v0.2.0

Changes:

1. New training sample prep strategy
   • For sequence network, a training set with equal size of positive and negative samples will be generated, positive samples are shifted chip-seq peaks augmented 5x by sampling with replacement, negative samples consist of flanking unbound/random unbound sites, the ratio of samples overlapped with ATAC-seq peaks is the same for pos/neg sets.
   • For bichrom network, positive samples are kept the same, negative samples are replaced with random unbound sites across the genome.
2. Train/Val/Test datasets are saved in TFRecord format now to speed up loading
3. MirroredStrategy has been employed to support multi-GPU training
4. Some bugs fixed

Full Changelog: v0.1.1...v0.2.0