fix typo in datasets/custom.py

[YimianDai]

Motivation

fix typo in datasets/custom.py.

Modification

Change the docs about results from list[tuple[torch.Tensor] to list[list[torch.Tensor]

The reason why it is list[list[torch.Tensor], not list[tuple[torch.Tensor] is listed as follows in parts of Chinese:

看一下脱水的 single_gpu_test 代码:

results = []
for batch_indices, data in zip(loader_indices, data_loader):
    with torch.no_grad():
        result = model(return_loss=False, **data)
    result = dataset.pre_eval(result, indices=batch_indices)
    results.extend(result)

也就是说 dataset.pre_eval 返回的 result 是 tuple[torch.Tensor]. 不过我们先来看下 model 返回的结果.

对于 mmseg 来说, model 就是 EncoderDecoder 了, 在 validation 的时候, model(・) 其实就是 EncoderDecoder.forward_test(・).

flowchart TD
%%{init: {'theme': 'base', 'themeVariables': { 'fontSize': '13px', 'fontFamily': 'Monaco'}}}%%

single_gpu_test --> EncoderDecoder.forward --> EncoderDecoder.forward_test --> EncoderDecoder.simple_test --> EncoderDecoder.inference --> EncoderDecoder.whole_inference --> EncoderDecoder.encode_decode --> EncoderDecoder._decode_head_forward_test

虽然如上图所示, forward_test 的调用链条很长, 但其实返回的 result 就是 simple_test 的返回结果. 而 simple_test 的代码如下

def simple_test(self, img, img_meta, rescale=True):
    seg_logit = self.inference(img, img_meta, rescale)
    seg_pred = seg_logit.argmax(dim=1)
    seg_pred = seg_pred.cpu().numpy()
    # unravel batch dim
    seg_pred = list(seg_pred)
    return seg_pred

• seg_logit 是 (N, C, H, W) 的 Tensor
• seg_logit.argmax(dim=1) 得到的 seg_pred 是 (N, H, W) 的 Tensor;
• 经过 seg_pred = list(seg_pred) 得到的 seg_pred 是 list[np.ndarray].

因此, model(・) 也就是 EncoderDecoder.forward_test(・) 返回的 result 是 list[np.ndarray]. 接下来要看一下 dataset.pre_eval 的代码.

• pre_eval 这个函数的功能本身就是 Collect eval result from each iteration. 从下面的代码可以看到, 返回的 pre_eval_results 还是一个 list[torch.Tensor].
• 注意, intersect_and_union 这个函数会将 np.ndarray 的输入变为 torch.Tensor 再计算, 所以返回的结果是 torch.Tensor, 因此, 返回的 pre_eval_results 就是 list[torch.Tensor].

def pre_eval(self, preds, indices):
    pre_eval_results = []

    for pred, index in zip(preds, indices):
        seg_map = self.get_gt_seg_map_by_idx(index)
        pre_eval_results.append(
            intersect_and_union(pred, seg_map, len(self.CLASSES),
                                self.ignore_index, self.label_map,
                                self.reduce_zero_label))

    return pre_eval_results

BC-breaking (Optional)

Does the modification introduce changes that break the backward-compatibility of the downstream repos?

No. It is only a modification to the comments.

[CLAassistant]

All committers have signed the CLA.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 89.50%. Comparing base (f0426c9) to head (16a8c8e).
Report is 293 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #1079   +/-   ##
=======================================
  Coverage   89.50%   89.50%           
=======================================
  Files         120      120           
  Lines        6670     6670           
  Branches     1040     1040           
=======================================
  Hits         5970     5970           
  Misses        496      496           
  Partials      204      204           

Flag	Coverage Δ
unittests	89.50% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[YimianDai]

I withdraw it. The original doc is right.