Error! Train with the pretrained model

[lamnt-nd994]

“Hello, I'm a beginner . I’m encountering an error when loading pretrained weights glint360k_r18 to continue training with my new dataset. The error message is: ‘ValueError: Layer count mismatch when loading weights from file. Model expected 63 layers, found 62 saved layers.’”

Thanks

[leondgarse]

Those shared weights are basic model only. Try loading using the basic_model without header, not providing to tt:

import models

mm = models.buildin_models(
      "r18",  # Or "r18" / "r34" / "r100"
      dropout=0,
      emb_shape=512,
      output_layer='E',
      bn_momentum=0.9,
      bn_epsilon=1e-5,
      use_bias=True,
      scale=True,
      activation='PReLU'
)
mm.load_weights('glint360k_cosface_r18_fp16_0.1.h5')
...

[lamnt-nd994]

Thank you for your response. I have successfully run it. Could you please explain why the accuracy is very low when training on the https://github.com/X-zhangyang/Asian-Face-Image-Dataset-AFD-dataset and testing on LFW, CFP_FP, and AgeDB_30? with lr= 0.1

[leondgarse]

Try freezing backbone and training header only first:

tt.train(
    [
        {"loss": losses.ArcfaceLoss(scale=64), "epoch": 2, "bottleneckOnly": True},
        {"loss": losses.ArcfaceLoss(scale=64), "epoch": 17},
    ]
)

And also use a smaller learning_rate like lr_base=0.025.

[lamnt-nd994]

I tried freezing backbone and training header only first but accuracy in "cfp_fp" was only 0.93 while Ported Models r18: 0.977143

[leondgarse]

Maybe just a difference in image encoding qulity when saving those bin files, refer Reproduce the results #110. While freezing backbone, the accuracy shouldn't change. May test the basic_model accuracy firstly.

[lamnt-nd994]

I got it. Thank a lot