Merge pull request #1459 from alexunderch:cleanup

PiperOrigin-RevId: 861173283
Change-Id: Id57d655e7ce13a9938316eb2fdb45b5e65f39b98

Author: lanctot

.github/workflows/wheels.yml

@@ -132,7 +132,7 @@ jobs:
 
       # Install the built wheel and run the full tests on this host. The full
       # tests include all the ones that use the machine learning libraries,
-      # such as Tensorflow, PyTorch, and JAX.
+      # such as PyTorch, and JAX.
       - name: Install bdist_wheel and full tests
         run: |
           source ./venv/bin/activate

docs/alpha_zero.md

@@ -1,14 +1,10 @@
 # AlphaZero
 
-OpenSpiel includes two implementations of AlphaZero, one based on Tensorflow (in
-Python). The other based on C++ LibTorch. This document covers mostly the
-TF-based implementation and common components. For the Libtorch-based
-implementation,
+This was an old implementation of OpenSpiel based on TF1. There is another other
+based on C++ LibTorch. This document covers mostly the old TF-based
+implementation and common components. For the Libtorch-based implementation,
 [see here](https://github.com/deepmind/open_spiel/tree/master/open_spiel/algorithms/alpha_zero_torch).
 
-**Note**: as of OpenSpiel 1.6, the Tensorflow version (based on TF1) is no
-longer maintained.
-
 **Disclaimer**: this is not the code that was used for the Go challenge matches
 or the AlphaZero paper results. It is a re-implementation for illustrative
 purposes, and although it can handle games like Connect Four, it is not designed

open_spiel/data/paper_data/routing_game_experiments/utils.py

@@ -21,7 +21,6 @@
 
 import matplotlib.pyplot as plt
 import numpy as np
-import tensorflow.compat.v1 as tf
 
 from open_spiel.python import policy as policy_module
 from open_spiel.python import rl_environment
@@ -30,7 +29,6 @@
 from open_spiel.python.algorithms import exploitability
 from open_spiel.python.algorithms import external_sampling_mccfr as external_mccfr
 from open_spiel.python.algorithms import fictitious_play
-from open_spiel.python.algorithms import nfsp
 from open_spiel.python.algorithms import noisy_policy
 from open_spiel.python.games import dynamic_routing
 from open_spiel.python.games import dynamic_routing_utils
@@ -41,6 +39,7 @@
 from open_spiel.python.mfg.algorithms import policy_value
 from open_spiel.python.mfg.games import dynamic_routing as mean_field_routing_game
 import pyspiel
+from open_spiel.python.pytorch import nfsp
 # pylint:enable=line-too-long
 
 
@@ -738,7 +737,6 @@ def action_probabilities(self, state, player_id=None):
 
 def neural_ficticious_self_play(seq_game,
                                 num_epoch,
-                                sess,
                                 compute_metrics=False):
   env = rl_environment.Environment(seq_game)
   # Parameters from the game.
@@ -769,13 +767,13 @@ def neural_ficticious_self_play(seq_game,
 
   # freq_epoch_printing = num_epoch // 10
   agents = [
-      nfsp.NFSP(sess, idx, info_state_size, num_actions, hidden_layers_sizes,
-                **kwargs) for idx in range(num_players)
+      nfsp.NFSP(
+          idx, info_state_size, num_actions, hidden_layers_sizes, **kwargs
+      )
+      for idx in range(num_players)
   ]
   joint_avg_policy = NFSPPolicies(env, agents, nfsp.MODE.average_policy)
 
-  sess.run(tf.global_variables_initializer())
-  # print("TF initialized.")
   tick_time = time.time()
   for _ in range(num_epoch):
     # if ep % freq_epoch_printing == 0: