更新日志¶
版本 0.15.2 - 2022.11.02¶
主要更新¶
Relate entrypoint & output to train task rather than algo
Bug修复¶
Sync algo name with Gitea: update and rules
版本 0.15.0 - 2022.09.29¶
主要更新¶
Support multiple dataset IDs in quick_train
Adapt to backend permission update in SDK
版本 0.14.0 - 2022.08.22¶
主要更新¶
Support PyTorch distributed training via Volcano Jobs
Adapt backend permission modification
版本 0.13.1 - 2022.06.21¶
主要更新¶
Add algorithm name verification
More intuitive way to specify resources in quick_train`
More intuitive err msg for uncommitted local algo in SDK quick_start
版本 0.12.2 - 2022.03.31¶
Bug修复¶
Fix version of urllib3 & requests
Discard custom algo default train_params
版本 0.12.0 - 2022.03.04¶
主要更新¶
Add basic retry on API requests
Bug修复¶
Adopt HTTP header naming convention Secret-Key instead of secret_key
Fix unexpected public projects by setting default visibility to private
版本 0.11.11 - 2022.01.11¶
主要更新¶
Support mounting pretrain task’s results dir in quick_train
Add task description in quick_train
Prepare for Gitea integration: use git to push algo
Bug修复¶
Make meaningful names for algo & dset in quick_start
Align return type for remote/local algo syncing
版本 0.11.10 - 2021.12.24¶
主要更新¶
Support (roughly) model mount in quick_train
Support both QuotaGroup’s ID and name in quick_train
版本 0.11.9 - 2021.12.17¶
主要更新¶
Use tqdm in SyncResourceUploader
文档与示例¶
Fix missing changes in docs since 0.10.8
Bug修复¶
Use if/else instead of try/catch in sync_algorithm
版本 0.11.8 - 2021.12.03¶
主要更新¶
Normalize internal logger with basicConfig & formats
Show besteffort quota group in CLI quota ls
文档与示例¶
Add notice in doc for flag support in quick_train
Bug修复¶
Add algorithm output directory path check
Use SyncResourceUploader by default, more stable across OSs
Sync metadata when using remote algo in quick_train
其他¶
Use GA to publish dist to PyPI
版本 0.11.7 - 2021.11.10¶
主要更新¶
Run/save training with/to all-in-one config file
Remake CLI log command with log streaming capability
Support resource request in HPO
Show QuotaGroup ID in quota ls command
Show config saved path in run training –save
文档与示例¶
Note for fake credentials in quickstart docs
Emphasize resource request for one training in HPO docs
Update some quickstart examples & docs with resource request
Bug修复¶
Eliminate ending slash in host URL
Ensure session close to avoid unexpected connection reset
Fix –save flag type & remove debugging print
版本 0.11.0 - 2021.09.15¶
主要更新¶
Support gpu_mem in quick_train
Possible to specify mirror’s name in quick_train
Adapt HPO feature to backend-integrated HPO API
Possible to omit dataset in quick_train
Load details when getting train tasks in HPO
Possible to change mirror on existing custom algo
Adapt to backend log API changes
Possible to change mirror name in HPO
Fetch HPO tasks explicitly from backend API
Handle quickstart default & custom project
Add QuotaGroup listing/detail APIs
CLI v2
文档与示例¶
Add PyTorch CNN example
Bug修复¶
Accept empty list in all get_list methods
Do not generate anylearn_tools.py for custom algo
Fix quickstart algo upload not aborting when error
HOTFIX: force state check when re-upload algo
其他¶
Yet another GA split-up for better balancing quota
版本 0.10.8 - 2021.07.07¶
注：为了与Anylearn其他组件版本保持相对一致，跳过0.10.7版本。
主要更新¶
Add unfinished train tasks’ resuming
训练任务里面封装模型转存接口
SDK用户接口封装
验证资源下载参数
Avoid duplicated resource upload by checksum
Add auto-HPO powered by NNI
添加设置用户计算资源数接口
添加训练任务描述
Add custom algo handling by name, optimize custom algo logic
文档与示例¶
Update docs outlines and contents
Reorganize jupyter notebook examples
完善API参考文档中接口封装的部分
Update resuming example with training object ref loss
Bug修复¶
Fix embedded return typing by future annotations
Fix user register request (send data by ‘data’ not ‘files’…)
其他¶
Add multi-os in testing action
Add script to clean remote resources tagged SDK
Split testing CI running only ubuntu-py37 in Dmagine
为默认值None添加Optional类型
版本 0.10.6 - 2021.05.11¶
主要更新¶
Add serving interfaces and an example in Jupyter notebook
Add project and train_task interfaces
Add quick training of non-standard algorithm (SDK “quick start”)
Finish eval basic interfaces with unit tests
Add quick evaluation of non-standard algorithm (SDK “quick start”)
Rename quick train/evaluate to quickstart
封装资源下载接口，训练任务新增模型导出
Add local DB-ORM with basic models and migration boilerplate
SDK汇报训练最终和中间结果指标
Formalize meta files for distribution
Add in-container tracking by metric reporting in training
Add GPU sharing for services
Make config static and friendly for testing
Add ID inputs and resource (created) outputs in quickstart
Add basic CLI (anylearnctl) required for HPO
文档与示例¶
增加数据集管理示例
Add multi-version docs publishing to gh-pages
Update README.md with public docs’ URL
Add example following CRISP-DM (wind power prediction)
Bug修复¶
Auto create anylearn_tools.py for local algo
DO NOT override data path passed in hyperparams
Fix quick eval params overriding and downloader return value
HOT FIX: add GPU num selection for quickstart
HOT FIX: send gpu_num when creating train tasks
其他¶
Add sync to thulab/Anylearn-sdk