junshan-kit 2.3.9__py2.py3-none-any.whl → 2.4.1__py2.py3-none-any.whl

junshan_kit/SPBM.py

@@ -0,0 +1,350 @@
+from junshan_kit import SPBM, SPBM_func
+import torch, time, os
+from torch.optim.optimizer import Optimizer
+from torch.nn.utils import parameters_to_vector, vector_to_parameters
+
+
+class PF(Optimizer):
+    def __init__(self, params, model, hyperparams, Paras):
+        defaults = dict()
+        super().__init__(params, defaults)
+        self.model = model
+        self.cutting_num = hyperparams['cutting_number']
+        self.M = hyperparams['M']
+        self.delta = hyperparams['delta']
+        self.Paras = Paras
+
+        self.x_his, self.g_his, self.f_his = [], [], []
+
+    def step(self, closure=None):
+        if closure is None:
+            raise RuntimeError("Closure required for CuttingPlaneOptimizer")
+        
+        # 清零梯度并前向计算
+        loss = closure()
+        
+        with torch.no_grad():
+            xk = parameters_to_vector(self.model.parameters())
+            # print(torch.norm(xk))
+            g_k = parameters_to_vector([p.grad if p.grad is not None else torch.zeros_like(p) for p in self.model.parameters()])
+
+            # Add cutting plane
+            x_his, f_his, g_his = SPBM_func.add_cutting(self.x_his, self.f_his, self.g_his,xk.detach().clone(), g_k.detach().clone(), loss.detach().clone(), self.cutting_num)
+
+            ## Cut selection
+            selected_x, selected_f, selected_g = SPBM_func.cut_selection(x_his, f_his, g_his, self.M)
+                    
+            # the coefficient of dual problem
+            Gk, rk, ek = SPBM_func.get_var(selected_x, selected_f, selected_g, self.delta)
+            
+            # SOVER (dual)
+            xk = SPBM_func.subproblem_pf(Gk, ek, xk, self.delta, self.Paras)
+            
+            # print(len(self.f_his))
+            vector_to_parameters(xk, self.model.parameters())
+
+
+        # 暂时返回 loss（tensor 类型）
+        return loss
+    
+# <SPBM-TR>
+class TR(Optimizer):
+    def __init__(self, params, model, hyperparams, Paras):
+        defaults = dict()
+        super().__init__(params, defaults)
+        self.model = model
+        self.cutting_num = hyperparams['cutting_number']
+        self.M = hyperparams['M']
+        self.delta = hyperparams['delta']
+        self.Paras = Paras
+
+        self.x_his, self.g_his, self.f_his = [], [], []
+
+    def step(self, closure=None):
+        if closure is None:
+            raise RuntimeError("Closure required for CuttingPlaneOptimizer")
+        
+        # Reset the gradient and perform forward computation
+        loss = closure()
+        
+        with torch.no_grad():
+            xk = parameters_to_vector(self.model.parameters())
+            # print(torch.norm(xk))
+            g_k = parameters_to_vector([p.grad if p.grad is not None else torch.zeros_like(p) for p in self.model.parameters()])
+
+            # Add cutting plane
+            x_his, f_his, g_his = SPBM_func.add_cutting(self.x_his, self.f_his, self.g_his,xk.detach().clone(), g_k.detach().clone(), loss.detach().clone(), self.cutting_num)
+
+            ## Cut selection
+            selected_x, selected_f, selected_g = SPBM_func.cut_selection(x_his, f_his, g_his, self.M)
+                    
+            # the coefficient of dual problem
+            Gk, rk, ek = SPBM_func.get_var(selected_x, selected_f, selected_g, self.delta)
+            
+            # SOVER (dual)
+            xk = SPBM_func.subproblem_tr_2(Gk, ek, xk, rk, self.Paras)
+            
+            # print(len(self.f_his))
+            vector_to_parameters(xk, self.model.parameters())
+
+        # tensor type
+        return loss
+# <SPBM-TR>  
+
+# <SPBM-TR_NoneSpecial> 
+class TR_NoneSpecial(Optimizer):
+    def __init__(self, params, model, hyperparams, Paras):
+        defaults = dict()
+        super().__init__(params, defaults)
+        self.model = model
+        self.cutting_num = hyperparams['cutting_number']
+        self.M = hyperparams['M']
+        self.delta = hyperparams['delta']
+        self.Paras = Paras
+
+        self.x_his, self.g_his, self.f_his = [], [], []
+
+    def step(self, closure=None):
+        if closure is None:
+            raise RuntimeError("Closure required for CuttingPlaneOptimizer")
+        
+        # Reset the gradient and perform forward computation
+        loss = closure()
+        
+        with torch.no_grad():
+            xk = parameters_to_vector(self.model.parameters())
+            # print(torch.norm(xk))
+            g_k = parameters_to_vector([p.grad if p.grad is not None else torch.zeros_like(p) for p in self.model.parameters()])
+
+            # Add cutting plane
