[Solved]: Finetuning model on custom dataset.

[JaninaMattes]

Hi @hpanwar08 I would like to use your pretrained Detectron2 model and fine tune it on my custom datset (same class labels as in the PubLayNet dataset). I use an adjusted from of the ''' tools/train_net_dla.py''' script and also adjusted the DLA_*.yml config file accordingly. If I build the docker file and run it, I keep running into

Error Message

File "finetune_net_dla.py", line 29, in <module> from detectron2.evaluation import ( ImportError: cannot import name 'CityscapesEvaluator' from 'detectron2.evaluation' (/anaconda/lib/python3.8/site-packages/detectron2/detectron2/evaluation/__init__.py)

What I have done so far

I am using Python 3.8.11 on Ubuntu 18.04 Bionic. The site-packages are installed and the command python -m site returns me the information:
sys.path = [ '/home/appuser/detectron2_repo', '/usr/local/lib/python38.zip', '/usr/local/lib/python3.8', '/usr/local/lib/python3.8/lib-dynload', '/home/appuser/.local/lib/python3.8/site-packages', '/usr/local/lib/python3.8/site-packages', ] USER_BASE: '/home/appuser/.local' (exists) USER_SITE: '/home/appuser/.local/lib/python3.8/site-packages' (exists)
And which python returns /usr/local/bin/python.

My Docker file

My Docker is not located in the ./docker , but directly under root directory, which is named detectron2_dla and is structured in accordance to the Docker file from Facebook Research repo:

`FROM nvidia/cuda:10.1-cudnn7-devel

ENV DEBIAN_FRONTEND noninteractive
RUN apt-get update && apt-get install -y
python3-opencv ca-certificates python3-dev git wget
sudo unzip ninja-build &&
rm -rf /var/lib/apt/lists/*

ARG USER_ID=1000
RUN useradd -m --no-log-init --system --uid ${USER_ID} appuser -g sudo
RUN echo '%sudo ALL=(ALL) NOPASSWD:ALL' >> /etc/sudoers
USER appuser
WORKDIR /home/appuser

ENV PATH="/home/appuser/.local/bin:${PATH}"
RUN wget https://bootstrap.pypa.io/get-pip.py &&
python3 get-pip.py --user &&
rm get-pip.py

RUN pip install --user torchvision torchtext==0.9.1 tensorboard cython cmake pyyaml==5.1 gdown
RUN pip install --user torch==1.8.1+cu101 torchvision==0.9.1+cu101 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
RUN pip install --user 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'

RUN pip install --user 'git+https://github.com/facebookresearch/fvcore'

RUN mkdir /home/appuser/detectron2_repo
RUN git clone https://github.com/facebookresearch/detectron2.git /home/appuser/detectron2_repo
ENV FORCE_CUDA="1"

ENV TORCH_CUDA_ARCH_LIST="Kepler;Kepler+Tesla;Maxwell;Maxwell+Tegra;Pascal;Volta;Turing"
RUN pip install --user -e /home/appuser/detectron2_repo
ENV FVCORE_CACHE="/tmp"

WORKDIR /home/appuser/detectron2_repo

COPY configs/ /home/appuser/detectron2_repo/configs
COPY finetune_net_dla.py /home/appuser/detectron2_repo

RUN gdown https://drive.google.com/uc?id=XXX
RUN unzip file.zip -d /home/appuser/detectron2_repo/datasets
RUN rm -r file.zip

RUN mkdir /home/appuser/detectron2_repo/pretrained_models
RUN wget https://www.dropbox.com/sh/wgt9skz67usliei/AADGw0h1y7K5vO0akulyXm-qa/model_final.pth -P /home/appuser/detectron2_repo/pretrained_models/

RUN python3 -m pip install 'git+https://github.com/facebookresearch/detectron2.git'

ENTRYPOINT [ "python3", "finetune_net_dla.py" ] `

My adjusted python script is the following:
`
import logging
import os
import datetime
from collections import OrderedDict
import torch

import detectron2.utils.comm as comm
from detectron2.checkpoint import DetectionCheckpointer
from detectron2.config import get_cfg
from detectron2.data import MetadataCatalog
from detectron2.engine import DefaultTrainer, default_argument_parser, default_setup, hooks, launch
from detectron2.evaluation import (
CityscapesEvaluator,
COCOEvaluator,
COCOPanopticEvaluator,
DatasetEvaluators,
LVISEvaluator,
PascalVOCDetectionEvaluator,
SemSegEvaluator,
verify_results,
)
from detectron2.modeling import GeneralizedRCNNWithTTA
from detectron2.data.datasets import register_coco_instances
from detectron2.utils.logger import setup_logger

logger = logging.getLogger("detectron2")

class FineTuner(DefaultTrainer):

