The project provide counting of sheets in stack of plywood based on photo of stack.
As base architecture for convolutional neural network the AlexNet architecture was choosen. The AlexNet was extended by adding several fully connected layers. In overall the 8 proposed models was selected and tested.
| Model Version | Additional layers | Train data | Loss function | Average error |
|---|---|---|---|---|
| 1 | nn.Linear(in_features=4096, out_features=4096), nn.Linear(in_features=4096, out_features=1024), nn.Linear(in_features=1024, out_features=1) |
Merged | MSE | 0.3102743761 |
| 2 | nn.Linear(in_features=4096, out_features=4096), nn.Linear(in_features=4096, out_features=1) |
Merged | MSE | 0.356755558 |
| 3 | nn.Linear(in_features=4096, out_features=4096), nn.Linear(in_features=4096, out_features=2048), nn.Linear(in_features=2048, out_features=1) |
Merged | MSE | 0.3254900442 |
| 4 | nn.Linear(in_features=4096, out_features=4096), nn.Linear(in_features=4096, out_features=1024), nn.Linear(in_features=1024, out_features=1) |
Merged | Smooth L1 | 0.3906039606 |
| 5 | nn.Linear(in_features=4096, out_features=4096), nn.Linear(in_features=4096, out_features=1024), nn.Linear(in_features=1024, out_features=1) |
Original | Smooth L1 | 0.1566808731 |
| 6 | nn.Linear(in_features=4096, out_features=1) | Merged | Smooth L1 | 0.4553910928 |
| 7 | nn.Linear(in_features=4096, out_features=4096), nn.Linear(in_features=4096, out_features=2048), nn.Linear(in_features=2048, out_features=1) |
Merged | Smooth L1 | 0.4294855982 |
| 8 | nn.Linear(in_features=4096, out_features=1024), nn.Linear(in_features=1024, out_features=1) |
Merged | Smooth L1 | 0.4225038808 |
Merged data - data containing cropped images of original images.
As result version 5 of proposed model was choosen.