| v1 |
Basic model trained with 10000 iterations |
0.2371 (last) |
0.276 |
link |
link |
config |
| v2 |
Same model trained with 10000 iterations with StepLR (lr: 0.0005, steps: 3000, 6000, 9000) |
0.2319 |
0.279 |
link |
link |
config |
| v3 |
Same model trained with no scheduler specified and no steps with lr: 0.001 |
0.24720 |
0.288 |
link |
link |
config |
| v4 |
Same model with step scheduler (3k, 6k, 9k) but not sure about eval_period |
0.24025 |
0.286 |
link |
link |
config |
| v5 |
Same model with cosine scheduler (warmup iter: 1000) |
0.24559 |
0.288 |
link |
link |
config |
| v6 |
Same model with OneCycleLR scheduler (iter: 10000) |
0.2435 |
0.287 |
link |
link |
config |
| v7 |
Same model with OneCycleLR scheduler - highlr (1e-2) (iter: 5000) |
0.2484 |
0.291 |
link |
link |
config |
| v8 |
Same model with OneCycleLR scheduler - highle (1e-2) (iter: 10000) |
0.2492 |
0.287 |
link |
link |
config |
| v9 |
v8 model with {THRESHOLD + 10} ([.25, .50, .65]) |
- |
0.288 |
|
|
|
| v10 |
v7 model with {THRESHOLD + 10} ([.25, .50, .65]) |
- |
0.292 |
|
|
|
| v11 |
Same model with OneCycleLR scheduler (iter: 15000) |
0.2481 |
0.286 |
link |
link |
config |
| v12 |
v11 submission with THRESHOLD as original ([.15, .35, .55]) |
|
0.287 |
|
|
|
| v13 |
v7 model modified -> Scored each pixel rather than each object |
|
0.291 |
|
|
|
| v14 |
v13 modification: Changed thresold to [.55, .75, .6] |
|
0.283 |
|
|
|
| v15 |
v14 modification -> Changed threshold to [.25, .50, .65] scoring each pixel |
|
0.292 |
|
|
|
| v16 |
High lr onecyclelr (1e-2), 15000it |
0.25 |
0.281 |
link |
link |
config |
| v18 |
Best cv so far model (v16) with custom calculated thresholds |
|
0.276 |
|
|
|
| v19 |
I suspect I messed up threshold sequence in previous version (tried fixing that) |
|
0.281 |
|
|
|
| v20 |
Best LB scored model with custom threshold calculation |
0.2484 |
0.292 |
|
|
|
| v21 |
Best cv model (v16) with per pixel score threshold calculation |
0.25 |
0.281 |
|
|
|
| v22 |
OneCycleLR 1e-3, roi-bs: 128, iter: 10000 |
0.2444 |
0.287 |
link |
|
config |
| v23 |
v22 model with random selected MIN_PIXELS = [25, 75, 100] |
0.2444 |
0.285 |
|
|
|
| v24 |
v22 model with MIN_PIXELS = [50 ,110, 60] |
0.2444 |
0.287 |
|
|
|
| v25 |
ERROR |
|
|
|
|
|
| v26 |
sartorius tuned model (2696 iter) [without each pixel scoring] |
|
0.294 |
|
|
|
| v27 |
v26 submission with each pixel being scored |
|
|
|
|
|