yolo-trainer / app.py
dharshanzeb's picture
Add main app.py β€” YOLO training + inference Space
d9f2fc5 verified
"""
🎯 YOLO Trainer & Detector
Train YOLOv8 on a custom dataset and run inference β€” all from a Gradio UI.
"""
import os
import io
import time
import queue
import threading
import zipfile
import yaml
import cv2
import numpy as np
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from PIL import Image as PILImage
from pathlib import Path
import gradio as gr
from ultralytics import YOLO
from huggingface_hub import hf_hub_download
# ═══════════════════════════════════════════════════════════════
# Constants
# ═══════════════════════════════════════════════════════════════
DATASET_REPO = "dharshanzeb/yolo-detection-dataset"
DATASET_DIR = "/tmp/yolo_dataset"
RUNS_DIR = "/tmp/yolo_runs"
BEST_MODEL_PATH = os.path.join(RUNS_DIR, "gradio_train", "weights", "best.pt")
LAST_MODEL_PATH = os.path.join(RUNS_DIR, "gradio_train", "weights", "last.pt")
CLASS_NAMES = ["car", "person", "dog", "cat", "bicycle"]
# Global state
trained_model_path = None
is_training = False
# ═══════════════════════════════════════════════════════════════
# Dataset download & preparation
# ═══════════════════════════════════════════════════════════════
def download_dataset():
"""Download YOLO dataset from HF Hub and prepare data.yaml."""
if os.path.exists(os.path.join(DATASET_DIR, "images", "train")):
yaml_path = os.path.join(DATASET_DIR, "data.yaml")
if os.path.exists(yaml_path):
return yaml_path, "βœ… Dataset already downloaded."
log = "πŸ“₯ Downloading dataset from HF Hub...\n"
os.makedirs(DATASET_DIR, exist_ok=True)
for split in ["train", "val", "test"]:
log += f" Downloading {split}.zip...\n"
zip_path = hf_hub_download(
repo_id=DATASET_REPO,
filename=f"yolo_format/{split}.zip",
repo_type="dataset",
)
with zipfile.ZipFile(zip_path) as zf:
zf.extractall(DATASET_DIR)
# Download and patch data.yaml
yaml_remote = hf_hub_download(
repo_id=DATASET_REPO,
filename="yolo_format/data.yaml",
repo_type="dataset",
)
with open(yaml_remote) as f:
cfg = yaml.safe_load(f)
cfg["path"] = DATASET_DIR
local_yaml = os.path.join(DATASET_DIR, "data.yaml")
with open(local_yaml, "w") as f:
yaml.dump(cfg, f)
# Count images
n_train = len(list(Path(DATASET_DIR, "images", "train").glob("*.jpg")))
n_val = len(list(Path(DATASET_DIR, "images", "val").glob("*.jpg")))
n_test = len(list(Path(DATASET_DIR, "images", "test").glob("*.jpg")))
log += f"\nβœ… Dataset ready!\n"
log += f" Train: {n_train} images\n"
log += f" Val: {n_val} images\n"
log += f" Test: {n_test} images\n"
log += f" Classes: {CLASS_NAMES}\n"
return local_yaml, log
# ═══════════════════════════════════════════════════════════════
# Metrics chart
# ═══════════════════════════════════════════════════════════════
def make_metrics_chart(history):
"""Create a loss + mAP chart from training history."""
