Grasp pixel wise 160x128 images, levine 2016 model, densenet_fcn shrink

[ahundt]

Implements #388

This implements levine 2016 pixel-wise training. I just fixed a critical bug where yx coordinates may not have been correct during pixel-wise training, so I'm performing a new run on [delta_depth, sin_theta, cos_theta] inputs.

Images are now 128x128, but to train the DenseNetFCN in reasonable time I needed to add an early transition down and transition up to shrink the data size so batches fit on GPU.

I've started two 100 epoch training runs on:

• costar-plan commit 6746888
• keras-contrib commit ahundt/keras-contrib@7807555

Pixel-wise DenseNetFCN model with early transitions

export CUDA_VISIBLE_DEVICES="1" && python2 grasp_train.py --batch_size=8 --epochs 100 --save_weights densenet_fcn_dataset_062_b_063_072_a_082_b_102_delta_depth_sin_cos_3 --grasp_model grasp_model_segmentation --optimizer SGD

example output filename:
2018-01-09-04-24-48_densenet_fcn_dataset_062_b_063_072_a_082_b_102_delta_depth_sin_cos_3-grasp_model_segmentation-epoch-001.h5

Pixel-wise Levine 2016 model:

export CUDA_VISIBLE_DEVICES="0" && python2 grasp_train.py --batch_size=6 --epochs 100 --save_weights levine_2016_segmentation_dataset_062_b_063_072_a_082_b_102_delta_depth_sin_cos_3 --grasp_model grasp_model_levine_2016_segmentation --optimizer SGD

Filename:
2018-01-09-03-58-56_levine_2016_segmentation_dataset_062_b_063_072_a_082_b_102_delta_depth_sin_cos_3-grasp_model_levine_2016_segmentation-epoch-002.h5

[ahundt]

These models haven't made any progress so I've created #401 and #402 to address the bugs. However, the actual pipeline changes in this PR are still good so it should be merged.