bodymocap/train/eftFitter.py [1037:1263]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - train_data_loader = CheckpointDataLoader(self.train_ds,checkpoint=self.checkpoint, batch_size=1, #Always o1 num_workers=self.options.num_workers, pin_memory=self.options.pin_memory, shuffle=False) #No Shuffle maxExemplarIter = self.options.maxExemplarIter # Iterate over all batches in an epoch outputList ={} for step, batch in enumerate(tqdm(train_data_loader)):#, desc='Epoch '+str(epoch), # total=len(self.train_ds) // self.options.batch_size, # initial=train_data_loader.checkpoint_batch_idx), # train_data_loader.checkpoint_batch_idx): #3DPW test # if 'downtown_bus_00' not in batch['imgname']: # continue #Only performed for 1/100 data (roughly hundred level) if self.options.bExemplar_analysis_testloss: sampleIdx = batch['sample_index'][0].item() if sampleIdx%100 !=0: continue if self.options.bExemplar_badsample_finder: sampleIdx = batch['sample_index'][0].item() # if sampleIdx%100 !=0: # continue bSkipExisting = self.options.bNotSkipExemplar==False #bNotSkipExemplar ===True --> bSkipExisting==False if bSkipExisting: if self.options.db_set =='panoptic': # fileNameOnly = os.path.basename(output['imageName'][0])[:-4] # fileNameOnly = (batch['pkl_save_name'][0])[:-4].replace("/","-") sampleIdx = batch['sample_index'][0].item() if self.options.bExemplar_dataLoaderStart>=0: sampleIdx +=self.options.bExemplar_dataLoaderStart sampleIdxSaveFrame = 100* (int(sampleIdx/100.0) + 1) fileName = '{:08d}.pkl'.format(sampleIdxSaveFrame) outputPath = os.path.join(exemplarOutputPath,fileName) # print(">> checking: {}".format(outputPath)) if os.path.exists(outputPath): print("Skipped: {}".format(outputPath)) continue elif '3dpw' in self.options.db_set: fileNameOnly = os.path.basename(batch['imgname'][0])[:-4] seqName = os.path.basename(os.path.dirname(batch['imgname'][0])) fileNameOnly = f"{seqName}_{fileNameOnly}" sampleIdx = batch['sample_index'][0].item() if self.options.bExemplar_dataLoaderStart>=0: sampleIdx +=self.options.bExemplar_dataLoaderStart fileName = '{}_{}.pkl'.format(fileNameOnly,sampleIdx) outputPath = os.path.join(exemplarOutputPath,fileName) if os.path.exists(outputPath): print("Skipped: {}".format(outputPath)) continue else: fileNameOnly = os.path.basename(batch['imgname'][0])[:-4] sampleIdx = batch['sample_index'][0].item() if self.options.bExemplar_dataLoaderStart>=0: sampleIdx +=self.options.bExemplar_dataLoaderStart fileName = '{}_{}.pkl'.format(fileNameOnly,sampleIdx) outputPath = os.path.join(exemplarOutputPath,fileName) if os.path.exists(outputPath): print("Skipped: {}".format(outputPath)) continue g_timer.tic() self.reloadModel() #For each sample #Freeze non resnet part model if self.options.ablation_layerteset_onlyLayer4: # self.model.conv1.requires_grad = False # self.model.bn1.requires_grad = False # self.model.relu.requires_grad = False # self.model.maxpool.requires_grad = False # self.model.layer1.requires_grad = False # self.model.layer2.requires_grad = False # self.model.layer3.requires_grad = False # self.model.layer4.requires_grad = False # self.model.fc1.requires_grad = False # self.model.drop1.requires_grad = False # self.model.fc2.requires_grad = False # self.model.drop2.requires_grad = False # self.model.decpose.requires_grad = False # self.model.decshape.requires_grad = False # self.model.deccam.requires_grad = False for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): #Optimize Layer 4 of resnet # print(name) # if 'fc' in name or 'decpose' in name or 'decshape' in name or 'deccam' in name: # print(f"activate {name}") # par.requires_grad = True if 'layer4' in name: # print(f">> Activate {name}") par.requires_grad = True if self.options.ablation_layerteset_onlyAfterRes: #Optimize HMR FC part for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'fc' in name or 'decpose' in name or 'decshape' in name or 'deccam' in name: # print(f"activate {name}") par.requires_grad = True if self.options.ablation_layerteset_Layer4Later: #Optimize Layer 4 of resent + HMR FC part for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'layer4' in name or 'fc' in name or 'decpose' in name or 'decshape' in name or 'deccam' in name: # print(f"activate {name}") par.requires_grad = True if self.options.ablation_layerteset_onlyRes: for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'layer' in name: # print(f"activate {name}") par.requires_grad = True if self.options.ablation_layerteset_Layer3Later: for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'layer3' in name or 'layer4' in name or 'fc' in name or 'decpose' in name or 'decshape' in name or 'deccam' in name: # print(f"activate {name}") par.requires_grad = True if self.options.ablation_layerteset_Layer2Later: for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'layer2' in name or 'layer3' in name or 