sxhxliang / detectron2_backbone Goto Github PK

View Code? Open in Web Editor NEW

209.0 3.0 40.0 90 KB

detectron2 backbone: resnet18, efficientnet, hrnet, mobilenet v2, resnest, bifpn

Python 100.00%

detectron2 backbone resnet18 efficientnet hrnet pytorch

detectron2_backbone's Introduction

Detectron2_Backbone

Installation

First install Detectron2 following the official guide: INSTALL.md. Then build detectron2_backbone with:

git clone https://github.com/sxhxliang/detectron2_backbone.git
cd detectron2_backbone
python setup.py build develop

Quick Start

install detectron2 and detectron2_backbone
import detectron2, import detectron2.
import detectron2_backbone import detectron2_backbone
add config to detectron2

# for example
# import detectron2
import detectron2 
from detectron2.config import get_cfg
# import detectron2_backbone
from detectron2_backbone import backbone
from detectron2_backbone.config import add_backbone_config

def setup(args):
    """
    Create configs and perform basic setups.
    """
    cfg = get_cfg()
    # add config to detectron2
    add_backbone_config(cfg)
    cfg.merge_from_file(args.config_file)
    cfg.merge_from_list(args.opts)
    cfg.freeze()
    default_setup(cfg, args)
    return cfg

Build Your Own Models

# your_config.yaml
MODEL:
  WEIGHTS: "your_path/resnet18_detectron2.pth"
  BACKBONE:
    NAME: "build_resnet18_fpn_backbone"
  ...

Backbones for Detectron2

resnet18:

build_resnet18_bacbkone
build_resnet18_fpn_backbone
build_fcos_resnet18_fpn_backbone

efficientnet:

build_efficientnet_backbone
build_efficientnet_fpn_backbone
build_fcos_efficientnet_fpn_backbone

# your_config.yaml
MODEL:
  WEIGHTS: "your_path/efficientnet_b0_detectron2.pth"
  BACKBONE:
    NAME: "build_efficientnet_fpn_backbone"
  EFFICIENTNET:
    NAME: "efficientnet_b0" # efficientnet_b1, efficientnet_2,  ..., efficientnet_b7
    OUT_FEATURES: ["stride4", "stride8", "stride16", "stride32"]
  FPN:
    IN_FEATURES: ["stride4", "stride8", "stride16", "stride32"]
  ...

dla:

build_dla_backbone
build_dla_fpn_backbone
build_fcos_dla_fpn_backbone

resnest:

build_resnest_backbone
build_resnest_fpn_backbone
build_fcos_resnest_fpn_backbone

vovnet:

build_vovnet_backbone
build_vovnet_fpn_backbone
build_fcos_vovnet_fpn_backbone

mobilenet v2:

build_mnv2_backbone
build_mnv2_fpn_backbone
build_fcos_mnv2_fpn_backbone

hrnet:

build_hrnet_backbone
build_hrnet_fpn_backbone

bifpn:

build_efficientnet_bifpn_backbone

# your_config.yaml
MODEL:
  WEIGHTS: "your_path/efficientnet_b0_detectron2.pth"
  BACKBONE:
    NAME: "build_efficientnet_bifpn_backbone"
  EFFICIENTNET:
    NAME: "efficientnet_b0" # efficientnet_b1, efficientnet_2,  ..., efficientnet_b7
    OUT_FEATURES: ["stride8", "stride16", "stride32"]
  FPN:
    IN_FEATURES: ["stride8", "stride16", "stride32"]
    OUT_CHANNELS: 64
    REPEAT: 3
  ...

Convert Model for Detectron2

python3 -m detectron2_backbone.tools.convert_weight --help

# download model
wget https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b0-355c32eb.pth -P your_path
# convert weight
python3 -m detectron2_backbone.tools.convert_weight --model your_path/efficientnet-b0-355c32eb.pth --output your_path/efficientnet_b0_detectron2.pth

your_config.yaml
MODEL:
  WEIGHTS: "your_path/efficientnet_b0_detectron2.pth"

detectron2_backbone's People

Contributors

Stargazers

Watchers

detectron2_backbone's Issues

config for resnet18

Hi,

Thanks for this amazing work! Could you (or maybe someone who's got this working) share the config for resnet18 please?

How can I use mobilenet as backbone, I put in detectron2/projects but it can't work

Question mAP and FPS

Dir sxhxliang,

Could I ask you the mAP and FPS you obtain with detectron_efficientNet on COCO dataset?

Thank you

yaml for MobilenetV2

Hi! Just wanted to ask what is the recommended config.yaml for the MobilenetV2 model. Thanks!

pretrained model issue "resnet18_detectron2.pth"

Hi sxhxliang,

May I ask about how you get the model "resnet18_detectron2.pth" you used in "WEIGHTS: "your_path/resnet18_detectron2.pth"".

