Bump up TensorFlow version to 2.7 in tests (NVIDIA#3475)

- removes out of the box support for TensorFlow 2.5.x in DALI (it is not tested and no prebuild plugin is provided but it still can be compiled during the installation and may work) - adjust TL1_tensorflow-dali_test test Signed-off-by: Janusz Lisiecki <[email protected]>
AbhijitManepatil · Nov 10, 2021 · 04d7b00 · 04d7b00
1 parent e51aa17
commit 04d7b00
Show file tree

Hide file tree

Showing 5 changed files with 9 additions and 10 deletions.
diff --git a/docs/examples/use_cases/tensorflow/resnet-n/nvutils/common.py b/docs/examples/use_cases/tensorflow/resnet-n/nvutils/common.py
@@ -24,6 +24,7 @@ def create_piecewise_constant_decay_with_warmup(batch_size, epoch_size,
                                           lr_values, warmup_steps,
                                           compute_lr_on_cpu, name)
 
+@tf.keras.utils.register_keras_serializable(package='Custom')
 class PiecewiseConstantDecayWithWarmup(
     tf.keras.optimizers.schedules.LearningRateSchedule):
   """Piecewise constant decay with warmup schedule."""

diff --git a/docs/examples/use_cases/tensorflow/resnet-n/nvutils/runner.py b/docs/examples/use_cases/tensorflow/resnet-n/nvutils/runner.py
@@ -17,7 +17,6 @@
 from builtins import range
 from nvutils import image_processing
 from nvutils import common
-from nvutils import hvd_patch
 
 import tensorflow as tf
 import tensorflow.keras as keras
@@ -163,7 +162,7 @@ def train(model_func, params):
   opt = keras.optimizers.SGD(learning_rate=lr_schedule, momentum=momentum)
   # Horovod: add Horovod DistributedOptimizer. We use a modified version to
   # support the custom learning rate schedule.
-  opt = hvd_patch.DistributedOptimizer(opt)
+  opt = hvd.DistributedOptimizer(opt)
   if tf.__version__ >= "2.4.0" and precision == 'fp16':
     opt = keras.mixed_precision.LossScaleOptimizer(opt, dynamic=False,
                                                    initial_scale=loss_scale)

diff --git a/docs/examples/use_cases/tensorflow/resnet-n/nvutils/runner_ctl.py b/docs/examples/use_cases/tensorflow/resnet-n/nvutils/runner_ctl.py
@@ -17,7 +17,6 @@
 from builtins import range
 from nvutils import image_processing
 from nvutils import common
-from nvutils import hvd_patch
 
 import tensorflow as tf
 import tensorflow.keras as keras

diff --git a/qa/TL1_tensorflow-dali_test/test.sh b/qa/TL1_tensorflow-dali_test/test.sh
@@ -1,5 +1,5 @@
 #!/bin/bash -e
-pip_packages="horovod==0.21.3"
+pip_packages="horovod==0.23.0"
 target_dir=./docs/examples/use_cases/tensorflow/resnet-n
 
 do_once() {
@@ -14,7 +14,7 @@ do_once() {
 
     # check if CUDA version is at least 11.x
     if [ "${CUDA_VERSION:0:2}" == "11" ]; then
-        # install TF 2.5.x for CUDA 11.x test
+        # install TF 2.6.x for CUDA 11.x test
         install_pip_pkg "pip install $($topdir/qa/setup_packages.py -i 0 -u tensorflow-gpu --cuda ${CUDA_VERSION}) -f /pip-packages"
     else
         # install TF 2.3.x for CUDA 10.x test
@@ -58,7 +58,7 @@ do_once() {
     export HOROVOD_WITHOUT_PYTORCH=1
     # horovod is added to `pip_packages` so it can be preloaded, but install it here when
     # TF is already available and we can set env variables
-    install_pip_pkg "pip install --force-reinstall horovod==0.21.3 -f /pip-packages"
+    install_pip_pkg "pip install --force-reinstall horovod==0.23.0 -f /pip-packages"
 
     for file in $(ls /data/imagenet/train-val-tfrecord-480-subset);
     do

diff --git a/qa/setup_packages.py b/qa/setup_packages.py
@@ -433,12 +433,12 @@ def get_pyvers_name(self, url, cuda_version):
                         "mxnet-cu{cuda_v}"),
                 CudaPackage("tensorflow-gpu",
                         { "100" : [
-                              PckgVer("1.15.5",  python_max_ver="3.7"),
+                              PckgVer("1.15.5", python_max_ver="3.7"),
                               "2.3.3"],
                           "110" : [
-                              PckgVer("1.15.5",  python_max_ver="3.7"),
-                              "2.5.1",
-                              PckgVer("2.6.0", dependencies=["tensorflow-estimator==2.6.0", "keras==2.6.0"]),
+                              PckgVer("1.15.5", python_max_ver="3.7"),
+                              PckgVer("2.6.2", dependencies=["tensorflow-estimator==2.6.0"]),
+                              PckgVer("2.7.0"),
                               PckgVer("1.15.5+nv21.09", python_min_ver="3.8", python_max_ver="3.8", alias="nvidia-tensorflow")]
                         }),
                 CudaPackageExtraIndex("torch",