if not history:
return None
epochs = [h["epoch"] for h in history]
fig, axes = plt.subplots(1, 2, figsize=(14, 5))
fig.patch.set_facecolor("#1a1a2e")
# ── Loss subplot ──
ax1 = axes[0]
ax1.set_facecolor("#16213e")
loss_keys = list(history[0].get("loss", {}).keys())
colors_loss = ["#e94560", "#f5a623", "#50fa7b"]
for i, k in enumerate(loss_keys):
vals = [h["loss"].get(k, 0) for h in history]
label = k.split("/")[-1] if "/" in k else k
color = colors_loss[i % len(colors_loss)]
ax1.plot(epochs, vals, marker="o", markersize=4, label=label,
color=color, linewidth=2)
ax1.set_title("Training Loss", color="white", fontsize=14, fontweight="bold")
ax1.set_xlabel("Epoch", color="white")
ax1.set_ylabel("Loss", color="white")
ax1.legend(facecolor="#16213e", edgecolor="white", labelcolor="white")
ax1.tick_params(colors="white")
ax1.grid(True, alpha=0.2, color="white")
for spine in ax1.spines.values():
spine.set_color("white")
# ── mAP subplot ──
ax2 = axes[1]
ax2.set_facecolor("#16213e")
map_keys = [
("metrics/mAP50(B)", "mAP@50", "#00d2ff"),
("metrics/mAP50-95(B)", "mAP@50-95", "#7b2ff7"),
]
for key, label, color in map_keys:
vals = [h["metrics"].get(key, 0) for h in history]
if any(v > 0 for v in vals):
ax2.plot(epochs, vals, marker="s", markersize=4, label=label,
color=color, linewidth=2)
ax2.set_title("Validation mAP", color="white", fontsize=14, fontweight="bold")
ax2.set_xlabel("Epoch", color="white")
ax2.set_ylabel("mAP", color="white")
ax2.set_ylim(0, 1)
ax2.legend(facecolor="#16213e", edgecolor="white", labelcolor="white")
ax2.tick_params(colors="white")
ax2.grid(True, alpha=0.2, color="white")
for spine in ax2.spines.values():
spine.set_color("white")
plt.tight_layout(pad=2)
buf = io.BytesIO()
plt.savefig(buf, format="png", dpi=120, bbox_inches="tight",
facecolor=fig.get_facecolor())
buf.seek(0)
chart = np.array(PILImage.open(buf).copy())
plt.close(fig)
return chart
# ═══════════════════════════════════════════════════════════════
# Training function (generator for streaming logs)
# ═══════════════════════════════════════════════════════════════
def train_yolo(model_size, epochs, batch_size, learning_rate, img_size):
"""
Generator function: trains YOLO and yields (log_text, metrics_chart, status)
after each epoch.
"""
global trained_model_path, is_training
if is_training:
yield "⚠️ Training already in progress. Please wait.", None, "⚠️ Busy"
return
is_training = True
log_queue_local = queue.Queue()
history = []
accumulated_log = ""
train_exception = [None]
try:
# Step 1: Download dataset
yield "πŸ“₯ Preparing dataset...", None, "πŸ“₯ Downloading..."
data_yaml, dl_log = download_dataset()
accumulated_log += dl_log + "\n"
yield accumulated_log, None, "πŸ“₯ Dataset ready"
# Step 2: Load model
model_variant = f"yolov8{model_size}.pt"
accumulated_log += f"πŸ”„ Loading {model_variant}...\n"
yield accumulated_log, None, f"πŸ”„ Loading {model_variant}"
model = YOLO(model_variant)
# Step 3: Attach callback
def on_fit_epoch_end(trainer):
try:
loss_dict = {}
if trainer.tloss is not None:
loss_dict = trainer.label_loss_items(trainer.tloss)
entry = {
"epoch": trainer.epoch + 1,
"epochs": trainer.epochs,
"metrics": dict(trainer.metrics) if trainer.metrics else {},
"fitness": float(trainer.fitness) if trainer.fitness else 0.0,
"loss": loss_dict,
}
log_queue_local.put(entry)
except Exception as e:
log_queue_local.put({"error": str(e)})
model.add_callback("on_fit_epoch_end", on_fit_epoch_end)
# Step 4: Run training in background thread
def run_training():
try:
device = 0 if __import__("torch").cuda.is_available() else "cpu"
model.train(
data=data_yaml,
epochs=int(epochs),
batch=int(batch_size),
lr0=float(learning_rate),
imgsz=int(img_size),
device=device,
workers=0,
project=RUNS_DIR,
name="gradio_train",
exist_ok=True,
pretrained=True,
mosaic=1.0,
mixup=0.1,
patience=50,
verbose=False,
)
except Exception as e:
train_exception[0] = e
finally:
log_queue_local.put(None) # sentinel
accumulated_log += f"\nπŸš€ Starting training: {model_variant} | {int(epochs)} epochs | batch={int(batch_size)} | lr={learning_rate}\n"
accumulated_log += f"{'─' * 60}\n"
yield accumulated_log, None, "πŸš€ Training started..."
