when I set the svhn’s pretrain weights for digit detection, i try to run the eval.py but a error occurs："ValueError: Shapes (1, 1, 1024, 27) and (1, 1, 1024, 45) are incompatible"

[little-carrrot-errr]

I change the pretrained part in configs/svhn.json file like below:

"pretrained" : { "keras_format": "configs/svhn/weights.h5", "darknet_format": "yolov3.weights" }

[little-carrrot-errr]

I use the command : python eval.py -c configs/svhn.json

[DanielNehemiah]

Facing the same issue... Anyone have a solution?

[akspad]

I changed the svhn.json file to the following:

{
    "model" : {
        "anchors":              [10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326],
        "labels":               ["1", "2", "3", "4", "5", "6", "7", "8", "9", "0"],
        "net_size":               288
    },
    "pretrained" : {
        "keras_format":             "configs/svhn/weights.h5",
        "darknet_format":           ""
    },
    "train" : {
        "min_size":             288,
        "max_size":             288,
        "num_epoch":            40,
        "train_image_folder":   "tests/dataset/svhn/imgs",
        "train_annot_folder":   "tests/dataset/svhn/anns",
        "valid_image_folder":   "",
        "valid_annot_folder":   "",
        "batch_size":           2,
        "learning_rate":        1e-4,
        "save_folder":         "configs/svhn",
        "jitter":               false
    }
}

[unknownSK]

@akspad Were you able to run it by making above mentioned changes?

[aberger7]

Has anyone gotten this to run?? I'm having the exact same issue and don't know what to do. @unknownSK @little-carrrot-errr @DanielNehemiah @akspad

[Murtaza-Kazmi]

Has anyone found the solution? I have been stuck on this for quite some time now 👎

[patrr2]

Has anyone found the solution? I have been stuck on this for quite some time now 👎

Did you try the solution proposed by @akspad? For me it worked

[Murtaza-Kazmi]

I did spend a good time resolving this back then, so I must have :)

[le-thanh-tung]

i added "skip_detect_layer=False" to create_model function