@classmethod
def build_evaluator(cls, cfg, dataset_name, output_folder=None):

    if output_folder is None:
        output_folder = os.path.join(cfg.OUTPUT_DIR, "inference")
    evaluator_list = []
    evaluator_type = MetadataCatalog.get(dataset_name).evaluator_type
    if evaluator_type in ["sem_seg", "coco_panoptic_seg"]:
        evaluator_list.append(
            SemSegEvaluator(
                dataset_name,
                distributed=False,
                num_classes=cfg.MODEL.SEM_SEG_HEAD.NUM_CLASSES,
                ignore_label=cfg.MODEL.SEM_SEG_HEAD.IGNORE_VALUE,
                output_dir=output_folder,
            )
        )
    if evaluator_type in ["coco", "coco_panoptic_seg"]:
        evaluator_list.append(COCOEvaluator(dataset_name, cfg, False, output_folder))
    if len(evaluator_list) == 1:
        return evaluator_list[0]
    return DatasetEvaluators(evaluator_list)

def setup(args):
cfg = get_cfg()
cfg.merge_from_file(args.config_file)
cfg.merge_from_list(args.opts)
cfg.freeze()
default_setup(cfg, args)
return cfg

def main(args):
cfg = setup(args)

if args.eval_only:
    model = FineTuner.build_model(cfg)
    DetectionCheckpointer(model, save_dir=cfg.OUTPUT_DIR).resume_or_load(
        cfg.MODEL.WEIGHTS, resume=args.resume
    )
    res = FineTuner.test(cfg, model)
    if comm.is_main_process():
        verify_results(cfg, res)
    if cfg.TEST.AUG.ENABLED:
        res.update(FineTuner.test_with_TTA(cfg, model))
    return res

fineTuner = FineTuner(cfg)
fineTuner.resume_or_load(resume=args.resume)
if cfg.TEST.AUG.ENABLED:
    fineTuner.register_hooks(
        [hooks.EvalHook(0, lambda: fineTuner.test_with_TTA(cfg, fineTuner.model))]
    )
return fineTuner.train()

if name == "main":
args = default_argument_parser().parse_args()
output_dir = os.path.join('./output', datetime.datetime.now().strftime('%Y%m%dT%H%M'))
os.makedirs(output_dir, exist_ok=True)
logger = setup_logger(output=output_dir)
logger.info("Command Line Args:", args)

register_coco_instances(
    "dla_train",
    {},
    "./datasets/custom/train/train2021.json",
    "./datasets/custom/train/images"
)

register_coco_instances(
    "dla_val",
    {},
    "./datasets/custom/val/val2021.json",
    "./datasets/custom/val/images"
)

metadata_train = MetadataCatalog.get("dla_train")
metadata_val = MetadataCatalog.get("dla_val")

cfg = get_cfg()
# mask rcnn resnet101
# cfg.merge_from_file("configs/DLA_mask_rcnn_R_101_FPN_3x.yaml")
# mask rcnn resnext
cfg.merge_from_file("configs/DLA_mask_rcnn_X_101_32x8d_FPN_3x_Finetune.yaml")

cfg.OUTPUT_DIR = output_dir

logger.info(cfg)
cfg_str = cfg.dump()
with open(os.path.join(cfg.OUTPUT_DIR, "finetune_config.yaml"), "w") as f:
    f.write(cfg_str)
f.close()

fineTuner = FineTuner(cfg)
fineTuner.resume_or_load(resume=False)
fineTuner.train()`

Do you have any idea what might cause this issue? I have tried fixing it, but couldn't find any working solution. Thank you in advance for any help.

[JaninaMattes]

Somehow from detectron2.evaluation import * instead of from detectron2.evaluation import ( CityscapesEvaluator, COCOEvaluator, COCOPanopticEvaluator, DatasetEvaluators, LVISEvaluator, PascalVOCDetectionEvaluator, SemSegEvaluator, verify_results, ) solved it for now.