'layer4' in name or 'fc' in name or 'decpose' in name or 'decshape' in name or 'deccam' in name: # print(f"activate {name}") par.requires_grad = True if self.options.ablation_layerteset_Layer1Later: for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'layer1' in name or 'layer2' in name or 'layer3' in name or 'layer4' in name or 'fc' in name or 'decpose' in name or 'decshape' in name or 'deccam' in name: # print(f"activate {name}") par.requires_grad = True if self.options.ablation_layerteset_all: #No Freeze. debugging purpose for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'conv1' in name or 'layer' in name or 'fc' in name or 'decpose' in name or 'decshape' in name or 'deccam' in name: # print(f"activate {name}") par.requires_grad = True if self.options.ablation_layerteset_onlyRes_withconv1: #Only use ResNet. Freeze HMR part all for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'conv1' in name or 'layer' in name: # print(f"activate {name}") par.requires_grad = True if self.options.ablation_layerteset_decOnly: for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'decpose' in name or 'decshape' in name or 'deccam' in name: # print(f"activate {name}") par.requires_grad = True if self.options.ablation_layerteset_fc2Later: for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'fc2' in name or 'decpose' in name or 'decshape' in name or 'deccam' in name: # print(f"activate {name}") par.requires_grad = True #Freeze all except the last layer of Resnet if self.options.ablation_layerteset_onlyRes50LastConv: for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'layer4.2.conv3' in name: # print(f"activate {name}") par.requires_grad = True # g_timer.toc(average =False, bPrint=True,title="reload") # self.exemplerTrainingMode() batch = {k: v.to(self.device) if isinstance(v, torch.Tensor) else v for k,v in batch.items()} output_backup={} for it in range(maxExemplarIter): ########################################################################################## ##### RUN EFT ########################################################################################## # g_timer.tic() if self.options.bUseHand3D: output, losses = self.run_eft_step_wHand(batch) - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - eft/train/eftFitter.py [1320:1550]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - train_data_loader = CheckpointDataLoader(self.train_ds,checkpoint=self.checkpoint, batch_size=1, #Always o1 num_workers=self.options.num_workers, pin_memory=self.options.pin_memory, shuffle=False) #No Shuffle maxExemplarIter = self.options.maxExemplarIter # Iterate over all batches in an epoch outputList ={} for step, batch in enumerate(tqdm(train_data_loader)):#, desc='Epoch '+str(epoch), # total=len(self.train_ds) // self.options.batch_size, # initial=train_data_loader.checkpoint_batch_idx), # train_data_loader.checkpoint_batch_idx): #3DPW test # if 'downtown_bus_00' not in batch['imgname']: # continue #Only performed for 1/100 data (roughly hundred level) if self.options.bExemplar_analysis_testloss: sampleIdx = batch['sample_index'][0].item() if sampleIdx%100 !=0: continue if self.options.bExemplar_badsample_finder: sampleIdx = batch['sample_index'][0].item() # if sampleIdx%100 !=0: # continue # if self.options.bDebug_visEFT and self.options.bUseHand2D: # is_valid = sum(batch['kp_leftHand_gt'][0,:,2]).clone().detach().cpu().numpy().item() + sum(batch['kp_rightHand_gt'][0,:,2]).clone().detach().cpu().numpy().item() # if is_valid<0.1: # continue bSkipExisting = self.options.bNotSkipExemplar==False #bNotSkipExemplar ===True --> bSkipExisting==False if bSkipExisting: if self.options.db_set =='panoptic': # fileNameOnly = os.path.basename(output['imageName'][0])[:-4] # fileNameOnly = (batch['pkl_save_name'][0])[:-4].replace("/","-") sampleIdx = batch['sample_index'][0].item() if self.options.bExemplar_dataLoaderStart>=0: sampleIdx +=self.options.bExemplar_dataLoaderStart sampleIdxSaveFrame = 100* (int(sampleIdx/100.0) + 1) fileName = '{:08d}.pkl'.format(sampleIdxSaveFrame) outputPath = os.path.join(exemplarOutputPath,fileName) # print(">> checking: {}".format(outputPath)) if os.path.exists(outputPath): print("Skipped: {}".format(outputPath)) continue elif '3dpw' in self.options.db_set: fileNameOnly = os.path.basename(batch['imgname'][0])[:-4] seqName = os.path.basename(os.path.dirname(batch['imgname'][0])) fileNameOnly = f"{seqName}_{fileNameOnly}" sampleIdx = batch['sample_index'][0].item() if self.options.bExemplar_dataLoaderStart>=0: sampleIdx +=self.options.bExemplar_dataLoaderStart fileName = '{}_{}.pkl'.format(fileNameOnly,sampleIdx) outputPath = os.path.join(exemplarOutputPath,fileName) if os.path.exists(outputPath): print("Skipped: {}".format(outputPath)) continue else: fileNameOnly = os.path.basename(batch['imgname'][0])[:-4] sampleIdx = batch['sample_index'][0].item() if