One issue in focal_loss.py

[souxun2015]

Hi,
Nice code for the memseg, concise and clear.

I found one issue in the focal loss part.
In the training script, the input of focal_loss is already through softmax function. But inside the focal loss, the softmax function is applied again.

# training part
            # predict
            outputs = model(inputs)
            outputs = F.softmax(outputs, dim=1)
            l1_loss = l1_criterion(outputs[:,1,:], masks)
            focal_loss = focal_criterion(outputs, masks)
            loss = (l1_weight * l1_loss) + (focal_weight * focal_loss)


# inside the focal loss
        logpt = F.log_softmax(input, dim=1)
        logpt = logpt.gather(1,target)
        logpt = logpt.view(-1)
        pt = logpt.exp()

BTW, I forked your repo, and I made some changes based on the paper and some code snippets released by the author.
You can have a look at my branch. https://github.com/souxun2015/MemSeg/tree/modifications

The main changes are located at:

1. focal_loss.py, add label_smoothing
2. memory_model.py, change the selection of samples from the memory bank
3. msff.py
4. decoder.py and coordatt.py

If you agree to these changes, I can make a pull request.

[TooTouch]

Hi, @souxun2015

I apologize for the delay in responding.
Thank you for sharing your modifications.

I ran new experiments with your advice about only focal loss.
I removed log softmax and added label smoothing into focal loss.
The results as follows:

	target	AUROC-image	AUROC-pixel	AUPRO-pixel
0	leather	100	99.2	99.7
1	pill	98.83	98.26	98.49
2	carpet	99.36	97.47	97.46
3	hazelnut	100	98.58	99.18
4	tile	100	99.53	99.47
5	cable	74.78	69.56	72.56
6	toothbrush	100	99.55	99.11
7	transistor	94.5	84.15	87.85
8	zipper	99.87	98.33	97.8
9	metal_nut	99.17	95.07	97.09
10	grid	100	99.05	98.97
11	bottle	100	98.94	98.77
12	capsule	91.74	97.26	97.1
13	screw	93.63	95.66	95.16
14	wood	99.82	96.75	98.03
15	Average	96.78	95.16	95.78

I think that other modifications does not need.
If you don't mind, I'd like to upload it with just the part about focal loss fixed.

Best regards,

[souxun2015]

Hi, @TooTouch
It's great to hear that you've conducted new experiments using the focal loss with label smoothing. Others are the dataset-specific hyperparameters, no need to spend time to modifty them.
Happy holiday!