Add grounding flops tool

[0xyangl]

Motivation

The existing tools/analysis_tools/get_flops.py does not support grounding / vision-language detection models (e.g., GroundingDINO) because these models require text inputs and have multi-modal architectures that cannot be traced end-to-end with mmengine.analysis.get_model_complexity_info.

This PR adds a dedicated FLOPs analysis tool that handles the unique architecture of grounding detection models, providing per-component FLOPs and parameter breakdowns.

Modification

New file: tools/analysis_tools/get_flops_grounding.py

A script that computes per-component FLOPs and parameter counts for grounding detection models:

• Vision Backbone: Accurate FLOPs via fvcore.nn.FlopCountAnalysis
• Text Encoder: Estimated FLOPs based on model type (CLIP, BERT, etc.)
• Neck (ChannelMapper): Estimated from config-driven channel/stride info
• Transformer Encoder/Decoder: Estimated from config-driven architecture params
• Detection Head: Parameter count

Key design choices:

• Automatically disables with_cp (gradient checkpointing) which is incompatible with JIT tracing, without modifying the original config
• Reads architecture parameters (channels, layers, embed_dim, etc.) dynamically from the model config instead of hardcoding
• Uses MMLogger consistent with existing mmdet tools

New file: tests/test_tools/test_get_flops_grounding.py

41 unit tests covering all helper functions and config readers.

BC-breaking

No. This PR only adds new files and does not modify any existing code.

Use cases

# Basic usage
python tools/analysis_tools/get_flops_grounding.py \
    configs/mm_grounding_dino/grounding_dino_swin-t_finetune_8xb4_20e_cat.py

# Custom input shape
python tools/analysis_tools/get_flops_grounding.py <config> --shape 640 640

# Specify number of classes for text encoder FLOPs estimation
python tools/analysis_tools/get_flops_grounding.py <config> --num-classes 80

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ 0xyangl
❌ lauriebax
_{You have signed the CLA already but the status is still pending? Let us recheck it.}