Use TensorFlow 2.3 for remote tests

This causes a regression in two benchmark tests:
- net4: Random network with RectifiedLinear neurons
- net6: LMU with eager mode off

Two other benchmarks are adjusted slightly to re-center them on the
times currently achieved under TensorFlow 2.2.

Here are the full results before this commit:

test_benchmarks.py::test_performance[net0-False-64-True-1.0-1.15]
  Exec times (15): 1.092 (min), 1.146 (max), 1.122 (mean), 0.016 (std)
test_benchmarks.py::test_performance[net1-False-64-True-2.25-2.55]
  Exec times (15): 2.468 (min), 2.507 (max), 2.488 (mean), 0.014 (std)
test_benchmarks.py::test_performance[net2-True-64-True-0.6-0.9]
  Exec times (15): 0.793 (min), 0.842 (max), 0.812 (mean), 0.015 (std)
test_benchmarks.py::test_performance[net3-True-64-True-0.95-1.15]
  Exec times (15): 1.034 (min), 1.082 (max), 1.057 (mean), 0.013 (std)
test_benchmarks.py::test_performance[net4-False-None-True-0.5-0.7]
  Exec times (15): 0.647 (min), 0.656 (max), 0.649 (mean), 0.002 (std)
test_benchmarks.py::test_performance[net5-True-100-True-1.25-1.45]
  Exec times (15): 1.345 (min), 1.399 (max), 1.367 (mean), 0.015 (std)
test_benchmarks.py::test_performance[net6-True-100-False-1.05-1.25]
  Exec times (15): 1.180 (min), 1.197 (max), 1.186 (mean), 0.005 (std)

Here are the full results after this commit:

test_benchmarks.py::test_performance[net0-False-64-True-1.0-1.15]
  Exec times (15): 1.109 (min), 1.186 (max), 1.134 (mean), 0.020 (std)
test_benchmarks.py::test_performance[net1-False-64-True-2.25-2.55]
  Exec times (15): 2.488 (min), 2.539 (max), 2.509 (mean), 0.015 (std)
test_benchmarks.py::test_performance[net2-True-64-True-0.6-0.9]
  Exec times (15): 0.684 (min), 0.728 (max), 0.700 (mean), 0.013 (std)
test_benchmarks.py::test_performance[net3-True-64-True-0.95-1.15]
  Exec times (15): 1.023 (min), 1.064 (max), 1.040 (mean), 0.011 (std)
test_benchmarks.py::test_performance[net4-False-None-True-0.5-0.7]
  Exec times (15): 0.700 (min), 0.717 (max), 0.705 (mean), 0.004 (std)
test_benchmarks.py::test_performance[net5-True-100-True-1.25-1.45]
  Exec times (15): 1.326 (min), 1.357 (max), 1.339 (mean), 0.009 (std)
test_benchmarks.py::test_performance[net6-True-100-False-1.05-1.25]
  Exec times (15): 1.237 (min), 1.262 (max), 1.245 (mean), 0.006 (std)

Author: hunse

.nengobones.yml

@@ -244,7 +244,7 @@ travis_yml:
     - script: remote-test
       env:
         NENGO_VERSION: git+https://github.com/nengo/nengo.git#egg=nengo[tests]
-        TF_VERSION: tensorflow<2.3.0
+        TF_VERSION: tensorflow==2.3.0
     - script: test-coverage
       env:
         NENGO_VERSION: git+https://github.com/nengo/nengo.git#egg=nengo[tests]

.travis.yml

@@ -40,7 +40,7 @@ jobs:
   -
     env:
       NENGO_VERSION="git+https://github.com/nengo/nengo.git#egg=nengo[tests]"
-      TF_VERSION="tensorflow<2.3.0"
+      TF_VERSION="tensorflow==2.3.0"
       SCRIPT="remote-test"
   -
     env:

nengo_dl/tests/test_benchmarks.py

@@ -207,9 +207,9 @@ def test_lmu(Simulator, native_nengo, pytestconfig):
             64,
             True,
             1.0,
-            1.15,
+            1.2,
         ),
-        (benchmarks.cconv(128, 64, nengo.LIF()), False, 64, True, 2.25, 2.55),
+        (benchmarks.cconv(128, 64, nengo.LIF()), False, 64, True, 2.3, 2.6),
         (
             benchmarks.integrator(128, 32, nengo.RectifiedLinear()),
             True,
@@ -231,11 +231,11 @@ def test_lmu(Simulator, native_nengo, pytestconfig):
             False,
             None,
             True,
-            0.5,
-            0.7,
+            0.6,
+            0.8,
         ),
         (benchmarks.lmu(1000, 1, native_nengo=True), True, 100, True, 1.25, 1.45),
-        (benchmarks.lmu(1000, 1, native_nengo=True), True, 100, False, 1.05, 1.25),
+        (benchmarks.lmu(1000, 1, native_nengo=True), True, 100, False, 1.15, 1.35),
     ],
 )
 def test_performance(net, train, minibatch_size, eager, min, max, capsys):