Performance when training a model on one dataset and testing in another?

[kaiyangzhou]

Performance when training a model on one dataset and testing in another?

[ortegatron]

Hi, seems like it is not possible because model last layer's size, the classifier on, is relative to the dataset, it has a size equal to possible classes.
I am thinking on taking out this last layer for the experiment, and classify using the previous layer, which should give a vector representing the person aspect.

[KaiyangZhou]

Hi, model's eval mode is set to ignore the classification layer. Pls refer to the model code for more details.

Thanks for adding support for Cross-Domain (Dataset) Person Re-identification.
For more info, please refer to this issue