LetsGoFir

> @Cynicsss I use 3 *1080ti to train on CityScapes without any change, but just reaching 70.23% mIoU. Is there something I should change in order to reach higher mIoU?...

I think so.

``` Number of instances ---------------- car: 4517 person: 3239 rider: 537 motorcycle: 148 bicycle: 1129 truck: 93 bus: 98 train: 23 ---------------- IOUs ---------------- car: MEAN: 0.7801381199070497 STD: 0.21180478329576186 person:...

And if I skip multi-component objects ```` IOUs ---------------- car: MEAN: 0.8120224407143601 STD: 0.17786273781972428 person: MEAN: 0.7364667286577709 STD: 0.1642026361339306 rider: MEAN: 0.6996582492880937 STD: 0.14832665964677744 motorcycle: MEAN: 0.6420486222188827 STD: 0.21690518847619442 bicycle:...

Do you skip the multi-component objects when evaluation @amlankar? Or maybe there are some problems with my env..

How about your performance with torch1.3?

Hi, does it work?

> It is the same model as the one the repo uses to extract but rather than dealing with caffe (can be painful) it is on pytorch. > Check that...

> @LetsGoFir Yes I have tried. The training seems to converge but the results are not as good as the ones with the original caffe feats. > Performance on Flickr30K...

> Well atm I don't have the npy file with the features. Nonetheless, in literature the most common split used in Flickr30K is the Karpathy split (https://cs.stanford.edu/people/karpathy/deepimagesent/). > You can...