def forward()

in ttw/models/landmark_classification.py [0:0]

• 33 lines of code
• 10 McCabe index (conditional complexity)

    def forward(self, batch):
        batchsize = batch['target'].size(0)
        logits = Variable(torch.FloatTensor(batchsize, self.num_classes)).zero_()

        if self.textrecog_features:
            embeddings = self.embed(batch['textrecog'])
            for i in range(batchsize):
                if self.pool == 'sum':
                    logits[i, :] += self.textrecog_linear(embeddings[i, :, :]).sum(dim=0)
                else:
                    logits[i, :] += self.textrecog_linear(embeddings[i, :, :]).max(dim=0)[0]

        if self.fasttext_features:
            for i in range(batchsize):
                if self.pool == 'sum':
                    logits[i, :] += self.fasttext_linear(batch['fasttext'][i, :, :]).sum(dim=0)
                else:
                    logits[i, :] += self.fasttext_linear(batch['fasttext'][i, :, :]).max(dim=0)[0]

        if self.resnet_features:
            for i in range(batchsize):
                if self.pool == 'sum':
                    logits[i, :] += self.resnet_linear(batch['resnet'][i, :, :]/10).sum(dim=0)
                else:
                    logits[i, :] += self.resnet_linear(batch['resnet'][i, :, :]/10).max(dim=0)[0]

        self.loss.weight = batch['weight'].view(-1).data
        out = dict()
        batch['target'] = batch['target'].float()
        target = batch['target'].view(-1)
        out['loss'] = self.loss(logits.view(-1), target)

        y_pred = torch.ge(self.sigmoid(logits), 0.5).float().data.numpy()
        y_true = batch['target'].data.numpy()

        out['f1'] = f1_score(y_true, y_pred, average='weighted')
        out['precision'] = precision_score(y_true, y_pred, average='weighted')
        out['recall'] = recall_score(y_true, y_pred, average='weighted')

        return out