-
Notifications
You must be signed in to change notification settings - Fork 2
Training Attempts
twongjirad edited this page Aug 27, 2016
·
8 revisions
- 9471 training images
- 4035 validation images
Allowing mirroring. Embedding 495x495 images into 512x512 image.
| Attempt | LearnRate | decay | mirror | class-weight | after 10 epochs | : notes : |
|---|---|---|---|---|---|---|
| 1 | 1.0e-5 | 0.0 | Y | 100 | acc=95,loss=0.5 | class weights too low (100) |
| 2 | 1.0e-4 | 0.0 | Y | 10000 | acc=80,loss=50-100 | after 8 epochs, was giving scores of 0 to all pixels. |
| 3 | 1.0e-5 | 0.0 | Y | 10000 | acc=75,loss=50-100 | init deconv layers to bilinear upsampling.(kernel=4,sride=2,pad=1) |
| 4 | 1.0e-4 | 0.0001 | Y | 10000 | (in progress) | removed biases from deconv. turn on batch norm. shrinking network to get bigger mini-batch size (2). More iters (50) for 100 batchsize. found another bug: final 1x1 convolution had a ReLU on it. Removed it. Also, changed lr and decay multipliers to be like FCN network. |