whathelll/LyftChallenge

Timings and performance

Standard UNet

1200/1200 [==============================] - 181s 151ms/step - loss: 0.0332 - Lyft_FScore: 0.9668 - Lyft_car_Fscore: 0.9388 - Lyft_road_Fscore: 0.9948
Training cycle took: 32.608065617084506
1000/1000 [==============================] - 40s 40ms/step
Validation: loss: 0.0968 - Lyft_FScore: 0.9032 - Lyft_car_Fscore: 0.8222 - Lyft_road_Fscore: 0.9843

1200/1200 [==============================] - 180s 150ms/step - loss: 0.0359 - Lyft_FScore: 0.9641 - Lyft_car_Fscore: 0.9332 - Lyft_road_Fscore: 0.9950
Training cycle took: 32.555992682774864
1000/1000 [==============================] - 39s 39ms/step
Validation: loss: 0.0960 - Lyft_FScore: 0.9040 - Lyft_car_Fscore: 0.8253 - Lyft_road_Fscore: 0.9827

1200/1200 [==============================] - 180s 150ms/step - loss: 0.0378 - Lyft_FScore: 0.9622 - Lyft_car_Fscore: 0.9295 - Lyft_road_Fscore: 0.9949
Training cycle took: 32.53776381413142
1000/1000 [==============================] - 40s 40ms/step
Validation: loss: 0.0950 - Lyft_FScore: 0.9050 - Lyft_car_Fscore: 0.8251 - Lyft_road_Fscore: 0.9849

UNet with Spatial Convolutions

1200/1200 [==============================] - 195s 162ms/step - loss: 0.0361 - Lyft_FScore: 0.9639 - Lyft_car_Fscore: 0.9337 - Lyft_road_Fscore: 0.9941
Training cycle took: 34.99065540631612
1000/1000 [==============================] - 44s 44ms/step
Validation: loss: 0.0949 - Lyft_FScore: 0.9051 - Lyft_car_Fscore: 0.8269 - Lyft_road_Fscore: 0.9833

UNet with Dilated Convolutions

1200/1200 [==============================] - 200s 167ms/step - loss: 0.0286 - Lyft_FScore: 0.9714 - Lyft_car_Fscore: 0.9469 - Lyft_road_Fscore: 0.9958
Training cycle took: 35.91497608423233
1000/1000 [==============================] - 43s 43ms/step
Validation: loss: 0.0928 - Lyft_FScore: 0.9072 - Lyft_car_Fscore: 0.8310 - Lyft_road_Fscore: 0.9834

Thoughts

These variations didn't seem to make too much difference

Speed comparison

model_v4

CPU: eager: 0.3854 graph: 0.3202 graph optimized: 0.3137

GPU: eager: 0.0435 graph: 0.0172 graph optimized: 0.0154

UNet

CPU: eager: 1.7789 graph: 1.7392 graph optimized: 1.7416

GPU: eager: 0.0494 graph: 0.0444 graph optimized: 0.041

More proper ENet

GPU: eager: 0.0623 graph: 0.0213 graph optimized: 0.019