Truncate numpy arrays with ndim >= 2 so that the total printed does not exceed max_seq_len

[crowsonkb]

When I tried to pretty-print an array I was working with, np.eye(100), it took a large number of lines! It turns out max_seq_len is the maximum subsequence length and that ndarrays are being considered simply as nested lists. I altered this behavior: for ndarrays where ndim >= 2, a temporary ctx_new.max_seq_len is determined for each print, which is the maximum subsequence length l such that np.prod(np.minimum(l, value.shape)) <= ctx.max_seq_len. It is determined through binary search (following the pseudocode here).

[tommikaikkonen]

Thanks!