base evolution model. needs different memory handling

This commit is contained in:
zomseffen 2022-03-11 14:19:55 +01:00
parent 33b5d9c83e
commit 4a05baa103
4 changed files with 218 additions and 48 deletions

View file

@ -146,7 +146,8 @@ class LabyrinthWorld(World):
# adding subjects
from labirinth_ai.Subject import Hunter, Herbivore
while len(self.subjects) < 2:
for _ in range(10):
while True:
px = random.randint(self.max_room_dim, self.board_shape[0] - self.max_room_dim)
py = random.randint(self.max_room_dim, self.board_shape[1] - self.max_room_dim)
if self.board[px, py] == 1:
@ -154,8 +155,10 @@ class LabyrinthWorld(World):
self.ins += self.subjects[-1].x_in
self.actions += self.subjects[-1].actions
self.targets += self.subjects[-1].target
break
while len(self.subjects) < 10:
for _ in range(40):
while True:
px = random.randint(self.max_room_dim, self.board_shape[0] - self.max_room_dim)
py = random.randint(self.max_room_dim, self.board_shape[1] - self.max_room_dim)
if self.board[px, py] == 1:
@ -163,6 +166,7 @@ class LabyrinthWorld(World):
self.ins += self.subjects[-1].x_in
self.actions += self.subjects[-1].actions
self.targets += self.subjects[-1].target
break
for x in range(self.board_shape[0]):
for y in range(self.board_shape[1]):
@ -173,36 +177,14 @@ class LabyrinthWorld(World):
def update(self):
# start = time.time()
if self.model is None:
for sub in self.subjects:
sub.calculateAction(self)
else:
states = list(map(lambda e: e.createState(self), self.subjects))
states = sum(list(map(lambda e: [e, e, e, e], states)), [])
vals = self.model.predict(states)
vals = np.reshape(np.transpose(np.reshape(vals, (len(self.subjects), 4, 2)), (0, 2, 1)),
(len(self.subjects), 1, 8))
list(map(lambda e: e[1].calculateAction(self, vals[e[0]], states[e[0]]), enumerate(self.subjects)))
for sub in self.subjects:
if sub.alive:
sub.update(self, doTrain=self.model is None)
sub.update(self)
sub.tick += 1
if self.model is not None:
if self.round >= self.nextTrain:
samples = list(map(lambda e: e.generateSamples(), self.subjects))
states = sum(list(map(lambda e: e[0], samples)), [])
targets = sum(list(map(lambda e: e[1], samples)), [])
self.model.fit(states, targets)
self.nextTrain = self.batchsize / 5
self.round = 0
for sub in self.subjects:
if len(sub.samples) > 20*self.batchsize:
sub.samples = sub.samples[:-20*self.batchsize]
else:
self.round += 1
new_subjects = []
kill_table = {}
live_table = {}

View file

@ -13,6 +13,7 @@ print(f"Using {device} device")
# Define model
class BaseModel(nn.Module):
evolutionary = False
def __init__(self, view_dimension, action_num, channels):
super(BaseModel, self).__init__()
self.flatten = nn.Flatten()
@ -39,6 +40,7 @@ class BaseModel(nn.Module):
actions.append(self.actions[action](x_flat))
return torch.stack(actions, dim=1)
class BaseDataSet(Dataset):
def __init__(self, states, targets):
assert len(states) == len(targets), "Needs to have as many states as targets!"
@ -87,7 +89,7 @@ def train(states, targets, model, optimizer):
# Backpropagation
optimizer.zero_grad()
loss.backward()
loss.backward(retain_graph=True)
optimizer.step()
if batch % 100 == 0:
@ -100,7 +102,7 @@ def train(states, targets, model, optimizer):
if __name__ == '__main__':
sample = np.random.random((1, 4, 11, 11))
sample = np.random.random((1, 486))
model = BaseModel(5, 4, 4).to(device)
print(model)
@ -109,7 +111,7 @@ if __name__ == '__main__':
# test = test.cpu().detach().numpy()
print(test)
state = np.random.random((4, 11, 11))
state = np.random.random((486,))
target = np.random.random((4, 2))
states = [
[state],