t = threading.Thread(target=run_training, daemon=True)
t.start()
# Step 5: Stream logs from queue
while True:
try:
item = log_queue_local.get(timeout=120)
except queue.Empty:
accumulated_log += "⏳ Waiting for update...\n"
yield accumulated_log, make_metrics_chart(history), "⏳ Waiting..."
continue
if item is None:
break
if "error" in item:
accumulated_log += f"⚠️ Callback error: {item['error']}\n"
yield accumulated_log, make_metrics_chart(history), "⚠️ Error"
continue
history.append(item)
e, E = item["epoch"], item["epochs"]
# Format loss
loss_parts = []
for k, v in item["loss"].items():
name = k.split("/")[-1] if "/" in k else k
loss_parts.append(f"{name}={v:.4f}")
loss_str = " | ".join(loss_parts) if loss_parts else "N/A"
# Format mAP
map50 = item["metrics"].get("metrics/mAP50(B)", 0)
map50_95 = item["metrics"].get("metrics/mAP50-95(B)", 0)
line = f"πŸ“ˆ Epoch {e:>3d}/{E} | {loss_str} | mAP50={map50:.4f} | mAP50-95={map50_95:.4f}\n"
accumulated_log += line
chart = make_metrics_chart(history)
status = f"πŸ‹οΈ Epoch {e}/{E} | mAP50={map50:.4f}"
yield accumulated_log, chart, status
t.join(timeout=10)
# Step 6: Check results
if train_exception[0]:
accumulated_log += f"\n❌ Training error: {train_exception[0]}\n"
yield accumulated_log, make_metrics_chart(history), "❌ Failed"
return
# Find best model
if os.path.exists(BEST_MODEL_PATH):
trained_model_path = BEST_MODEL_PATH
elif os.path.exists(LAST_MODEL_PATH):
trained_model_path = LAST_MODEL_PATH
accumulated_log += f"\n{'═' * 60}\n"
accumulated_log += f"πŸŽ‰ TRAINING COMPLETE!\n"
accumulated_log += f"{'═' * 60}\n"
if trained_model_path:
accumulated_log += f"πŸ“ Model saved: {trained_model_path}\n"
accumulated_log += f"πŸ‘‰ Switch to the Inference tab to test your model!\n"
else:
accumulated_log += f"⚠️ No model file found after training.\n"
if history:
final_map = history[-1]["metrics"].get("metrics/mAP50(B)", 0)
accumulated_log += f"\nπŸ“Š Final mAP@50: {final_map:.4f}\n"
chart = make_metrics_chart(history)
yield accumulated_log, chart, "βœ… Training complete!"
except Exception as e:
accumulated_log += f"\n❌ Error: {str(e)}\n"
yield accumulated_log, make_metrics_chart(history) if history else None, "❌ Error"
finally:
is_training = False
# ═══════════════════════════════════════════════════════════════
# Inference function
# ═══════════════════════════════════════════════════════════════
def run_inference(image, conf_threshold, use_pretrained):
"""Run YOLO detection on an uploaded image."""
if image is None:
gr.Warning("Please upload an image first!")
return None, "⚠️ No image uploaded"
# Select model
if use_pretrained:
model_path = "yolov8n.pt"
model_label = "YOLOv8n (COCO pretrained)"
else:
if trained_model_path and os.path.exists(trained_model_path):
model_path = trained_model_path
model_label = f"Custom trained ({os.path.basename(trained_model_path)})"
else:
gr.Warning("No trained model found! Train first or use pretrained COCO model.")
return None, "⚠️ No trained model. Train first or check 'Use Pretrained'."
