Remove nightly dependencies for pip install.

Nightly can often break which causes usage of Robustness Metrics itself to be unstable. Let's remove the explicit dependence and require users to manually install either the stable or nightly for now.

See also google/uncertainty-baselines#530 for Uncertainty Baselines and the GitHub issues raised about this (google/uncertainty-baselines#407, google/uncertainty-baselines#525).

PiperOrigin-RevId: 397788272

README.md

@@ -22,6 +22,11 @@ or directly from the repository as
 pip install "git+https://github.com/google-research/robustness_metrics.git#egg=robustness_metrics"
 ```
 
+Note that `robustness_metrics` requires TensorFlow and TensorFlow Probability,
+but installing `robustness_metrics` does not automatically install them. You
+will need to install TensorFlow ( `tensorflow` or `tf-nightly`) and TensorFlow
+Probability (`tensorflow_probability` or `tfp-nightly`).
+
 There are three steps to evaluate a model: 1. import the model; 2. launch an
 experiment; and 3. examine results.
 

setup.py

@@ -34,10 +34,13 @@
         "tabulate",
         "tensorflow_datasets",
         "tensorflow_hub",
-        "tf-nightly",
-        "tfp-nightly",
     ],
-    extras_require={},
+    extras_require={
+        "tensorflow": [
+            "tf-nightly",
+            "tfp-nightly",
+        ],
+    },
     classifiers=[
         "Development Status :: 4 - Beta",
         "Intended Audience :: Developers",