate\n",
"[codecarbon INFO @ 17:12:24] Energy consumed for RAM : 0.003142 kWh. RAM Power : 6.0 W\n",
"[codecarbon INFO @ 17:12:24] Energy consumed for all CPUs : 0.002618 kWh. Total CPU Power : 5.0 W\n",
"[codecarbon INFO @ 17:12:24] 0.005760 kWh of electricity used since the beginning.\n"
]
}
],
"source": [
"# Stop tracking emissions\n",
"emissions_data = tracker.stop_task()"
]
},
{
"cell_type": "code",
"execution_count": 12,
"metadata": {},
"outputs": [],
"source": [
"import numpy as np\n",
"\n",
"# Calculate classification metrics\n",
"classification_accuracy = accuracy_score(true_labels, predictions)\n",
"classification_precision = precision_score(true_labels, predictions)\n",
"classification_recall = recall_score(true_labels, predictions)\n",
"\n",
"# Calculate mean IoU for object detection (only for images with smoke)\n",
"# For each image, we compute the max IoU between the predicted box and all true boxes\n",
"ious = []\n",
"for true_boxes, pred_box in zip(true_boxes_list, pred_boxes):\n",
" max_iou = compute_max_iou(true_boxes, pred_box)\n",
" ious.append(max_iou)\n",
"\n",
"mean_iou = float(np.mean(ious)) if ious else 0.0"
]
},
{
"cell_type": "code",
"execution_count": 13,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"{'submission_timestamp': '2025-01-23T17:13:47.158903',\n",
" 'classification_accuracy': 0.4974610697359513,\n",
" 'classification_precision': 0.8362892223738063,\n",
" 'classification_recall': 0.49625581866019025,\n",
" 'mean_iou': 0.0026954029097350594,\n",
" 'energy_consumed_wh': 5.759879923426909,\n",
" 'emissions_gco2eq': 0.2006914961719638,\n",
" 'emissions_data': {'run_id': 'fbab9dd9-2893-4216-91c4-232be358d4dd',\n",
" 'duration': 1885.054949500016,\n",
" 'emissions': 0.0002006914961719638,\n",
" 'emissions_rate': 1.0646457428260931e-07,\n",
" 'cpu_power': 5.0,\n",
" 'gpu_power': 0.0,\n",
" 'ram_power': 6.0,\n",
" 'cpu_energy': 0.002618128800231918,\n",
" 'gpu_energy': 0,\n",
" 'ram_energy': 0.0031417511231949906,\n",
" 'energy_consumed': 0.005759879923426909,\n",
" 'country_name': 'Switzerland',\n",
" 'country_iso_code': 'CHE',\n",
" 'region': 'zurich',\n",
" 'cloud_provider': '',\n",
" 'cloud_region': '',\n",
" 'os': 'macOS-15.2-arm64-arm-64bit',\n",
" 'python_version': '3.12.7',\n",
" 'codecarbon_version': '2.8.3',\n",
" 'cpu_count': 8,\n",
" 'cpu_model': 'Apple M1',\n",
" 'gpu_count': None,\n",
" 'gpu_model': None,\n",
" 'ram_total_size': 16.0,\n",
" 'tracking_mode': 'machine',\n",
" 'on_cloud': 'N',\n",
" 'pue': 1.0},\n",
" 'dataset_config': {'dataset_name': 'pyronear/pyro-sdis',\n",
" 'test_size': 0.2,\n",
" 'test_seed': 42}}"
]
},
"execution_count": 13,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# Prepare results dictionary\n",
"results = {\n",
" \"submission_timestamp\": datetime.now().isoformat(),\n",
" \"classification_accuracy\": float(classification_accuracy),\n",
" \"classification_precision\": float(classification_precision),\n",
" \"classification_recall\": float(classification_recall),\n",
" \"mean_iou\": mean_iou,\n",
" \"energy_consumed_wh\": emissions_data.energy_consumed * 1000,\n",
" \"emissions_gco2eq\": emissions_data.emissions * 1000,\n",
" \"emissions_data\": clean_emissions_data(emissions_data),\n",
" \"dataset_config\": {\n",
" \"dataset_name\": request.dataset_name,\n",
" \"test_size\": request.test_size,\n",
" \"test_seed\": request.test_seed,\n",
" },\n",
"}\n",
"results"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": []
}
],
"metadata": {
"kernelspec": {
"display_name": ".venv",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.12.7"
}
},
"nbformat": 4,
"nbformat_minor": 2
}