The experimental results of TResnet_M cannot be reproduced

[limengyang1992]

Hello, I used the pre-training model of TResnet_M, the image input size is 224, the learning rate is 2e-4 in the paper, the rest of the parameters are consistent with the training file, I only got 78.86 results instead of 84.2 in MODEL_ZERO.Do I need to make any changes?