Code of segmentation on MSCOCO

[zhaozhengChen]

Hi authors, deeplab-pytorch only support VOC and COCO-stuff. Could you please release your code for segmentation on MSCOCO2014? Thanks!

[PengtaoJiang]

@jbeomlee93, I have the same request. Could you provide the deeplab-pytorch code for MS COCO? I want to know the hyper-parameters in your experiments.

[jbeomlee93]

Dear @zhaozhengChen and @PengtaoJiang,

Sorry for the late reply.

We used the below configs for training a segmentation network.

EXP:
    ID: coco
    OUTPUT_DIR: data_coco_RIB

DATASET:
    NAME: coco
    ROOT: ./data/datasets/coco_2014
    LABELS: ./data_coco/datasets/coco/labels.txt
    N_CLASSES: 81
    IGNORE_LABEL: 255
    SCALES: [0.5, 0.75, 1.0, 1.25, 1.5, 1.75, 2.0]
    SPLIT:
        TRAIN: train_2014
        VAL: val_2014
        TEST: test

DATALOADER:
    NUM_WORKERS: 8

IMAGE:
    MEAN:
        R: 122.675
        G: 116.669
        B: 104.008
    SIZE:
        BASE: # None
        TRAIN: 481
        TEST: 513

MODEL:
    NAME: DeepLabV2_ResNet101_MSC
    N_BLOCKS: [3, 4, 23, 3]
    ATROUS_RATES: [6, 12, 18, 24]
    INIT_MODEL: data/models/imagenet/deeplabv1_resnet101-imagenet.pth

SOLVER:
    BATCH_SIZE:
        TRAIN: 10
        TEST: 1
    ITER_MAX: 100000
    ITER_SIZE: 1
    ITER_SAVE: 2500
    ITER_TB: 20
    LR_DECAY: 10
    LR: 2.5e-4
    MOMENTUM: 0.9
    OPTIMIZER: sgd
    POLY_POWER: 0.9
    WEIGHT_DECAY: 5.0e-4
    AVERAGE_LOSS: 20
    FREEZE_BN: True
    BALANCE_LOSS: True

CRF:
    ITER_MAX: 10
    POS_W: 3
    POS_XY_STD: 1
    BI_W: 4
    BI_XY_STD: 67
    BI_RGB_STD: 3

We also used balanced cross-entropy loss. Please refer to this.

Thanks!

[Italy2006]

Dear @zhaozhengChen and @PengtaoJiang,

Sorry for the late reply.

We used the below configs for training a segmentation network.

EXP:
    ID: coco
    OUTPUT_DIR: data_coco_RIB

DATASET:
    NAME: coco
    ROOT: ./data/datasets/coco_2014
    LABELS: ./data_coco/datasets/coco/labels.txt
    N_CLASSES: 81
    IGNORE_LABEL: 255
    SCALES: [0.5, 0.75, 1.0, 1.25, 1.5, 1.75, 2.0]
    SPLIT:
        TRAIN: train_2014
        VAL: val_2014
        TEST: test

DATALOADER:
    NUM_WORKERS: 8

IMAGE:
    MEAN:
        R: 122.675
        G: 116.669
        B: 104.008
    SIZE:
        BASE: # None
        TRAIN: 481
        TEST: 513

MODEL:
    NAME: DeepLabV2_ResNet101_MSC
    N_BLOCKS: [3, 4, 23, 3]
    ATROUS_RATES: [6, 12, 18, 24]
    INIT_MODEL: data/models/imagenet/deeplabv1_resnet101-imagenet.pth

SOLVER:
    BATCH_SIZE:
        TRAIN: 10
        TEST: 1
    ITER_MAX: 100000
    ITER_SIZE: 1
    ITER_SAVE: 2500
    ITER_TB: 20
    LR_DECAY: 10
    LR: 2.5e-4
    MOMENTUM: 0.9
    OPTIMIZER: sgd
    POLY_POWER: 0.9
    WEIGHT_DECAY: 5.0e-4
    AVERAGE_LOSS: 20
    FREEZE_BN: True
    BALANCE_LOSS: True

CRF:
    ITER_MAX: 10
    POS_W: 3
    POS_XY_STD: 1
    BI_W: 4
    BI_XY_STD: 67
    BI_RGB_STD: 3

We also used balanced cross-entropy loss. Please refer to this.

Thanks!

Thanks for your work, but i want to know how many GPU you use for coco training? And how to use the freeze bn in the network?