优化器类的定义和实现

This commit is contained in:
2025-11-14 11:34:48 +08:00
parent a37ee38369
commit 79b09b2578
11 changed files with 4004 additions and 0 deletions

314
opt/predictor.py Normal file
View File

@@ -0,0 +1,314 @@
from core.common import *
from opt.smm.basis import BaseOpt
import torch
class Predictor:
def __init__(self):
pass
@staticmethod
def training(self, params):
pass
@staticmethod
def testing(self, params):
pass
@staticmethod
def predict(self, cp_points, cp_nozzle, board_width=None, board_height=None):
pass
class Net(torch.nn.Module):
def __init__(self, input_size, hidden_size=1000, output_size=1):
super(Net, self).__init__()
self.fc1 = torch.nn.Linear(input_size, hidden_size)
self.relu = torch.nn.ReLU() # <20><><EFBFBD><EFBFBD><EEBAAF>
self.fc2 = torch.nn.Linear(hidden_size, hidden_size)
# self.relu1 = torch.nn.ReLU() # <20><><EFBFBD><EFBFBD><EEBAAF>
self.fc3 = torch.nn.Linear(hidden_size, output_size)
def forward(self, x):
x = self.fc1(x)
# x = self.relu(x)
x = self.fc2(x)
x = self.relu(x)
x = self.fc3(x)
return x
class NeuralPredictor(Predictor, BaseOpt):
def __init__(self):
super().__init__()
self.min_placement_points = 10
self.max_placement_points = 1000
self.max_component_types = 30
self.default_feeder_limit = 1
self.max_nozzle_types = 4
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.net = Net(input_size=self.get_feature(), output_size=1).to(self.device)
self.net_file = 'opt/param.pth'
try:
self.net.load_state_dict(torch.load(self.net_file, map_location=self.device))
except:
warnings.warn('the parameters of neural net model load failed', UserWarning)
def init_weights(self):
for m in self.net.modules():
if isinstance(m, torch.nn.Linear):
torch.nn.init.xavier_uniform_(m.weight)
torch.nn.init.zeros_(m.bias)
def subobjective(self, cp_points, cp_nozzle, config):
if len(cp_points.keys()) or sum(cp_points.values()) == 0:
return 0, 0, 0, 0
nozzle_heads, nozzle_points = defaultdict(int), defaultdict(int)
for idx, points in cp_points.items():
if points == 0:
continue
nozzle = cp_nozzle[idx]
nozzle_points[nozzle] += points
nozzle_heads[nozzle] = 1
anc_round_counter = 0
while sum(nozzle_heads.values()) != config.head_num:
max_cycle_nozzle = None
for nozzle, head_num in nozzle_heads.items():
if max_cycle_nozzle is None or nozzle_points[nozzle] / head_num > nozzle_points[max_cycle_nozzle] / \
nozzle_heads[max_cycle_nozzle]:
max_cycle_nozzle = nozzle
assert max_cycle_nozzle is not None
nozzle_heads[max_cycle_nozzle] += 1
head_nozzle_assignment, min_cost = None, None
# generate initial nozzle group
nozzle_group = []
# averagely assign for the same type of nozzles, and generate nozzle group
nozzle_points_cpy = copy.deepcopy(nozzle_points)
for nozzle, heads in nozzle_heads.items():
points = nozzle_points_cpy[nozzle] // heads
for _ in range(heads):
nozzle_group.append([nozzle, points])
nozzle_points_cpy[nozzle] -= heads * points
for idx, [nozzle, _] in enumerate(nozzle_group):
if nozzle_points_cpy[nozzle]:
nozzle_group[idx][1] += 1
nozzle_points_cpy[nozzle] -= 1
while True:
# assign nozzle group to each head
nozzle_group.sort(key=lambda x: -x[1])
tmp_head_nozzle_assignment = []
head_total_points = [0 for _ in range(config.head_num)]
for idx, nozzle_item in enumerate(nozzle_group):
if idx < config.head_num:
tmp_head_nozzle_assignment.append([nozzle_item.copy()])
head_total_points[idx] += nozzle_item[1]
else:
min_head = np.argmin(head_total_points)
tmp_head_nozzle_assignment[min_head].append(nozzle_item.copy())
head_total_points[min_head] += nozzle_item[1]
cost = config.cycle_time * max(head_total_points)
for head in range(config.head_num):
for cycle in range(len(tmp_head_nozzle_assignment[head])):
if cycle + 1 == len(tmp_head_nozzle_assignment[head]):
if tmp_head_nozzle_assignment[head][cycle][0] != tmp_head_nozzle_assignment[head][-1][0]:
