# nch=csm.shape[1] # nbas=x.shape[0] # tmp1=torch.zeros((nch,nbas,2,NF*nx*nintl)).cuda() # for i in range(nch): # tmp=sf.complex_mult(x,csm[:,i].repeat(nbas,1,1,1,1),dim=2) # tmp=torch.reshape(tmp,(nbas,1,2,nx,nx)) # #xx=ktrajT[0].unsqueeze(0) # tmp1[i]=nufft_ob(tmp,traj.repeat(nbas,1,1).cuda()).squeeze(1) # tmp1=tmp1.permute(1,0,2,3) # return tmp1 #%% G = UnetClass().to(gpu) G = SmallModel1().to(gpu) GV = SmallModel().to(gpu) #G.load_state_dict(torch.load('wts-10U2.pt')) #GV.load_state_dict(torch.load('wts-10V2.pt')) #GV.load_state_dict(torch.load('./PTmodels/27Oct_112451am_500ep_27oct/wts-500.pt')) #optimizer=torch.optim.SGD([{'params':G.parameters(),'lr':5e-3,'momentum':0.9}]) optimizer = torch.optim.AdamW([{'params': G.parameters(), 'lr': 1e-3}]) #optimizer=torch.optim.AdamW([{'params':G.parameters(),'lr':1e-3},{'params':GV.parameters(),'lr':1e-3}]) scheduler = ReduceLROnPlateau(optimizer, mode='min', factor=0.7, patience=6, verbose=True, min_lr=5e-5)
#wtsFname1='wts-2U'+str(60+1)+'.pt' G=UnetClass().to(gpu) G.load_state_dict(torch.load(wtsFname1)) #atbT=atbT*W z=atbT.permute(3,0,1,2) z=torch.reshape(z,(1,2*nbasis,512,512)).to(gpu) #z = torch.randn((1,60,512,512),device=gpu, dtype=dtype) z = Variable(z,requires_grad=True) wtsFname2='wts-3V'+str(62+1)+'.pt' #wtsFname2='wts-2V'+str(60+1)+'.pt' GV=SmallModel().to(gpu) GV.load_state_dict(torch.load(wtsFname2)) #z1 = torch.randn((1,30,900),device=gpu, dtype=dtype) z1=torch.reshape(VT,(1,1,nbasis,NF)) z1 = Variable(z1,requires_grad=False) #optimizer=torch.optim.SGD([{'params':G.parameters(),'lr':5e-3,'momentum':0.9}]) optimizer1=torch.optim.AdamW([{'params':z,'lr':1e-3},{'params':G.parameters(),'lr':1e-3}]) scheduler = ReduceLROnPlateau(optimizer1, mode='min', factor=0.7, patience=6, verbose=True, min_lr=5e-5) #z1=z sf.tic() for ep1 in range(5): for bat in range(1): u1=G(z)
NF = 900 #100#900 nx = 512 #N=400 N1 = 300 N2 = 50 nch = 4 thres = 0.05 nbasis = 30 lam = 0.01 st = 0 batch_sz = 100 TF = 900 #%% G = UnetClass().to(gpu) GV = SmallModel().to(gpu) #G.load_state_dict(torch.load('wts-4U96.pt')) #GV.load_state_dict(torch.load('wts-4V96.pt')) #optimizer=torch.optim.SGD([{'params':G.parameters(),'lr':5e-3,'momentum':0.9}]) optimizer = torch.optim.AdamW([{ 'params': G.parameters(), 'lr': 1e-4 }, { 'params': GV.parameters(), 'lr': 1e-4 }]) scheduler = ReduceLROnPlateau(optimizer, mode='min', factor=0.7, patience=6, verbose=True,
