Saving and loading trained classifiers

[anayurg]

Hey there,

I'm playing around with the Tsetlin classifier and I'm having some trouble figuring out how to save a trained classifier so that it can be loaded later. Do you have a go-to method that you can recommend? I couldn't do it with pickle and joblib. So I tried pickling the clause banks and weight banks separately but that seems to work only if I pickle them before the classifier is used for any prediction which in my case is not an option.

I'm attaching a minimal example just in case but I imagine there must be another way to do it that I haven't figured out.

saving.py

import pickle
from tmu.models.classification.vanilla_classifier import TMClassifier
from tmu.data import TMUDatasetSource

data = TMUDatasetSource().get_dataset(
        "XOR_biased",
        cache=True,
        cache_max_age=1,
        features=["X", "Y"],
        labels=["xor"],
        shuffle=True,
        train_ratio=1000,
        test_ratio=1000,
        return_type=dict
    )

tm = TMClassifier(20,10,2)
tm.fit(data["x_train"], data["y_train"])

# This method works only if these two lines are commented out:
# y_pred = tm.predict(data["x_test"])
# print((y_pred == data["y_test"]).mean())

for i in range(len(tm.weight_banks)):
    with open('wb%s.pkl' % i, 'wb') as f:
        pickle.dump(tm.weight_banks[i], f)
for i in range(len(tm.clause_banks)):
    with open('cb%s.pkl' % i, 'wb') as f:
        pickle.dump(tm.clause_banks[i], f)

loading.py

from tmu.models.classification.vanilla_classifier import TMClassifier
from tmu.data import TMUDatasetSource
import os
import pickle

data = TMUDatasetSource().get_dataset(
        "XOR_biased",
        cache=True,
        cache_max_age=1,
        features=["X", "Y"],
        labels=["xor"],
        shuffle=True,
        train_ratio=1000,
        test_ratio=1000,
        return_type=dict
    )

tm_new = TMClassifier(20,10,2)
tm_new.clause_banks = []
tm_new.number_of_classes = 2
tm_new.weight_banks = []

files = os.listdir('/home/folder/')
cb_count = sum(1 for file in files if file.startswith('cb'))
wb_count = sum(1 for file in files if file.startswith('wb'))

for i in range(wb_count):
    with open('wb%s.pkl' % i, 'rb') as f:
        wb = pickle.load(f)
        tm_new.weight_banks.append(wb)
for i in range(cb_count):
    with open('cb%s.pkl' % i, 'rb') as f:
        cb = pickle.load(f)
        tm_new.clause_banks.append(cb)

y_pred_new = tm_new.predict(data["x_test"])
print((y_pred_new == data["y_test"]).mean())

The error that I get when I first predict something and then save the clause and weight banks is the following:

Traceback (most recent call last):
  File "loading.py", line 36, in <module>
    y_pred_new = tm_new.predict(data["x_test"])
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/envs/tmtest/lib/python3.11/site-packages/tmu/models/classification/vanilla_classifier.py", line 311, in predict
    self.clause_banks[i].calculate_clause_outputs_predict(self.encoded_X_test,
  File "/home/envs/tmtest/lib/python3.11/site-packages/tmu/clause_bank/clause_bank.py", line 168, in calculate_clause_outputs_predict
    self.lcm_p,
    ^^^^^^^^^^
AttributeError: 'ClauseBank' object has no attribute 'lcm_p'. Did you mean: 'lcc_p'?

Thanks for the help!

[satunheim]

Please also see Issue #46 which is related to this.

[perara]

Dear @anayurg,

May I suggest just loading the whole TM directly?

saving:

import pickle
from tmu.models.classification.vanilla_classifier import TMClassifier
from tmu.data.tmu_datasource import TMUDatasetSource

data = TMUDatasetSource().get_dataset(
    "XOR_biased",
    cache=True,
    cache_max_age=1,
    features=["X", "Y"],
    labels=["xor"],
    shuffle=True,
    train_ratio=1000,
    test_ratio=1000,
    return_type=dict
)

tm = TMClassifier(20,10,2)
tm.fit(data["x_train"], data["y_train"])

with open('tm.pkl', 'wb') as f:
    pickle.dump(tm, f)

loading:

from tmu.models.classification.vanilla_classifier import TMClassifier
from tmu.data.tmu_datasource import TMUDatasetSource
import os
import pickle

data = TMUDatasetSource().get_dataset(
    "XOR_biased",
    cache=True,
    cache_max_age=1,
    features=["X", "Y"],
    labels=["xor"],
    shuffle=True,
    train_ratio=1000,
    test_ratio=1000,
    return_type=dict
)

with open('tm.pkl', 'rb') as f:
    tm_new = pickle.load(f)

y_pred_new = tm_new.predict(data["x_test"])
print((y_pred_new == data["y_test"]).mean())

[perara]

I've also corrected AttributeError: 'ClauseBank' object has no attribute 'lcm_p'. Did you mean: 'lcc_p'?

Thanks for finding this :)

Reopen if this does not satisfy your requirements.