KeyError: '_subspace' when using results_as_dataframe and ThompsonSampling+CMAES #40
Description
Calling results_as_dataframe works fine if I use a MongoDBConnection and QuasiRandom sampler. However, changing the sampler to ThompsonSampling causes results_as_dataframe to throw an exception KeyError: '_subspace'. Here is some example code that demonstrates the problem. Uncommenting the line that uses ThompsonSamplingand commenting out the line that usesQuasiRandom` results in the error.
from chocolate import Space, ThompsonSampling, CMAES, SQLiteConnection, QuasiRandom, log, quantized_uniform
s = Space([
{
"algo": "svm",
"C": log(low=-3, high=5, base=10),
"kernel": {
"linear": None,
"rbf": {
"gamma": log(low=-2, high=3, base=10)
}
}
},
{
"algo": "knn",
"n_neighbors": quantized_uniform(low=1, high=20, step=1)
}])
conn = SQLiteConnection(url="sqlite:///db.db")
sampler = QuasiRandom(conn, s)
# sampler = ThompsonSampling(CMAES, conn, s)
token, params = sampler.next()
print(f'Token: {token}')
print(f'Parameters: {params}')
results = conn.results_as_dataframe()
print(results)
The output, exception and stack trace when using ThompsonSampling are:
Token: {'_chocolate_id': 0, '_arm_id': 1}
Parameters: {'C': 80716.84865011052, 'gamma': 3.7193589528638826, 'kernel': 'rbf', 'algo': 'svm'}
Traceback (most recent call last):
File "C:\Users\williams\Anaconda3\envs\deeplearning\lib\runpy.py", line 193, in _run_module_as_main
"__main__", mod_spec)
File "C:\Users\williams\Anaconda3\envs\deeplearning\lib\runpy.py", line 85, in _run_code
exec(code, run_globals)
File "c:\Users\williams\.vscode\extensions\ms-python.python-2020.9.114305\pythonFiles\lib\python\debugpy\__main__.py", line 45, in <module>
cli.main()
File "c:\Users\williams\.vscode\extensions\ms-python.python-2020.9.114305\pythonFiles\lib\python\debugpy/..\debugpy\server\cli.py", line 430, in main
run()
File "c:\Users\williams\.vscode\extensions\ms-python.python-2020.9.114305\pythonFiles\lib\python\debugpy/..\debugpy\server\cli.py", line 267, in run_file
runpy.run_path(options.target, run_name=compat.force_str("__main__"))
File "C:\Users\williams\Anaconda3\envs\deeplearning\lib\runpy.py", line 263, in run_path
pkg_name=pkg_name, script_name=fname)
File "C:\Users\williams\Anaconda3\envs\deeplearning\lib\runpy.py", line 96, in _run_module_code
mod_name, mod_spec, pkg_name, script_name)
File "C:\Users\williams\Anaconda3\envs\deeplearning\lib\runpy.py", line 85, in _run_code
exec(code, run_globals)
File "g:\Research\WS\DeepRL\experimental\test_thompsonsampling_bug.py", line 26, in <module>
results = conn.results_as_dataframe()
File "C:\Users\williams\Anaconda3\envs\deeplearning\lib\site-packages\chocolate\base.py", line 65, in results_as_dataframe
result = s([r[k] for k in s.names()])
File "C:\Users\williams\Anaconda3\envs\deeplearning\lib\site-packages\chocolate\base.py", line 65, in <listcomp>
result = s([r[k] for k in s.names()])
KeyError: '_subspace'
Looking at the database that is generated, I can see that the results table is lacking a _subspace column. Note that sampling new parameters works fine, as does storing losses, but I can't extract the results.
When everything works, I expect the output to look something like the following:
Token: {'_chocolate_id': 0}
Parameters: {'n_neighbors': 7, 'algo': 'knn'}
n_neighbors algo
id
0 7 knn