about retrain shufflenetv2 question

First of all, thanks for your perfect projects.

Environments

pyhton: 3.7
pytorch: 1.7+cpu
torchvison: 0.8.1+cpu
system-os: ubuntu18.04

Hyperparameters

lr: 0.001
momentum: 0.9
weights_decay: 0.0001
batch_size: 16

Question introduction

Recently, I was learning the source code your provided in torchvision about shufflenetv2.
But when I was fine-training the network(only training fc layer), I had a problem that network convergence is very slow. like this:

[epoch 0] accuracy: 0.246
[epoch 1] accuracy: 0.253
[epoch 2] accuracy: 0.28
[epoch 3] accuracy: 0.305
[epoch 4] accuracy: 0.338
[epoch 5] accuracy: 0.353

I have read this document https://pytorch.org/docs/stable/torchvision/models.html#classification
According to this document, I downloaded the weights https://download.pytorch.org/models/shufflenetv2_x1-5666bf0f80.pth, and use same preprocessing method.

    data_transform = {
        "train": transforms.Compose([transforms.RandomResizedCrop(224),
                                     transforms.RandomHorizontalFlip(),
                                     transforms.ToTensor(),
                                     transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]),
        "val": transforms.Compose([transforms.Resize(256),
                                   transforms.CenterCrop(224),
                                   transforms.ToTensor(),
                                   transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])}

But with conditions unchanged, I just replace the model with resnet34 your provided in torchvision, and I can get great results. like this:

[epoch 0] accuracy: 0.968

Strangely, When fine-training shfflenetv2 if I change the learning rate from 0.001 to 0.1, I can get the following results:

[epoch 0] accuracy: 0.85
[epoch 1] accuracy: 0.848
.....
[epoch 29] accuracy: 0.899

Does fine-training shufflenet network need such a large learning rate?

I guess the preprocessing algorithm is not like that. Because if I use the mobilenetv2 network, I can get better results under the same conditions. Could you help me find out what's wrong? Thank you very much.

Code

https://github.com/WZMIAOMIAO/deep-learning-for-image-processing/blob/master/pytorch_classification/Test7_shufflenet/train.py

Start with pytorch_classification/Test7_shufflenet/train.py and compare its ShuffleNetV2 setup with the ResNet34 and MobileNetV2 runs described in the issue. Run the training using the listed preprocessing, weights, and learning rates, then determine whether the expected convergence difference can be reproduced and documented.