Have you ever looked at your computer and died inside a bit seeing thousands if not million of folders full of JSON files with results of experiments? If only storing them inside a nice SQL databse was as easy as json.dumps(). What if it could be that easy? After all python comes equipped with it's own nifty sqlite3.
None. You are good to go :)
Only python >= 3.10
Doc are WIP, but in the meanwhile tests show everything you need:
db_file = f"my_run.db" # maybe a better name than this? up to you.
# Let's log some hyper-params per run, and a simple steps/loss for the actual data
# This only creates the database file and the tables!
create_db(db_file, info_cols={"init": str, "steps": int, "lr": float}, data_cols={"step": int, "loss": float})Now you are ready to launch all the experiments you want and all you need to do is pass them the db_file (god bless SQLite):
... in some other worker ...
# This creates a new entry in the INFO_TABLE and creates a unique (primary key) run_id
db = runDB(emptydb_file, {"init": "xavier", "lr": 0.001, "steps": 10})
...
for i in range(run["steps"]):
# This creates a new row in the DATA_TABLE that containis the logged information and the run_id (foreign key)
db.insert({"step": i, "loss": 1 / (i + 1.0)})These functions are just wrappers around sqlite3, you can just query the underlying database as you see fit:
sql = """
SELECT runid
FROM INFO
WHERE init = :init AND lr = :lr AND steps = :steps
"""
args = ({"init":"xavier", "lr":0.001, "steps":20})
# Use sqlite3 directly
with sqlite3.connect(db_file) as con:
cur = con.execute(sql, args)
results = cur.fetchall()
con.close() # yup you still need to close it
# or the provided utility, with some extra fancyness
from rundb import query
results = query(db_file, sql=sql, args=args, as_dict=True)
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent