Resources

logs.zip: Pre-trained models for MNIST, CLEVR bounding box prediction, and CLEVR state prediction.
out.zip: Exported predictions of trained models. These are usable for output visualisation.

To use these, place them into the dspn folder and unzip them there.

Here is the table of results produced by python summarise.py for these particular pre-trained models:

                               run    ap
dataset model iters threshold
box     base  10    0.500000     1  99.5
                    0.900000     1  95.9
                    0.950000     1  69.0
                    0.980000     1   1.7
                    0.990000     1   0.0
        dspn  10    0.500000     1  99.0
                    0.900000     1  95.5
                    0.950000     1  88.5
                    0.980000     1  47.2
                    0.990000     1   6.0
              20    0.500000     1  99.9
                    0.900000     1  99.6
                    0.950000     1  94.1
                    0.980000     1  41.6
                    0.990000     1   3.8
              30    0.500000     1  99.9
                    0.900000     1  99.1
                    0.950000     1  86.5
                    0.980000     1  31.6
                    0.990000     1   2.7
state   base  10    0.125000     1   0.0
                    0.250000     1   0.2
                    0.500000     1   0.9
                    1.000000     1   1.4
                    inf          1   3.3
        dspn  10    0.125000     1   1.1
                    0.250000     1   9.7
                    0.500000     1  33.4
                    1.000000     1  56.0
                    inf          1  74.7
              20    0.125000     1   3.4
                    0.250000     1  28.5
                    0.500000     1  70.8
                    1.000000     1  87.0
                    inf          1  90.4
              30    0.125000     1   2.0
                    0.250000     1  22.7
                    0.500000     1  69.4
                    1.000000     1  90.1
                    inf          1  92.2