Welcome to the world of FCOS (Fully Convolutional One-Stage Object Detection)! This blog post will take you through the process of enhancing your FCOS models with new features and optimizations. Get ready to embark on a journey of object detection excellence!

Installation

Before we dive into the enhancements, ensure you have properly set up your environment. Please check the INSTALL.md for installation instructions, as the setup is similar to the original FCOS project.

Understanding the Results

Let’s take a look at some results from your FCOS training sessions. Here’s a breakdown of the models:

Model                         Total training mem (GB)  Multi-scale training  Testing time  AP (minival)  link
FCOS_R_50_FPN_1x                29.3                          No                           71ms          37.0          [model](https://pan.baidu.com/s1Xcbx7EfOGvwnexXAuovM0A)
FCOS_R_50_FPN_1x_center          30.61                         No                           71ms          37.8          [model](https://pan.baidu.com/s1Gs7AzmJRmeYhXUPDQZuSLA)
FCOS_R_50_FPN_1x_center_liou     30.61                         No                           71ms          38.1          [model](https://pan.baidu.com/s1HpYrkAsVXNvXRFTd06SGgA)
FCOS_R_50_FPN_1x_center_giou     30.61                         No                           71ms          38.2          [model](https://pan.baidu.com/s13_o6343Ikg4td01kVXxGSw)
FCOS_R_101_FPN_2x                44.1                          Yes                          74ms          41.4          [model](https://pan.baidu.com/s1u_5OD5NURYe1EYFWnohgEA)
FCOS_R_101_FPN_2x_center_giou     44.1                          Yes                          74ms          42.5          [model](https://pan.baidu.com/s1qhHM067ywwlEXfamaFq23g)

These results highlight various models along with their memory requirements, training configurations, and average precision (AP). You can see the impact of configurations—like using center sampling and different loss functions—on performance.

Training Your FCOS Model

Ready to train your own model? Here’s the command line you’ll need for training the FCOS_R_50_FPN_1x on 8 GPUs with Synchronous Stochastic Gradient Descent (SGD):

python -m torch.distributed.launch
     --nproc_per_node=8
     --master_port=$((RANDOM + 10000))
     tools/train_net.py
     --skip-test
     --config-file configs/fcos/fcos_R_50_FPN_1x_center_giou.yaml
     DATALOADER.NUM_WORKERS 2
     OUTPUT_DIR training_dir/fcos_R_50_FPN_1x_center_giou

This command will do the following:

• –nproc_per_node=8: Launch the training on 8 GPUs. If you want to use fewer GPUs, just change this number.
• OUTPUT_DIR: This is where your model will be saved.
• If you wish to modify other aspects, like batch size or backbone models, make sure to alter the relevant config files.

Dealing with Common Issues

As you venture into this exciting enhancement journey, you might encounter a few bumps along the way. Here are some troubleshooting ideas:

• Issue: High memory usage during training.
• Solution: Consider reducing the batch size by adjusting SOLVER.IMS_PER_BATCH in your configuration file.
• Issue: Training does not seem to improve.
• Solution: Verify your loss function and configurations. You can experiment with different loss functions, as indicated in the results above.
• Issue: Training takes too long.
• Solution: Ensure you are utilizing all GPUs efficiently. Adjust –nproc_per_node accordingly if you have fewer available.

For more insights, updates, or to collaborate on AI development projects, stay connected with fxis.ai.

Final Words

Congratulations on expanding your abilities with FCOS! Mastering object detection tools lays the groundwork for developing advanced AI systems. At fxis.ai, we believe that such advancements are crucial for the future of AI, as they enable more comprehensive and effective solutions. Our team is continually exploring new methodologies to push the envelope in artificial intelligence, ensuring that our clients benefit from the latest technological innovations.