初始化 #

import volcengine_ml_platform
from volcengine_ml_platform import wandb 

project = "ci"                   # 项目名称
id = "run_20230714_bb4b99f4"     # run_id
volcengine_ml_platform.init()
api = wandb.TrackingApi()        
run = api.run(project=project, run_id=id)

导出summary/config数据 #

>>> run.config
{'batch_size': 32, 'channels': 16, 'epochs': 4, 'lr': 0.1}

>>> run.summary
{'best_acc': 0.99, 'best_loss': 0.01, 'final_acc': 0.88, 'final_loss': 0.11}

导出Scalar数据 #

# 导出所有图表
>>> h = run.history()
>>> pd.DataFrame(h)
      step  train_loss      loss  custom_step  validation_loss
0        0         NaN  0.479459          NaN              NaN
1        1         NaN  0.764807          NaN              NaN
2        2         NaN  0.561282          NaN              NaN
3        3         NaN  0.117248          NaN              NaN
4        4         NaN  0.492853          NaN              NaN
...    ...         ...       ...          ...              ...
1095  1095    0.010417       NaN       9025.0         0.010417
1096  1096    0.010309       NaN       9216.0         0.010309
1097  1097    0.010204       NaN       9409.0         0.010204
1098  1098    0.010101       NaN       9604.0         0.010101
1099  1099    0.010000       NaN       9801.0         0.010000

[1100 rows x 5 columns]

# 指定图表名称
>>> names = run.list_entity_names()
>>> h = run.history(name=[names[0], names[1]])
>>> pd.DataFrame(h)
          loss  custom_step  step
0     0.479459          NaN     0
1     0.764807          NaN     1
2     0.561282          NaN     2
3     0.117248          NaN     3
4     0.492853          NaN     4
...        ...          ...   ...
1095       NaN       9025.0  1095
1096       NaN       9216.0  1096
1097       NaN       9409.0  1097
1098       NaN       9604.0  1098
1099       NaN       9801.0  1099

[1100 rows x 3 columns]

run.history()方法返回的数据与平台界面展示的数据完全一致，但是平台界面为了兼顾前端性能，返回的是经过采样的数据。如果需要看全量数据，需要使用run.scan_history()方法

导出表格数据 #

>>> table_names = run.list_table_names() # 获取所有表格的名称
>>> t = run.get_table(table_names[0])    # 指定其中一个表格，获取数据
>>> pd.DataFrame(t)
  round  gpt-4  gpt-3.5  llama
0     a    1.0      0.8    0.9
1     b    0.5      0.6    0.9
2     c    0.1      0.6    0.3
3     d    0.3      0.5    1.5
4     e    0.4      0.4    0.1