Open kunjing96 opened 3 years ago
Following README.md, I tried reproducing the results of 2.42% error. But I get the 2.69% error.
The settings: args = Namespace(arch='PDARTS', auxiliary=True, auxiliary_weight=0.4, batch_size=128, cifar100=False, cutout=True, cutout_length=16, drop_path_prob=0.3, epochs=600, grad_clip=5, init_channels=36, layers=20, learning_rate=0.025, momentum=0.9, note='train_pdarts', report_freq=50, save='saveeval-train_pdarts-20201224-104142', seed=0, tmp_data_dir='/home/jingkun/dataset/cifar10', weight_decay=0.0003, workers=4)
The architecture is provided by you. Genotype(normal=[('skip_connect', 0), ('dil_conv_3x3', 1), ('skip_connect', 0), ('sep_conv_3x3', 1), ('sep_conv_3x3', 1), ('sep_conv_3x3', 3), ('sep_conv_3x3', 0), ('dil_conv_5x5', 4)], normal_concat=range(2, 6), reduce=[('avg_pool_3x3', 0), ('sep_conv_5x5', 1), ('sep_conv_3x3', 0), ('dil_conv_5x5', 2), ('max_pool_3x3', 0), ('dil_conv_3x3', 1), ('dil_conv_3x3', 1), ('dil_conv_5x5', 3)], reduce_concat=range(2, 6))
The result: 2020-12-25 23:26:02,246 Epoch: 550 lr 3.930396e-04 2020-12-25 23:26:03,187 Train Step: 000 Objs: 8.522417e-02 Acc: 98.437500 2020-12-25 23:26:32,968 Train Step: 050 Objs: 6.838607e-02 Acc: 98.697917 2020-12-25 23:27:02,576 Train Step: 100 Objs: 6.075401e-02 Acc: 98.870668 2020-12-25 23:27:32,054 Train Step: 150 Objs: 5.905080e-02 Acc: 98.918667 2020-12-25 23:28:01,535 Train Step: 200 Objs: 5.893938e-02 Acc: 98.931126 2020-12-25 23:28:31,117 Train Step: 250 Objs: 5.942694e-02 Acc: 98.919945 2020-12-25 23:29:00,611 Train Step: 300 Objs: 5.887211e-02 Acc: 98.909884 2020-12-25 23:29:30,305 Train Step: 350 Objs: 6.007082e-02 Acc: 98.842593 2020-12-25 23:29:53,998 Train_acc: 98.848000 2020-12-25 23:29:54,409 Valid Step: 000 Objs: 1.527611e-01 Acc: 96.093750 2020-12-25 23:30:00,970 Valid Step: 050 Objs: 1.155871e-01 Acc: 97.288603 2020-12-25 23:30:04,584 Valid_acc: 97.310000
The gap of 2.42% error and 2.69% error is so big. What's the problem? Can you help me? @chenxin061
Following README.md, I tried reproducing the results of 2.42% error. But I get the 2.69% error.
The settings: args = Namespace(arch='PDARTS', auxiliary=True, auxiliary_weight=0.4, batch_size=128, cifar100=False, cutout=True, cutout_length=16, drop_path_prob=0.3, epochs=600, grad_clip=5, init_channels=36, layers=20, learning_rate=0.025, momentum=0.9, note='train_pdarts', report_freq=50, save='saveeval-train_pdarts-20201224-104142', seed=0, tmp_data_dir='/home/jingkun/dataset/cifar10', weight_decay=0.0003, workers=4)
The architecture is provided by you. Genotype(normal=[('skip_connect', 0), ('dil_conv_3x3', 1), ('skip_connect', 0), ('sep_conv_3x3', 1), ('sep_conv_3x3', 1), ('sep_conv_3x3', 3), ('sep_conv_3x3', 0), ('dil_conv_5x5', 4)], normal_concat=range(2, 6), reduce=[('avg_pool_3x3', 0), ('sep_conv_5x5', 1), ('sep_conv_3x3', 0), ('dil_conv_5x5', 2), ('max_pool_3x3', 0), ('dil_conv_3x3', 1), ('dil_conv_3x3', 1), ('dil_conv_5x5', 3)], reduce_concat=range(2, 6))
The result: 2020-12-25 23:26:02,246 Epoch: 550 lr 3.930396e-04 2020-12-25 23:26:03,187 Train Step: 000 Objs: 8.522417e-02 Acc: 98.437500 2020-12-25 23:26:32,968 Train Step: 050 Objs: 6.838607e-02 Acc: 98.697917 2020-12-25 23:27:02,576 Train Step: 100 Objs: 6.075401e-02 Acc: 98.870668 2020-12-25 23:27:32,054 Train Step: 150 Objs: 5.905080e-02 Acc: 98.918667 2020-12-25 23:28:01,535 Train Step: 200 Objs: 5.893938e-02 Acc: 98.931126 2020-12-25 23:28:31,117 Train Step: 250 Objs: 5.942694e-02 Acc: 98.919945 2020-12-25 23:29:00,611 Train Step: 300 Objs: 5.887211e-02 Acc: 98.909884 2020-12-25 23:29:30,305 Train Step: 350 Objs: 6.007082e-02 Acc: 98.842593 2020-12-25 23:29:53,998 Train_acc: 98.848000 2020-12-25 23:29:54,409 Valid Step: 000 Objs: 1.527611e-01 Acc: 96.093750 2020-12-25 23:30:00,970 Valid Step: 050 Objs: 1.155871e-01 Acc: 97.288603 2020-12-25 23:30:04,584 Valid_acc: 97.310000
The gap of 2.42% error and 2.69% error is so big. What's the problem? Can you help me? @chenxin061