tuning parameters about tim

[Yanfei-Qin]

for example pre-training model

[mboudiaf]

Hi, I've tried re-training from scratch with num_workers = 0, but haven't noticed any significant difference. Are you still facing this issue ? (when you've re-trained a model from scratch, you must delete the "best/" folder at the directory where your model is saved before launching the test experiment. This folder contains saved features that are used to speed-up the experiments, but that must be deleted every time you have a new model. You can do it by hand, or simply specify the option eval.fresh_start=True when launching)

[Yanfei-Qin]

when I run resnet18.sh，only in the date of mini-ImageNet，after training，using the Tim-GD to val，the accuracy of 1-shot 72.96 and 5-shot 84.17，Less than 1%. Do you have any other skills？