cost += self.nozzle_change_weight
else:
if tmp_head_nozzle_assignment[head][cycle][0] != tmp_head_nozzle_assignment[head][cycle + 1][0]:
cost += self.nozzle_change_weight
while True:
min_head, max_head = np.argmin(head_total_points), np.argmax(head_total_points)
min_head_nozzle, max_head_nozzle = tmp_head_nozzle_assignment[min_head][-1][0], \
tmp_head_nozzle_assignment[max_head][-1][0]
if min_head_nozzle == max_head_nozzle:
break
min_head_list, max_head_list = [min_head], [max_head]
minmax_head_points = 0
for head in range(config.head_num):
if head in min_head_list or head in max_head_list:
minmax_head_points += head_total_points[head]
continue
# the max/min heads with the sum nozzle type
if tmp_head_nozzle_assignment[head][-1][0] == tmp_head_nozzle_assignment[min_head][-1][0]:
min_head_list.append(head)
minmax_head_points += head_total_points[head]
if tmp_head_nozzle_assignment[head][-1][0] == tmp_head_nozzle_assignment[max_head][-1][0]:
max_head_list.append(head)
minmax_head_points += head_total_points[head]
# todo: restriction of available nozzle
# the reduction of cycles is not offset the cost of nozzle change
average_points = minmax_head_points // (len(min_head_list) + len(max_head_list))
reminder_points = minmax_head_points % (len(min_head_list) + len(max_head_list))
max_cycle = average_points + (1 if reminder_points > 0 else 0)
for head in range(config.head_num):
if head in min_head_list or head in max_head_list:
continue
max_cycle = max(max_cycle, head_total_points[head])
nozzle_change_counter = 0
for head in min_head_list:
if tmp_head_nozzle_assignment[head][0] == tmp_head_nozzle_assignment[head][-1]:
nozzle_change_counter += 2
else:
nozzle_change_counter += 1
if self.cycle_weight * (max(head_total_points) - max_cycle) < self.nozzle_change_weight * nozzle_change_counter:
break
cost -= self.cycle_weight * (max(head_total_points) - max_cycle) - self.nozzle_change_weight * nozzle_change_counter
required_points = 0 # <20><><EFBFBD><EFBFBD>̯<EFBFBD><CCAF><EFBFBD><EFBFBD>װ<EFBFBD><D7B0><EFBFBD><EFBFBD><EFBFBD>϶<EFBFBD><CFB6><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>
for head in min_head_list:
points = average_points - head_total_points[head]
tmp_head_nozzle_assignment[head].append([max_head_nozzle, points])
head_total_points[head] = average_points
required_points += points
for head in max_head_list:
tmp_head_nozzle_assignment[head][-1][1] -= required_points // len(max_head_list)
head_total_points[head] -= required_points // len(max_head_list)
required_points -= (required_points // len(max_head_list)) * len(max_head_list)
for head in max_head_list:
if required_points <= 0:
break
tmp_head_nozzle_assignment[head][-1][1] -= 1
head_total_points[head] -= 1
required_points -= 1
if min_cost is None or cost < min_cost:
min_cost = cost
head_nozzle_assignment = copy.deepcopy(tmp_head_nozzle_assignment)
else:
break
# <20><><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>һ<EFBFBD><D2BB><EFBFBD><EFBFBD><EFBFBD><EFBFBD>
idx, nozzle = 0, nozzle_group[0][0]
for idx, [nozzle_, _] in enumerate(nozzle_group):
if nozzle_ != nozzle:
break
average_points, remainder_points = nozzle_points[nozzle] // (idx + 1), nozzle_points[nozzle] % (idx + 1)
nozzle_group.append([nozzle, 0])
for idx, [nozzle_, _] in enumerate(nozzle_group):
if nozzle_ == nozzle:
nozzle_group[idx][1] = average_points + (1 if remainder_points > 0 else 0)
remainder_points -= 1
cycle_counter, nozzle_change_counter = 0, 0
for head in range(config.head_num):
head_cycle_counter = 0
for cycle in range(len(head_nozzle_assignment[head])):
if cycle + 1 == len(head_nozzle_assignment[head]):
if head_nozzle_assignment[head][0][0] != head_nozzle_assignment[head][-1][0]:
nozzle_change_counter += 1
else:
