Minor fixes for GitHub workflow action and the mean/std values of DataTransforms

[xieyxclack]

This pr includes two minor fixes:

• Change the rule for triggering the GitHub workflow action test_act: from pull_request: types: [opened, synchronize, edited] to schedule: cron: '0 8 * * 0'. It implies that test_act would no longer be triggered when someone gives a pull request, it would only be triggered once a week. The reasons for this modification are:
  • It needs lots of time for running test_act and causes the unnecessary resource cost;
  • test_atc is just a unit-test for one specific FL-NLP application but not a general functionality, therefore most of the developments would be changed its correctness.
  • Running test_atc requires more packages such as Rouge and METEOR. Considering that these packages are only used in NLP applications, we don't include these packages in the environment. Thus we perform this unit-test via GitHub workflow action (which is easier to include the package installation during testing) rather than running it via Jenkins. Later test_atc can be moved to our server when these requirements have been installed.
• Change the mean/std values for DataTransforms
  • For FEMNIST: mean=[0.9637], std=[0.1592]. These values are calculated by ourselves and are similar to those provided by other researchers.
  • For CIFAR10: mean= [0.4914, 0.4822, 0.4465], std= [0.2470, 0.2435, 0.2616], which is borrowed from the value of mean, std in cifar-10  facebookarchive/fb.resnet.torch#180
  • For other cv datasets containing 3-channel figures: mean= [0.485, 0.456, 0.406], std= [0.229, 0.224, 0.225], which are standard values calculated from ImageNet

We point out that when BatchNorm is adopted, the mean/std values of DataTransforms make little effect on the model performance. And we haven't changed the yaml files of benchmarking (those under FederatedScope/benchmark/).

[rayrayraykk]

LGTM, @DavdGao, please double-check the mean/std value.

[DavdGao]

LGTM