Hi, thanks for sharing the code of such a good work. This PR mainly refines codes related to Body UV head network by formatting the code in the canonical style and fixes some minor bugs as shown in related commits. An additional function of visualizing training metrics in a tensorboard is added which is compatible with current framework and easy to be enabled by a simple command-line argument.

Major modifications include:

1. Fix a typo in config.py;

2. Refine code style in vis.py and add more comments for understanding more easily;

3. Refine densepose evaluation in densepose_cocoeval.py by simplifying the extraction of geodesic distance between two vertices from the pre-computed matrix in a compact representaion. Code style is also refined;

4. Fix broken links in INSTALL.md and README.md (similar PR in Fix broken link to detectron in Install.md #211);

5. Refine the implementation of building body_uv_rcnn head network before the real training process starts. Specifically,

• Add a Body UV constant parameter BODY_UV_RCNN.NUM_SEMANTIC_PARTS = 14 in config.py to represent the number of semantic body parts which is used to build AnnIndex_lowres layer;
• Fix a small python syntax error in roi_db.py;
• Refine the implementation of constructing the outputs and losses of Body UV head branch in body_uv_rcnn_heads.py;
• Refine the codes of adding DensePose specific input blobs for body UV targets supervision in body_uv_rcnn.py, involved functions in body_uv_rcnn_heads.py, fast_rcnn.py and pool_points_interp.cu are also adjusted at the same time;
• Refine DensePose specific utilized functions in densepose_methods.py.

6. Refine loading a file of RoI proposals or model weights in io.py and add an additional function of visualizing training metrics (classification accuracy and losses) during the training process by using tensorboardX without hurting the compatibility of the whole framework.