try:
device = 0 if __import__("torch").cuda.is_available() else "cpu"
model = YOLO(model_path)
results = model.predict(
source=image,
conf=float(conf_threshold),
iou=0.45,
device=device,
verbose=False,
)
result = results[0]
# Draw bounding boxes
plotted_bgr = result.plot(conf=True, labels=True, line_width=2)
plotted_rgb = cv2.cvtColor(plotted_bgr, cv2.COLOR_BGR2RGB)
# Build detection summary
n_detections = len(result.boxes)
if n_detections == 0:
summary = f"πŸ” **{model_label}**\n\nNo objects detected (conf > {conf_threshold})"
else:
lines = [f"πŸ” **{model_label}** β€” Found **{n_detections}** objects:\n"]
lines.append("| # | Class | Confidence | Bbox (x1,y1,x2,y2) |")
lines.append("|---|-------|-----------|---------------------|")
for i, box in enumerate(result.boxes):
cls_id = int(box.cls[0].item())
cls_name = result.names[cls_id]
conf = box.conf[0].item()
coords = [round(v, 1) for v in box.xyxy[0].tolist()]
lines.append(f"| {i+1} | **{cls_name}** | {conf:.2f} | {coords} |")
summary = "\n".join(lines)
return plotted_rgb, summary
except Exception as e:
return None, f"❌ Error: {str(e)}"
# ═══════════════════════════════════════════════════════════════
# Sample images from dataset for inference demo
# ═══════════════════════════════════════════════════════════════
def load_sample_image():
"""Load a random sample from the test set."""
test_dir = Path(DATASET_DIR) / "images" / "test"
if not test_dir.exists():
# Download dataset first
download_dataset()
if test_dir.exists():
images = list(test_dir.glob("*.jpg"))
if images:
import random
img_path = random.choice(images)
return np.array(PILImage.open(img_path))
return None
# ═══════════════════════════════════════════════════════════════
# Build Gradio UI
# ═══════════════════════════════════════════════════════════════
css = """
.gradio-container { max-width: 1100px !important; margin: 0 auto !important; }
.train-log { font-family: 'Courier New', monospace !important; font-size: 13px !important; }
"""
with gr.Blocks(css=css, title="🎯 YOLO Trainer & Detector", theme=gr.themes.Soft()) as demo:
gr.Markdown("""
# 🎯 YOLO Trainer & Detector
**Train** a YOLOv8 model on the [yolo-detection-dataset](https://huggingface.co/datasets/dharshanzeb/yolo-detection-dataset)
and **run inference** on any image β€” all from this UI.
""")
with gr.Tabs():
# ══════════════════════════════════════════════════════
# TAB 1: TRAINING
# ══════════════════════════════════════════════════════
with gr.Tab("πŸ‹οΈ Train"):
with gr.Row():
# Left column: controls
with gr.Column(scale=1, min_width=280):
gr.Markdown("### βš™οΈ Training Configuration")
model_size = gr.Dropdown(
choices=[
("YOLOv8 Nano (fastest)", "n"),
("YOLOv8 Small", "s"),
("YOLOv8 Medium", "m"),
],
value="n",
label="Model Size",
)
epochs = gr.Slider(
minimum=1, maximum=100, value=20, step=1,
label="Epochs",
)
batch_size = gr.Slider(
minimum=4, maximum=64, value=16, step=4,
label="Batch Size",
)
learning_rate = gr.Slider(
minimum=0.0001, maximum=0.1, value=0.01, step=0.0001,
label="Learning Rate",
)
img_size = gr.Dropdown(
choices=[320, 416, 512, 640],
value=640,
label="Image Size",
)
train_btn = gr.Button(
"πŸš€ Start Training", variant="primary", size="lg"
)
train_status = gr.Markdown("*Ready to train*")
gr.Markdown("""
---
### πŸ“‹ Dataset Info
- **5 classes**: car, person, dog, cat, bicycle
- **500** train / **100** val / **50** test images
