load a trained model only for evaluation about ast

[YuanGongND]

It seems that your are using audio_model.mlp_heads rather than audio_model.module.mlp_heads for classification, which indicates your audio_model is not an torch.nn.dataparallel object, if so, not only your mlp_heads, but also all other parts of AST model should not in the audio_model.module dict but the audio_model dict.

One thing you could try is setting strict=True when you load the model and see how it says.

[YuanGongND]

Hi there,

I am not sure if it is a training/test mismatch or model loading issue. Are you using our code and get TEST mAPs every epoch during training? And that doesn't match with the test mAP if you load the model and do inference separately? If that is the case, is there a difference in data loading (especially the norm stats) in these two processes?

Or you only get mAP on the training set during training, and that doesn't match with the test mAP? I think it is normal that the training/test mAPs are different.

I didn't see an issue with your model loading if you also save your model as dataparallel object.

-Yuan

[hbellafkir]

this issue was solved by initializing the mlp head correctly using:

with torch.no_grad():
                self.mlp_head[0].weight = nn.Parameter(sd["module.mlp_head.0.weight"])
                self.mlp_head[0].bias = nn.Parameter(sd["module.mlp_head.0.bias"])
                self.mlp_head[1].weight = nn.Parameter(sd["module.mlp_head.1.weight"])
                self.mlp_head[1].bias = nn.Parameter(sd["module.mlp_head.1.bias"])