question about number of parameters

[huchinlp]

Hi, I trained some models using the pre-defined configurations but
the number of parameters is much larger than what you reported (55.1M vs. 31.5M).

Configuration:
HAT_iwslt14deen_titanxp@168.8ms_bleu@34.8.yml

Here is the code I used to calculate the number of parameters:
(embedding layers are excluded)

import torch

m = torch.load('checkpoints/iwslt14.de-en/subtransformer/HAT_iwslt14deen_titanxp@\
168.8ms_bleu@34.8/checkpoint_best.pt', map_location='cpu')
m = m['model']

n = 0
for k in m:
    if 'emb' not in k:
        n += m[k].numel()

print(n)

[Hanrui-Wang]

Hi Huchi,

Sorry for my late reply, I was too busy in the past several weeks.
It is required to extract the SubTransformer weights from the checkpoints we shared to get the correct model size. The reason Is that we finetuned a SubTransformer by always sampling that SubTransformer from the SuperTransformer. So the checkpoint contains all the weights of a SuperTransformer, but we need to only use the SubTransformer part to do testing and profiling.

Please refer to train.py line 61 to 64 for how to get SubTransformer model size:

# Log model size
if args.train_subtransformer:
    print(f"| SubTransformer size (without embedding weights): {model.get_sampled_params_numel(utils.get_subtransformer_config(args))}")
    embed_size = args.decoder_embed_dim_subtransformer * len(task.tgt_dict)
    print(f"| Embedding layer size: {embed_size} \n")

Thanks!
Hanrui

[huchinlp]

Awsome, thanks ! Hanrui Wang <notifications@github.com> 于 2020年11月10日周二 上午5:57写道：

…

Hi Huchi, Sorry for my late reply, I was too busy in the past several weeks. It is required to extract the SubTransformer weights from the checkpoints we shared to get the correct model size. The reason Is that we finetuned a SubTransformer by always sampling that SubTransformer from the SuperTransformer. So the checkpoint contains all the weights of a SuperTransformer, but we need to only use the SubTransformer part to do testing and profiling. Please refer to train.py line 61 to 64 for how to get SubTransformer model size: # Log model sizeif args.train_subtransformer: print(f"| SubTransformer size (without embedding weights): {model.get_sampled_params_numel(utils.get_subtransformer_config(args))}") embed_size = args.decoder_embed_dim_subtransformer * len(task.tgt_dict) print(f"| Embedding layer size: {embed_size} \n") Thanks! Hanrui — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#5 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AJXERISK56RZWHGVLTAP53DSPBQUVANCNFSM4SOML3BQ> .