YOLO-Coffee-Object-Detection/3_train.py
2026-05-03 14:52:49 +01:00

152 lines
5.0 KiB
Python

"""
STEP 3 - train.py
================
Fine-tunes YOLOv8-nano on your augmented dataset.
Pre-trained COCO weights give you a huve head-start - the backbone
already knows edges, textures, and shapes; it only needs to learn
what Famico (and other products) look like.
Install:
pip install ultralytics
Usage:
python train.py
Outputs:
runs/detect/coffee_v1/weights/best.pt <- use this for detection
"""
import os
import torch
from pathlib import Path
from ultralytics import YOLO
# ── Configuration ───────────────────────────────────────────────────
os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"
DATASET_YAML = "dataset.yml" # Generated by prepare_dataset.py
MODEL = "yolov8n.pt" # nano - fast, good for few-shot
# alternativse to try: yolov8s.pt (better acc)
# yolov8m.pt (even better, needs GPU)
PROJECT = "runs/detect"
RUN_NAME = "coffee_v1"
IMG_SIZE = 640
# ── Training hyperparameters ───────────────────────────────────────────────────
#
# Tuned for: small dataset (few-shot), packaged goods, retail shelf.
#
EPOCHS = 150
PATIENCE = 30 # Early Stopping - stops if no improvement
BATCH = 8 # Lower if you get OOM on CPU/small GPU.
LR0 = 0.005 # Initial LR (Learing Rate) - lower than default (0.01)
LRF = 0.01 # final LR = LR0 * LRF
MOMENTUM = 0.938
WEIGHT_DECAY = 0.0005
WARMUP_EPOCHS = 5 # Longer warmup stabilises few-shot training.
IOU_THRESH = 0.5 # IoU threshold for NMS.
CONF_THRESH = 0.25 # confidence threshold at val time.
# ── YOLOv8 built-in augmentation (on top of our albumentatios pass) ────────────
#
# These run during tarining batches - complementary to augment.py.
# Mosaic is the single most impactful augmentation for small datasets.
#
MOSAIC = 1.0 # mosaic probability
MIXUP = 0.1 # mixup probability
COPY_PASTE = 0.1
DEGREES = 5.0 # degrees ±5°
TRANSLATE = 0.1
SCALE = 0.5
FLIPUD = 0.0
FLIPLR = 0.5
HSV_H = 0.015
HSV_S = 0.7
HSV_V = 0.4
def check_enviroment():
device = "cuda" if torch.cuda.is_available() else "cpu"
if device == "cpu":
print("[INFO] No GPU detected - trainning on CPU")
print(" Expected time: ~60-90min for 150 epochs on a typical laptop.")
print(" Tips: use Google Colab or Kaggle they got free GPU, if training is too slow.\n")
else:
gpu= torch.cuda.get_device_name(0)
print(f"[INFO] GPU detected: {gpu}")
print(" Expected time: ~5-15min.\n")
return device
def train():
device = check_enviroment()
if not Path(DATASET_YAML).exists():
print(f"[ERROR] {DATASET_YAML} not found.")
print(" Run prepare_dataset.py first.")
return
model = YOLO(MODEL)
print(f"Starting training: {EPOCHS} epochs, img size {IMG_SIZE}")
print(f"Dataset: {DATASET_YAML}")
print(f"Output: {PROJECT}/{RUN_NAME}/weights/best.pt\n")
results = model.train(
data = DATASET_YAML,
epochs = EPOCHS,
patience = PATIENCE,
imgsz = IMG_SIZE,
batch = BATCH,
device = device,
project = PROJECT,
name = RUN_NAME,
exist_ok = True,
lr0 = LR0,
lrf = LRF,
momentum = MOMENTUM,
weight_decay = WEIGHT_DECAY,
warmup_epochs = WARMUP_EPOCHS,
iou = IOU_THRESH,
conf = CONF_THRESH,
# Built-in augmentation
mosaic = MOSAIC,
mixup = MIXUP,
copy_paste = COPY_PASTE,
degrees = DEGREES,
translate = TRANSLATE,
scale = SCALE,
flipud = FLIPUD,
fliplr = FLIPLR,
hsv_h = HSV_H,
hsv_s = HSV_S,
hsv_v = HSV_V,
save = True,
save_period = 10, # Checkpoint every 10 epochs
plots = True, # Saves training curvse as PNGs
verbose = True,
seed = 42,
)
best_weights = Path(PROJECT) / RUN_NAME / "weights" / "best.pt"
print("\n" + "-" * 60)
print(f"Training complete.")
print(f"Best weights saved -> {best_weights}")
print(f"mAP50: {results.results_dict.get('metrics/mAP50(B)', 'N/A'):.3f}")
print(f"mAP50-95: {results.results_dict.get('metrics/mAP50-95(B)', 'N/A'):.3f}")
print("-" * 60)
print("\nRun next: python detect.py --source scene.jpg")
return best_weights
if __name__ == "__main__":
train()