Increasing --num-gpus from 1 to 2 and cfg.SOLVER.IMS_PER_BATCH from 1 to 2 gave higher ETA not lower

Instructions To Reproduce the Issue:

CUDA_VISIBLE_DEVICES=0,1,2,3 python plain_train_net.py --num-gpus=1 --config-file=/..../detectron2/configs/COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml

CUDA_VISIBLE_DEVICES=0,1,2,3 python plain_train_net.py --num-gpus=2 --config-file=/..../detectron2/configs/COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml

full code you wrote or full changes you made (git diff) based on plain_train_net.py

#!/usr/bin/env python
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
Detectron2 training script with a plain training loop.

This script reads a given config file and runs the training or evaluation.
It is an entry point that is able to train standard models in detectron2.

In order to let one script support training of many models,
this script contains logic that are specific to these built-in models and therefore
may not be suitable for your own project.
For example, your research project perhaps only needs a single "evaluator".

Therefore, we recommend you to use detectron2 as a library and take
this file as an example of how to use the library.
You may want to write your own script with your datasets and other customizations.

Compared to "train_net.py", this script supports fewer default features.
It also includes fewer abstraction, therefore is easier to add custom logic.
"""

import logging
import os
from collections import OrderedDict
import torch
from torch.nn.parallel import DistributedDataParallel

import detectron2.utils.comm as comm
from detectron2.checkpoint import DetectionCheckpointer, PeriodicCheckpointer
from detectron2.config import get_cfg
from detectron2.data import (
    MetadataCatalog,
    build_detection_test_loader,
    build_detection_train_loader,
)
from detectron2.engine import default_argument_parser, default_setup, launch
from detectron2.evaluation import (
    CityscapesInstanceEvaluator,
    CityscapesSemSegEvaluator,
    COCOEvaluator,
    COCOPanopticEvaluator,
    DatasetEvaluators,
    LVISEvaluator,
    PascalVOCDetectionEvaluator,
    SemSegEvaluator,
    inference_on_dataset,
    print_csv_format,
)
from detectron2.modeling import build_model
from detectron2.solver import build_lr_scheduler, build_optimizer
from detectron2.utils.events import (
    CommonMetricPrinter,
    EventStorage,
    JSONWriter,
    TensorboardXWriter,
)

import pickle
from detectron2.data import DatasetCatalog, MetadataCatalog

logger = logging.getLogger("detectron2")

def get_evaluator(cfg, dataset_name, output_folder=None):
    """
    Create evaluator(s) for a given dataset.
    This uses the special metadata "evaluator_type" associated with each builtin dataset.
    For your own dataset, you can simply create an evaluator manually in your
    script and do not have to worry about the hacky if-else logic here.
    """
    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 ["MY"]:
        evaluator_list.append(
            SemSegEvaluator(
                dataset_name,
                distributed=True,
                num_classes=cfg.MODEL.SEM_SEG_HEAD.NUM_CLASSES,
                ignore_label=cfg.MODEL.SEM_SEG_HEAD.IGNORE_VALUE,
                output_dir=output_folder,
            )
        )
        evaluator_list.append(COCOEvaluator(dataset_name, cfg, True, output_folder))
    if evaluator_type in ["sem_seg", "coco_panoptic_seg"]:
        evaluator_list.append(
            SemSegEvaluator(
                dataset_name,
                distributed=True,
                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, True, output_folder))
    if evaluator_type == "coco_panoptic_seg":
        evaluator_list.append(COCOPanopticEvaluator(dataset_name, output_folder))
    if evaluator_type == "cityscapes_instance":
        assert (
            torch.cuda.device_count() >= comm.get_rank()
        ), "CityscapesEvaluator currently do not work with multiple machines."
        return CityscapesInstanceEvaluator(dataset_name)
    if evaluator_type == "cityscapes_sem_seg":
        assert (
            torch.cuda.device_count() >= comm.get_rank()
        ), "CityscapesEvaluator currently do not work with multiple machines."
        return CityscapesSemSegEvaluator(dataset_name)
    if evaluator_type == "pascal_voc":
        return PascalVOCDetectionEvaluator(dataset_name)
    if evaluator_type == "lvis":
        return LVISEvaluator(dataset_name, cfg, True, output_folder)
    if len(evaluator_list) == 0:
        raise NotImplementedError(
            "no Evaluator for the dataset {} with the type {}".format(dataset_name, evaluator_type)
        )
    if len(evaluator_list) == 1:
        return evaluator_list[0]
    return DatasetEvaluators(evaluator_list)

