How to evaluate model with current weights?

Hi,

While I am running a pretrained CRDNN model, I finetune it on some data with .fit() method. Then, I would like to see its performance with .evaluate() function.

However, when I run the evaluate() function the log is something like this:

  0%|          | 0/16 [00:00<?, ?it/s]
  6%|▋         | 1/16 [00:22<05:42, 22.84s/it]
 12%|█▎        | 2/16 [00:29<03:03, 13.12s/it]
 19%|█▉        | 3/16 [00:54<04:04, 18.84s/it]
 25%|██▌       | 4/16 [01:21<04:23, 21.98s/it]
 31%|███▏      | 5/16 [01:47<04:18, 23.45s/it]
 38%|███▊      | 6/16 [02:17<04:17, 25.71s/it]
 44%|████▍     | 7/16 [02:43<03:50, 25.65s/it]
 50%|█████     | 8/16 [03:11<03:31, 26.41s/it]
 56%|█████▋    | 9/16 [03:20<02:27, 21.08s/it]
 62%|██████▎   | 10/16 [03:26<01:37, 16.31s/it]
 69%|██████▉   | 11/16 [03:40<01:17, 15.56s/it]
 75%|███████▌  | 12/16 [04:08<01:18, 19.58s/it]
 81%|████████▏ | 13/16 [04:38<01:08, 22.67s/it]
 88%|████████▊ | 14/16 [04:50<00:38, 19.27s/it]
 94%|█████████▍| 15/16 [04:57<00:15, 15.83s/it]
100%|██████████| 16/16 [05:35<00:00, 22.38s/it]
100%|██████████| 16/16 [05:35<00:00, 20.98s/it]
speechbrain.utils.train_logger - Epoch loaded: 15 - test loss: 1.50, test CER: 31.43, test WER: 37.69

As you can see, it seems that model is evaluated on Epoch 15 which probably indicates that it loads weights from that epoch instead of using the weights which have just been finetuned. Whenever I do this finetuning and then evaluate(), the scores are exactly the same (up to 8 decimal digits). How do I overcome that?

TLDR; I want to use current weights with .evaluate() function, not with weights loaded from best epoch. How do I achieve that?

[TParcollet]

(In practice, you can, I believe, just pass an argument that forces to use the latest instead of the best using the keys).

[KacperKubara]

brain.checkpointer=None works. I looked over the source code for brain.checkpointer.delete_checkpoints(num_to_keep=0) but this delets the checkpoints from the directory which is not exactly what I want to do.