panjianning/cbnet

implement cbnet with mmdetection

cbnet

Implement cbnet with mmdetection. Please clone mmdet1.1.0 (other versions may need modifications), place the files（cbnet.py, resnet_vd.py, non_local.py, etc.） in the right directory.

Model download

arch	backbone	Lr schd	box AP	download
cascade_rcnn	dual resnet_vd200 + dcnv2 + nonlocal	2.5x	52.2(soft-nms)	Google Drive

Result

On COCO val2017

1. cascade_rcnn + dual res2net101 + dcnv2 + fpn + 1x: 48.8 mAP

• Single scale training and testing

2. cascade_rcnn + dual resnet_vd200 + dcnv2 + nonlocal + fpn + 2.5x + soft-nms: 52.2 mAP

• Multiscale-training, single-scale testing
• Weights are transfered from CBResNet200-vd-FPN-Nonlocal in paddle detection
• mAP with vanilla nms： 50.8

---

non_local.py is modified to adapt NonLocal module from paddledetection to mmdetection

backbone config example

backbone=dict(
    type='CBNet',
    backbone_type='ResNet_VD',
    num_repeat=2,
    connect_norm_eval=True,
    depth=200,
    out_indices=(0, 1, 2, 3),
    norm_cfg=dict(type='BN', requires_grad=True),
    norm_eval=True,
    dcn=dict(type='DCNv2', deformable_groups=1, fallback_on_stride=False),
    stage_with_dcn=(False, True, True, True),
    stage_with_non_local=(False, False, True, False),
    frozen_stages=1,
    style='pytorch')

To get the pretrained model on imagenet like dual_xxx.pth:

def make_pretrained_model(input_path, output_path, repeat_num=2):
    cp = torch.load(input_path)
    keys = list(cp['state_dict'].keys())
    for key in keys:
        for i in range(repeat_num):
            cp['state_dict']['cb{}.{}'.format(i + 1, key)] = cp['state_dict'][key]
        cp['state_dict'].pop(key)
    torch.save(cp, output_path)

FAQ

Paddle to PyTorch?

1. get the numpy-array-style weights in paddle detection

from paddle import fluid
import os
from ppdet.core.workspace import load_config, create
import ppdet.utils.checkpoint as checkpoint
import numpy as np
from collections import OrderedDict
import torch


def load_var_dict(cfg_path, weight_path, device_id=-1):
    place = fluid.CUDAPlace(0) if device_id >= 0 else fluid.CPUPlace()
    exe = fluid.Executor(place)

    cfg = load_config(cfg_path)
    main_arch = cfg.architecture
    model = create(main_arch)
    startup_prog = fluid.Program()
    infer_prog = fluid.Program()
    with fluid.program_guard(infer_prog, startup_prog):
        with fluid.unique_name.guard():
            inputs_def = cfg['TestReader']['inputs_def']
            inputs_def['iterable'] = True
            feed_vars, loader = model.build_inputs(**inputs_def)
            _ = model.test(feed_vars)
    checkpoint.load_params(exe, infer_prog, weight_path)
    var_dict = {}
    for var in startup_prog.list_vars():
        fluid_tensor = fluid.global_scope().find_var(var.name).get_tensor()
        var_dict[var.name] = np.array(fluid_tensor)

    return var_dict

2. Implement same models in mmdet, and do name mapping for variables.