Update gligen/ldm/models/diffusion/plms.py

gligen/ldm/models/diffusion/plms.py

@@ -3,9 +3,10 @@ import numpy as np
 from tqdm import tqdm
 from functools import partial
 from copy import deepcopy
+from diffusers import AutoencoderKL, LMSDiscreteScheduler
 from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like
 import math
-from ldm.models.diffusion.loss import  caculate_loss_att_fixed_cnt, caculate_loss_self_att
+from ldm.models.diffusion.loss import  caculate_loss_att_fixed_cnt, caculate_loss_self_att, caculate_loss_LoCo_V2
 class PLMSSampler(object):
     def __init__(self, diffusion, model, schedule="linear", alpha_generator_func=None, set_alpha_scale=None):
         super().__init__()
@@ -57,14 +58,14 @@ class PLMSSampler(object):
 
 
     # @torch.no_grad()
-    def sample(self, S, shape, input, uc=None, guidance_scale=1, mask=None, x0=None, loss_type='SAR_CAR'):
+    def sample(self, S, shape, input, uc=None, guidance_scale=1, mask=None, x0=None, loss_type=None):
         self.make_schedule(ddim_num_steps=S)
         # import pdb; pdb.set_trace()
         return self.plms_sampling(shape, input, uc, guidance_scale, mask=mask, x0=x0, loss_type=loss_type)
 
 
     # @torch.no_grad()
-    def plms_sampling(self, shape, input, uc=None, guidance_scale=1, mask=None, x0=None, loss_type='SAR_CAR'):
+    def plms_sampling(self, shape, input, uc=None, guidance_scale=1, mask=None, x0=None, loss_type=None):
 
         b = shape[0]
         
@@ -81,6 +82,7 @@ class PLMSSampler(object):
         if self.alpha_generator_func != None:
             alphas = self.alpha_generator_func(len(time_range))
 
+
         for i, step in enumerate(time_range):
 
             # set alpha and restore first conv layer 
@@ -102,12 +104,7 @@ class PLMSSampler(object):
             # three loss types
             if loss_type !=None and loss_type!='standard':
                 if input['object_position'] != []:
-                    if loss_type=='SAR_CAR':
-                        x = self.update_loss_self_cross( input,i, index, ts )
-                    elif loss_type=='SAR':
-                        x = self.update_only_self( input,i, index, ts )
-                    elif loss_type=='CAR':
-                        x = self.update_loss_only_cross( input,i, index, ts )
+                    x = self.update_loss_LoCo( input,i, index, ts,  time_factor = time_factor)              
                     input["x"] = x
             img, pred_x0, e_t = self.p_sample_plms(input, ts, index=index, uc=uc, guidance_scale=guidance_scale, old_eps=old_eps, t_next=ts_next)
             input["x"] = img
@@ -119,11 +116,11 @@ class PLMSSampler(object):
      
     def update_loss_self_cross(self, input,index1, index, ts,type_loss='self_accross' ):
         if index1 < 10:
-            loss_scale = 4
-            max_iter = 1
-        elif index1 < 20:
             loss_scale = 3
-            max_iter = 1
+            max_iter = 5
+        elif index1 < 20:
+            loss_scale = 2
+            max_iter = 3
         else:
             loss_scale = 1
             max_iter = 1
@@ -136,29 +133,25 @@ class PLMSSampler(object):
         input["timesteps"] = ts
         
         print("optimize", index1)
-        self.model.train()
         while loss.item() > loss_threshold and iteration < max_iter and (index1 < max_index) :
             print('iter', iteration)
-            # import pdb; pdb.set_trace()
             x = x.requires_grad_(True)
             input['x'] = x
             e_t,  att_first, att_second, att_third, self_first, self_second, self_third = self.model(input)
-            bboxes = input['boxes_att']
+            bboxes = input['boxes']
             object_positions = input['object_position'] 
             loss1 = caculate_loss_self_att(self_first, self_second, self_third, bboxes=bboxes,
                                 object_positions=object_positions, t = index1)*loss_scale 
             loss2 = caculate_loss_att_fixed_cnt(att_second,att_first,att_third, bboxes=bboxes,
                                 object_positions=object_positions, t = index1)*loss_scale
             loss = loss1 + loss2
-            print('loss', loss, loss1, loss2)  
-            # hh = torch.autograd.backward(loss, retain_graph=True)
-            grad_cond = torch.autograd.grad(loss.requires_grad_(True), [x])[0]
-            # grad_cond = x.grad 
+            print('AR loss:', loss, 'SAR:', loss1, 'CAR:', loss2)  
+            hh = torch.autograd.backward(loss)
+            grad_cond = x.grad 
             x = x - grad_cond
             x = x.detach()
             iteration += 1
-            
-        
+            torch.cuda.empty_cache()
         return x
 
