-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmulti_stage_ppo_tuning.py
299 lines (284 loc) · 19.1 KB
/
multi_stage_ppo_tuning.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
import pickle
import os
from model.instance import Instance
from tools.common import directory
import torch
torch.autograd.set_detect_anomaly(True)
import pandas as pd
from torch.nn import Module
from torch.optim import Optimizer
from typing import Callable
from model.agent import MultiAgents_Batch, MAPPO_Loss, MAPPO_Losses, MultiAgent_OneInstance
import time as systime
from tools.common import load_instance, objective_value, freeze, unfreeze, unfreeze_all, freeze_several_and_unfreeze_others
from translators.graph2solution_translator import translate_solution
from debug.loss_tracker import LossTracker
import math
from model.reward_memory import Memory
# ==================================================================
# =*= MULTI-STAGE PROXIMAL POLICY OPTIMIZATION (PPO) FINE TUNING =*=
# ==================================================================
__author__ = "Anas Neumann - [email protected]"
__version__ = "1.0.0"
__license__ = "Apache 2.0 License"
PPO_CONF = {
"train_iterations": [300, 150, 150, 150, 300],
"opt_epochs": 3,
"clip_ratio": 0.1,
"policy_loss": 1.0,
"batch": 16,
"value_loss": 0.1,
"entropy": 0.02,
"discount_factor": 1.0,
"bias_variance_tradeoff": 1.0
}
LEARNING_RATE = 1e-3
AGENTS = ["outsourcing", "scheduling", "material_use"]
AGENT = 0
OUTSOURCING = 0
SCHEDULING = 1
MATERIAL_USE = 2
def save_models(agents: list[(Module, str)], embedding_stack: Module, shared_critic: Module, optimizer: Optimizer, run_number:int, complete_path: str):
index = str(run_number)
torch.save(embedding_stack.state_dict(), complete_path+'/gnn_weights_'+index+'.pth')
torch.save(shared_critic.state_dict(), complete_path+'/critic_weights_'+index+'.pth')
torch.save(optimizer.state_dict(), complete_path+'/adam_'+index+'.pth')
for agent, name in agents:
torch.save(agent.state_dict(), complete_path+'/'+name+'_weights_'+index+'.pth')
def scheduling_stage(target_instance : Instance, agents: list[(Module, str)], solve_function: Callable, path: str, epochs: int, iterations:int, batch_size: int, device: str, id: str, size: str, interactive: bool = False):
"""
First (1st) optimization stage: scheduling agent alone
"""
scheduling_optimizer = torch.optim.Adam(list(agents[SCHEDULING][AGENT].parameters()), lr=LEARNING_RATE)
losses = MAPPO_Losses(agent_names=[AGENTS[SCHEDULING]])
_vloss_TRACKER: LossTracker = LossTracker(xlabel="Training epochs (3 per batch of episodes)", ylabel="Value loss", title="Value loss", color="blue", show=interactive)
_scheduling_loss_TRACKER: LossTracker = LossTracker(xlabel="Training epochs (3 per batch of episodes)", ylabel="Scheduling loss (policy)", title="Scheduling loss (policy)", color="green", show=interactive)
_Cmax_TRACKER: LossTracker = LossTracker(xlabel="Solving episode", ylabel="Makespan", title="Final Makespan by episode", color="red", show=interactive)
_batch_replay_memory: list[MultiAgent_OneInstance] = []
reward_MEMORY: Memory = Memory(target_instance.size + "_" + str(target_instance.id))
for episode in range(iterations):
print(f"PPO solving episode: {episode+1}/{iterations}...")
transitions, reward_MEMORY, _, cmax, _ = solve_function(instance=target_instance, agents=agents, train=True, trainable=[0,1,0], device=device, fixed_alpha=1.0, reward_MEMORY=reward_MEMORY)
_batch_replay_memory.append(transitions)
_Cmax_TRACKER.update(cmax)
if (episode+1) % batch_size == 0:
print(f"PPO time for optimization after episode: {episode+1}/{iterations}!")
