How to retrain the NER and Relational model from last checkpoint?

[karndeepsingh]

Hi,
I have to train 15000 sentences for NER and the RELATIONAL model. I am planning to train it on 11GB GPU batch-wise. So, I can retrain the model always from the last checkpoint or best weights.

Thanks

[ljvmiranda921]

Hi @karndeepsingh ,

When you first train via spaCy, you'd always get the model-best and model-last in the output directory that you supplied. If you want to resume training from them, you can supply their paths to the source parameter of your components in the configuration file. Something like this:

[components.ner]
source = "./output/model-last"  # or model-best

[components.tok2vec]
source = "./output/model-last"  # or model-best

Note that, it won't always be 100% equivalent because the rest of the training settings (like the optimizer, batcher, etc.) will be reset.

[karndeepsingh]

Hi @karndeepsingh ,

When you first train via spaCy, you'd always get the model-best and model-last in the output directory that you supplied. If you want to resume training from them, you can supply their paths to the source parameter of your components in the configuration file. Something like this:

[components.ner]
source = "./output/model-last"  # or model-best

[components.tok2vec]
source = "./output/model-last"  # or model-best

Note that, it won't always be 100% equivalent because the rest of the training settings (like the optimizer, batcher, etc.) will be reset.

But the training will intiate considering the last trained weights right?

[karndeepsingh]

Hi @karndeepsingh ,

When you first train via spaCy, you'd always get the model-best and model-last in the output directory that you supplied. If you want to resume training from them, you can supply their paths to the source parameter of your components in the configuration file. Something like this:

[components.ner]
source = "./output/model-last"  # or model-best

[components.tok2vec]
source = "./output/model-last"  # or model-best

Note that, it won't always be 100% equivalent because the rest of the training settings (like the optimizer, batcher, etc.) will be reset.

Thanks For answering. I will do the changes.

I have one more query please help me to understand. I have been trying to train NER+ RELATION model together and I am getting the following error shown below, please help me to understand what is the problem and what needs to change? Thanks again.