Skip to content

Commit

Permalink
bug fix in the mp implementation
Browse files Browse the repository at this point in the history
  • Loading branch information
@ParsaGhadermazi
ParsaGhadermazi committed Jul 30, 2024
1 parent 9b37490 commit 4b58b37
Showing 1 changed file with 40 additions and 42 deletions.
82 changes: 40 additions & 42 deletions adtoolbox/optimize.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -414,55 +414,53 @@ def optimize(self, perturbation_method:str="random",ode_method="LSODA",parallel:
self._aquired["cost"]=self._aquired["cost"][-50:]
return self.history
elif kwargs.get("parallel_framework","ray")=="native":
if not self._initialized:
logging.info("Generating initial population:")
with mp.Pool(NUM_CORES) as pool:
with mp.Pool(NUM_CORES) as pool:
if not self._initialized:
logging.info("Generating initial population:")
costs=[pool.apply_async(_single_cost,args=(self.base_model,dict(zip(self.tunables.keys(),pop)),exp,self.var_type,ode_method)) for exp in self.train_data for pop in self._popuplation]
costs=[c.get() for c in costs]
logging.info("All initial population costs calculated.")
costs=list(np.array(costs).reshape(-1,len(self.train_data)).sum(axis=1))
self._aquired={"parameters":self._popuplation,"cost":[c for c in costs]}

self._best_tensor=self._aquired["parameters"][np.argmin(self._aquired["cost"])]
self._best_cost=np.min(self._aquired["cost"])
for i in range(self.n_steps):
loss=self._train_surrogate()
logging.info(f"Training Loss: {loss}")
new_params=self._suggest_parameters()
new_params[new_params<self._param_space[:,0]]=self._param_space[:,0][new_params<self._param_space[:,0]]
new_params[new_params>self._param_space[:,1]]=self._param_space[:,1][new_params>self._param_space[:,1]]
if (len(self._aquired["cost"])>1) and (abs(self._aquired["cost"][-1]-self._aquired["cost"][-2])<1e-3):
logging.info("Local optima reached: Perturbing the best solution and continuing.")
if perturbation_method=="random":
new_params=np.random.normal(self._best_tensor,np.abs(self._best_tensor/self.exp_std))
elif perturbation_method=="estimate_gradient_directions":
diff=self._aquired["parameters"]-self._best_tensor
cost_diff=(np.array(self._aquired["cost"])-self._best_cost).reshape(1,-1).repeat(diff.shape[1],axis=0).T
diff=-np.sign(np.mean(np.sign(np.multiply(diff,cost_diff)),axis=0))
new_params=self._best_tensor+np.random.normal(np.multiply(self._best_tensor,diff)/self.exp_std,np.abs(self._best_tensor/self.exp_std),size=diff.shape[0])



new_params[new_params<self._param_space[:,0]]=self._param_space[:,0][new_params<self._param_space[:,0]]
new_params[new_params>self._param_space[:,1]]=self._param_space[:,1][new_params>self._param_space[:,1]]
## make sure not in the local optima
new_cost=[pool.apply_async(_single_cost,args=(self.base_model,dict(zip(self.tunables.keys(),new_params)),exp,self.var_type,ode_method)) for exp in self.train_data]
new_cost=np.sum([c.get() for c in new_cost])
self._aquired["parameters"]=np.vstack((self._aquired["parameters"],new_params))
self._aquired["cost"].append(new_cost)
self._best_tensor=self._aquired["parameters"][np.argmin(self._aquired["cost"])]
self._best_cost=np.min(self._aquired["cost"])
for i in range(self.n_steps):
loss=self._train_surrogate()
logging.info(f"Training Loss: {loss}")
new_params=self._suggest_parameters()
new_params[new_params<self._param_space[:,0]]=self._param_space[:,0][new_params<self._param_space[:,0]]
new_params[new_params>self._param_space[:,1]]=self._param_space[:,1][new_params>self._param_space[:,1]]

if (len(self._aquired["cost"])>1) and (abs(self._aquired["cost"][-1]-self._aquired["cost"][-2])<1e-3):
logging.info("Local optima reached: Perturbing the best solution and continuing.")
if perturbation_method=="random":
new_params=np.random.normal(self._best_tensor,np.abs(self._best_tensor/self.exp_std))
elif perturbation_method=="estimate_gradient_directions":
diff=self._aquired["parameters"]-self._best_tensor
cost_diff=(np.array(self._aquired["cost"])-self._best_cost).reshape(1,-1).repeat(diff.shape[1],axis=0).T
diff=-np.sign(np.mean(np.sign(np.multiply(diff,cost_diff)),axis=0))
new_params=self._best_tensor+np.random.normal(np.multiply(self._best_tensor,diff)/self.exp_std,np.abs(self._best_tensor/self.exp_std),size=diff.shape[0])




new_params[new_params<self._param_space[:,0]]=self._param_space[:,0][new_params<self._param_space[:,0]]
new_params[new_params>self._param_space[:,1]]=self._param_space[:,1][new_params>self._param_space[:,1]]
## make sure not in the local optima

new_cost=[pool.apply_async(_single_cost,args=(self.base_model,dict(zip(self.tunables.keys(),new_params)),exp,self.var_type,ode_method)) for exp in self.train_data]
new_cost=np.sum([c.get() for c in new_cost])
self._aquired["parameters"]=np.vstack((self._aquired["parameters"],new_params))
self._aquired["cost"].append(new_cost)
self._best_tensor=self._aquired["parameters"][np.argmin(self._aquired["cost"])]
self._best_cost=np.min(self._aquired["cost"])
logging.info(f"Step {i+1}/{self.n_steps} completed.Current cost:{new_cost} Best cost: {self._best_cost}")
if i%self.save_every==0:
self.history={"parameters":self._aquired["parameters"],"cost":self._aquired["cost"],"tunable_parameters":list(self.tunables.keys())}
with open(f"{str(self.history_file_path.absolute())}","wb") as file:
pickle.dump(self.history,file)
logging.info(f"optimization results saved to {self.history_file_path.absolute()}")
if i>51:
self._aquired["parameters"]=self._aquired["parameters"][-50:,:]
self._aquired["cost"]=self._aquired["cost"][-50:]
logging.info(f"Step {i+1}/{self.n_steps} completed.Current cost:{new_cost} Best cost: {self._best_cost}")
if i%self.save_every==0:
self.history={"parameters":self._aquired["parameters"],"cost":self._aquired["cost"],"tunable_parameters":list(self.tunables.keys())}
with open(f"{str(self.history_file_path.absolute())}","wb") as file:
pickle.dump(self.history,file)
logging.info(f"optimization results saved to {self.history_file_path.absolute()}")
if i>51:
self._aquired["parameters"]=self._aquired["parameters"][-50:,:]
self._aquired["cost"]=self._aquired["cost"][-50:]
return self.history
else:
raise ValueError("Parallel framework not supported.")
Expand Down

0 comments on commit 4b58b37

Please sign in to comment.