Is there any problem with the current code running by default settings?

[luoxuandi970319]

Thank you for your contributions to the CIL field!
When I download the current code of PyCIL, run the foster code with the default settings, and set "init_cls = 10" and "increment = 10", the ACC is 64.78.
Is there any problem with the current code running by default settings?

[luoxuandi970319]

This is the json:
{
"prefix": "cil",
"dataset": "cifar100",
"memory_size": 2000,
"memory_per_class": 20,
"fixed_memory": true,
"shuffle": true,
"init_cls": 10,
"increment": 10,
"model_name": "foster",
"convnet_type": "resnet32",
"device": ["0"],
"seed": [1993],
"beta1":0.96,
"beta2":0.97,
"oofc":"ft",
"is_teacher_wa":false,
"is_student_wa":false,
"lambda_okd":1,
"wa_value":1,
"init_epochs": 200,
"init_lr" : 0.1,
"init_weight_decay" : 5e-4,
"boosting_epochs" : 170,
"compression_epochs" : 130,
"lr" : 0.1,
"batch_size" : 128,
"weight_decay" : 5e-4,
"num_workers" : 8,
"T" : 2
}

And this is the result.

2024-05-06 14:41:19,433 [trainer.py] => CNN: {'total': 51.59, '00-09': 48.6, '10-19': 34.0, '20-29': 51.8, '30-39': 48.4, '40-49': 57.2, '50-59': 47.1, '60-69': 55.6, '70-79': 51.6, '80-89': 59.3, '90-99': 62.3, 'old': 50.4, 'new': 62.3}
2024-05-06 14:41:19,433 [trainer.py] => NME: {'total': 43.32, '00-09': 39.9, '10-19': 27.6, '20-29': 43.4, '30-39': 37.8, '40-49': 50.0, '50-59': 34.3, '60-69': 45.9, '70-79': 40.4, '80-89': 42.6, '90-99': 71.3, 'old': 40.21, 'new': 71.3}
2024-05-06 14:41:19,433 [trainer.py] => CNN top1 curve: [90.6, 76.65, 72.53, 66.42, 62.88, 60.15, 58.94, 54.76, 53.28, 51.59]
2024-05-06 14:41:19,433 [trainer.py] => CNN top5 curve: [99.5, 96.05, 94.53, 91.9, 90.18, 88.3, 86.56, 84.75, 83.13, 81.47]
2024-05-06 14:41:19,433 [trainer.py] => NME top1 curve: [90.7, 71.5, 64.83, 55.35, 53.6, 50.9, 49.4, 44.7, 44.49, 43.32]
2024-05-06 14:41:19,433 [trainer.py] => NME top5 curve: [99.4, 94.2, 89.63, 83.42, 83.12, 79.55, 78.74, 76.21, 76.18, 72.7]

Average Accuracy (CNN): 64.78
Average Accuracy (NME): 56.879
2024-05-06 14:41:19,433 [trainer.py] => Average Accuracy (CNN): 64.78
2024-05-06 14:41:19,433 [trainer.py] => Average Accuracy (NME): 56.879

If the parameter 'fix_memory = false', the Average Accuracy (CNN): 66.57

I also run the release version 0.2.1, and got the same result.

[zhoudw-zdw]

Please check the source code of FOSTER for more details.

https://github.com/G-U-N/ECCV22-FOSTER