Unable to obtain the same average error as Table 1 in the paper

[ghost]

The average error on celeba is 8.25 in the paper, but my result is 10.14, which is obtained by directly using the codes in this repo. I list the configurations I used below, is there any problem?

in configs.json

    "celeba": {
        "path": "data/celeba",
        "all_tasks": ["0", "1", "2", "3", "4", "5", "6", "7", "8", "9", "10", "11", "12", "13", "14", "15", "16", "17", "18", "19", "20",
                      "21", "22", "23", "24", "25", "26", "27", "28", "29", "30", "31", "32", "33", "34", "35", "36", "37", "38", "39"],
        "img_rows": 64,
        "img_cols": 64
    }

in sample.json

{
    "optimizer": "Adam", 
    "batch_size": 256, 
    "lr": 0.0005,
    "dataset": "celeba", 
    "tasks": ["0", "1", "2", "3", "4", "5", "6", "7", "8", "9", "10", "11", "12", "13", "14", "15", "16", "17", "18", "19", "20",
                  "21", "22", "23", "24", "25", "26", "27", "28", "29", "30", "31", "32", "33", "34", "35", "36", "37", "38", "39"],
    "normalization_type": "loss+",
    "algorithm": "mgda",
    "use_approximation": true,
    "scales": {"0":0.025, "1":0.025, "2":0.025, "3":0.025, "4":0.025, "5":0.025, "6":0.025, "7":0.025, "8":0.025, "9":0.025, "10":0.025, 
               "11":0.025, "12":0.025, "13":0.025, "14":0.025, "15":0.025, "16":0.025, "17":0.025, "18":0.025, "19":0.025, "20":0.025, 
               "21":0.025, "22":0.025, "23":0.025, "24":0.025, "25":0.025, "26":0.025, "27":0.025, "28":0.025, "29":0.025, "30":0.025, 
               "31":0.025, "32":0.025, "33":0.025, "34":0.025, "35":0.025, "36":0.025, "37":0.025, "38":0.025, "39":0.025}
}

and the learning rate is multiplied by 0.85 every 10 epochs:

    if (epoch+1) % 10 == 0:
        # Every 50 epoch, half the LR
        for param_group in optimizer.param_groups:
            param_group['lr'] *= 0.85
        logger.info('Multiply the learning rate by {} [{} steps]'.format(0.85, n_iter))

and I train the model on two GPUs.

[keil555]

The average error on celeba is 8.25 in the paper, but my result is 10.14, which is obtained by directly using the codes in this repo. I list the configurations I used below, is there any problem?

in configs.json

    "celeba": {
        "path": "data/celeba",
        "all_tasks": ["0", "1", "2", "3", "4", "5", "6", "7", "8", "9", "10", "11", "12", "13", "14", "15", "16", "17", "18", "19", "20",
                      "21", "22", "23", "24", "25", "26", "27", "28", "29", "30", "31", "32", "33", "34", "35", "36", "37", "38", "39"],
        "img_rows": 64,
        "img_cols": 64
    }

in sample.json

{
    "optimizer": "Adam", 
    "batch_size": 256, 
    "lr": 0.0005,
    "dataset": "celeba", 
    "tasks": ["0", "1", "2", "3", "4", "5", "6", "7", "8", "9", "10", "11", "12", "13", "14", "15", "16", "17", "18", "19", "20",
                  "21", "22", "23", "24", "25", "26", "27", "28", "29", "30", "31", "32", "33", "34", "35", "36", "37", "38", "39"],
    "normalization_type": "loss+",
    "algorithm": "mgda",
    "use_approximation": true,
    "scales": {"0":0.025, "1":0.025, "2":0.025, "3":0.025, "4":0.025, "5":0.025, "6":0.025, "7":0.025, "8":0.025, "9":0.025, "10":0.025, 
               "11":0.025, "12":0.025, "13":0.025, "14":0.025, "15":0.025, "16":0.025, "17":0.025, "18":0.025, "19":0.025, "20":0.025, 
               "21":0.025, "22":0.025, "23":0.025, "24":0.025, "25":0.025, "26":0.025, "27":0.025, "28":0.025, "29":0.025, "30":0.025, 
               "31":0.025, "32":0.025, "33":0.025, "34":0.025, "35":0.025, "36":0.025, "37":0.025, "38":0.025, "39":0.025}
}

and the learning rate is multiplied by 0.85 every 10 epochs:

    if (epoch+1) % 10 == 0:
        # Every 50 epoch, half the LR
        for param_group in optimizer.param_groups:
            param_group['lr'] *= 0.85
        logger.info('Multiply the learning rate by {} [{} steps]'.format(0.85, n_iter))

and I train the model on two GPUs.

Hi, Ghost! I also failed to obtain the same results in Table 1. Have you solved this problem yet?