torshie
torshie
@MayaPosch Yes. I'm not sure whether it's possible to do that, considering all those restrictions on Android TV.
Yes, Android TV apps are standard Android apps, in fact you can install and run many Android apks directly on an Android TV, e.g. a browser app. However TV apps...
It seems `encode()` in `data_generators/tokenizer.py` doesn't support Chinese. It cannot tokenize a Chinese sentence. The attached patch could make it support character based tokenization. [cjk.txt](https://github.com/tensorflow/tensor2tensor/files/2974078/cjk.txt)
Same problem here, even with a Titan V 32G x 8 system, I run into the OOM problem.
@tjruwase Just tested, zero stage 2 is fine.
使用了partitioner后，问题仍然存在，用下面的代码可以复现 (2302版本） 测试代码说明： 测试模型经过设计使得它具有以下几个特点(具体实现方法参见model_fn函数) * embedding的初始值为 1 * 一个key每训练一步，其权重值下降 0.01 * train/eval loss跟embedding的权重相同 训练的key依次为： 20步0, 40步1, 30步2, 10步3 训练命令： ./bare_minimum.py train 评测命令： ./bare_minimum.py eval --value 1 # 评测 key 为...
经验证，https://github.com/candyzone/DeepRec/commit/261ccfbc45afda6d4b92facc8131537a5b53b4c8 能解决这个问题