My goal is to find the maximum length of consecutive zeros in each row. So for instance, if I have a tensor like input = torch.tensor([[0, 1, 0, 0, 0, 1],[0, 0, 1, 0, 1, 0],[1, 0, 0, 0, 0, 0]]) I expect to get the result tensor([3, 2, 5]) I’ve done this using numpy (listed below assuming “input” is a numpy binary matrix) and it tends to be very efficient, but I cannot seem to find a way using tensors that is at least as efficient. I’ve tried to follow a similar logic using torch operations but the performance is always worse. Thanks! import numpy as np input = np.array([[0, 1, 0, 0, 0, 1],[0, 0, 1, 0, 1, 0],[1, 0, 0, 0, 0, 0]]) # Pad the matrix with ones padded_matrix = np.pad(input, ((0, 0), (1, 1)), constant_values=1) # Compute differences diffs = np.diff(padded_matrix, axis=1) # Identify start and end of zero runs start_indices = np.where(diffs == -1) end_indices = np.where(diffs == 1) #Compute lengths of zero runs run_lengths = end_indices[1] - start_indices[1] # Create a result array initialized with zeros max_zeros = np.zeros(binaryGrid.shape[0], dtype=int) # Use np.maximum.at to find the maximum run length for each row np.maximum.at(max_zeros, start_indices[0], run_lengths) print(max_zeros) Continue reading...
