fix(integrations): Older TF versions compatibility issues with WandbM…

…odelCheckpoint (wandb#4432)

fix(integrations): Older TF versions compatibility issues with WandbModelCheckpoint

Co-authored-by: Dmitry Duev <[email protected]>
Co-authored-by: Ayush Thakur <[email protected]>
Co-authored-by: Dmitry Duev <[email protected]>

.circleci/config.yml

@@ -1226,6 +1226,7 @@ workflows:
                 - "metaflow"
                 - "tf115"
                 - "tf21"
+                - "tf24"
                 - "tf25"
                 - "tf26"
                 - "ray112"

.codecov.yml

@@ -4,14 +4,14 @@ codecov:
     # To calculate after_n_builds use
     # ./tools/coverage-tool.py jobs | wc -l
     # also change comment block after_n_builds just below
-    after_n_builds: 37
+    after_n_builds: 38
     wait_for_ci: no
 
 comment:
   layout: "reach, diff, flags, files"
   behavior: default
   require_changes: no
-  after_n_builds: 37
+  after_n_builds: 38
 
 ignore:
   - "wandb/vendor"

.coveragerc

@@ -12,6 +12,7 @@ canonicalsrc =
     .tox/func-s_sklearn-py37/lib/python3.7/site-packages/wandb/
     .tox/func-s_tf115-py37/lib/python3.7/site-packages/wandb/
     .tox/func-s_tf21-py37/lib/python3.7/site-packages/wandb/
+    .tox/func-s_tf24-py37/lib/python3.7/site-packages/wandb/
     .tox/func-s_tf25-py37/lib/python3.7/site-packages/wandb/
     .tox/func-s_tf26-py37/lib/python3.7/site-packages/wandb/
 

tests/functional_tests/t0_main/keras/keras_model_checkpoint_tf_2_4.yea

@@ -0,0 +1,16 @@
+id: 0.keras.modelcheckpoint.tf24
+tag:
+    shard: tf24
+plugin:
+    - wandb
+command:
+    program: test_keras_model_checkpoint.py
+depend:
+    requirements:
+        - tensorflow==2.4.1
+assert:
+    - :wandb:runs_len: 1
+    - :op:contains:
+        - :wandb:runs[0][telemetry][3]  # feature
+        - 39  # keras_model_checkpoint
+    - :wandb:runs[0][exitcode]: 0

tests/functional_tests/t0_main/keras/test_keras_model_checkpoint.py

@@ -3,8 +3,6 @@
 import wandb
 from wandb.keras import WandbModelCheckpoint
 
-tf.keras.utils.set_random_seed(1234)
-
 run = wandb.init(project="keras")
 
 x = np.random.randint(255, size=(100, 28, 28, 1))
@@ -37,9 +35,9 @@ def get_model():
         WandbModelCheckpoint(
             filepath="wandb/model/model_{epoch}",
             monitor="accuracy",
-            save_best_only=True,
+            save_best_only=False,
             save_weights_only=False,
-            save_freq=1,
+            save_freq=2,
         )
     ],
 )

tox.ini

@@ -7,7 +7,7 @@ envlist=
     flake8,
     docstrings,
     py{36,37,38,39,launch,launch38},
-    func-s_{base,sklearn,metaflow,tf115,tf21,tf25,tf26,ray112,ray2,service,docs,
+    func-s_{base,sklearn,metaflow,tf115,tf21,tf24,tf25,tf26,ray112,ray2,service,docs,
             imports{1,2,3,4,5,6,7,8,9,10,11,12},noml,grpc}-py37,
     standalone-{cpu,gpu,tpu,local}-py38,
     func-cover,
@@ -409,7 +409,7 @@ commands =
     cp .coverage coverage.xml cover-results/
     coverage report -m --ignore-errors --skip-covered --omit "wandb/vendor/*"
 
-[testenv:func-s_{base,sklearn,metaflow,tf115,tf21,tf25,tf26,ray112,ray2,service,py310,docs,imports1,imports2,imports3,imports4,imports5,imports6,imports7,imports8,imports9,imports10,imports11,imports12,noml,grpc,kfp}-{py36,py37,py38,py39,py310}]
+[testenv:func-s_{base,sklearn,metaflow,tf115,tf21,tf24,tf25,tf26,ray112,ray2,service,py310,docs,imports1,imports2,imports3,imports4,imports5,imports6,imports7,imports8,imports9,imports10,imports11,imports12,noml,grpc,kfp}-{py36,py37,py38,py39,py310}]
 install_command = pip install --extra-index-url https://download.pytorch.org/whl/cpu {opts} {packages}
 commands_pre =
 setenv =
@@ -437,6 +437,7 @@ commands =
     func-s_metaflow-py{36,37,38,39,310}: yea {env:CI_PYTEST_SPLIT_ARGS:} --strict --shard metaflow run {posargs:--all}
     func-s_tf115-py{36,37,38,39,310}: yea {env:CI_PYTEST_SPLIT_ARGS:} --strict --shard tf115 run {posargs:--all}
     func-s_tf21-py{36,37,38,39,310}: yea {env:CI_PYTEST_SPLIT_ARGS:} --strict --shard tf21 run {posargs:--all}
+    func-s_tf24-py{36,37,38,39,310}: yea {env:CI_PYTEST_SPLIT_ARGS:} --strict --shard tf24 run {posargs:--all}
     func-s_tf25-py{36,37,38,39,310}: yea {env:CI_PYTEST_SPLIT_ARGS:} --strict --shard tf25 run {posargs:--all}
     func-s_tf26-py{36,37,38,39,310}: yea {env:CI_PYTEST_SPLIT_ARGS:} --strict --shard tf26 run {posargs:--all}
     func-s_ray112-py{36,37,38,39,310}: yea {env:CI_PYTEST_SPLIT_ARGS:} --strict --shard ray112 run {posargs:--all}

wandb/integration/keras/callbacks/model_checkpoint.py

@@ -4,6 +4,7 @@
 import sys
 from typing import Any, Dict, List, Optional, Union
 
+import tensorflow as tf  # type: ignore
 from tensorflow.keras import callbacks  # type: ignore
 
 import wandb
@@ -108,16 +109,22 @@ def __init__(
         if self.save_best_only:
             self._check_filepath()
 
+        self._is_old_tf_keras_version: Optional[bool] = None
+
     def on_train_batch_end(
         self, batch: int, logs: Optional[Dict[str, float]] = None
     ) -> None:
         if self._should_save_on_batch(batch):
-            # Save the model
-            self._save_model(epoch=self._current_epoch, batch=batch, logs=logs)
-            # Get filepath where the model checkpoint is saved.
-            filepath = self._get_file_path(
-                epoch=self._current_epoch, batch=batch, logs=logs
-            )
+            if self.is_old_tf_keras_version:
+                # Save the model and get filepath
+                self._save_model(epoch=self._current_epoch, logs=logs)
+                filepath = self._get_file_path(epoch=self._current_epoch, logs=logs)
+            else:
+                # Save the model and get filepath
+                self._save_model(epoch=self._current_epoch, batch=batch, logs=logs)
+                filepath = self._get_file_path(
+                    epoch=self._current_epoch, batch=batch, logs=logs
+                )
             # Log the model as artifact
             aliases = ["latest", f"epoch_{self._current_epoch}_batch_{batch}"]
             self._log_ckpt_as_artifact(filepath, aliases=aliases)
@@ -127,7 +134,10 @@ def on_epoch_end(self, epoch: int, logs: Optional[Dict[str, float]] = None) -> N
         # Check if model checkpoint is created at the end of epoch.
         if self.save_freq == "epoch":
             # Get filepath where the model checkpoint is saved.
-            filepath = self._get_file_path(epoch=epoch, batch=None, logs=logs)
+            if self.is_old_tf_keras_version:
+                filepath = self._get_file_path(epoch=epoch, logs=logs)
+            else:
+                filepath = self._get_file_path(epoch=epoch, batch=None, logs=logs)
             # Log the model as artifact
             aliases = ["latest", f"epoch_{epoch}"]
             self._log_ckpt_as_artifact(filepath, aliases=aliases)
@@ -173,3 +183,15 @@ def _check_filepath(self) -> None:
                 "This ensures correct interpretation of the logged artifacts.",
                 repeat=False,
             )
+
+    @property
+    def is_old_tf_keras_version(self) -> Optional[bool]:
+        if self._is_old_tf_keras_version is None:
+            from pkg_resources import parse_version
+
+            if parse_version(tf.keras.__version__) < parse_version("2.6.0"):
+                self._is_old_tf_keras_version = True
+            else:
+                self._is_old_tf_keras_version = False
+
+        return self._is_old_tf_keras_version