mnist benchmark

[mousecpn]

for sharing the test set performance in mnist and the training experience

[mousecpn]

【Baseline】

模型结构：

1. conv 1->64 3*3
2. bn(64)
3. relu
4. conv 64->128
5. bn(128)
6. relu
7. maxpooling
8. linear

优化器: SGD(lr=0.01, momentum=0.5)
batch_size = 64
test_acc: 98.93%

[choushunn]

优化器: SGD(lr=0.01, momentum=0.5)
batch_size = 64
test_acc: 98.47%

====================================
1.Conv2D(filters=32, kernel_size=(5, 5), padding='same')
2.Conv2D(filters=64, kernel_size=(5, 5), padding='same')
3.BatchNormalization()
4.Activation('relu')
5.MaxPool2D(pool_size=(2,2), strides=2, padding='same')
6.Dropout(0.1)
7.Flatten()
8.Dense(1024, activation='relu')
9.Dropout(0.2)
10.Dense( 10, activation='softmax')

optimizer='sgd'
batch_size=64,
epochs=4,
test_acc: 99.53%

[NOTGOOOOD]

1.Conv1 1->10 5×5
2.Relu
3.Maxpooling
4.Conv2 10->20 3×3
5.Linear1 2000->500
6.Relu
7.Linear2 500->10

优化器: Adam
batch_size = 64
test_acc: 99.41%

[hakunacc]

1.Conv2D(1-->6)
2.relu
3.maxpooling
4.Conv2D(6-->16)
5.relu
6.maxpooling
7.liner(320-->84)
8.liner(84-->10)

优化器: Adam
batch_size = 64
test_acc: 99.18%

[eanson023]

模型结构:

1. Conv2d(1–>32,kernel_size=(5, 5), stride=(1, 1), padding=(2, 2),activation=‘relu’)
2. Conv2d(32—>64, kernel_size=(3, 3))
3. bn(64)
4. Relu
5. maxpooling
6. DeepWise_PointWise_Conv(64–>8,activation=‘relu’)
7. linear(8x11x11–>10,dropout=0.2,activation='softmax')

优化器:SGD(lr=0.02)
batch_size = 64
test_acc: 98.33%