Reloading model and params from Checkpoint about xlm HOT 5 CLOSED

Ah yes, I also had this because I tried to reload on a single GPU a model trained on multiple GPU. Problem in that case is that with multi-GPU, the model is encapsulated in a module (this is why you have all the extra .module in the reloaded checkpoint parameters).

See 34825ea#diff-e750911d9404a6f817e2015251a4a654R458
I added a commented line. Comment out:
getattr(self, name).load_state_dict(data[name])
and uncomment:
getattr(self, name).load_state_dict({k[len('module.'):]: v for k, v in data[name].items()})
it should solve the issue.

from xlm.

34825ea
should do the trick. You still have to provide the parameters though. What you can do is simply copy paste the "running command" at the beginning of the train.log of the experiment with the checkpoint you want to reload, and simply add --reload_checkpoint EXP_PATH/checkpoint.pth

from xlm.

Thank you for the quick response.
Now I'm getting runtime error ;

Traceback (most recent call last):
File "train.py", line 330, in
main(params)
File "train.py", line 250, in main
trainer = SingleTrainer(model, data, params)
File "/NMT/XLM/src/trainer.py", line 704, in init
super().init(data, params)
File "/NMT/XLM/src/trainer.py", line 94, in init
self.reload_checkpoint()
File "/NMT/XLM/src/trainer.py", line 457, in reload_checkpoint
getattr(self, name).load_state_dict(data[name])
File "/usr/local/lib/python3.5/dist-packages/torch/nn/modules/module.py", line 769, in load_state_dict
self.class.name, "\n\t".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for TransformerModel:
Missing key(s) in state_dict: "position_embeddings.weight", "lang_embeddings.weight", "embeddings.weight", "layer_norm_emb.bias", "layer_norm_emb.weight", "attentions.0.q_lin.bias", "attentions.0.q_lin.weight", "attentions.0.k_lin.bias", "attentions.0.k_lin.weight", "attentions.0.v_lin.bias", "attentions.0.v_lin.weight", "attentions.0.out_lin.bias", "attentions.0.out_lin.weight", "attentions.1.q_lin.bias", "attentions.1.q_lin.weight", "attentions.1.k_lin.bias", "attentions.1.k_lin.weight", "attentions.1.v_lin.bias", "attentions.1.v_lin.weight", "attentions.1.out_lin.bias", "attentions.1.out_lin.weight", "attentions.2.q_lin.bias", "attentions.2.q_lin.weight", "attentions.2.k_lin.bias", "attentions.2.k_lin.weight", "attentions.2.v_lin.bias", "attentions.2.v_lin.weight", "attentions.2.out_lin.bias", "attentions.2.out_lin.weight", "attentions.3.q_lin.bias", "attentions.3.q_lin.weight", "attentions.3.k_lin.bias", "attentions.3.k_lin.weight", "attentions.3.v_lin.bias", "attentions.3.v_lin.weight", "attentions.3.out_lin.bias", "attentions.3.out_lin.weight", "attentions.4.q_lin.bias", "attentions.4.q_lin.weight", "attentions.4.k_lin.bias", "attentions.4.k_lin.weight", "attentions.4.v_lin.bias", "attentions.4.v_lin.weight", "attentions.4.out_lin.bias", "attentions.4.out_lin.weight", "attentions.5.q_lin.bias", "attentions.5.q_lin.weight", "attentions.5.k_lin.bias", "attentions.5.k_lin.weight", "attentions.5.v_lin.bias", "attentions.5.v_lin.weight", "attentions.5.out_lin.bias", "attentions.5.out_lin.weight", "layer_norm1.0.bias", "layer_norm1.0.weight", "layer_norm1.1.bias", "layer_norm1.1.weight", "layer_norm1.2.bias", "layer_norm1.2.weight", "layer_norm1.3.bias", "layer_norm1.3.weight", "layer_norm1.4.bias", "layer_norm1.4.weight", "layer_norm1.5.bias", "layer_norm1.5.weight", "ffns.0.lin1.bias", "ffns.0.lin1.weight", "ffns.0.lin2.bias", "ffns.0.lin2.weight", "ffns.1.lin1.bias", "ffns.1.lin1.weight", "ffns.1.lin2.bias", "ffns.1.lin2.weight", "ffns.2.lin1.bias", "ffns.2.lin1.weight", "ffns.2.lin2.bias", "ffns.2.lin2.weight", "ffns.3.lin1.bias", "ffns.3.lin1.weight", "ffns.3.lin2.bias", "ffns.3.lin2.weight", "ffns.4.lin1.bias", "ffns.4.lin1.weight", "ffns.4.lin2.bias", "ffns.4.lin2.weight", "ffns.5.lin1.bias", "ffns.5.lin1.weight", "ffns.5.lin2.bias", "ffns.5.lin2.weight", "layer_norm2.0.bias", "layer_norm2.0.weight", "layer_norm2.1.bias", "layer_norm2.1.weight", "layer_norm2.2.bias", "layer_norm2.2.weight", "layer_norm2.3.bias", "layer_norm2.3.weight", "layer_norm2.4.bias", "layer_norm2.4.weight", "layer_norm2.5.bias", "layer_norm2.5.weight", "pred_layer.proj.bias", "pred_layer.proj.weight".
