实验跟踪模块
track-ml的Python项目详细描述
音轨
安装
只需使用：
pip install track-ml
现在这需要python 3。
用法
报告各种感兴趣的指标，并自动配置和持久化日志记录。
importtrackdeftraining_function(param1=0.01,param2=10):local="~/track/myproject"remote="s3://my-track-bucket/myproject"withtrack.trial(local,remote,param_map={"param1":param1,"param2":param2}):model=create_model()forepochinrange(100):model.train()loss=model.get_loss()track.metric(iteration=epoch,loss=loss)track.debug("epoch {} just finished with loss {}",epoch,loss)model.save(os.path.join(track.trial_dir(),"model{}.ckpt".format(epoch)))<>检查现有实验
$ python -m track.trials --local_dir ~/track/myproject trial_id "start_time>2018-06-28" param2 trial_id start_time param2 8424fb387a 2018-06-28 11:17:28.752259 10
绘图结果
importtrackimportmatplotlibmatplotlib.use('Agg')importmatplotlib.pyplotaspltproj=track.Project("~/track/myproject","s3://my-track-bucket/myproject")most_recent=proj.ids["start_time"].idxmax()most_recent_id=proj.ids["trial_id"].iloc[[most_recent]]res=proj.results(most_recent_id)plt.plot(res[["iteration","loss"]].dropna())plt.savefig("loss.png")
恢复保存的工件
model.load(proj.fetch_artifact(most_recent_id[0],'model10.ckpt'))model.serve_predictions()
欢迎加入QQ群-->： 979659372
