About reproducing the proto net results with ConvNet

[flexibility2]

@Han-Jia , Thanks for your source code! But I have some troubles in reproducing them. I have achieved only 51.97% for the 1-shot 5-way of ConvNet case and 70.60% for 5-shot 5-way of ConvNet case. The hyper-parameters I used is:

• --lr 0.0001 --temperature 32 --max_epoch 200 step_size (schedulr) = 10, gamma=0.2 for 1-shot 5-way of ConvNet

• --lr 0.0001 --temperature 32 --max_epoch 200 step_size (schedulr) = 10, gamma=0.2 for 5-shot 5-way of ConvNet

Should I make any changes to other default parameters to reproduce the result in Table1 (52.61% and 71.33% test acc)?

any information will be greatly appreciated

[Han-Jia]

Hello,

For the ConvNet backbone, please try

• --lr 0.0001 --temperature 64 --max_epoch 200 step_size (schedulr) = 50, gamma=0.1 for 1-shot 5-way of ConvNet

• --lr 0.0001 --temperature 16 --max_epoch 200 step_size (schedulr) = 30, gamma=0.1 for 5-shot 5-way of ConvNet

1. For the temperature, it seems high-temperature value will not degrade the results. It's better to try larger temperature values.
2. You can also increase the number of validation tasks to select the best model.