fix bug in crossover: the child from two normal networks should always be normal.

tensorneat/.idea/.gitignore

@@ -0,0 +1,8 @@
+# 默认忽略的文件
+/shelf/
+/workspace.xml
+# 基于编辑器的 HTTP 客户端请求
+/httpRequests/
+# Datasource local storage ignored files
+/dataSources/
+/dataSources.local.xml

tensorneat/algorithm/neat/__init__.py

@@ -1,3 +1,4 @@
+from .ga import *
 from .gene import *
 from .genome import *
 from .species import *

tensorneat/algorithm/neat/ga/crossover/default.py

@@ -2,6 +2,7 @@ import jax, jax.numpy as jnp
 
 from .base import BaseCrossover
 
+
 class DefaultCrossover(BaseCrossover):
 
     def __call__(self, randkey, genome, nodes1, conns1, nodes2, conns2):
@@ -14,17 +15,19 @@ class DefaultCrossover(BaseCrossover):
         # crossover nodes
         keys1, keys2 = nodes1[:, 0], nodes2[:, 0]
         # make homologous genes align in nodes2 align with nodes1
-        nodes2 = self.align_array(keys1, keys2, nodes2, False)
+        nodes2 = self.align_array(keys1, keys2, nodes2, is_conn=False)
 
         # For not homologous genes, use the value of nodes1(winner)
         # For homologous genes, use the crossover result between nodes1 and nodes2
-        new_nodes = jnp.where(jnp.isnan(nodes1) | jnp.isnan(nodes2), nodes1, self.crossover_gene(randkey_1, nodes1, nodes2))
+        new_nodes = jnp.where(jnp.isnan(nodes1) | jnp.isnan(nodes2), nodes1,
+                              self.crossover_gene(randkey_1, nodes1, nodes2, is_conn=False))
 
         # crossover connections
         con_keys1, con_keys2 = conns1[:, :2], conns2[:, :2]
-        conns2 = self.align_array(con_keys1, con_keys2, conns2, True)
+        conns2 = self.align_array(con_keys1, con_keys2, conns2, is_conn=True)
 
-        new_conns = jnp.where(jnp.isnan(conns1) | jnp.isnan(conns2), conns1, self.crossover_gene(randkey_2, conns1, conns2))
+        new_conns = jnp.where(jnp.isnan(conns1) | jnp.isnan(conns2), conns1,
+                              self.crossover_gene(randkey_2, conns1, conns2, is_conn=True))
 
         return new_nodes, new_conns
 
@@ -54,14 +57,14 @@ class DefaultCrossover(BaseCrossover):
 
         return refactor_ar2
 
-    def crossover_gene(self, rand_key, g1, g2):
+    def crossover_gene(self, rand_key, g1, g2, is_conn):
-        """
-        crossover two genes
-        :param rand_key:
-        :param g1:
-        :param g2:
-        :return:
-        only gene with the same key will be crossover, thus don't need to consider change key
-        """
         r = jax.random.uniform(rand_key, shape=g1.shape)
-        return jnp.where(r > 0.5, g1, g2)
+        new_gene = jnp.where(r > 0.5, g1, g2)
+        if is_conn:  # fix enabled
+            enabled = jnp.where(
+                g1[:, 2] + g2[:, 2] > 0,  # any of them is enabled
+                1,
+                0
+            )
+            new_gene = new_gene.at[:, 2].set(enabled)
+        return new_gene

tensorneat/algorithm/neat/ga/mutation/default.py

@@ -154,8 +154,8 @@ class DefaultMutation(BaseMutation):
         nodes_keys = jax.random.split(k1, num=nodes.shape[0])
         conns_keys = jax.random.split(k2, num=conns.shape[0])
 
-        new_nodes = jax.vmap(genome.node_gene.mutate, in_axes=(0, 0))(nodes_keys, nodes)
+        new_nodes = jax.vmap(genome.node_gene.mutate)(nodes_keys, nodes)
-        new_conns = jax.vmap(genome.conn_gene.mutate, in_axes=(0, 0))(conns_keys, conns)
+        new_conns = jax.vmap(genome.conn_gene.mutate)(conns_keys, conns)
 
         # nan nodes not changed
         new_nodes = jnp.where(jnp.isnan(nodes), jnp.nan, new_nodes)

tensorneat/algorithm/neat/gene/node/base.py

@@ -8,5 +8,5 @@ class BaseNodeGene(BaseGene):
     def __init__(self):
         super().__init__()
 
