Open Zhengyu-Li opened 6 years ago
What about changing the parameter of first convolution(from 3 to 1)? But, I'm not sure that is proper... Actually, i just changed that parameter to conv_bn(1, 32, 2), and it works. But.. it results low accuracy(about 90%) on test set. Please tell me if you have same problem or solve this problem.
Can we train the MobileNet on Mnist? I've tried it but failed. It seems that the input of MobileNet must have 3 channels, but Mnist data only has 1 channel. How can I modify the codes to fit Mnist data?