Unable to use Xgboost model trained locally

Hi Dahui

Sorry but we can't reproduce your issue😢. I've tried with the sklearn example model with our documented base64 encode/decode method. Everything works fine. Here is how I did it locally:

import base64
import pickle
from sklearn.calibration import CalibratedClassifierCV
from sklearn.datasets import make_classification
from sklearn.naive_bayes import GaussianNB

# Just create a dummy model
X, y = make_classification(n_samples=100, n_features=13,
                           n_redundant=0, random_state=42)
base_clf = GaussianNB()
calibrated_clf = CalibratedClassifierCV(base_estimator=base_clf, cv=3)
calibrated_clf.fit(X, y)

# encode by base64 and get output string
# follow step 1-3 in https://www.quantconnect.com/docs/v2/writing-algorithms/importing-data/bulk-downloads#04-Transport-Binary-Data
pickle_bytes = pickle.dumps(calibrated_clf)
base64_str = base64.b64encode(pickle_bytes).decode('ascii')
### Just save the string in a text file and upload to dropbox

# Reload the model
# follow step 4-5 in https://www.quantconnect.com/docs/v2/writing-algorithms/importing-data/bulk-downloads#04-Transport-Binary-Data
base64_bytes = base64_str.encode('ascii')
model = base64.b64decode(base64_bytes)
restored_model = pickle.loads(model)

# Test the recovered model
restored_model.predict_proba(X)[:5, :]
### returned the below
### array([[0.28232815, 0.71767185],
###        [0.91912966, 0.08087034],
###        [0.28008689, 0.71991311],
###        [0.14125231, 0.85874769],
###        [0.92129938, 0.07870062]])

I've also tested the exact code in your attached backtest, but replaced with my testing dropbox file's link. It also worked.

Actually, your dropbox link inside the snippet failed to recover the model when decoding locally either. Would you mind also providing the code on how do you encode and output the base64 string for further assistance? That would be the part most likely breaking things.

Best
Louis

Thanks a lot for the help Louis. Below is the code I use for encoding and outputting the base64 string:

calibrated = pickle.load(open('/Users/16105/Documents/Stocks/models/xgboost_13features_grpcd-5_mar-oct20222_apr2021_2.pkl', 'rb'))
pickle_bytes = pickle.dumps(calibrated)
base64_str = base64.b64encode(pickle_bytes).decode('ascii')
text_file = open("/Users/16105/Documents/Stocks/models/xgboost_13features_grpcd-5_mar-oct20222_apr2021_2.base64", "w")
text_file.write(base64_str)
text_file.close()
#Upload 'text_file' to Dropbox

I am actually able to decode and use the model locally, directly downloading the file on the Dropbox link. I am also able to decode/use an example uncalibrated model into QuantConnect and use it successfully. So potentially I suppose I could just upload the uncalibrated version of the model as well as the data used to calibrate the model, then calibrate the model on QuantConnect directly (so long as using the calibration package on QuantConnect doesn't pose an issue). I'm using Python 3.8 if that makes a difference.

I just tried calibrating the model directly within QuantConnect with some dummy data and it seemed to work. So I suppose uploading the calibration data and recalibrating could be a potential workaround although I would much prefer to just upload the calibrated model directly.