实验记录resnet20/cifar100

Cifar100 / resnet20：

1、Baseline

Namespace:(batch_size=128, decay=0.0003, epoch=200, gammas=[0.1, 0.1, 0.5],  learning_rate=0.1, momentum=0.9, optimizer='SGD',  schedule=[80, 120, 160])

Best acc: 68.85%

80 和 120 是拐点

2、batch_size, gammas

Namespace(batch_size=512,  decay=0.0003, epoch=200, gammas=[0.1, 0.1, 0.1],  learning_rate=0.1, momentum=0.9, optimizer='SGD',schedule=[55, 110, 160])

best：66.51%

batchsize增大会导致准确率下降

3、lr * 0.1

Namespace(batch_size=128,  decay=0.0003, epoch=200, gammas=[0.1, 0.1, 0.5],  learning_rate=0.01, momentum=0.9,optimizer='SGD',schedule=[80, 120, 160])

66.67%、

初始lr改为0.01会导致准确率下降

4、调整schedule & gammas

Namespace(batch_size=128,  decay=0.0003, epoch=200, gammas=[0.1, 0.1, 0.1],  learning_rate=0.1, momentum=0.9,  optimizer='SGD', schedule=[55, 110, 160])

68.85%

不变

5、

epoch=100，baseline

Namespace(batch_size=64,  decay=0.0003, epoch=100, gammas=[0.1, 0.1, 0.5], lambda1=0.0, learning_rate=0.1, momentum=0.9, , optimizer='SGD'schedule=[20, 40, 60], time=0)

EPOCH=98,best_acc= 64.310%

改了dataloader：

6、

epoch200, batch=64

Namespace(batch_size=64,  decay=0.0003, epoch=200, gammas=[0.1, 0.1, 0.1], lambda1=0.0, learning_rate=0.1, momentum=0.9, noptimizer='SGD', schedule=[55, 110, 160])

69.200%

test_acc：68.15%

        继续训：

        0：EPOCH=271,best_acc= 69.380%

             test acuuracy 68.470%

        1：EPOCH=282,best_acc= 69.460%

        test acuuracy 68.550%

        2：EPOCH=213,best_acc= 69.450%

        test acuuracy 68.560%

        3：EPOCH=213,best_acc= 69.450%

        test acuuracy 68.570%

---

改验证集：

Namespace(batch_size=64, decay=0.0003, epoch=150, gammas=[0.1, 0.1, 0.1], learning_rate=0.1, momentum=0.9,num_workers=16, optimizer='SGD',  schedule=[40, 80, 120])

EPOCH=138,best_acc= 83.920%

Test_acc: 66.30