Shape error in IOU calculation

[farhadinima75]

Hi there.
In the evaluation step, I've got this error:

	  File "train.py", line 119, in <module>
	    intr, unn = calMetric_iou(gt_value, result)
	  File "./DSAMNet/data_utils.py", line 17, in calMetric_iou
	    tp = np.sum(np.logical_and(predict == 1, label == 1))
	ValueError: operands could not be broadcast together with shapes (16,256,256) (16,64,256,256)

BatchSize is 16 and I use default parameters.
The output of the model (prob) is with the shape: torch.Size([16, 64, 256, 256]) and this is the cause of error.

[DA-II]

I had the same problem，this code is Not reproducible。

[csh1668]

I had the same problem, but
pip install torch==1.2.0 torchvision==0.4.0
downgrading librarys solved the issue.

[1998wq]

I had the same problem，this code is Not reproducible。

[1998wq]

我也有同样的问题，这个代码是不可复制的。

[SenpaiSepehr]

Code is reproducible! However it requires minor tweaking. Before I share the solution, please take a second look at section "C. Metric Module" of the paper. It explains why the result (label) argument is (16,64,256,256).

The Fix
calMetric_iou needs input parameters of same dimension. In line 115, we perform .unsqueeze(1) to add a 4-th dimension to label (16,1,256,256). However this operation is undone in line 121: gt_value = np.squeeze(gt_value), giving us a 3D array. Thus, neglecting this operation will allow you to successfully run the script inside calMetric_iou. This solution is independent of which torch version you use!
Furthermore, the argument positions passed to calMetric_iou are flipped! Make sure to swap gt_value and predict.

~ hope this helps.