Add EvolvedAttention: A transformer-based neural network strategy for Prisoner's Dilemma (#1471)

* Adding attention strategy

* Adding deadline=None

* Adding verbosity to see which test is stalling the CI on ubuntu

* Use spawn instead of fork for new process

* Removing pytest-xdist -n auto

* Long run time

* Adding evolved attentio to long run strategies

* Adding min_value for s = 0

* Changing max_examples=5 to max_examples=2

* Adding test against strategies

* Removing stochastic strategy in test

* Fixing randomize test

* Lazy loading of model weights

* mypy compliance

* Shorter run time for test_meta

* Adding coverage to frequency analyzer

Author: moderouin

axelrod/data/all_classifiers.yml

@@ -1,12 +1,12 @@
-$\phi$:
+$\\phi$:
   inspects_source: false
   long_run_time: false
   makes_use_of: !!set {}
   manipulates_source: false
   manipulates_state: false
   memory_depth: .inf
   stochastic: false
-$\pi$:
+$\\pi$:
   inspects_source: false
   long_run_time: false
   makes_use_of: !!set {}
@@ -439,6 +439,14 @@ Evolved ANN 5 Noise 05:
   manipulates_state: false
   memory_depth: .inf
   stochastic: false
+EvolvedAttention:
+  inspects_source: false
+  long_run_time: True
+  makes_use_of: !!set {}
+  manipulates_source: false
+  manipulates_state: false
+  memory_depth: 200
+  stochastic: false
 Evolved FSM 16:
   inspects_source: false
   long_run_time: false

axelrod/data/model_attention.pth

@@ -2,6 +2,8 @@
 import pkgutil
 from typing import Callable, Dict, List, Optional, Tuple
 
+import torch
+
 
 def axl_filename(path: pathlib.Path) -> pathlib.Path:
     """Given a path under Axelrod/, return absolute filepath.
@@ -77,3 +79,14 @@ def load_pso_tables(filename="pso_gambler.csv", directory="data"):
         values = list(map(float, row[4:]))
         d[(name, int(a), int(b), int(c))] = values
     return d
+
+
+def load_attention_model_weights(
+    filename="model_attention.pth", directory="axelrod/data"
+):
+    """Load attention model weights."""
+    path = str(axl_filename(pathlib.Path(directory) / filename))
+    weights = torch.load(
+        path, map_location=torch.device("cpu"), weights_only=True
+    )
+    return weights

axelrod/load_data_.py

@@ -30,6 +30,7 @@
 from .ann import ANN, EvolvableANN  # pylint: disable=unused-import
 from .apavlov import APavlov2006, APavlov2011
 from .appeaser import Appeaser
+from .attention import EvolvedAttention
 from .averagecopier import AverageCopier, NiceAverageCopier
 from .axelrod_first import (
     FirstByDavis,
@@ -348,6 +349,7 @@
     EvolvedHMM5,
     EvolvedLookerUp1_1_1,
     EvolvedLookerUp2_2_2,
+    EvolvedAttention,
     FirmButFair,
     FirstByAnonymous,
     FirstByDavis,