+            x_his, f_his, g_his = SPBM_func.add_cutting(self.x_his, self.f_his, self.g_his,xk.detach().clone(), g_k.detach().clone(), loss.detach().clone(), self.cutting_num)
+
+            ## Cut selection
+            selected_x, selected_f, selected_g = SPBM_func.cut_selection(x_his, f_his, g_his, self.M)
+                    
+            # the coefficient of dual problem
+            Gk, rk, ek = SPBM_func.get_var(selected_x, selected_f, selected_g, self.delta)
+            
+            # SOVER (dual)
+            xk = SPBM_func.subproblem_tr_NoneSpecial(Gk, ek, xk, rk, self.Paras)
+            
+            # print(len(self.f_his))
+            vector_to_parameters(xk, self.model.parameters())
+
+        # tensor type
+        return loss
+# <SPBM-TR_NoneSpecial>  
+
+class TR_primal(Optimizer):
+    def __init__(self, params, model, hyperparams, Paras):
+        defaults = dict()
+        super().__init__(params, defaults)
+        self.model = model
+        self.cutting_num = hyperparams['cutting_number']
+        self.M = hyperparams['M']
+        self.delta = hyperparams['delta']
+        self.Paras = Paras
+
+        self.x_his, self.g_his, self.f_his = [], [], []
+
+    def step(self, closure=None):
+        if closure is None:
+            raise RuntimeError("Closure required for CuttingPlaneOptimizer")
+        
+        # Reset the gradient and perform forward computation
+        loss = closure()
+        
+        with torch.no_grad():
+            xk = parameters_to_vector(self.model.parameters())
+            # print(torch.norm(xk))
+            g_k = parameters_to_vector([p.grad if p.grad is not None else torch.zeros_like(p) for p in self.model.parameters()])
+
+            # Add cutting plane
+            x_his, f_his, g_his = SPBM_func.add_cutting(self.x_his, self.f_his, self.g_his,xk.detach().clone(), g_k.detach().clone(), loss.detach().clone(), self.cutting_num)
+
+            ## Cut selection
+            selected_x, selected_f, selected_g = SPBM_func.cut_selection(x_his, f_his, g_his, self.M)
+                    
+            # the coefficient of dual problem
+            Gk, rk, ek = SPBM_func.get_var(selected_x, selected_f, selected_g, self.delta)
+            
+            # SOVER (dual)
+            xk = SPBM_func.subproblem_tr_primal(Gk, ek, xk, rk, self.Paras)
+            
+            # print(len(self.f_his))
+            vector_to_parameters(xk, self.model.parameters())
+
+        # tensor type
+        return loss
+    
+
+class TR_NoneLower(Optimizer):
+    def __init__(self, params, model, hyperparams, Paras):
+        defaults = dict()
+        super().__init__(params, defaults)
+        self.model = model
+        self.cutting_num = hyperparams['cutting_number']
+        self.M = hyperparams['M']
+        self.delta = hyperparams['delta']
+        self.Paras = Paras
+
+        self.x_his, self.g_his, self.f_his = [], [], []
+
+    def step(self, closure=None):
+        if closure is None:
+            raise RuntimeError("Closure required for CuttingPlaneOptimizer")
+        
+        # Reset the gradient and perform forward computation
+        loss = closure()
+        
+        with torch.no_grad():
+            xk = parameters_to_vector(self.model.parameters())
+            # print(torch.norm(xk))
+            g_k = parameters_to_vector([p.grad if p.grad is not None else torch.zeros_like(p) for p in self.model.parameters()])
+
+            # Add cutting plane
+            x_his, f_his, g_his = SPBM_func.add_cutting(self.x_his, self.f_his, self.g_his,xk.detach().clone(), g_k.detach().clone(), loss.detach().clone(), self.cutting_num)
+
+            ## Cut selection
+            selected_x, selected_f, selected_g = SPBM_func.cut_selection(x_his, f_his, g_his, self.M)
+                    
+            # the coefficient of dual problem
+            Gk, rk, ek = SPBM_func.get_var(selected_x, selected_f, selected_g, self.delta)
+            
+            # SOVER (dual)
+            xk = SPBM_func.subproblem_tr_NoneLower(Gk, ek, xk, rk, self.Paras)
+            
+            # print(len(self.f_his))
+            vector_to_parameters(xk, self.model.parameters())
+
+        # tensor type
+        return loss
+
+class TR_NoneCut(Optimizer):
+    def __init__(self, params, model, hyperparams, Paras):
+        defaults = dict()
+        super().__init__(params, defaults)
+        self.model = model
+        self.cutting_num = hyperparams['cutting_number']
+        self.M = hyperparams['M']
+        self.delta = hyperparams['delta']
+        self.Paras = Paras
+
+        self.x_his, self.g_his, self.f_his = [], [], []
+
+    def step(self, closure=None):
+        if closure is None:
+            raise RuntimeError("Closure required for CuttingPlaneOptimizer")
+        
+        # Reset the gradient and perform forward computation
+        loss = closure()
+        
+        with torch.no_grad():
+            xk = parameters_to_vector(self.model.parameters())
+            # print(torch.norm(xk))
+            g_k = parameters_to_vector([p.grad if p.grad is not None else torch.zeros_like(p) for p in self.model.parameters()])
+
+            # Add cutting plane
+            x_his, f_his, g_his = SPBM_func.add_cutting(self.x_his, self.f_his, self.g_his,xk.detach().clone(), g_k.detach().clone(), loss.detach().clone(), self.cutting_num)
+
+            # ## Cut selection
+            # selected_x, selected_f, selected_g = SPBM_func.cut_selection(x_his, f_his, g_his, self.M)
+                    
+            # the coefficient of dual problem
+            Gk, rk, ek = SPBM_func.get_var(x_his, f_his, g_his, self.delta)
+            
+            # SOVER (dual)
+            xk = SPBM_func.subproblem_tr_NoneLower(Gk, ek, xk, rk, self.Paras)
+            
+            # print(len(self.f_his))
+            vector_to_parameters(xk, self.model.parameters())
+
+        # tensor type
+        return loss
+    
+# ************************** SPBM-PF **************************
+class PF_NoneLower(Optimizer):
+    def __init__(self, params, model, hyperparams, Paras):
+        defaults = dict()
+        super().__init__(params, defaults)
+        self.model = model
+        self.cutting_num = hyperparams['cutting_number']
+        self.M = hyperparams['M']
+        self.delta = hyperparams['delta']
+        self.Paras = Paras
+
+        self.x_his, self.g_his, self.f_his = [], [], []
+
+    def step(self, closure=None):
+        if closure is None:
+            raise RuntimeError("Closure required for CuttingPlaneOptimizer")
+        
+        # Reset the gradient and perform forward computation
+        loss = closure()
+        
+        with torch.no_grad():
+            xk = parameters_to_vector(self.model.parameters())
+            # print(torch.norm(xk))
+            g_k = parameters_to_vector([p.grad if p.grad is not None else torch.zeros_like(p) for p in self.model.parameters()])
+
+            # Add cutting plane
+            x_his, f_his, g_his = SPBM_func.add_cutting(self.x_his, self.f_his, self.g_his,xk.detach().clone(), g_k.detach().clone(), loss.detach().clone(), self.cutting_num)
+
+            ## Cut selection
+            selected_x, selected_f, selected_g = SPBM_func.cut_selection(x_his, f_his, g_his, self.M)
+                    
+            # the coefficient of dual problem
+            Gk, rk, ek = SPBM_func.get_var(selected_x, selected_f, selected_g, self.delta)
+            
+            # SOVER (dual)
+            xk = SPBM_func.subproblem_pf_NoneLower(Gk, ek, xk, self.delta, self.Paras)
+            
+            # print(len(self.f_his))
+            vector_to_parameters(xk, self.model.parameters())
+
+        # tensor type
+        return loss
+    
+
+class PF_NoneCut(Optimizer):
+    def __init__(self, params, model, hyperparams, Paras):
+        defaults = dict()
+        super().__init__(params, defaults)
+        self.model = model
+        self.cutting_num = hyperparams['cutting_number']
+        self.M = hyperparams['M']
+        self.delta = hyperparams['delta']
+        self.Paras = Paras
+
+        self.x_his, self.g_his, self.f_his = [], [], []
+
+    def step(self, closure=None):
+        if closure is None:
+            raise RuntimeError("Closure required for CuttingPlaneOptimizer")
+        
+        # Reset the gradient and perform forward computation
+        loss = closure()
+        
+        with torch.no_grad():
+            xk = parameters_to_vector(self.model.parameters())
+            # print(torch.norm(xk))
+            g_k = parameters_to_vector([p.grad if p.grad is not None else torch.zeros_like(p) for p in self.model.parameters()])
+
+            # Add cutting plane
+            x_his, f_his, g_his = SPBM_func.add_cutting(self.x_his, self.f_his, self.g_his,xk.detach().clone(), g_k.detach().clone(), loss.detach().clone(), self.cutting_num)
+
+            # ## Cut selection
+            # selected_x, selected_f, selected_g = SPBM_func.cut_selection(x_his, f_his, g_his, self.M)
+                    
+            # the coefficient of dual problem
+            Gk, rk, ek = SPBM_func.get_var(x_his, f_his, g_his, self.delta)
+            
+            # SOVER (dual)
+            xk = SPBM_func.subproblem_pf_NoneLower(Gk, ek, xk, self.delta, self.Paras)
+            
+            # print(len(self.f_his))
+            vector_to_parameters(xk, self.model.parameters())
+
+        # tensor type
+        return loss
+