EfficientDet for PyTorch

Warnings

• breaking The bounding box output format for 'Predict' and 'Train' benches changed from `XYWH` (x, y, width, height) to `XYXY` (x1, y1, x2, y2). Users upgrading from older versions or following outdated tutorials must adjust their parsing logic.
  Fix: Update your code to expect `XYXY` format for bounding boxes. This format is `[x_min, y_min, x_max, y_max]`.
• breaking `effdet` has tight dependencies on `timm` (PyTorch Image Models) versions. Upgrading `effdet` often requires updating `timm` to a specific version (e.g., `>=0.3` or `>=0.9`) due to API changes in `timm`'s helper functions and model backbones.
  Fix: Ensure your `timm` installation matches the requirements of your `effdet` version. Check `effdet`'s `requirements.txt` or GitHub README for the exact `timm` version compatibility.
• gotcha Input image dimensions must be divisible by 128 due to the EfficientDet's BiFPN (Bi-directional Feature Pyramid Network) architecture, which processes features at various scales (P3 to P7).
  Fix: When resizing input images, ensure both width and height are multiples of 128 (e.g., 512x512, 640x640, 768x768). Non-compliant sizes might lead to errors or unexpected behavior.
• deprecated The default focal loss implementation changed. The older version, which might have more numerical stability issues but potentially lower memory usage, can be explicitly enabled during training.
  Fix: If reproducing old results or encountering training issues, use the `--legacy-focal` argument in training scripts to revert to the previous focal loss implementation. Otherwise, the new focal loss is used by default.

Install

• pip install effdet Install latest version

Imports

• create_model

  from effdet import create_model

  The `create_model` function is the recommended high-level API for instantiating models, including loading pretrained weights. Direct import of `EfficientDet` and `DetBenchTrain` is less common for typical usage and mostly for advanced customization.

Quickstart

This quickstart demonstrates how to load a pre-trained EfficientDet model, prepare a dummy image, and perform inference to detect objects. It utilizes the high-level `create_model` function and standard PyTorch image transformations.

import torch
from effdet import create_model
from effdet.data import resolve_input_config
from torchvision import transforms
from PIL import Image
import os

# Create a dummy image for a runnable example
dummy_image_path = "dummy_image.png"
if not os.path.exists(dummy_image_path):
    img = Image.new('RGB', (640, 640), color = 'red')
    img.save(dummy_image_path)

# 1. Load a pre-trained EfficientDet model
# Use 'tf_efficientdet_d0' for a small, fast model.
# bench_task='predict' is crucial for inference mode.
model_name = 'tf_efficientdet_d0'
model = create_model(model_name, pretrained=True, bench_task='predict')
model.eval()

# 2. Prepare the image for inference
img = Image.open(dummy_image_path).convert('RGB')

# Resolve input configuration from the model's pretrained_cfg
input_config = resolve_input_config(model.pretrained_cfg)

# Define image transformation pipeline
transform = transforms.Compose([
    transforms.Resize(input_config['input_size']),
    transforms.ToTensor(),
    transforms.Normalize(mean=input_config['mean'], std=input_config['std'])
])

# Apply transformations and add a batch dimension
input_tensor = transform(img).unsqueeze(0)

# 3. Perform inference
with torch.no_grad():
    output = model(input_tensor)

# The output format is typically [x1, y1, x2, y2, score, class]
# Print top 5 detected objects (if any)
if output.numel() > 0:
    print(f"Detected objects (top 5, if available):\n{output[0][:5]}")
else:
    print("No objects detected.")

# Clean up dummy image
os.remove(dummy_image_path)