I have tried the model.pth, python.exe main.py --dataset "food101" --image_path "./food" --train_path "./images.txt" --test_path "./images.txt" --weight_path "food2k_resnet50_0.0001.pth" --test --use_checkpoint true --checkpoint "model.pth"
but there are many errors, like below:

RuntimeError: Error(s) in loading state_dict for PRENet:
        Unexpected key(s) in state_dict: "attn1_2.q_conv.0.weight", "attn1_2.q_conv.0.bias", "attn1_2.kv_conv.0.weight", "attn1_2.kv_conv.0.bias", "attn1_2.attn.weight", "attn1_2.attn.bias", "attn1_3.q_conv.0.weight", "attn1_3.q_conv.0.bias", "attn1_3.kv_conv.0.weight", "attn1_3.kv_conv.0.bias", "attn1_3.attn.weight", "attn1_3.attn.bias", "attn2_3.q_conv.0.weight", "attn2_3.q_conv.0.bias", "attn2_3.kv_conv.0.weight", "attn2_3.kv_conv.0.bias", "attn2_3.attn.weight", "attn2_3.attn.bias", "attn2_1.q_conv.0.weight", "attn2_1.q_conv.0.bias", "attn2_1.kv_conv.0.weight", "attn2_1.kv_conv.0.bias", "attn2_1.attn.weight", "attn2_1.attn.bias", "attn3_1.q_conv.0.weight", "attn3_1.q_conv.0.bias", "attn3_1.kv_conv.0.weight", "attn3_1.kv_conv.0.bias", "attn3_1.attn.weight", "attn3_1.attn.bias", "attn3_2.q_conv.0.weight", "attn3_2.q_conv.0.bias", "attn3_2.kv_conv.0.weight", "attn3_2.kv_conv.0.bias", "attn3_2.attn.weight", "attn3_2.attn.bias".
        size mismatch for sconv1.weight: copying a param with shape torch.Size([1024, 3072, 3, 3]) from checkpoint, the shape in current model is torch.Size([1024, 1024, 3, 3]).
        size mismatch for sconv2.weight: copying a param with shape torch.Size([1024, 3072, 3, 3]) from checkpoint, the shape in current model is torch.Size([1024, 1024, 3, 3]).
        size mismatch for sconv3.weight: copying a param with shape torch.Size([1024, 3072, 3, 3]) from checkpoint, the shape in current model is torch.Size([1024, 1024, 3, 3]).

There are 2 points, first the model.pth may be trained at the food101 datasets, second the model may be not suitable for the resnet50 although the default value recommends this pth.

And last, I have tried food2k_resnet50_0.0001.pth as the pretrained model, but the result is not good and I don't know how to get good results.

Can I just use the pretrained model to inference #10

Description

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions