Caching preprocessing in online input module

[dirkweissenborn]

The online input module used to first preprocess and then start training. To be "online" it should however preprocess on the fly s.t. training can start immediately (this is just nicer from a user perspective).

Of course preprocessing is cached for epochs after the first. For large datasets we cache preprocessing on file to not run into memory issues.

will also resolve #342

[codecov-io]

Codecov Report

Merging #348 into master will increase coverage by 0.22%. The diff coverage is 64.48%.

@@            Coverage Diff             @@
##           master     #348      +/-   ##
==========================================
+ Coverage    48.7%   48.93%   +0.22%     
==========================================
  Files          88       88              
  Lines        5061     5121      +60     
==========================================
+ Hits         2465     2506      +41     
- Misses       2596     2615      +19

Impacted Files	Coverage Δ
...aders/extractive_qa/tensorflow/modular_qa_model.py	0% <0%> (ø)	:arrow_up:
jack/core/shared_resources.py	100% <100%> (ø)	:arrow_up:
jack/tfutil/xqa.py	93.33% <100%> (ø)	:arrow_up:
jack/train_reader.py	50.8% <100%> (+0.39%)	:arrow_up:
jack/readers/extractive_qa/shared.py	86.04% <26.66%> (+4.37%)	:arrow_up:
jack/util/preprocessing.py	66.38% <30.76%> (-4.37%)	:arrow_down:
jack/io/load.py	82.6% <33.33%> (ø)	:arrow_up:
jack/core/data_structures.py	81.48% <58.33%> (-18.52%)	:arrow_down:
jack/core/input_module.py	86.07% <87.5%> (-1.43%)	:arrow_down:
jack/readers/multiple_choice/shared.py	82.23% <0%> (-1.98%)	:arrow_down:
... and 1 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update a045085...d88ee97. Read the comment docs.