Sometimes you want to transfer certain weights from multiple models into a single model or just want to merge multiple models. There are at least two ways to do that: init_from_checkpoint gen...

记录另外一次 Estimator 的 core 排查。 背景 复现代码 Lessons learned 参考 背景 我们用户使用 Estimator 都是使用 chief + worker + ps + evaluator 的形式。在线上偶发的 evaluator 这个角色会 core 掉。 NotFoundError: ** open file failed *...

最近用户反馈一个 Estimator 的问题，OOM (out of memory)。由于完全使用的是 Estimator 的逻辑，出错问题 排查了很久。最终使用多个工具找到问题的根源。ps_util 打印内存，objgraph 寻找内存泄露。 背景 序列化问题 模型导出 最小复现样例 objgraph 再战 reset defau...

特征列 feature column 深度神经网络输入 Feature columns 引入 Tensorflow 和其它库 构造示例数据 示例数据 演示几种不同类型的特征列 数值列 (Numeric columns) 分桶列 (Bucketized columns) 分类特征 (Categorical columns) 指示和向量列（...

介绍 Estiamtor 相关知识。 背景 非 Estimator 基于 Parameter server 架构的分布式学习 Estimator 行为控制 构造 Estimator 使用创建的 Estimator 分布式 Estimator 角色 驱动分布式训练 ...

Each language has its own known testing framework which makes writing test cases easier. In Python, the standard and built-in are unittest module. Together with pytest or coverage, we can get a sta...

Introduction Show me the code Checkpoint 格式 SavedModel Directory structure Cpp CheckpointReader implementation tensor_bundle.proto ...

Enum is very commonly used both in cpp and python. In this blog, we will introduce enumeration in cpp and python. Cpp Unscoped enumeration Scoped enumeration Python ...

This is one of the series of describing how tensorflow works. I will do my best to understand and experiment with tensorflow. In this blog I will introduce how an op and its corresponding kernel is...

Building tensorflow from source on mac is really not an easy thing at first. This blog introduces problems met during building. Bazel installation Build command Configure Buil...