batch_result: MultiAgents_Batch = MultiAgents_Batch(
batch=_batch_replay_memory,
agent_names=[AGENTS[SCHEDULING]],
gamma=PPO_CONF['discount_factor'],
lam=PPO_CONF['bias_variance_tradeoff'],
weight_policy_loss=PPO_CONF['policy_loss'],
weight_value_loss=PPO_CONF['value_loss'],
weight_entropy_bonus=PPO_CONF['entropy'],
clipping_ratio=PPO_CONF['clip_ratio'])
for e in range(epochs):
print(f"\t Optimization epoch: {e+1}/{epochs}")
scheduling_optimizer.zero_grad()
training_loss, details = batch_result.compute_losses([agents[SCHEDULING]], return_details=True)
print(f"\t Multi-agent batch loss: {training_loss} - Differentiable computation graph = {training_loss.requires_grad}!")
training_loss.backward(retain_graph=False)
scheduling_optimizer.step()
details: MAPPO_Loss
_vloss_TRACKER.update(details.value_loss)
_scheduling_loss_TRACKER.update(details.get(AGENTS[SCHEDULING]).policy_loss)
losses.add(details)
_batch_replay_memory = []
_vloss_TRACKER.save(path+directory.solutions+'/'+size+'/stage_1_value_loss_'+id)
_scheduling_loss_TRACKER.save(path+directory.solutions+'/'+size+'/stage_1_scheduling_loss_'+id)
_Cmax_TRACKER.save(path+directory.solutions+'/'+size+'/stage_1_cmax_'+id)
return agents
def material_use_stage(target_instance : Instance, agents: list[(Module, str)], solve_function: Callable, path: str, epochs: int, iterations:int, batch_size: int, device: str, id: str, size: str, interactive: bool = False):
"""
Second (2nd) optimization stage: material use agent alone
"""
material_optimizer = torch.optim.Adam(list(agents[MATERIAL_USE][AGENT].parameters()), lr=LEARNING_RATE)
losses = MAPPO_Losses(agent_names=[AGENTS[MATERIAL_USE]])
_vloss_TRACKER: LossTracker = LossTracker(xlabel="Training epochs (3 per batch of episodes)", ylabel="Value loss", title="Value loss", color="blue", show=interactive)
_material_loss_TRACKER: LossTracker = LossTracker(xlabel="Training epochs (3 per batch of episodes)", ylabel="Material use loss (policy)", title="Material use loss (policy)", color="green", show=interactive)
_Cmax_TRACKER: LossTracker = LossTracker(xlabel="Solving episode", ylabel="Makespan", title="Final Makespan by episode", color="red", show=interactive)
_batch_replay_memory: list[MultiAgent_OneInstance] = []
reward_MEMORY: Memory = Memory(target_instance.size + "_" + str(target_instance.id))
for episode in range(iterations):
print(f"PPO solving episode: {episode+1}/{iterations}...")
transitions, reward_MEMORY, _, cmax, _ = solve_function(instance=target_instance, agents=agents, train=True, trainable=[0,0,1], device=device, fixed_alpha=1.0, reward_MEMORY=reward_MEMORY)
_batch_replay_memory.append(transitions)
_Cmax_TRACKER.update(cmax)
if (episode+1) % batch_size == 0:
print(f"PPO time for optimization after episode: {episode+1}/{iterations}!")
batch_result: MultiAgents_Batch = MultiAgents_Batch(
batch=_batch_replay_memory,
agent_names=[AGENTS[MATERIAL_USE]],
gamma=PPO_CONF['discount_factor'],
lam=PPO_CONF['bias_variance_tradeoff'],
weight_policy_loss=PPO_CONF['policy_loss'],
weight_value_loss=PPO_CONF['value_loss'],
weight_entropy_bonus=PPO_CONF['entropy'],
clipping_ratio=PPO_CONF['clip_ratio'])
for e in range(epochs):
print(f"\t Optimization epoch: {e+1}/{epochs}")
material_optimizer.zero_grad()
training_loss, details = batch_result.compute_losses([agents[MATERIAL_USE]], return_details=True)
print(f"\t Multi-agent batch loss: {training_loss} - Differentiable computation graph = {training_loss.requires_grad}!")