- **640Γ—640** resolution
- 15% hard negatives included
""")
# Right column: logs + chart
with gr.Column(scale=2):
gr.Markdown("### πŸ“Š Training Progress")
metrics_chart = gr.Image(
label="Loss & mAP Curves",
interactive=False,
height=320,
)
train_log = gr.Textbox(
label="Training Log",
lines=18,
max_lines=30,
interactive=False,
autoscroll=True,
elem_classes=["train-log"],
)
# Wire training button
train_btn.click(
fn=train_yolo,
inputs=[model_size, epochs, batch_size, learning_rate, img_size],
outputs=[train_log, metrics_chart, train_status],
)
# ══════════════════════════════════════════════════════
# TAB 2: INFERENCE
# ══════════════════════════════════════════════════════
with gr.Tab("πŸ” Detect"):
gr.Markdown("### Upload an image to detect objects")
with gr.Row():
# Left: input
with gr.Column(scale=1):
input_image = gr.Image(
label="πŸ“€ Upload Image",
type="numpy",
sources=["upload", "clipboard"],
height=400,
)
with gr.Row():
conf_threshold = gr.Slider(
minimum=0.05, maximum=0.95, value=0.25, step=0.05,
label="Confidence Threshold",
)
with gr.Row():
use_pretrained = gr.Checkbox(
value=False,
label="Use Pretrained COCO Model (YOLOv8n)",
info="Check this if you haven't trained yet",
)
with gr.Row():
detect_btn = gr.Button(
"πŸ” Detect Objects", variant="primary", size="lg"
)
sample_btn = gr.Button(
"🎲 Load Sample", variant="secondary", size="lg"
)
# Right: output
with gr.Column(scale=1):
output_image = gr.Image(
label="πŸ“Έ Detection Result",
type="numpy",
interactive=False,
height=400,
)
detection_summary = gr.Markdown("*Upload an image and click Detect*")
# Wire inference
detect_btn.click(
fn=run_inference,
inputs=[input_image, conf_threshold, use_pretrained],
outputs=[output_image, detection_summary],
)
sample_btn.click(
fn=load_sample_image,
outputs=[input_image],
)
# ══════════════════════════════════════════════════════
# TAB 3: ABOUT
# ══════════════════════════════════════════════════════
with gr.Tab("ℹ️ About"):
gr.Markdown("""
## How It Works
### πŸ‹οΈ Training
1. The dataset ([dharshanzeb/yolo-detection-dataset](https://huggingface.co/datasets/dharshanzeb/yolo-detection-dataset))
is auto-downloaded from the HF Hub
2. YOLOv8 is initialized with COCO-pretrained weights (transfer learning)
3. Training runs with your configured hyperparameters
4. Real-time metrics (loss + mAP) are displayed after each epoch
5. The best model (by mAP) is saved automatically
### πŸ” Inference
- **Custom model**: Uses the model you just trained
- **Pretrained COCO**: Uses YOLOv8n trained on 80 COCO classes (good for real photos)
### 🎯 Dataset Classes
| ID | Class | Description |
|---|---|---|
| 0 | Car | Red car shapes with windows & wheels |
| 1 | Person | Blue stick figures |
| 2 | Dog | Brown dog shapes |
| 3 | Cat | Orange cat shapes with ears |
| 4 | Bicycle | Green bicycles with wheels |
### πŸ’‘ Tips
- **First time?** Start with `YOLOv8 Nano` + `20 epochs` β€” trains in ~5 min on GPU
- **Better accuracy?** Try `YOLOv8 Small` + `50 epochs` + `lr=0.01`
- **No GPU?** Training works on CPU too (just slower). Use pretrained COCO for instant inference.
- **Low mAP?** Increase epochs or try a larger model size
### πŸ”— Links
- [Dataset](https://huggingface.co/datasets/dharshanzeb/yolo-detection-dataset)
- [Ultralytics YOLOv8](https://docs.ultralytics.com/)
- [Gradio](https://gradio.app/)
""")
if __name__ == "__main__":
demo.queue(max_size=5).launch()