v1.5.0

Restructure as pip-installable package with v2 attention exports

What's New

• Pip-installable: pip install actionformer
• v2 Modern Attention Components:
  • MaskedMHAv2, TransformerBlockv2 (flash attention + RoPE)
  • RotaryPositionEmbedding, SwiGLU, apply_rotary_pos_emb
  • HAS_FLASH_ATTN flag
• Cross-platform: Python fallback for NMS when C++ extension unavailable

Usage

from actionformer import (
    load_config,
    make_meta_arch,
    RotaryPositionEmbedding,
    RMSNorm,
    SwiGLU,
    HAS_FLASH_ATTN,
)

Breaking Change

Update imports from libs.* to actionformer:

# Before
from libs.core import load_config

# After
from actionformer import load_config

v1.4.0

Highlights

ActionFormer is now a proper pip-installable Python package.

Breaking Changes

• Imports changed from libs.* to actionformer.*

Migration

# Before
from libs.core import load_config
from libs.modeling import make_meta_arch

# After
from actionformer import load_config, make_meta_arch

Installation

pip install actionformer

v1.3.1 - CLI Entry Points

CLI Entry Points for pip-installed package

New CLI Commands

After pip install, the following commands are available:

actionformer-train configs/thumos_i3d.yaml --output my_exp
actionformer-train-ddp configs/thumos_i3d.yaml --amp
actionformer-eval configs/thumos_i3d.yaml ./ckpt/my_exp

Fix

• Training scripts now accessible from installed package
• No need to locate script files manually

v1.3.0 - TBTFormer

TBTFormer with Boundary Distribution Regression

New Features

• TBTFormer meta-arch: probabilistic boundary prediction for noisy annotations
• Distribution Focal Loss (DFL): learns over discrete boundary bins
• BDR Head: predicts distribution instead of scalar offsets

Configuration

model:
  meta_arch: "TBTFormer"
  fpn_type: "cs_fpn"
  reg_max: 16
  dfl_weight: 0.25

Backward compatible with existing configs.

v1.2.0 - SnapFormer

SnapFormer for Point Detection

New architecture for detecting instant/point events (duration ≈ 0).

New Features

• SnapFormer meta-arch: heatmap-based detection for zero-duration events
• CrossScaleFPN (cs_fpn): bidirectional PANet-style feature fusion
• Gaussian focal loss: CornerNet-style loss for heatmap regression

Configuration

model:
  meta_arch: "SnapFormer"
  fpn_type: "cs_fpn"

v1.1.0 - Custom Dataset Loader

Custom Dataset Loader Support

New Features

• External dataset classes: Load custom dataset classes via custom_class config
• Cleaner library structure for generic ActionFormer usage

Configuration

dataset:
  custom_class: "my_project.datasets.MyDataset"

Allows using ActionFormer as a library with your own dataset implementations.

v1.0.0

ActionFormer with pip installation support.

Installation

pip install git+https://github.com/sreevadde/actionformer.git@v1.0.0

Features

• Pip installable package
• C++ NMS extension (auto-compiled during install)
• All dependencies included (torch, numpy, pandas, etc.)