Equivalent kitti_multiloss settings for fcn

[skinnersplace]

Hi newbie question, apologies!

I have a workflow currently running with kitti_multiloss.py as the decoder, which produces the following output:

2018-01-02 14:20:36,506 root INFO Raw Results: 2018-01-02 14:20:36,506 root INFO [train] MaxF1 (raw) : 95.8023 2018-01-02 14:20:36,506 root INFO [train] BestThresh (raw) : 48.2353 2018-01-02 14:20:36,507 root INFO [train] Average Precision (raw) : 90.6480 2018-01-02 14:20:36,507 root INFO [val] MaxF1 (raw) : 96.0370 2018-01-02 14:20:36,507 root INFO [val] BestThresh (raw) : 48.6275 2018-01-02 14:20:36,507 root INFO [val] Average Precision (raw) : 90.9239 2018-01-02 14:20:36,507 root INFO Speed (msec) (raw) : 268.7356 2018-01-02 14:20:36,507 root INFO Speed (fps) (raw) : 3.7211 2018-01-02 14:20:36,508 root INFO Smooth Results: 2018-01-02 14:20:36,509 root INFO [train] MaxF1 (smooth) : 94.4523 2018-01-02 14:20:36,509 root INFO [train] BestThresh (smooth) : 55.0980 2018-01-02 14:20:36,509 root INFO [train] Average Precision (smooth) : 90.2423 2018-01-02 14:20:36,509 root INFO [val] MaxF1 (smooth) : 94.8248 2018-01-02 14:20:36,509 root INFO [val] BestThresh (smooth) : 52.5490 2018-01-02 14:20:36,509 root INFO [val] Average Precision (smooth) : 90.5789 2018-01-02 14:20:36,509 root INFO Speed (msec) (smooth) : 268.6646 2018-01-02 14:20:36,510 root INFO Speed (fps) (smooth) : 3.7221

I wanted to retrain using the pool5 layer, for speed. As kittli_mulltiloss does not pull its segmentations from the hypes.json file, I have switched to the decoder fcn.py.

When i run the same data set with fcn.py I get :

018-01-02 18:50:45,713 root INFO Raw Results: 2018-01-02 18:50:45,713 root INFO [train] MaxF1 (raw) : 6.1611 2018-01-02 18:50:45,713 root INFO [train] BestThresh (raw) : 0.0000 2018-01-02 18:50:45,713 root INFO [train] Average Precision (raw) : 3.1784 2018-01-02 18:50:45,713 root INFO [val] MaxF1 (raw) : 6.2836 2018-01-02 18:50:45,713 root INFO [val] BestThresh (raw) : 0.0000 2018-01-02 18:50:45,713 root INFO [val] Average Precision (raw) : 3.2437 2018-01-02 18:50:45,713 root INFO Speed (msec) (raw) : 273.3912 2018-01-02 18:50:45,714 root INFO Speed (fps) (raw) : 3.6578 2018-01-02 18:50:45,714 root INFO Smooth Results: 2018-01-02 18:50:45,715 root INFO [train] MaxF1 (smooth) : 6.1611 2018-01-02 18:50:45,715 root INFO [train] BestThresh (smooth) : 0.0000 2018-01-02 18:50:45,715 root INFO [train] Average Precision (smooth) : 3.1784 2018-01-02 18:50:45,716 root INFO [val] MaxF1 (smooth) : 6.2836 2018-01-02 18:50:45,716 root INFO [val] BestThresh (smooth) : 0.0000 2018-01-02 18:50:45,716 root INFO [val] Average Precision (smooth) : 3.2437 2018-01-02 18:50:45,716 root INFO Speed (msec) (smooth) : 274.9923 2018-01-02 18:50:45,716 root INFO Speed (fps) (smooth) : 3.6365

So I believe the segmentation model implemented in fcn8_vgg.py is not the same as that implemented when I use fcn. py

Can someone explain how I need to set up the convolutional layers to get the same results using fcn.py ?

I am currenly configured as follows: }, "loss": "xentropy", "model": { "architecture_file": "encoder\fcn8_vgg.py", "evaluator_file": "evals\kitti_eval.py", "input_file": "inputs\fom_seg_input.py", "objective_file": "decoder\fcn.py", "optimizer_file": "optimizer\generic_optimizer.py" }, "arch": { "fcn_in": "fc7", "image_size": 50, "num_channels": 3, "num_classes": 2, "weight": [1,2], "whitening": false }, ], "solver": { "adam_eps": 1e-05, "batch_size": 1, "epsilon": 1e-09, "learning_rate": 1e-05, "learning_rate_step": null, "max_steps": 2000, "opt": "Adam", "threads": 8 }, "use_fc_wd": true, "wd": 0.0005

Many thanks in advance - Skinner

[skinnersplace]

would still like some help understanding how to create a hypens config for fcn.py decoder that would recreate the same results that i get using fcn8_vgg.py with kittimultiloss. is it possible to reverse engineer the orignal config?