View file

@ -0,0 +1,176 @@
import torch
from torch import nn
import numpy as np
import tqdm
from torch.utils.data import Dataset, DataLoader
from labirinth_ai.Models.BaseModel import device
class NodeGene:
valid_types = ['sensor', 'hidden', 'output']
def __init__(self, node_id, node_type, bias=None):
assert node_type in self.valid_types, 'Unknown node type!'
self.node_id = node_id
self.node_type = node_type
if node_type == 'hidden':
assert bias is not None, 'Expected a bias for hidden node types!'
self.bias = bias
else:
self.bias = None
class ConnectionGene:
def __init__(self, start, end, enabled, innovation_num, weight=None, recurrent=False):
self.start = start
self.end = end
self.enabled = enabled
self.innvovation_num = innovation_num
self.recurrent = recurrent
if weight is None:
self.weight = np.random.random(1)[0] * 2 - 1.0
else:
self.weight = weight
class EvolutionModel(nn.Module):
evolutionary = True
def __init__(self, view_dimension, action_num, channels, genes=None):
super(EvolutionModel, self).__init__()
self.flatten = nn.Flatten()
self.action_num = action_num
self.viewD = view_dimension
self.channels = channels
if genes is None:
self.num_input_nodes = channels * (2 * self.viewD + 1) * (2 * self.viewD + 1) + 2
self.genes = {'nodes': {}, 'connections': []}
node_id = 0
for _ in range(self.num_input_nodes):
self.genes['nodes'][node_id] = NodeGene(node_id, 'sensor')
node_id += 1
first_action = node_id
for _ in range(action_num * 2):
self.genes['nodes'][node_id] = NodeGene(node_id, 'output')
node_id += 1
for index in range(self.num_input_nodes):
for action in range(action_num * 2):
self.genes['connections'].append(
ConnectionGene(index, first_action + action, True, index*(action_num * 2) + action)
)
self.incoming_connections = {}
for connection in self.genes['connections']:
if connection.end not in self.incoming_connections.keys():
self.incoming_connections[connection.end] = []
self.incoming_connections[connection.end].append(connection)
self.layers = {}
self.indices = {}
self.has_recurrent = False
non_recurrent_indices = {}
with torch.no_grad():
for key, value in self.incoming_connections.items():
value.sort(key=lambda element: element.start)
lin = nn.Linear(len(value), 1, bias=self.genes['nodes'][key].bias is not None)
for index, connection in enumerate(value):
lin.weight[0, index] = value[index].weight
if self.genes['nodes'][key].bias is not None:
lin.bias[0] = self.genes['nodes'][key].bias
non_lin = nn.ELU()
sequence = nn.Sequential(
lin,
non_lin
)
self.add_module('layer_' + str(key), sequence)
self.layers[key] = sequence
self.indices[key] = list(map(lambda element: element.start, value))
non_recurrent_indices[key] = list(filter(lambda element: not element.recurrent, value))
if not self.has_recurrent and len(non_recurrent_indices[key]) != len(self.indices[key]):
self.has_recurrent = True
non_recurrent_indices[key] = list(map(lambda element: element.start, non_recurrent_indices[key]))
rank_of_node = {}
for i in range(self.num_input_nodes):
rank_of_node[i] = 0
layers_to_add = list(non_recurrent_indices.items())
while len(layers_to_add) > 0:
for index, (key, incoming_nodes) in enumerate(list(layers_to_add)):
max_rank = -1
all_ranks_found = True
for incoming_node in incoming_nodes:
if incoming_node in rank_of_node.keys():
max_rank = max(max_rank, rank_of_node[incoming_node])
else:
all_ranks_found = False
if all_ranks_found:
rank_of_node[key] = max_rank + 1
layers_to_add = list(filter(lambda element: element[0] not in rank_of_node.keys(), layers_to_add))
ranked_layers = list(rank_of_node.items())
ranked_layers.sort(key=lambda element: element[1])
ranked_layers = list(filter(lambda element: element[1] > 0, ranked_layers))
self.layer_order = list(map(lambda element: element[0], ranked_layers))
self.memory = torch.Tensor((max(map(lambda element: element[1].node_id, self.genes['nodes'].items())) + 1))
def forward(self, x, memory=None):
x_flat = self.flatten(x)
if memory is None:
memory = torch.Tensor(self.memory)
outs = []
for batch_element in x_flat:
memory[0:self.num_input_nodes] = batch_element
for layer_index in self.layer_order:
memory[layer_index] = self.layers[layer_index](memory[self.indices[layer_index]])
outs.append(memory[self.num_input_nodes: self.num_input_nodes + self.action_num * 2])
outs = torch.stack(outs)
self.memory = torch.Tensor(memory)
return torch.reshape(outs, (x.shape[0], 4, 2))
else:
memory[:, 0:self.num_input_nodes] = x
for layer_index in self.layer_order:
memory[:, layer_index] = self.layers[layer_index](memory[:, self.indices[layer_index]])
return torch.reshape(
memory[:, self.num_input_nodes: self.num_input_nodes + self.action_num * 2],
(x.shape[0], 4, 2))
if __name__ == '__main__':
sample = np.random.random((1, 486))
model = EvolutionModel(5, 4, 4).to(device)
print(model)
print(model.has_recurrent)
test = model(torch.tensor(sample, dtype=torch.float32))
# test = test.cpu().detach().numpy()
print(test)
state = np.random.random((1, 486))
target = np.random.random((4, 2))
states = [
[state],
[state],
[state],
[state],
]
targets = [
[target],
[target],
[target],
[target],
]
optimizer = torch.optim.RMSprop(model.parameters(), lr=1e-3)
from labirinth_ai.Models.BaseModel import train
train(states, targets, model, optimizer)

View file

@ -382,6 +382,8 @@ class NetLearner(Subject):
self.lastRewards = []
self.accumulated_rewards = 0
def visualize(self):
print(self.name)
layers = self.model.get_weights()
@ -542,6 +544,8 @@ class NetLearner(Subject):
self.train()
self.nextTrain = min(self.batchsize + self.nextTrain, (self.historySizeMul + 1) * self.batchsize)
self.accumulated_rewards += self.lastReward
self.lastAction = self.action
self.lastState = self.state
self.lastReward = 0
@ -728,10 +732,12 @@ class Herbivore(NetLearner):
if len(action) == 2:
if len(world.subjectDict[(self.x + action[0], self.y + action[1])]) > 0:
for sub in world.subjectDict[(self.x + action[0], self.y + action[1])]:
if isinstance(sub, Hunter):
if sub.alive:
self.kills += 1
sub.alive = False
self.alive = True
sub.kills += 1
sub.alive = True
sub.lastReward += 10
self.alive = False
self.lastRewards = []
if right in directions:
@ -795,6 +801,10 @@ class Herbivore(NetLearner):
return action
def respawnUpdate(self, x, y, world: LabyrinthWorld):
super(Herbivore, self).respawnUpdate(x, y, world)
self.lastReward -= 1
class Hunter(NetLearner):
name = 'Hunter'