training_loss.backward(retain_graph=False)
material_optimizer.step()
details: MAPPO_Loss
_vloss_TRACKER.update(details.value_loss)
_material_loss_TRACKER.update(details.get(AGENTS[MATERIAL_USE]).policy_loss)
losses.add(details)
_batch_replay_memory = []
_vloss_TRACKER.save(path+directory.solutions+'/'+size+'/stage_2_value_loss_'+id)
_material_loss_TRACKER.save(path+directory.solutions+'/'+size+'/stage_2_material_use_loss_'+id)
_Cmax_TRACKER.save(path+directory.solutions+'/'+size+'/stage_2_cmax_'+id)
return agents
def outsourcing_stage(target_instance : Instance, agents: list[(Module, str)], solve_function: Callable, path: str, epochs: int, iterations:int, batch_size: int, device: str, id: str, size: str, substage: int, interactive: bool = False):
"""
Third (3rd) optimization stage: outsourcing agent alone
"""
outsourcing_optimizer = torch.optim.Adam(list(agents[OUTSOURCING][AGENT].parameters()), lr=LEARNING_RATE)
losses = MAPPO_Losses(agent_names=[AGENTS[OUTSOURCING]])
_vloss_TRACKER: LossTracker = LossTracker(xlabel="Training epochs (3 per batch of episodes)", ylabel="Value loss", title="Value loss", color="blue", show=interactive)
_outsourcing_loss_TRACKER: LossTracker = LossTracker(xlabel="Training epochs (3 per batch of episodes)", ylabel="Outsourcing loss (policy)", title="Outsourcing loss (policy)", color="pink", show=interactive)
_Cmax_TRACKER: LossTracker = LossTracker(xlabel="Solving episode", ylabel="Makespan", title="Final Makespan by episode", color="red", show=interactive)
_cost_TRACKER: LossTracker = LossTracker(xlabel="Solving episode", ylabel="Cost", title="Final Cost by episode", color="orange", show=interactive)
_batch_replay_memory: list[MultiAgent_OneInstance] = []
reward_MEMORY: Memory = Memory(target_instance.size + "_" + str(target_instance.id))
for episode in range(iterations):
print(f"PPO solving episode: {episode+1}/{iterations}...")
transitions, reward_MEMORY, _, cmax, cost = solve_function(instance=target_instance, agents=agents, train=True, trainable=[1,0,0], device=device, reward_MEMORY=reward_MEMORY)
_Cmax_TRACKER.update(cmax)
_cost_TRACKER.update(cost)
_batch_replay_memory.append(transitions)
if (episode+1) % batch_size == 0:
print(f"PPO time for optimization after episode: {episode+1}/{iterations}!")
batch_result: MultiAgents_Batch = MultiAgents_Batch(
batch=_batch_replay_memory,
agent_names=[AGENTS[OUTSOURCING]],
gamma=PPO_CONF['discount_factor'],
lam=PPO_CONF['bias_variance_tradeoff'],
weight_policy_loss=PPO_CONF['policy_loss'],
weight_value_loss=PPO_CONF['value_loss'],
weight_entropy_bonus=PPO_CONF['entropy'],
clipping_ratio=PPO_CONF['clip_ratio'])
for e in range(epochs):
print(f"\t Optimization epoch: {e+1}/{epochs}")
outsourcing_optimizer.zero_grad()
training_loss, details = batch_result.compute_losses([agents[OUTSOURCING]], return_details=True)
print(f"\t Multi-agent batch loss: {training_loss} - Differentiable computation graph = {training_loss.requires_grad}!")
