How to output feature maps

[greattBILL]

I use resnet101 and now i want to output feature maps. i write the linear_head.py but it doest work

 def __init__(self,
                 num_classes,
                 in_channels,
                 init_cfg=dict(type='Normal', layer='Linear', std=0.01),
                 *args,
                 **kwargs):
        super(LinearClsHead, self).__init__(init_cfg=init_cfg, *args, **kwargs)

        self.in_channels = in_channels
        self.num_classes = num_classes

        if self.num_classes <= 0:
            raise ValueError(
                f'num_classes={num_classes} must be a positive integer')

        self._init_layers()

    def _init_layers(self):
        self.fc = nn.Linear(self.in_channels, self.num_classes)

    def init_weights(self):
        normal_init(self.fc, mean=0, std=0.01, bias=0)

    def simple_test(self, img):
        """Test without augmentation."""
        cls_score = self.fc(img)
        x = self.extract_feat(img)
        print(type(x))
        #mmcv.imwrite("test.img",img)
        #x = self.extract_feat(img)

        if isinstance(cls_score, list):
            cls_score = sum(cls_score) / float(len(cls_score))
        pred = F.softmax(cls_score, dim=1) if cls_score is not None else None

        on_trace = is_tracing()
        if torch.onnx.is_in_onnx_export() or on_trace:
            return pred
        pred = list(pred.detach().cpu().numpy())
        return pred

how to do?thz

[mzr1996]

Hello, if you just want the feature map, you can build the backbone only.
For example,

>>> import torch
>>> from mmcv import Config
>>> from mmcls.models import build_classifier
>>> 
>>> cfg = Config.fromfile("./configs/_base_/models/resnet101.py")
>>> cfg.model.backbone.out_indices = (0, 1, 2, 3)  # Output the feature map of all stages.
>>> 
>>> inputs = torch.rand(8, 3, 224, 224)
>>> resnet = build_classifier(cfg.model)
>>> outputs = resnet.extract_feat(inputs)  # with neck
>>> for out in outputs:
...     print(out.shape)
torch.Size([8, 256])
torch.Size([8, 512])
torch.Size([8, 1024])
torch.Size([8, 2048])
>>> outputs = resnet.backbone(inputs)      # without neck
>>> for out in outputs:
...     print(out.shape)
torch.Size([8, 256, 56, 56])
torch.Size([8, 512, 28, 28])
torch.Size([8, 1024, 14, 14])
torch.Size([8, 2048, 7, 7])

[greattBILL]

hello sir thank you very much but i still have some question
1.inputs = torch.rand(8, 3, 224, 224) what the 8 means?
2.i still want the Activation value（after fc,before softmax). do you have the ways?
wish u have a happy day

[mzr1996]

hello sir thank you very much but i still have some question 1.inputs = torch.rand(8, 3, 224, 224) what the 8 means? 2.i still want the Activation value（after fc,before softmax). do you have the ways? wish u have a happy day

1. The 8 is just batch size.
2. Well, we don't have the final feature extraction API by now, we will add it later. For now, you can modify the simple_test function like:

    def simple_test(self, x):
        """Test without augmentation."""
        if isinstance(x, tuple):
            x = x[-1]
        cls_score = self.fc(x)
        return cls_score

[greattBILL]

The part of final feature extraction is too hard for me .I really look forward to this API！Thank you for your contribution

[greattBILL]

@mzr1996
i still have some questions about your backbone.

cfg = Config.fromfile("./configs/base/models/resnet101.py")
abs it onlyhave cfg，i want to use my checkpoints(such as epoch_100.pth),how do i do?
thank！

[mzr1996]

Please try:

cfg.model.init_cfg = dict(type='Pretrained', checkpoint='./epoch_100.pth')
resnet = build_classifier(cfg.model)

[greattBILL]

这是来自QQ邮箱的假期自动回复邮件。   您好，我最近正在休假中，无法亲自回复您的邮件。我将在假期结束后，尽快给您回复。