def do_test(cfg, model):
    results = OrderedDict()
    for dataset_name in cfg.DATASETS.TEST:
        data_loader = build_detection_test_loader(cfg, dataset_name)
        evaluator = get_evaluator(
            cfg, dataset_name, os.path.join(cfg.OUTPUT_DIR, "inference", dataset_name)
        )
        results_i = inference_on_dataset(model, data_loader, evaluator)
        results[dataset_name] = results_i
        if comm.is_main_process():
            logger.info("Evaluation results for {} in csv format:".format(dataset_name))
            print_csv_format(results_i)
    if len(results) == 1:
        results = list(results.values())[0]
    return results

def do_train(cfg, model, resume=False):
    model.train()
    optimizer = build_optimizer(cfg, model)
    scheduler = build_lr_scheduler(cfg, optimizer)

    checkpointer = DetectionCheckpointer(
        model, cfg.OUTPUT_DIR, optimizer=optimizer, scheduler=scheduler
    )
    start_iter = (
        checkpointer.resume_or_load(cfg.MODEL.WEIGHTS, resume=resume).get("iteration", -1) + 1
    )
    max_iter = cfg.SOLVER.MAX_ITER

    periodic_checkpointer = PeriodicCheckpointer(
        checkpointer, cfg.SOLVER.CHECKPOINT_PERIOD, max_iter=max_iter
    )

    writers = (
        [
            CommonMetricPrinter(max_iter),
            JSONWriter(os.path.join(cfg.OUTPUT_DIR, "metrics.json")),
            TensorboardXWriter(cfg.OUTPUT_DIR),
        ]
        if comm.is_main_process()
        else []
    )

    # compared to "train_net.py", we do not support accurate timing and
    # precise BN here, because they are not trivial to implement in a small training loop
    data_loader = build_detection_train_loader(cfg)
    logger.info("Starting training from iteration {}".format(start_iter))
    with EventStorage(start_iter) as storage:
        for data, iteration in zip(data_loader, range(start_iter, max_iter)):
            iteration = iteration + 1
            storage.step()

            loss_dict = model(data)
            losses = sum(loss_dict.values())
            assert torch.isfinite(losses).all(), loss_dict

            loss_dict_reduced = {k: v.item() for k, v in comm.reduce_dict(loss_dict).items()}
            losses_reduced = sum(loss for loss in loss_dict_reduced.values())
            if comm.is_main_process():
                storage.put_scalars(total_loss=losses_reduced, **loss_dict_reduced)

            optimizer.zero_grad()
            losses.backward()
            optimizer.step()
            storage.put_scalar("lr", optimizer.param_groups[0]["lr"], smoothing_hint=False)
            scheduler.step()

            if (
                cfg.TEST.EVAL_PERIOD > 0
                and iteration % cfg.TEST.EVAL_PERIOD == 0
                and iteration != max_iter
            ):
                do_test(cfg, model)
                # Compared to "train_net.py", the test results are not dumped to EventStorage
                comm.synchronize()

            if iteration - start_iter > 5 and (iteration % 20 == 0 or iteration == max_iter):
                for writer in writers:
                    writer.write()
            periodic_checkpointer.step(iteration)

def setup(args):
    if 'MY_train' not in DatasetCatalog.list():
        fileRoot = "/home/skeptomist/train/wet/detectron2/"
        filename = "MY_dataset.pkl"
        file = open(fileRoot + filename, 'rb')
        data = pickle.load(file)
        file.close()

        classes = ["Core"]
        for d in ["train"]:
            DatasetCatalog.register("MY_" + d, lambda d=d: data)
            MetadataCatalog.get("MY_" + d).set(thing_classes=classes)
            MetadataCatalog.get("MY_" + d).set(evaluator_type="MY")

