How to resume training

[abhishekbu]

How do i resume training, what changes shiuld i make to continue training from where it left

[molimat]

You can just start from the new checkpoint you generated from before. If you take a look on "train.py" you see that every epoch is generating a check point on /checkpoints folder. You can use it as a new weight to start your new train from.

[Tomiha]

Hmmm, got the same question.

Already tried using the ModelCheckpoint weights, but I receive the following error when doing that:

ValueError: Shapes (255,) and (75,) are incompatible

1. Trained the network with the original weights (yolov3.tf) for 10 epochs [colab_gpu.ipynb step 5]
2. On epoch 10 'yolov3_train_10.tf' is created
3. Restart the training, but instead of passing 'yolov3.tf', I gave it 'yolov3_train_10.tf'
4. Throws above error..

Am I missing something?

[amh28]

Hmmm, got the same question.

Already tried using the ModelCheckpoint weights, but I receive the following error when doing that:

ValueError: Shapes (255,) and (75,) are incompatible

1. Trained the network with the original weights (yolov3.tf) for 10 epochs [colab_gpu.ipynb step 5]
2. On epoch 10 'yolov3_train_10.tf' is created
3. Restart the training, but instead of passing 'yolov3.tf', I gave it 'yolov3_train_10.tf'
4. Throws above error..

Am I missing something?

@Tomiha Did you solve it?

Regards

[Tomiha]

Hi @amh28,

Yes, I got it working. If I remember correctly, there was an issue with the number of classes I passed through when training from previous weights.