training_loss.backward(retain_graph=False)
outsourcing_optimizer.step()
details: MAPPO_Loss
_vloss_TRACKER.update(details.value_loss)
_outsourcing_loss_TRACKER.update(details.get(AGENTS[OUTSOURCING]).policy_loss)
losses.add(details)
_batch_replay_memory = []
_vloss_TRACKER.save(path+directory.solutions+'/'+size+'/stage_3_'+str(substage)+'_value_loss_'+id)
_outsourcing_loss_TRACKER.save(path+directory.solutions+'/'+size+'/stage_3_'+str(substage)+'_outsourcing_loss_'+id)
_cost_TRACKER.save(path+directory.solutions+'/'+size+'/stage_3_'+str(substage)+'_cost_'+id)
_Cmax_TRACKER.save(path+directory.solutions+'/'+size+'/stage_3_'+str(substage)+'_cmax_'+id)
return agents
def multi_agent_stage(target_instance : Instance, agents: list[(Module, str)], solve_function: Callable, path: str, epochs: int, iterations:int, batch_size: int, device: str, id: str, size: str, start_time: int, interactive: bool = False):
"""
Last (4th) optimization stage: all three agents together
"""
multi_agent_optimizer = torch.optim.Adam(list(agents[SCHEDULING][AGENT].parameters()) + list(agents[OUTSOURCING][AGENT].parameters()) + list(agents[MATERIAL_USE][AGENT].parameters()), lr=LEARNING_RATE)
losses = MAPPO_Losses(agent_names=[name for _,name in agents])
_vloss_TRACKER: LossTracker = LossTracker(xlabel="Training epochs (3 per batch of episodes)", ylabel="Value loss", title="Value loss", color="blue", show=interactive)
_scheduling_loss_TRACKER: LossTracker = LossTracker(xlabel="Training epochs (3 per batch of episodes)", ylabel="Scheduling loss (policy)", title="Scheduling loss (policy)", color="green", show=interactive)
_outsourcing_loss_TRACKER: LossTracker = LossTracker(xlabel="Training epochs (3 per batch of episodes)", ylabel="Outsourcing loss (policy)", title="Outsourcing loss (policy)", color="pink", show=interactive)
_Cmax_TRACKER: LossTracker = LossTracker(xlabel="Solving episode", ylabel="Makespan", title="Final Makespan by episode", color="red", show=interactive)
_cost_TRACKER: LossTracker = LossTracker(xlabel="Solving episode", ylabel="Cost", title="Final Cost by episode", color="orange", show=interactive)
_best_obj: float = math.inf
_best_episode: int = 0
_time_to_best: float = 0
_batch_replay_memory: list[MultiAgent_OneInstance] = []
reward_MEMORY: Memory = Memory(target_instance.size + "_" + str(target_instance.id))
for episode in range(iterations):
print(f"PPO solving episode: {episode+1}/{iterations}...")
transitions, reward_MEMORY, graph, cmax, cost = solve_function(instance=target_instance, agents=agents, train=True, trainable=[1,1,1], device=device, reward_MEMORY=reward_MEMORY)
_batch_replay_memory.append(transitions)
_current_obj = objective_value(cmax, cost, target_instance.w_makespan)
if _current_obj < _best_obj:
_best_obj = _current_obj
_best_episode = episode
_time_to_best = systime.time()-start_time
_Cmax_TRACKER.update(cmax)
_cost_TRACKER.update(cost)
if (episode+1) % batch_size == 0:
print(f"PPO time for optimization after episode: {episode+1}/{iterations}!")
batch_result: MultiAgents_Batch = MultiAgents_Batch(
batch=_batch_replay_memory,
agent_names=[name for _,name in agents],
gamma=PPO_CONF['discount_factor'],
lam=PPO_CONF['bias_variance_tradeoff'],
weight_policy_loss=PPO_CONF['policy_loss'],
weight_value_loss=PPO_CONF['value_loss'],
weight_entropy_bonus=PPO_CONF['entropy'],
clipping_ratio=PPO_CONF['clip_ratio'])
for e in range(epochs):
print(f"\t Optimization epoch: {e+1}/{epochs}")
multi_agent_optimizer.zero_grad()
training_loss, details = batch_result.compute_losses(agents, return_details=True)
print(f"\t Multi-agent batch loss: {training_loss} - Differentiable computation graph = {training_loss.requires_grad}!")
training_loss.backward(retain_graph=False)
multi_agent_optimizer.step()
details: MAPPO_Loss
_vloss_TRACKER.update(details.value_loss)
_scheduling_loss_TRACKER.update(details.get(AGENTS[SCHEDULING]).policy_loss)
_outsourcing_loss_TRACKER.update(details.get(AGENTS[OUTSOURCING]).policy_loss)
losses.add(details)
_batch_replay_memory = []
_vloss_TRACKER.save(path+directory.solutions+'/'+size+'/final_value_loss_'+id)
_scheduling_loss_TRACKER.save(path+directory.solutions+'/'+size+'/final_scheduling_loss_'+id)
_outsourcing_loss_TRACKER.save(path+directory.solutions+'/'+size+'/final_outsourcing_loss_'+id)
_cost_TRACKER.save(path+directory.solutions+'/'+size+'/final_cost_'+id)
_Cmax_TRACKER.save(path+directory.solutions+'/'+size+'/final_cmax_'+id)
with open(path+directory.solutions+'/'+size+'/fine_tune_gns_graph_'+id+'.pkl', 'wb') as f:
pickle.dump(graph, f)
with open(path+directory.solutions+'/'+size+'/fine_tune_gns_solution_'+id+'.pkl', 'wb') as f:
pickle.dump(translate_solution(graph, target_instance), f)
with open(path+directory.solutions+'/'+size+'/fine_tune_losses_'+id+'.pkl', 'wb') as f:
pickle.dump(losses, f)
return _best_obj, _best_episode, _time_to_best
def multi_stage_fine_tuning(agents: list[(Module, str)], embedding_stack: Module, shared_critic: Module, path: str, solve_function: Callable, device: str, id: str, size: str, interactive: bool = False, debug_mode: bool=False):
"""
Multi-stage PPO function to fine-tune agents on the target instance
"""
_itrs: int = PPO_CONF['train_iterations']
_epochs: int = PPO_CONF['opt_epochs']
_batch_size: int = PPO_CONF['batch']
_instance: Instance = load_instance(path+directory.instances+'/test/'+size+'/instance_'+id+'.pkl')
print(f"Target instance {size}_{id} loaded....")
_start_time = systime.time()
embedding_stack.train()
shared_critic.train()
for agent,_ in agents:
agent.train()
print("I. FINE TUNING STAGE 1: scheduling agent")
freeze_several_and_unfreeze_others(agents, [AGENTS[OUTSOURCING], AGENTS[MATERIAL_USE]])
agents = scheduling_stage(_instance, agents, solve_function, path, _epochs, _itrs[0], _batch_size, device, id, size, interactive)
print("II. FINE TUNING STAGE 2: material use agent with freezed embedding and critic layers")
freeze_several_and_unfreeze_others(agents, [AGENTS[SCHEDULING], AGENTS[OUTSOURCING]])
freeze(embedding_stack)
freeze(shared_critic)
agents = material_use_stage(_instance, agents, solve_function, path, _epochs, _itrs[1], _batch_size, device, id, size, interactive)
print("III. FINE TUNING STAGE 3.1: outsourcing agent with freezed embedding and critic layers")
freeze_several_and_unfreeze_others(agents, [AGENTS[SCHEDULING], AGENTS[MATERIAL_USE]])
freeze(embedding_stack)
freeze(shared_critic)
agents = outsourcing_stage(_instance, agents, solve_function, path, _epochs, _itrs[2], _batch_size, device, id, size, 1, interactive)
print("III. FINE TUNING STAGE 3.2: outsourcing agent and all layers")
unfreeze(embedding_stack)
unfreeze(shared_critic)
agents = outsourcing_stage(_instance, agents, solve_function, path, _epochs, _itrs[3], _batch_size, device, id, size, 2, interactive)
print("IV. FINE TUNING STAGE 4: multi-agent with all layers")
unfreeze_all(agents)
_best_obj, _best_episode, _time_to_best = multi_agent_stage(_instance, agents, solve_function, path, _epochs, _itrs[4], _batch_size, device, id, size, _start_time, interactive)
final_metrics = pd.DataFrame({
'index': [_instance.id],
'value': [_best_obj],
'episode': [_best_episode],
'time_to_best': [_time_to_best],
'computing_time': [systime.time()-_start_time],
'device_used': [device]})
final_metrics.to_csv(path+directory.instances+'/test/'+size+'/solution_fine_tuned_gns_'+id+'.csv', index=False)