Major changes - hamf integration. Reductions are now first class, type array lists are faster. All
container types now support fast reduction via clojure.core/reduce.
JSON parsing in this library is deprecated. Please use the charred library for fast json and
csv parsing.
The Tensor deftype is changed to DataTensor as the auto-generated ->Tensor function causes
issues with the compilation of the ->tensor public API function. Also the function to
create the data classes concretely have now been private.
:trim-leading-whitespace?, :trim-trailing-whitespace?, and :nil-empty-values? are
all supported in csv parsing to bring feature set up to par with univocity. These changes
allow all tech.ml.dataset csv-based unit tests to succeed.
Double/NaN, when casted to a boolean, evaluates to false. This make the behavior of
(ds/filter-column col identity) and (dtype/elemwise-cast item :boolean) consistent
across float, double and all object datatypes.
insn is not loaded until required with requiring-resolve from the FFI system. This
allows you to pre-generate FFI bindings during build time but not require insn or
org.ow2.asm/asm at runtime enabling greater compatibilty with hadoop-based systems
which include a legacy version of org.ow2.asm/asm which is incompatible with
insn.
tech.v3.parallel - adds ForkJoinPool/commonPool-based pmap, and upmap parallelism
constructs that will bail and use map variants if already in a fork join thread.
Changed API to rolling so that you can pass in a normal subtraction operator
and get the result you expect. This is reverse of how it worked in version
8.064.