def interpolate(x, scale=2, dhwc=None):
    def interpolate_torch(x):
        input = torch.Tensor(x)
        b, d, h, w, c = input.shape
        input = input.permute(0, 4, 1, 2, 3).contiguous()
        input_resized = F.interpolate(input, size=(int(d * scale), int(h * scale), int(w * scale)), mode='trilinear')
        input_resized = input_resized.permute(0, 2, 3, 4, 1).contiguous().numpy()
        return input_resized

    if dhwc is not None:
        d, h, w, c = dhwc[0], dhwc[1], dhwc[2], dhwc[3]  # dhwc stand for input size instead of output size!
    else:
        b, d, h, w, c = x.get_shape().as_list()[0], x.get_shape().as_list()[1], x.get_shape().as_list()[2], \
                        x.get_shape().as_list()[3], x.get_shape().as_list()[4]
    output = tf.py_func(lambda x: interpolate_torch(x), [x], tf.float32)
    output = tf.reshape(output, [-1, int(d * scale), int(h * scale), int(w * scale), int(c)])
    return output
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17

provided by Bo Hu