    """
    Create configs and perform basic setups.
    """
    cfg = get_cfg()
    cfg.merge_from_file(args.config_file)
    cfg.merge_from_list(args.opts)

    #cfg.merge_from_file(model_zoo.get_config_file("COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml"))
    cfg.DATASETS.TRAIN = ("MY_train",)
    cfg.DATASETS.TEST = ()   # no metrics implemented for this dataset
    cfg.DATALOADER.NUM_WORKERS = 4 #2
    cfg.MODEL.WEIGHTS = ""
    cfg.SOLVER.IMS_PER_BATCH = 1 #16 #2
    cfg.SOLVER.BASE_LR = 0.0002 #0.00025  # pick a good LR
    cfg.SOLVER.MAX_ITER = 1000 # 300 iterations seems good enough for this dataset; you will need to train longer for a practical dataset like 1000
    cfg.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE = 512   # faster, and good enough for this toy dataset (default: 512)
    cfg.MODEL.ROI_HEADS.NUM_CLASSES = 1 # only has one class
    cfg.MODEL.RETINANET.NUM_CLASSES = cfg.MODEL.ROI_HEADS.NUM_CLASSES
    cfg.MODEL.DEVICE = 'cuda'  # cpu or cuda

    cfg.freeze()
    default_setup(
        cfg, args
    )  # if you don't like any of the default setup, write your own setup code
    return cfg

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

    model = build_model(cfg)
    logger.info("Model:\n{}".format(model))
    if args.eval_only:
        DetectionCheckpointer(model, save_dir=cfg.OUTPUT_DIR).resume_or_load(
            cfg.MODEL.WEIGHTS, resume=args.resume
        )
        return do_test(cfg, model)

    distributed = comm.get_world_size() > 1
    if distributed:
        model = DistributedDataParallel(
            model, device_ids=[comm.get_local_rank()], broadcast_buffers=False
        )

    do_train(cfg, model, resume=args.resume)
    return do_test(cfg, model)

if __name__ == "__main__":
    args = default_argument_parser().parse_args()
    print("Command Line Args:", args)
    launch(
        main,
        args.num_gpus,
        num_machines=args.num_machines,
        machine_rank=args.machine_rank,
        dist_url=args.dist_url,
        args=(args,),
    )

what exact command you run:

CUDA_VISIBLE_DEVICES=0,1,2,3 python plain_train_net.py --num-gpus=1 --config-file=/..../detectron2/configs/COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml

CUDA_VISIBLE_DEVICES=0,1,2,3 python plain_train_net.py --num-gpus=2 --config-file=/..../detectron2/configs/COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml

full logs you observed:

for --num-gpus=1:

[09/26 00:39:39 d2.utils.events]: iter: 20 total_loss: 60.715 loss_cls: 13.017 loss_box_reg: 13.655 loss_rpn_cls: 2.048 loss_rpn_loc: 27.307 lr: 0.000004 max_mem: 1512M [09/26 00:39:52 d2.utils.events]: eta: 0:10:27 iter: 40 total_loss: 12.671 loss_cls: 2.315 loss_box_reg: 3.246 loss_rpn_cls: 0.282 loss_rpn_loc: 6.025 lr: 0.000008 max_mem: 1512M

+-----------------------------------------------------------------------------+ | Processes: GPU Memory | | GPU PID Type Process name Usage | |=============================================================================| | 0 10529 C python 2450MiB | +-----------------------------------------------------------------------------+

for --num-gpus=2 and changing cfg.SOLVER.IMS_PER_BATCH = 2 (from cfg.SOLVER.IMS_PER_BATCH = 1):

[09/26 00:42:51 d2.utils.events]: iter: 20 total_loss: 80.743 loss_box_reg: 2.978 loss_cls: 17.136 loss_rpn_cls: 2.445 loss_rpn_loc: 50.026 lr: 0.000004 max_mem: 1668M [09/26 00:43:06 d2.utils.events]: eta: 0:12:05 iter: 40 total_loss: 8.762 loss_box_reg: 0.633 loss_cls: 2.570 loss_rpn_cls: 0.515 loss_rpn_loc: 4.560 lr: 0.000008 max_mem: 1668M