     def update_loss_only_cross(self, input,index1, index, ts,type_loss='self_accross'):
@@ -184,6 +177,7 @@ class PLMSSampler(object):
         while loss.item() > loss_threshold and iteration < max_iter and (index1 < max_index) :
             print('iter', iteration)
             x = x.requires_grad_(True)
+            print('x shape', x.shape)
             input['x'] = x
             e_t,  att_first, att_second, att_third, self_first, self_second, self_third = self.model(input)
             
@@ -193,7 +187,55 @@ class PLMSSampler(object):
                         object_positions=object_positions, t = index1)*loss_scale
             loss = loss2
             print('loss', loss) 
-            hh = torch.autograd.backward(loss)
+            hh = torch.autograd.backward(loss,  retain_graph=True)
+            grad_cond = x.grad
+            x = x - grad_cond 
+            x = x.detach()
+            iteration += 1
+            torch.cuda.empty_cache()
+        return x
+
+    def update_loss_LoCo(self, input,index1, index, ts, time_factor, type_loss='self_accross'):
+        
+        # loss_scale = 30
+        # max_iter = 5
+        #print('time_factor is: ', time_factor)
+        if index1 < 10:
+            loss_scale = 8
+            max_iter = 5
+        elif index1 < 20:
+            loss_scale = 5
+            max_iter = 5
+        else:
+            loss_scale = 1
+            max_iter = 1
+        loss_threshold = 0.1
+        
+        max_index = 30
+        x = deepcopy(input["x"]) 
+        iteration = 0
+        loss = torch.tensor(10000)
+        input["timesteps"] = ts
+        
+        # print("optimize", index1)
+        while loss.item() > loss_threshold and iteration < max_iter and (index1 < max_index) :
+            # print('iter', iteration)
+            x = x.requires_grad_(True)
+            # print('x shape', x.shape)
+            input['x'] = x
+            e_t,  att_first, att_second, att_third, self_first, self_second, self_third = self.model(input)
+            
+            bboxes = input['boxes']
+            object_positions = input['object_position'] 
+            loss2 = caculate_loss_LoCo_V2(att_second,att_first,att_third, bboxes=bboxes,
+                        object_positions=object_positions, t = index1)*loss_scale
+            # loss = loss2
+            # loss.requires_grad_(True)
+            #print('LoCo loss', loss) 
+
+
+            
+            hh = torch.autograd.backward(loss2,  retain_graph=True)
             grad_cond = x.grad
             x = x - grad_cond 
             x = x.detach()
@@ -244,13 +286,12 @@ class PLMSSampler(object):
     def p_sample_plms(self, input, t, index, guidance_scale=1., uc=None, old_eps=None, t_next=None):
         x = deepcopy(input["x"]) 
         b = x.shape[0]
-        self.model.eval()
+
         def get_model_output(input):
             e_t, first, second, third,_,_,_ = self.model(input) 
             if uc is not None and guidance_scale != 1:
-                unconditional_input = dict(x=input["x"], timesteps=input["timesteps"], context=uc, inpainting_extra_input=None, grounding_extra_input=None)
-                # unconditional_input=input
-                e_t_uncond, _, _, _, _, _, _ = self.model( unconditional_input) 
+                unconditional_input = dict(x=input["x"], timesteps=input["timesteps"], context=uc, inpainting_extra_input=input["inpainting_extra_input"], grounding_extra_input=input['grounding_extra_input'])
+                e_t_uncond, _, _, _, _, _, _ = self.model( unconditional_input ) 
                 e_t = e_t_uncond + guidance_scale * (e_t - e_t_uncond)
             return e_t