( FIXED )WHY FPS is so low on my datasets ??

[g110056012]

Hi, I'm wondeing why the fps on so low??? Is is because my mask are too much? Due to the datasets. I have set my config: cfg.MODEL.SPARSE_INST.DECODER.NUM_MASKS = 200 cfg.MODEL.SPARSE_INST.DECODER.NUM_CLASSES = 1

I use config:sparse_inst_r50_giam_fp16.yaml to train my model, whitch outcome result is bad, FPS is so low. and my final.model is so big ( 395,974.225 bytes )

Test : Due to my val dataset : 2 pics so I change the test_net.py ( durations[100:] -> durations[1:] ) , I'm wonder if is it correct?

def test_sparseinst_speed(cfg, fp16=False): ......

        if idx % 1 == 0:
            print("process: [{}/{}] fps: {:.3f}".format(idx,
                                                        len(data_loader), 1/np.mean(durations[1:])))
        evaluator.process(inputs, [{"instances": output}])

#evaluate results = evaluator.evaluate() print_csv_format(results)

latency = np.mean(durations[1:]) fps = 1 / latency print("speed: {:.4f}s FPS: {:.2f}".format(latency, fps)) Result : speed : 0.2337s , FPS : 4.09

[116022017144]

could you solve it, buddy? I think it's so hard

[g110056012]

I find out the reason is that the first image and second image always take lots of inference time, so the FPS calculation should ignore the first and second image, but i dont know how to explain it and why is durations[ 100 : ] instead of durations[ 2 : ]. Perhaps Author could help explain? tks.

I print the durations from the test_net.py code, you can notice that all the inference time are low (0.0..) except the first and second image (0.29, 0.24). [0.2986535410163924, 0.243336574989371, 0.03537970897741616, 0.03437506000045687, 0.03500638995319605, 0.018007997889071703, 0.017738210037350655, 0.03445056697819382, 0.017488110926933587, .................. ]

So, if you calculate FPS with first and second image durations, it will bring low FPS result.