CNN MNIST Example is minimizing instead of maximizing

[DuaneNielsen]

In the CNN MNIST example ...

from evosax import FitnessShaper
fit_shaper = FitnessShaper(centered_rank=True,
                           z_score=False,
                           w_decay=0.1,
                           maximize=False)

Should be

from evosax import FitnessShaper
fit_shaper = FitnessShaper(centered_rank=True,
                           z_score=False,
                           w_decay=0.1,
                           maximize=True)

With this i obtain

ParameterReshaper: 11274 parameters detected for optimization.
ParameterReshaper: 11274 parameters detected for optimization.
Generation: 0 {'train_acc': 0.09507812559604645, 'test_acc': 0.10459999740123749}
Generation: 100 {'train_acc': 0.5151562094688416, 'test_acc': 0.5392000079154968}
Generation: 200 {'train_acc': 0.4524902105331421, 'test_acc': 0.47589999437332153}
Generation: 300 {'train_acc': 0.517333984375, 'test_acc': 0.5645999908447266}
Generation: 400 {'train_acc': 0.5697656273841858, 'test_acc': 0.585099995136261}
Generation: 500 {'train_acc': 0.5973535180091858, 'test_acc': 0.635699987411499}
Generation: 600 {'train_acc': 0.6465820074081421, 'test_acc': 0.670199990272522}
Generation: 700 {'train_acc': 0.571484386920929, 'test_acc': 0.592199981212616}
Generation: 800 {'train_acc': 0.5779687166213989, 'test_acc': 0.6055999994277954}
Generation: 900 {'train_acc': 0.5424999594688416, 'test_acc': 0.5643999576568604}
Generation: 1000 {'train_acc': 0.4705566167831421, 'test_acc': 0.47669997811317444}
Generation: 1100 {'train_acc': 0.47603514790534973, 'test_acc': 0.4980999827384949}
Generation: 1200 {'train_acc': 0.4590234160423279, 'test_acc': 0.5117999911308289}
Generation: 1300 {'train_acc': 0.5096582174301147, 'test_acc': 0.5390999913215637}
Generation: 1400 {'train_acc': 0.5160546898841858, 'test_acc': 0.5342000126838684}
Generation: 1500 {'train_acc': 0.4954589605331421, 'test_acc': 0.5008999705314636}
Generation: 1600 {'train_acc': 0.5203222632408142, 'test_acc': 0.5051000118255615}
Generation: 1700 {'train_acc': 0.4918164014816284, 'test_acc': 0.5573999881744385}
Generation: 1800 {'train_acc': 0.4961327910423279, 'test_acc': 0.505299985408783}
Generation: 1900 {'train_acc': 0.48252928256988525, 'test_acc': 0.501800000667572}
Generation: 2000 {'train_acc': 0.43098631501197815, 'test_acc': 0.4576999843120575}
Generation: 2100 {'train_acc': 0.4217871129512787, 'test_acc': 0.44529998302459717}
Generation: 2200 {'train_acc': 0.419990211725235, 'test_acc': 0.44599997997283936}
Generation: 2300 {'train_acc': 0.42906248569488525, 'test_acc': 0.4562999904155731}
Generation: 2400 {'train_acc': 0.4432031214237213, 'test_acc': 0.4359000027179718}

[DuaneNielsen]

Also, good default parameter settings are.. train batch_size 256 popsize 400, reaches test acc score of 0.84 in 1000 generations. There are bigger settings that can go higher, but for a quick demo those settings seem OK.

Generation: 0 {'train_acc': 0.10082031041383743, 'test_acc': 0.08980000019073486}
Generation: 100 {'train_acc': 0.774658203125, 'test_acc': 0.832099974155426}
Generation: 200 {'train_acc': 0.7950195074081421, 'test_acc': 0.8150999546051025}
Generation: 300 {'train_acc': 0.8118066191673279, 'test_acc': 0.826200008392334}
Generation: 400 {'train_acc': 0.82177734375, 'test_acc': 0.8105999827384949}
Generation: 500 {'train_acc': 0.7590429782867432, 'test_acc': 0.802299976348877}
Generation: 600 {'train_acc': 0.8229589462280273, 'test_acc': 0.809499979019165}
Generation: 700 {'train_acc': 0.7835742235183716, 'test_acc': 0.8240999579429626}
Generation: 800 {'train_acc': 0.8318847417831421, 'test_acc': 0.8420999646186829}
Generation: 900 {'train_acc': 0.7520995736122131, 'test_acc': 0.8055999875068665}