MultiObjectiveOptimization

MultiObjectiveOptimization copied to clipboard

Unable to obtain the same average error as Table 1 in the paper

1

The average error on celeba is 8.25 in the paper, but my result is 10.14, which is obtained by directly using the codes in this repo. I list the configurations...

ghost

Cityscapes experiment

24

Hi, Thanks for open sourcing the code, this is great! Could you share your json parameter file for cityscapes? Also, I think it is is missing the file `depth_mean.npy` to...

maria8899

Both the `find_min_norm_element` (https://github.com/isl-org/MultiObjectiveOptimization/blob/master/multi_task/min_norm_solvers.py#L120) and `find_min_norm_element_FW` (https://github.com/isl-org/MultiObjectiveOptimization/blob/master/multi_task/min_norm_solvers.py#L166) functions define a while loop based on an `iter_count` variable. Line 13 of Algorithm 2 in the paper (https://arxiv.org/pdf/1810.04650.pdf) notes that optimization is...

newalexander

Hi Ozen, I notice that your function encoding_instancemap in cityscapes_loaders.py is strange. The ignore pixel in variable ins is set to self.ignore_index, as default 250. But those pixel is not...

tzjtatata

I use " "algorithm": "mgda", "use_approximation": true,", That's mean the MGDA_UB is used。 When I trained the task，At the begin time, It can run true. But when it run some...

zshy1205

I find the input size of CityScapes is 256x512, but in the transform, the target is resize into 32x64. what's more, the measure metrix is based on 32x64. So, why...

iris0329

Connection between fmp and task branches

4

May I ask what should be the correct dimension of grads? Suppose the output feature map size is 100, and the batch_size is 512, should the rep and grads be...

milliema

Hi, @ozansener, I find that there may be some mismatch among the depth **loss**, **metric** and **reported results** in the paper. In the depth loss, those pixels with

liyangliu

the time of "loss_data[t] = loss.data[0]" is so long

2

I find that "loss_data[t] = loss.data[0]" will cost 0.5s. Is there any way to decrease the cost?

peterytian

inconsistent implementation of update for task specific weights with the description in paper

3

Hi, thanks for releasing this awesome code! Currently, i am working on reproducing the result on cityscapes in paper. I found that in paper the description of mtl update equation...

bsaint