abhipanda4 / feature-generating-networks Goto Github PK

There is a big mistake in the code. The classifiers are trained on a concatenation of image features and the class attributes. Obviously, it should be just the image features.

Hi, Abhipanda4!
Thanks for sharing codes. I find something wrong about the input variables for self.final_classifier in Class Trainer. I think this classifier should't be trained by using any attribute information of unseen classes, otherwise it will violate the zero-shot learning setting. Is that so?

`

def fit_final_classifier(self, img_features, label_attr, label_idx):

    img_features = autograd.Variable(img_features.float()).to(self.device)
    label_attr = autograd.Variable(label_attr.float()).to(self.device)
    label_idx = label_idx.to(self.device)

    X_inp = self.get_conditional_input(img_features, label_attr) **# ?**
    Y_pred = self.final_classifier(X_inp)

    self.optim_final_cls.zero_grad()
    loss = self.criterion_cls(Y_pred, label_idx)
    loss.backward()
    self.optim_final_cls.step()

    return loss.item()

`

hello, I am confused with the gzsl_dataset. When train the final classifier, the syn_dataset contain the gzsl_dataset. Dose it mean that gzsl_dataset = real seen dataset?

Hi, can I train with custom dataset?
How can I create res101.mat & att_splits.mat files?