np.nditer iterator element

[kanglcn]

This works:

arr = da.ones((100,100),chunks=(10,10),dtype=bool)
arr_delayed = arr.to_delayed()

num_true_delayed = np.empty_like(arr_delayed,dtype=object)
with np.nditer(arr_delayed,flags=['multi_index','refs_ok'], op_flags=['readwrite']) as arr_it:
    for arr_block in arr_it:
        idx = arr_it.multi_index
        num_true_delayed[idx] = delayed(np.count_nonzero)(arr_delayed[idx],keepdims=True)
        num_true_delayed[idx] = da.from_delayed(num_true_delayed[idx],shape=(1,1),meta=np.array((),dtype=int))

num_true = da.block(num_true_delayed.tolist())

num_true.compute()

but this doesn't work:

arr = da.ones((100,100),chunks=(10,10),dtype=bool)
arr_delayed = arr.to_delayed()

num_true_delayed = np.empty_like(arr_delayed,dtype=object)
with np.nditer(arr_delayed,flags=['multi_index','refs_ok'], op_flags=['readwrite']) as arr_it:
    for arr_block in arr_it:
        idx = arr_it.multi_index
        num_true_delayed[idx] = delayed(np.count_nonzero)(arr_block,keepdims=True)
        num_true_delayed[idx] = da.from_delayed(num_true_delayed[idx],shape=(1,1),meta=np.array((),dtype=int))

num_true = da.block(num_true_delayed.tolist())

num_true.compute()

I have no idea with this.