Hi,
On 2023-07-30 18:35, Rebecca N. Palmer wrote:
theano has been mostly abandoned upstream since 2018. (The Aesara fork
is not abandoned, but includes interface changes including the import
name, so would break reverse dependencies not specifically altered to
use it. Neither of the two in Debian are so altered.)
Since numpy 1.24, theano has been completely broken (won't even import).
Some parts of this are fixable (#1033589), but other parts have no
known fix (#1027215). Note that the "OK" status on Salsa CI is *not* an
actual fix, but is because theano skips most of its tests in Salsa CI
because they take several hours.
theano's reverse dependencies (keras and deepnano) are both also broken
by this. keras _also_ has apparently unrelated problems, #1026738, and
is orphaned, #1027938.
This was previously discussed in #1027215 (on, among others, deepnano's
team list), where it was noted that removing keras would also block the
addition of qmean (ITP #976981), but attempts to fix theano failed.
This is sad news for qmean. I got the package up and running and I use
it quite a lot (built locally currently due to unresolved issues with
pretrained network weights). The only viable solution I see right now is
to ask qmean's upstream to migrate away from keras.
Best,
Andrius