Yang An

@brightmart 谢谢您回复！还有个细节想确认下，我看到clue榜单上面写"TNEWS默认使用RoBERTa-wwm-large模型分数作为初始化"，而榜单上面初始化的分数我看到是57.42，这个指的就是说RoBERTa-wwm-large baseline在test 1.1测试集上面分数为57.42吧

@brightmart 您好，想再问一个今天tnews1.1提交分数异常的问题。我今天提交了1版tnews1.1的模型预测结果，与我的上一版模型在tnews1.1的结果对比，我统计了下有2339个sample预测标签不同。我的上一版模型于8.20日提交tnews1.1，得到了57.81分，但是这版提交只有28.95分，这个超出了此次提交最低可能的分数下界（1w测试样例，最低只可能是57.81-23.39=34.42分）。请问是不是tnews1.1的分数计算存在异常？麻烦您帮忙check下。我的两个提交文件分别是： 旧的提交结果 https://yangan2.oss-cn-beijing.aliyuncs.com/tnews11_predict.old.json 新的提交结果 https://yangan2.oss-cn-beijing.aliyuncs.com/tnews11_predict.json

@brightmart 请问老师，分数异常的问题有进展吗？麻烦了

Hi @renatoviolin, how to activate FP16 training for your modified code?

@renatoviolin Thank you very much! A further question please~ If I changed the settings of training and warmup steps in the config, should I change the params of `tf.contrib.mixed_precision.ExponentialUpdateLossScaleManager` (in...

Hi @renatoviolin, it seems the modification of FP16 is not compatible with multi-GPU setting? The following error occurs (platform TF1.12, CUDNN v7, CUDA 9): ``` I0703 12:01:55.457263 140202310260480 tf_logging.py:159] batch_all_reduce...

Hi, could you please share the exact script you ran?

@eugfomitcheva Hi, could you please share the exact script you ran? Meanwhile, are you using the fairseq codebase included in this OFA repo or using the official fairseq?

Hi, currently the VQA task code supports beam-search inference during validation and testing (in contrast with all-candidate inference, please refer to readme), but the finetuning objective still must be constrained...

Hi, a pull request related to this issue #124 has been proposed recently, which will add a new config to activate unconstrained finetuning. However, we find bugs are still existing...