if head_nozzle_assignment[head][cycle][0] != head_nozzle_assignment[head][cycle + 1][0]:
nozzle_change_counter += 1
head_cycle_counter += head_nozzle_assignment[head][cycle][1]
cycle_counter = max(cycle_counter, head_cycle_counter)
# === Ԫ<><D4AA>ʰȡ<CAB0><C8A1><EFBFBD><EFBFBD>Ԥ<EFBFBD><D4A4> ===
cp_info = []
for idx, points in cp_points.items():
if points == 0:
continue
feeder_limit = 1 # todo: <20><>ʱ<EFBFBD><CAB1><EFBFBD><EFBFBD><EFBFBD><EFBFBD>һ<EFBFBD><D2BB><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD><EFBFBD>
reminder_points = points % feeder_limit
for _ in range(feeder_limit):
cp_info.append([idx, points // feeder_limit + (1 if reminder_points > 0 else 0), cp_nozzle[idx]])
reminder_points -= 1
cp_info.sort(key=lambda x: -x[1])
nozzle_level, nozzle_counter = defaultdict(int), defaultdict(int)
level_points = defaultdict(int)
for info in cp_info:
nozzle = info[2]
if nozzle_counter[nozzle] and nozzle_counter[nozzle] % nozzle_heads[nozzle] == 0:
nozzle_level[nozzle] += 1
level = nozzle_level[nozzle]
level_points[level] = max(level_points[level], info[1])
nozzle_counter[nozzle] += 1
pickup_counter = sum(points for points in level_points.values())
return cycle_counter, nozzle_change_counter, anc_round_counter, pickup_counter
def encode(self, cp_points: defaultdict[str], cp_nozzle: defaultdict[str], board_width, board_height, config):
assert len(cp_points.keys()) == len(cp_nozzle.keys())
# === general info ===
total_points = sum(points for points in cp_points.values())
total_component_types, total_nozzle_types = len(cp_points.keys()), len(set(cp_nozzle.values()))
data = [total_points, total_component_types, total_nozzle_types]
data.extend([board_width, board_height])
# === heuristic info ===
cycle, nozzle_change, anc_move, pickup = self.subobjective(cp_points, cp_nozzle, config)
data.extend([cycle, nozzle_change, anc_move, pickup])
# === nozzle info ===
nozzle_points = defaultdict(int)
for cp_idx, nozzle in cp_nozzle.items():
nozzle_points[cp_nozzle[cp_idx]] += cp_points[cp_idx] # points for different nozzle type
nozzle_items = [[nozzle, points] for nozzle, points in nozzle_points.items()]
nozzle_items = sorted(nozzle_items, key=lambda x: x[1], reverse=True)
nz2idx = defaultdict(int)
nozzle_slice = [0 for _ in range(self.max_nozzle_types)]
for idx, [nozzle, points] in enumerate(nozzle_items):
nz2idx[nozzle] = idx if idx < self.max_nozzle_types else self.max_nozzle_types - 1
nozzle_slice[idx if idx < self.max_nozzle_types else -1] += points
data.extend(nozzle_slice)
# === part info ===
part_data_slice = defaultdict(list)
for idx in range(self.max_nozzle_types):
part_data_slice[idx] = []
cp_items = [[component, points] for component, points in cp_points.items()]
cp_items = sorted(cp_items, key=lambda x: (x[1], nz2idx[cp_nozzle[x[0]]] * 0.1 + x[1]), reverse=True)
for component, points in cp_items:
nozzle = cp_nozzle[component]
part_data_slice[nz2idx[nozzle]].append(points)
data_slice = [0 for _ in range(self.max_nozzle_types)]
for idx, part_list in part_data_slice.items():
data_slice[idx] = len(part_list)
data.extend(data_slice)
for idx in range(self.max_nozzle_types):
if len(part_data_slice[idx]) <= self.max_component_types:
part_data_slice[idx].extend([0 for _ in range(self.max_component_types - len(part_data_slice[idx]))])
else:
part_data_slice[idx] = part_data_slice[idx][:self.max_component_types]
data.extend(part_data_slice[idx])
return data
def get_feature(self):
return (self.max_component_types + 2) * self.max_nozzle_types + 5 + 4
def eval(self, cp_points, cp_nozzle, board_width, board_height, config):
encoding = np.array(self.encode(cp_points, cp_nozzle, board_width, board_height, config))
encoding = torch.from_numpy(encoding.reshape((-1, np.shape(encoding)[0]))).float().to(self.device)
return self.net(encoding)[0, 0].item()