-    def forward(self, attrs, inputs):
+    def forward(self, attrs, inputs, is_output_node=False):
         raise NotImplementedError

tensorneat/algorithm/neat/gene/node/default.py

@@ -85,11 +85,14 @@ class DefaultNodeGene(BaseNodeGene):
                 (node1[4] != node2[4])
         )
 
-    def forward(self, attrs, inputs):
+    def forward(self, attrs, inputs, is_output_node=False):
         bias, res, act_idx, agg_idx = attrs
 
         z = agg(agg_idx, inputs, self.aggregation_options)
         z = bias + res * z
-        z = act(act_idx, z, self.activation_options)
+
+        # the last output node should not be activated
+        if not is_output_node:
+            z = act(act_idx, z, self.activation_options)
 
         return z

tensorneat/algorithm/neat/genome/default.py

@@ -25,19 +25,13 @@ class DefaultGenome(BaseGenome):
 
         if output_transform is not None:
             try:
-                aux = output_transform(jnp.zeros(num_outputs))
+                _ = output_transform(jnp.zeros(num_outputs))
             except Exception as e:
                 raise ValueError(f"Output transform function failed: {e}")
         self.output_transform = output_transform
 
     def transform(self, nodes, conns):
         u_conns = unflatten_conns(nodes, conns)
-
-        # DONE: Seems like there is a bug in this line
-        # conn_enable = jnp.where(~jnp.isnan(u_conns[0]), True, False)
-        # modified: exist conn and enable is true
-        # conn_enable = jnp.where( (~jnp.isnan(u_conns[0])) & (u_conns[0] == 1), True, False)
-        # advanced modified: when and only when enabled is True
         conn_enable = u_conns[0] == 1
 
         # remove enable attr
@@ -64,13 +58,7 @@ class DefaultGenome(BaseGenome):
 
             def hit():
                 ins = jax.vmap(self.conn_gene.forward, in_axes=(1, 0))(conns[:, :, i], values)
-                # ins = values * weights[:, i]
-
                 z = self.node_gene.forward(nodes_attrs[i], ins)
-                # z = agg(nodes[i, 4], ins, self.config.aggregation_options)  # z = agg(ins)
-                # z = z * nodes[i, 2] + nodes[i, 1]  # z = z * response + bias
-                # z = act(nodes[i, 3], z, self.config.activation_options)  # z = act(z)
-
                 new_values = values.at[i].set(z)
                 return new_values
 
@@ -78,7 +66,11 @@ class DefaultGenome(BaseGenome):
                 return values
 
             # the val of input nodes is obtained by the task, not by calculation
-            values = jax.lax.cond(jnp.isin(i, self.input_idx), miss, hit)
+            values = jax.lax.cond(
+                jnp.isin(i, self.input_idx),
+                miss,
+                hit
+            )
 
             return values, idx + 1
 

tensorneat/examples/brax/walker.py

@@ -18,7 +18,7 @@ if __name__ == '__main__':
                         activation_default=Act.tanh,
                     )
                 ),
-                pop_size=10,
+                pop_size=10000,
                 species_size=10,
             ),
         ),

tensorneat/pipeline.py

@@ -69,7 +69,7 @@ class Pipeline:
                         pop_transformed
                     )
 
-        fitnesses = jnp.where(jnp.isnan(fitnesses), -1e6, fitnesses)
+        # fitnesses = jnp.where(jnp.isnan(fitnesses), -1e6, fitnesses)
 
         alg_state = self.algorithm.tell(state.alg, fitnesses)
 
@@ -80,8 +80,10 @@ class Pipeline:
 
     def auto_run(self, ini_state):
         state = ini_state
+        print("start compile")
+        tic = time.time()
         compiled_step = jax.jit(self.step).lower(ini_state).compile()
-
+        print(f"compile finished, cost time: {time.time() - tic:.6f}s", )
         for _ in range(self.generation_limit):
 
             self.generation_timestamp = time.time()
@@ -91,11 +93,6 @@ class Pipeline:
             state, fitnesses = compiled_step(state)
 
             fitnesses = jax.device_get(fitnesses)
-            for idx, fitnesses_i in enumerate(fitnesses):
-                if np.isnan(fitnesses_i):
-                    print("Fitness is nan")
-                    print(previous_pop[0][idx], previous_pop[1][idx])
-                    assert False
 
             self.analysis(state, previous_pop, fitnesses)
 

tensorneat/problem/rl_env/rl_jit.py

@@ -8,7 +8,6 @@ from .. import BaseProblem
 class RLEnv(BaseProblem):
     jitable = True
 
-    # TODO: move output transform to algorithm
     def __init__(self):
         super().__init__()