Unexpected key(s) in state_dict: "module.position_embeddings.weight", "module.lang_embeddings.weight", "module.embeddings.weight", "module.layer_norm_emb.weight", "module.layer_norm_emb.bias", "module.attentions.0.q_lin.weight", "module.attentions.0.q_lin.bias", "module.attentions.0.k_lin.weight", "module.attentions.0.k_lin.bias", "module.attentions.0.v_lin.weight", "module.attentions.0.v_lin.bias", "module.attentions.0.out_lin.weight", "module.attentions.0.out_lin.bias", "module.attentions.1.q_lin.weight", "module.attentions.1.q_lin.bias", "module.attentions.1.k_lin.weight", "module.attentions.1.k_lin.bias", "module.attentions.1.v_lin.weight", "module.attentions.1.v_lin.bias", "module.attentions.1.out_lin.weight", "module.attentions.1.out_lin.bias", "module.attentions.2.q_lin.weight", "module.attentions.2.q_lin.bias", "module.attentions.2.k_lin.weight", "module.attentions.2.k_lin.bias", "module.attentions.2.v_lin.weight", "module.attentions.2.v_lin.bias", "module.attentions.2.out_lin.weight", "module.attentions.2.out_lin.bias", "module.attentions.3.q_lin.weight", "module.attentions.3.q_lin.bias", "module.attentions.3.k_lin.weight", "module.attentions.3.k_lin.bias", "module.attentions.3.v_lin.weight", "module.attentions.3.v_lin.bias", "module.attentions.3.out_lin.weight", "module.attentions.3.out_lin.bias", "module.attentions.4.q_lin.weight", "module.attentions.4.q_lin.bias", "module.attentions.4.k_lin.weight", "module.attentions.4.k_lin.bias", "module.attentions.4.v_lin.weight", "module.attentions.4.v_lin.bias", "module.attentions.4.out_lin.weight", "module.attentions.4.out_lin.bias", "module.attentions.5.q_lin.weight", "module.attentions.5.q_lin.bias", "module.attentions.5.k_lin.weight", "module.attentions.5.k_lin.bias", "module.attentions.5.v_lin.weight", "module.attentions.5.v_lin.bias", "module.attentions.5.out_lin.weight", "module.attentions.5.out_lin.bias", "module.layer_norm1.0.weight", "module.layer_norm1.0.bias", "module.layer_norm1.1.weight", "module.layer_norm1.1.bias", "module.layer_norm1.2.weight", "module.layer_norm1.2.bias", "module.layer_norm1.3.weight", "module.layer_norm1.3.bias", "module.layer_norm1.4.weight", "module.layer_norm1.4.bias", "module.layer_norm1.5.weight", "module.layer_norm1.5.bias", "module.ffns.0.lin1.weight", "module.ffns.0.lin1.bias", "module.ffns.0.lin2.weight", "module.ffns.0.lin2.bias", "module.ffns.1.lin1.weight", "module.ffns.1.lin1.bias", "module.ffns.1.lin2.weight", "module.ffns.1.lin2.bias", "module.ffns.2.lin1.weight", "module.ffns.2.lin1.bias", "module.ffns.2.lin2.weight", "module.ffns.2.lin2.bias", "module.ffns.3.lin1.weight", "module.ffns.3.lin1.bias", "module.ffns.3.lin2.weight", "module.ffns.3.lin2.bias", "module.ffns.4.lin1.weight", "module.ffns.4.lin1.bias", "module.ffns.4.lin2.weight", "module.ffns.4.lin2.bias", "module.ffns.5.lin1.weight", "module.ffns.5.lin1.bias", "module.ffns.5.lin2.weight", "module.ffns.5.lin2.bias", "module.layer_norm2.0.weight", "module.layer_norm2.0.bias", "module.layer_norm2.1.weight", "module.layer_norm2.1.bias", "module.layer_norm2.2.weight", "module.layer_norm2.2.bias", "module.layer_norm2.3.weight", "module.layer_norm2.3.bias", "module.layer_norm2.4.weight", "module.layer_norm2.4.bias", "module.layer_norm2.5.weight", "module.layer_norm2.5.bias", "module.pred_layer.proj.weight", "module.pred_layer.proj.bias".

from xlm.

Thanks !!

from xlm.

Ah yes, I also had this because I tried to reload on a single GPU a model trained on multiple GPU. Problem in that case is that with multi-GPU, the model is encapsulated in a module (this is why you have all the extra .module in the reloaded checkpoint parameters).

See 34825ea#diff-e750911d9404a6f817e2015251a4a654R458
I added a commented line. Comment out:
getattr(self, name).load_state_dict(data[name])
and uncomment:
getattr(self, name).load_state_dict({k[len('module.'):]: v for k, v in data[name].items()})
it should solve the issue.

This solved my issues, when I trained TLM on multi gpu's and translating using just 1 gpu.

from xlm.