def run_large()

in web_tool/ModelSessionPyTorchCycle.py [0:0]

• 39 lines of code
• 11 McCabe index (conditional complexity)

    def run_large(self,naip_data):
        eval_size = 256
        batch_size = 128
        _,w,h = naip_data.shape

        features_out = np.zeros((1,64, w, h))
        preds_out = [ np.zeros((w,h,21)) for _ in range(self.num_models) ]

        x_coords, y_coords = [],[]
        x, y = 0, 0
        while x+eval_size<w:
            x_coords.append(x)
            x += eval_size-10
        x_coords.append(w-eval_size)
        while y+eval_size<h:
            y_coords.append(y)
            y += eval_size-10
        y_coords.append(h-eval_size)

        def evaluate():
            inputs = T.from_numpy(batch[:len(batch_coords)]).float().to(self.device)
            features = self.core_model(inputs)
            preds = [ model(features) for model in self.augment_models ]
            for j,c in enumerate(batch_coords):
                xj,yj = c
                features_out[0,:,xj+5:xj+eval_size-5,yj+5:yj+eval_size-5] = features[j,:,5:-5,5:-5].cpu().numpy()
                for m in range(self.num_models):
                    preds_out[m][xj+5:xj+eval_size-5,yj+5:yj+eval_size-5,:] = np.rollaxis(preds[m][j,1:,5:-5,5:-5].softmax(0).cpu().numpy(), 0, 3)

        batch = np.zeros((batch_size,4,eval_size,eval_size))
        i = 0
        batch_coords = []
        for x in x_coords:
            for y in y_coords:
                batch_coords.append((x,y))
                batch[i] = naip_data[:,x:x+eval_size,y:y+eval_size]
                i += 1
                if i == batch_size:
                    evaluate()
                    i = 0
                    batch_coords = []
        if i>0: evaluate()

        return features_out, preds_out