[Bug] Privacy issue: dataset and metadata of dataset about plato HOT 4 CLOSED

This is a known issue that at the time of writing this code, I am not sure how to solve this appropriately. TensorFlow (Keras) models need to be built before being used, but to build the model, one needs the input shape of the dataset. The DataSource class then needs to include an input_shape() function to return the input shape of the dataset. The datasource instance, however, doesn't need to load any data when initializing itself. Conceptually, one can detach all the data loading code from init(), and put them in a separate function.

from plato.

Similar things happen when using yolov5 in Pytorch. The input shape of a model should be better defined in a model configuration, instead of being dynamicly adjusted based on the given dataset.

It is a little bit strange to see that the server loads coco128 dataset (should be kept at the client side):

[INFO][03:15:52]: [Server #8] Loading a pre-trained model.
[INFO][03:15:52]: [Server #8] Loading a model from /home/data/pretrained/yolov5.pth.
[INFO][03:15:52]: Data source: YOLO
[INFO][03:15:52]: Downloading the YOLO dataset. This may take a while.
[INFO][03:15:52]: Downloading https://github.com/ultralytics/yolov5/releases/download/v1.0/coco128.zip.
[███[INFO][03:15:52]: Extracting ./data/coco128.zip to ./data.█████████████████████████████████████ ] 99.24%
[████████████████████████████████████████████████████████████████████████████████████████████████████] 100.00%

from plato.

It makes sense to say that the model should somehow provide the input shape, however since we are using vanilla models from PyTorch (nn.Model), MindSpore (nn.cell), and TensorFlow/Keras, I am not sure how this can be obtained. One possibility, which is a bit less elegant, is to explicitly add the data shape into the configuration file.

The Yolov5 + coco128 work is done jointly with Qintao. Perhaps he can take a look at this issue -- assigned.

from plato.

Perhaps define a configuration to solve this issue.

from plato.