+-----------------------------------------------------------------------------+ | Processes: GPU Memory | | GPU PID Type Process name Usage | |=============================================================================| | 0 10074 C /opt/conda/envs/d2-10.2-1.6/bin/python 2531MiB | | 1 10075 C /opt/conda/envs/d2-10.2-1.6/bin/python 2635MiB | +-----------------------------------------------------------------------------+

Expected behavior:

When using 2 GPUs instead of 1 and not changing anything else except cfg.SOLVER.IMS_PER_BATCH from 1 to 2, expected the ETA to decrease but instead it increased from eta: 0:10:27 to eta: 0:12:05

Environment:

----------------------  ------------------------------------------------------------------------------
sys.platform            linux
Python                  3.8.5 | packaged by conda-forge | (default, Sep 24 2020, 16:55:52) [GCC 7.5.0]
numpy                   1.19.2
detectron2              0.2.1 @/opt/conda/envs/d2-10.2-1.6/lib/python3.8/site-packages/detectron2
Compiler                GCC 7.3
CUDA compiler           CUDA 10.2
detectron2 arch flags   sm_35, sm_37, sm_50, sm_52, sm_60, sm_61, sm_70, sm_75
DETECTRON2_ENV_MODULE   <not set>
PyTorch                 1.6.0 @/opt/conda/envs/d2-10.2-1.6/lib/python3.8/site-packages/torch
PyTorch debug build     False
GPU available           True
GPU 0,1                 Tesla K80
CUDA_HOME               /usr/local/cuda
Pillow                  7.2.0
torchvision             0.7.0 @/opt/conda/envs/d2-10.2-1.6/lib/python3.8/site-packages/torchvision
torchvision arch flags  sm_35, sm_50, sm_60, sm_70, sm_75
fvcore                  0.1.2.post20200926
----------------------  ------------------------------------------------------------------------------
PyTorch built with:
  - GCC 7.3
  - C++ Version: 201402
  - Intel(R) Math Kernel Library Version 2019.0.5 Product Build 20190808 for Intel(R) 64 architecture applications
  - Intel(R) MKL-DNN v1.5.0 (Git Hash e2ac1fac44c5078ca927cb9b90e1b3066a0b2ed0)
  - OpenMP 201511 (a.k.a. OpenMP 4.5)
  - NNPACK is enabled
  - CPU capability usage: AVX2
  - CUDA Runtime 10.2
  - NVCC architecture flags: -gencode;arch=compute_37,code=sm_37;-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75
  - CuDNN 7.6.5
  - Magma 2.5.2
  - Build settings: BLAS=MKL, BUILD_TYPE=Release, CXX_FLAGS= -Wno-deprecated -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -fopenmp -DNDEBUG -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -DUSE_VULKAN_WRAPPER -O2 -fPIC -Wno-narrowing -Wall -Wextra -Werror=return-type -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wno-sign-compare -Wno-unused-parameter -Wno-unused-variable -Wno-unused-function -Wno-unused-result -Wno-unused-local-typedefs -Wno-strict-overflow -Wno-strict-aliasing -Wno-error=deprecated-declarations -Wno-stringop-overflow -Wno-error=pedantic -Wno-error=redundant-decls -Wno-error=old-style-cast -fdiagnostics-color=always -faligned-new -Wno-unused-but-set-variable -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math -Werror=format -Wno-stringop-overflow, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, USE_CUDA=ON, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_MKL=ON, USE_MKLDNN=ON, USE_MPI=OFF, USE_NCCL=ON, USE_NNPACK=ON, USE_OPENMP=ON, USE_STATIC_DISPATCH=OFF,

If your issue looks like an installation issue / environment issue, please first try to solve it with the instructions in https://detectron2.readthedocs.io/tutorials/install.html#common-installation-issues

facebookresearch / detectron2