The confidence score

[martinhoang11]

Hi kprokofi, thanks for your great and very details training pipeline project.
I want to ask about the confidence score: it returns like [array([-0.04811478, -0.20894697, 0.00486435, ..., 0.3441045 , 0.04338361, -0.12098037], dtype=float32)] when i test on an image. And the label prediction is confidence[i][1], negative numbers sometimes returned.

• Is it correct for the confidence score return ?
• Do I have to sort the list ?
  It seems not the same as your video demo. Looking forward to hearing from you. Thank you

[stakanich]

Hi. For these who will have this problem later:
In trainer.py there is row:
output = self.make_output(input_, hot_target)
and in the definition of this function later:

features = self.model(input_)
...
all_tasks_output = model1.make_logits(features, all=True)

In models/model_tools.py you can find it:

def forward(self, x):
        x = self.features(x)
        x = self.conv_last(x)
        x = self.avgpool(x)
        return x
def make_logits(self, features, all=False):
        all = all if self.multi_heads else False
        output = features.view(features.size(0), -1)
        spoof_out = self.spoofer(output)
        if all:
            type_spoof = self.spoof_type(output)
            lightning_type = self.lightning(output)
            real_atr = torch.sigmoid(self.real_atr(output))
            return spoof_out, type_spoof, lightning_type, real_atr
        return spoof_out

So, during a training, this model uses not only forward method, but one additional too. And in utils.py in build_model you can find these rows multiple times:

elif mode == 'convert':
    model.forward = model.forward_to_onnx

which is:

def forward_to_onnx(self,x):
        x = self.features(x)
        x = self.conv_last(x)
        x = self.avgpool(x)
        x = x.view(x.size(0), -1)
        spoof_out = self.spoofer(x)
        if isinstance(spoof_out, tuple):
            spoof_out = spoof_out[0]
        probab = F.softmax(spoof_out*self.scaling, dim=-1)
        return probab

With using one of these methods you will get an output with a shape [N, 2] where N - batch-size. First element is the confidence that image is real, and the second is the score for 'spoof' answer.