Question FPS about simple_detectron2_efficientdet HOT 3 OPEN

Hi @wenjun90

The inference time of EfficientDet-D0 is ~54ms/img on a Tesla V100 GPU, I think there is still room for optimization.

If you want to train D1, according to the official config:

            name='efficientdet-d1',
            backbone_name='efficientnet-b1',
            image_size=640,
            fpn_num_filters=88,
            fpn_cell_repeats=4,
            box_class_repeats=3,

just modify the configs/EfficientDet_D0.yaml:

MODEL:
    WEIGHTS: "/home/b1_detectron2.pth"
    EfficientNet:
        VERSION: 1
    FPN:
        OUT_CHANNELS: 88
        REPEAT: 4
    RETINANET:
        NUM_CONVS: 3

and of course, you need to adjust the input size accordingly.

from simple_detectron2_efficientdet.

Hi @zzzxxxttt
I tried the training with D3 with config like that:
BASE: "Base-RetinaNet.yaml"
MODEL:
BACKBONE:
NAME: "build_retinanet_efficientnet_bifpn_backbone"
WEIGHTS: "efficientdet-d3.pth"
MASK_ON: False
EfficientNet:
VERSION: 3
NORM: 'SyncBN'
FREEZE_AT: -1
FPN:
IN_FEATURES: ["stride-8", "stride-16", "stride-32"]
OUT_CHANNELS: 160
IN_FEATURE_P6P7: 'stride-32'
REPEAT: 6
NORM: 'SyncBN'
RETINANET:
NUM_CONVS: 4
NORM: 'SyncBN'
FOCAL_LOSS_GAMMA: 1.5
SOLVER:
LR_SCHEDULER_NAME: "WarmupCosineLR"
BASE_LR: 0.02
IMS_PER_BATCH: 4
STEPS: (210000, 250000)
MAX_ITER: 270000
TEST:
EVAL_PERIOD: 7500

and I met input_size comme paper but, result is not better d0.

Have you tried with d2,d3?

Thanks

from simple_detectron2_efficientdet.

Currently I have to focus on some personal affairs, so the validation of other models is delayed. I saw you use batchsize=4 with lr=0.02, according to the linear scaling rule, batchsize=4 corresponds to lr=0.02*(4/16)=0.005, and the MAX_ITER should be set to 300*117287/4~=8800000.

from simple_detectron2_efficientdet.