Transformer Pre-trained weight

[hshen14]

I would like to run the reproduce the bleu for transformer model. May I know whether there is pre-trained weight for sharing? Thanks. @Superjomn @panyx0718 @luotao1

https://github.com/PaddlePaddle/models/blob/develop/fluid/PaddleNLP/neural_machine_translation/transformer/README_cn.md

[panyx0718]

@guoshengCS @kuke

[guoshengCS]

We upload the pre-trained weight and data used in Transformer base model just now. Please download from http://transformer-model-data.bj.bcebos.com/wmt16_ende_data_bpe_clean.tar.gz for data and http://transformer-model-data.bj.bcebos.com/iter_100000.infer.model.tar.gz for pre-trained weight.

[hshen14]

Thanks @guoshengCS. I tried the inference on CPU and the speed seems very slow. I have some questions:

1. Is there any instructions to measure the performance? Do you have the total inference time on CPU?
2. Some sentences looks abnormal during inference, with @ and 乱码. Is it okay?
3. Would you please show the corresponding BLEU with the model you provided?
4. Would you please share gen_data/mosesdecoder/scripts/generic/multi-bleu.perl?

Er konnte nicht wieder@@ beleb@@ t werden , und er starb am nächsten Morgen .
Den@@ g besch@@ wert sich schlie乱码lich (schließlich) , dass sein Kopf verletzt dann un@@ bewusst .

[guoshengCS]

1. I haven't tried training and inference on CPU. Sorry for that I can't give a benchmark for the performance on CPU, and I haven't heard any released Transformer performance on CPU.
2. @ is introduce by BPE(byte-pair encoding, which is used in the paper) and can be removed by ed -r 's/(@@ )|(@@ ?$)//g' . The examples seem same as my result, and it might work correctly.

[paddle@yq01-gpu-v110-255-100-01 transformer_1.1]$ grep 'Er konnte nicht wieder\|Den@@ g besch@@ wert sich' ~/guosheng/transformer_test_1.1/models/fluid/neural_machine_translation/transformer/results_2016/predict_iter100000.txt
Den@@ g besch@@ wert sich schließlich , dass sein Kopf verletzt dann un@@ bewusst .
Er konnte nicht wieder@@ beleb@@ t werden , und er starb am nächsten Morgen .

3. Testing with multi-bleu.perl and the pre-trained weight, the BLEU score is 33.64
4. You can get multi-bleu.perl according to this https://github.com/PaddlePaddle/models/blob/develop/fluid/PaddleNLP/neural_machine_translation/transformer/gen_data.sh#L110

[hshen14]

Thanks @guoshengCS. Reproduced.