self.options.bExemplar_dataLoaderStart>=0: sampleIdx +=self.options.bExemplar_dataLoaderStart fileName = '{}_{}.pkl'.format(fileNameOnly,sampleIdx) outputPath = os.path.join(exemplarOutputPath,fileName) if os.path.exists(outputPath): print("Skipped: {}".format(outputPath)) continue g_timer.tic() self.reloadModel() #For each sample #Freeze non resnet part model if self.options.ablation_layerteset_onlyLayer4: # self.model.conv1.requires_grad = False # self.model.bn1.requires_grad = False # self.model.relu.requires_grad = False # self.model.maxpool.requires_grad = False # self.model.layer1.requires_grad = False # self.model.layer2.requires_grad = False # self.model.layer3.requires_grad = False # self.model.layer4.requires_grad = False # self.model.fc1.requires_grad = False # self.model.drop1.requires_grad = False # self.model.fc2.requires_grad = False # self.model.drop2.requires_grad = False # self.model.decpose.requires_grad = False # self.model.decshape.requires_grad = False # self.model.deccam.requires_grad = False for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): #Optimize Layer 4 of resnet # print(name) # if 'fc' in name or 'decpose' in name or 'decshape' in name or 'deccam' in name: # print(f"activate {name}") # par.requires_grad = True if 'layer4' in name: # print(f">> Activate {name}") par.requires_grad = True if self.options.ablation_layerteset_onlyAfterRes: #Optimize HMR Decoder part for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'fc' in name or 'decpose' in name or 'decshape' in name or 'deccam' in name: # print(f"activate {name}") par.requires_grad = True if self.options.ablation_layerteset_Layer4Later: #Optimize Layer 4 of resent + HMR FC part for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'layer4' in name or 'fc' in name or 'decpose' in name or 'decshape' in name or 'deccam' in name: # print(f"activate {name}") par.requires_grad = True if self.options.ablation_layerteset_onlyRes: for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'layer' in name: # print(f"activate {name}") par.requires_grad = True if self.options.ablation_layerteset_Layer3Later: for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'layer3' in name or 'layer4' in name or 'fc' in name or 'decpose' in name or 'decshape' in name or 'deccam' in name: # print(f"activate {name}") par.requires_grad = True if self.options.ablation_layerteset_Layer2Later: for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'layer2' in name or 'layer3' in name or 'layer4' in name or 'fc' in name or 'decpose' in name or 'decshape' in name or 'deccam' in name: # print(f"activate {name}") par.requires_grad = True if self.options.ablation_layerteset_Layer1Later: for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'layer1' in name or 'layer2' in name or 'layer3' in name or 'layer4' in name or 'fc' in name or 'decpose' in name or 'decshape' in name or 'deccam' in name: # print(f"activate {name}") par.requires_grad = True if self.options.ablation_layerteset_all: #No Freeze. debugging purpose for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'conv1' in name or 'layer' in name or 'fc' in name or 'decpose' in name or 'decshape' in name or 'deccam' in name: # print(f"activate {name}") par.requires_grad = True if self.options.ablation_layerteset_onlyRes_withconv1: #Only use ResNet. Freeze HMR part all for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'conv1' in name or 'layer' in name: # print(f"activate {name}") par.requires_grad = True if self.options.ablation_layerteset_decOnly: #Optimize the last layer of hmr decoder for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'decpose' in name or 'decshape' in name or 'deccam' in name: # print(f"activate {name}") par.requires_grad = True if self.options.ablation_layerteset_fc2Later: for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'fc2' in name or 'decpose' in name or 'decshape' in name or 'deccam' in name: # print(f"activate {name}") par.requires_grad = True #Freeze all except the last layer of Resnet if self.options.ablation_layerteset_onlyRes50LastConv: for par in self.model.parameters(): par.requires_grad = False for name, par in self.model.named_parameters(): if 'layer4.2.conv3' in name: # print(f"activate {name}") par.requires_grad = True # g_timer.toc(average =False, bPrint=True,title="reload") # self.exemplerTrainingMode() batch = {k: v.to(self.device) if isinstance(v, torch.Tensor) else v for k,v in batch.items()} output_backup={} for it in range(maxExemplarIter): ########################################################################################## ##### RUN EFT ########################################################################################## # g_timer.tic() if self.options.bUseHand3D: output, losses = self.run_eft_step_wHand(batch) - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -