Open ZYNORl opened 2 months ago
In self-supervised learning, there was an overlap between the two patches in the code implementation, but I saw that the picture on the paper showed that there was no overlap, what's the consideration behind this difference?
def generate_coords_position(H, W, K): with task('P1'): p1 = generate_coords(H, W, K) h1, w1 = p1 pos = np.random.randint(8) with task('P2'): J = K // 4 K3_4 = 3 * K // 4 h_dir, w_dir = pos_to_diff[pos] h_del, w_del = np.random.randint(J, size=2) h_diff = h_dir * (h_del + K3_4) w_diff = w_dir * (w_del + K3_4) h2 = h1 + h_diff w2 = w1 + w_diff h2 = np.clip(h2, 0, H - K) w2 = np.clip(w2, 0, W - K) p2 = (h2, w2) return p1, p2, pos
In self-supervised learning, there was an overlap between the two patches in the code implementation, but I saw that the picture on the paper showed that there was no overlap, what's the consideration behind this difference?