fix(zero): default to non-recurrent architecture

 - the batching breaks up the temporal link between examples that would normally be processed by the time_lstm in A3C.
 - I guess this could be fixed somehow, but I'm not sure how at the moment.

Author: justindujardin

libraries/mathy_python/mathy/agents/densenet.py

@@ -17,6 +17,7 @@ def __init__(
         **kwargs,
     ):
         super(DenseNetBlock, self).__init__(**kwargs)
+        self.units = units
         self.use_shared = use_shared
         self.activate = tf.keras.layers.Activation("relu", name="relu")
         if normalization == "batch":

libraries/mathy_python/mathy/agents/zero/config.py

@@ -8,7 +8,11 @@ class SelfPlayConfig(BaseConfig):
     temperature_threshold: float = 0.5
     self_play_problems: int = 64
     training_iterations: int = 100
-    cpuct: float = 1.0 
+    cpuct: float = 1.0
     normalization_style: str = "batch"
     # When profile is true and workers == 1 the main thread will output worker_0.profile on exit
     profile: bool = False
+    # Don't use the LSTM with Zero agents because the random sampling breaks
+    # timestep correlations across the batch.
+    use_lstm: bool = False
+

libraries/mathy_python/mathy/agents/zero/self_play_runner.py

@@ -39,7 +39,7 @@ def get_predictor(self, game):
             )
 
             model: PolicyValueModel = get_or_create_policy_model(
-                args=config, env_actions=game.action_space.n,
+                args=config, env_actions=game.action_space.n, is_main=True
             )
             return model
 

libraries/mathy_python/tests/test_densenet.py

@@ -0,0 +1,56 @@
+import pytest
+import numpy as np
+from ..mathy.agents.densenet import DenseNetBlock, DenseNetStack
+import tensorflow as tf
+
+
+def test_densenet_construction():
+    layer: DenseNetStack = DenseNetStack()
+    assert layer is not None
+    model = tf.keras.Sequential([layer])
+    model(np.zeros(shape=(128, 1)))
+
+
+def test_densenet_errors():
+    with pytest.raises(ValueError):
+        DenseNetStack(normalization_style="invalid")
+
+
+@pytest.mark.parametrize("norm_type", ("batch", "layer"))
+def test_densenet_normalization_types(norm_type: str):
+    model = tf.keras.Sequential(
+        [DenseNetStack(units=24, num_layers=4, normalization_style=norm_type)]
+    )
+    y = model(np.zeros(shape=(10, 1)))
+
+
+def test_densenet_no_activation():
+    model = tf.keras.Sequential(
+        [DenseNetStack(units=24, num_layers=4, activation=None)]
+    )
+    y = model(np.zeros(shape=(10, 1)))
+    assert y.shape[0] == 10
+
+
+def test_densenet_output_transform():
+    layer: DenseNetStack = DenseNetStack(
+        units=10,
+        num_layers=4,
+        activation=None,
+        output_transform=tf.keras.layers.Dense(128),
+    )
+    assert layer is not None
+    model = tf.keras.Sequential([layer])
+    x = np.zeros(shape=(10, 1))
+    y = model(x)
+    assert y.shape[1] == 128
+
+
+def test_densenet_share_weights():
+    layer: DenseNetStack = DenseNetStack(share_weights=True)
+    layer_two: DenseNetStack = DenseNetStack(share_weights=True)
+    assert layer.get_weights() == layer_two.get_weights()
+    model = tf.keras.Sequential([layer, layer_two])
+    x = np.zeros(shape=(10, 1))
+    y = model(x)
+