Best regards,
Haoren

efficientnet with FPN can not be run

efficient output shape are:

EEE:  0 torch.Size([4, 16, 512, 400])
EEE:  1 torch.Size([4, 24, 256, 200])
EEE:  2 torch.Size([4, 80, 64, 50])
EEE:  3 torch.Size([4, 112, 64, 50])

this is impossible to run on FPN......

config.yaml

Hi @sxhxliang ,

Could you give me your config.yaml with D0 and D1 please?

Thank you so much

Can dla be used？

I run dla be backbone, but model do not have load_pretrained_model function!

Is this repo really needed for RESNET18 ?

Hello,

I recently started using detectron2, if i want to train a custom dataset on resnet18, don't I just need to take RESNET50 model, set DEPTH to 18 and set RES2_OUT_CHANNELS to 64 ?

This way :

        data["MODEL"]["RESNETS"]["DEPTH"] = 18
        data["MODEL"]["RESNETS"]["RES2_OUT_CHANNELS"] = 64

And obviously, without any trained model to import weights from, set weights to "" and unfreeze first two layers

        cfg.MODEL.WEIGHTS = ""
        cfg.MODEL.BACKBONE.FREEZE_AT = 0

As only for RESNET18, what does this repository brings compared to what I just showed above ?

Apologies if it's a naive question, I'm new to this whole deeplearning stuff.

Thanks

Could you share your config file as a example?

Hi,

I'm impressed by your work.

Could you share your full EfficientDet config file as an example?

Recently, I also implemented bifpn with vovnet but the model result is not as good as the fpn-one.

I expect the reason that is related to the normalization in the bifpn (I tested GN, BN, and SyncBN).

Configs for EfficientNet B0x FPN

Hi there! First I have to say thank you for your amazing work and contribution 😁.

Here is the configuration file that I use for the Faster RCNN EfficientNet B0 FPN;

File faster_rcnn_EfficientNet_B0_FPN.yaml:

MODEL:
  META_ARCHITECTURE: "GeneralizedRCNN"
  MASK_ON: False
  WEIGHTS: "https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b0-355c32eb.pth"
  BACKBONE:
    NAME: "build_efficientnet_fpn_backbone"
    FREEZE_AT: 0
  EFFICIENTNET:
    NAME: "efficientnet_b0" # efficientnet_b1, efficientnet_2,  ..., efficientnet_b7
    FEATURE_INDICES: [1, 4, 10, 15]
    OUT_FEATURES: ["stride4", "stride8", "stride16", "stride32"]
  FPN:
    IN_FEATURES: ["stride4", "stride8", "stride16", "stride32"]
  ANCHOR_GENERATOR:
    SIZES: [[32], [64], [128], [256], [512]]  # One size for each in feature map
    ASPECT_RATIOS: [[0.5, 1.0, 2.0]]  # Three aspect ratios (same for all in feature maps)
  RPN:
    IN_FEATURES: ["p2", "p3", "p4", "p5", "p6"]
    PRE_NMS_TOPK_TRAIN: 2000  # Per FPN level
    PRE_NMS_TOPK_TEST: 1000  # Per FPN level
    # Detectron1 uses 2000 proposals per-batch,
    # (See "modeling/rpn/rpn_outputs.py" for details of this legacy issue)
    # which is approximately 1000 proposals per-image since the default batch size for FPN is 2.
    POST_NMS_TOPK_TRAIN: 1000
    POST_NMS_TOPK_TEST: 1000
  ROI_HEADS:
    NAME: "StandardROIHeads"
    IN_FEATURES: ["p2", "p3", "p4", "p5"]
  ROI_BOX_HEAD:
    NAME: "FastRCNNConvFCHead"
    NUM_FC: 2
    POOLER_RESOLUTION: 7
  ROI_MASK_HEAD:
    NAME: "MaskRCNNConvUpsampleHead"
    NUM_CONV: 4
    POOLER_RESOLUTION: 14
DATASETS:
  TRAIN: ("coco_2017_train",)
  TEST: ("coco_2017_val",)
SOLVER:
  IMS_PER_BATCH: 16
  BASE_LR: 0.02
  STEPS: (60000, 80000)
  MAX_ITER: 90000
INPUT:
  MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800)
VERSION: 2

However, replacing the NAME with, for example, efficientnet_b1, It yields a mismatch on the dimensions during training, as the out features of the backbone does not much exactly those of the FPN.

I will really appreciate if you (or anyone!) can help me with that, thanks!

Error in bifpn.py

File "/home/user/detectron2_backbone/detectron2_backbone/backbone/bifpn.py", line 328
def init(self, in_channels, out_channels, norm=''):
^

It seems there are a lot of issues regarding indentation (3 chars instead of 3 etc.).

If I can find the time I will try to fix and make a PR for it.

Pretrained models problem

Appreciate your great work.

In terms of your README, I can only find the pretrained efficientnet model. So where can I download other pretrained models. For example, resnet18, and use the convert_weight.py to convert it for training?