# nch=csm.shape[1] # nbas=x.shape[0] # tmp1=torch.zeros((nch,nbas,2,NF*nx*nintl)).cuda() # for i in range(nch): # tmp=sf.complex_mult(x,csm[:,i].repeat(nbas,1,1,1,1),dim=2) # tmp=torch.reshape(tmp,(nbas,1,2,nx,nx)) # #xx=ktrajT[0].unsqueeze(0) # tmp1[i]=nufft_ob(tmp,traj.repeat(nbas,1,1).cuda()).squeeze(1) # tmp1=tmp1.permute(1,0,2,3) # return tmp1 #%% #G=UnetClass().to(gpu) G = SmallModel1().to(gpu) GV = SmallModel().to(gpu) G.load_state_dict(torch.load('wtsN-U3.pt')) GV.load_state_dict(torch.load('wtsN-V3.pt')) #GV.load_state_dict(torch.load('./PTmodels/27Oct_112451am_500ep_27oct/wts-500.pt')) #optimizer=torch.optim.SGD([{'params':G.parameters(),'lr':5e-3,'momentum':0.9}]) #optimizer=torch.optim.AdamW([{'params':G.parameters(),'lr':1e-4}]) optimizer = torch.optim.AdamW([{ 'params': G.parameters(), 'lr': 1e-4 }, { 'params': GV.parameters(), 'lr': 1e-4 }, { 'params': z,
for i in range(nch): tmp2 = sf.pt_fft2c( sf.pt_cpx_multipy(x, self.csmT[i].repeat(nbas, 1, 1, 1))) tmp2 = torch.reshape(tmp2, (nbas, self.NX)) tmp2 = tmp2.repeat(self.NF, 1, 1) * mask tmp5[i] = tmp2.sum(axis=1) del tmp2, x return tmp5.cpu() #%% #G=UnetClass().to(gpu) G = SmallModel1().to(gpu) GV = SmallModel().to(gpu) #G.load_state_dict(torch.load('wts-10U1.pt')) #GV.load_state_dict(torch.load('wts-10V5.pt')) #GV.load_state_dict(torch.load('./PTmodels/27Oct_112451am_500ep_27oct/wts-500.pt')) #optimizer=torch.optim.SGD([{'params':G.parameters(),'lr':5e-3,'momentum':0.9}]) optimizer = torch.optim.AdamW([{ 'params': G.parameters(), 'lr': 1e-4 }, { 'params': GV.parameters(), 'lr': 1e-4 }]) scheduler = ReduceLROnPlateau(optimizer, mode='min', factor=0.7,
#%% # sf.tic() #z1=torch.normal(0,0.01,(1,1,2,nf1)).cuda() z1 = 0.01 * torch.ones((1, 1, 2, nf1)).to(gpu) #z1[0,0,1]=-0.01 z1[0, 0, 1, :] = torch.tensor(V[22, 0:nf1]) z1[0, 0, 0, :] = torch.tensor(V[27, 0:nf1]) #vv=V[30:32] #z1=torch.reshape(torch.tensor(V[29:31,0:nf1]),(1,1,2,nf1)).to(gpu) z1 = Variable(z1, requires_grad=True) V1 = torch.tensor( V[:, 0:nf1]).cuda() #+torch.normal(0,0.01,(V.shape[0],V.shape[1])).cuda() #V1=torch.reshape(V1,(1,1,nbasis,nf1)) GV = SmallModel(16).to(gpu) #GV=generatorV(2).to(gpu) optimizer = torch.optim.Adam([{ 'params': GV.parameters(), 'lr': 1e-4 }, { 'params': z1, 'lr': 1e-4 }]) scheduler = ReduceLROnPlateau(optimizer, mode='min', factor=0.7, patience=6, verbose=True, min_lr=1e-7) pp = np.array([])
def myPSNR(org,recon): sqrError=abs(org-recon)**2 N=torch.prod(torch.tensor(org.shape[-2:])) mse=sqrError.sum(dim=(-1,-2)) mse=mse/N #maxval=np.max(org,axis=(-1,-2)) + 1e-15 psnr=10*torch.log10(mx**2/(mse+1e-15)) return psnr #%% #G=UnetClass().to(gpu) G=SmallModel1().to(gpu) G.load_state_dict(torch.load('tempp_10000.pt')) optimizer=torch.optim.Adam([{'params':G.parameters(),'lr':1e-4}]) scheduler = ReduceLROnPlateau(optimizer, mode='min', factor=0.7, patience=6, verbose=True, min_lr=1e-5) #z1=recT+torch.normal(0,0.02,(recT1.shape[0],recT1.shape[1],recT1.shape[2],recT1.shape[3],recT1.shape[4])).cuda() z=recT1 z=torch.reshape(z,(1,2*nbasis,nx,nx)).to(gpu) #z1=torch.cuda.FloatTensor(3,2*nbasis,nx,nx).fill_(0) #noi=[0.0003,0.05,0.01] #noi=[0.003,0.0005,0.01] #noi=[0.05,0.0005,0.01] #noi=[0.05,0.01,0.1] # for it in range(3): # z1[it]=z+torch.normal(0,noi[it],(z.shape[0],z.shape[1],z.shape[2],z.shape[3])).cuda() # z=z.repeat(3,1,1,1) pp=np.array([]) sf.tic() for rn in range(40): #z=atb#+torch.normal(0,noi[it],(atb.shape[0],atb.shape[1],atb.shape[2],atb.shape[3])).cuda() #z=torch.reshape(z,(1,2*nbasis,nx,nx)) #z=atb+torch.normal(0,0.005,(atb.shape[0],atb.shape[1],atb.shape[2],atb.shape[3])).cuda() #z=torch.reshape(z,(1,2*nbasis,nx,nx)) u1=G(z)#+z.cuda() u1=torch.reshape(u1,(nbasis,1,2,nx,nx)) l1_reg=0 for param in G.parameters(): l1_reg += param.abs().sum() loss=((abs(u1-recT1))).pow(2).sum()+1.0*l1_reg if rn%20==0: # print(rn,loss.item()) #torch.cuda.empty_cache() plt.figure(rn) dd=(v2[0,0,:,20].unsqueeze(1).unsqueeze(1)*u1.squeeze(1)).sum(dim=0).detach().cpu().numpy() plt.imshow(np.abs((dd[0,100:400,100:400]+dd[1,100:400,100:400]*1j)),cmap='gray') plt.show() plt.pause(1) dd=(v2[0,0,:,10:50].unsqueeze(-1).unsqueeze(-1).detach().cpu()*u1[:,:,:,100:400,100:400].detach().cpu()).sum(dim=0).detach().cpu() org1=(v2[0,0,:,10:50].unsqueeze(-1).unsqueeze(-1).detach().cpu()*recT[:,:,:,100:400,100:400].detach().cpu()).sum(dim=0) psnr=torch.mean(myPSNR(org1,dd)) optimizer.zero_grad() loss.backward() optimizer.step() scheduler.step(loss.detach()) print(rn,loss.item(),psnr.item()) pp=np.append(pp,psnr) sf.toc() #%% # sf.tic() #z1=torch.normal(0,0.01,(1,1,2,nf1)).cuda() z1=0.01*torch.ones((1,1,2,nf1)).to(gpu) z1[0,0,1,:]=torch.tensor(V[22,0:nf1]) z1[0,0,0,:]=torch.tensor(V[27,0:nf1]) #vv=V[30:32] #z1=torch.reshape(torch.tensor(V[29:31,0:nf1]),(1,1,2,nf1)).to(gpu) z1 = Variable(z1,requires_grad=True) V1=torch.tensor(V[:,0:nf1]).cuda()#+torch.normal(0,0.01,(V.shape[0],V.shape[1])).cuda() #V1=torch.reshape(V1,(1,1,nbasis,nf1)) GV=SmallModel(16).to(gpu) #GV=generatorV(2).to(gpu) optimizer=torch.optim.Adam([{'params':GV.parameters(),'lr':1e-4},{'params':z1,'lr':1e-6}]) scheduler = ReduceLROnPlateau(optimizer, mode='min', factor=0.7, patience=6, verbose=True, min_lr=1e-7) pp=np.array([]) for ep1 in range(25000): for bat in range(1): #z1=torch.reshape(V1,(1,1,nbasis,nf1)) l1_reg=0. for param in GV.parameters(): l1_reg += param.abs().sum() v1=GV(z1) #v1=v1.permute(1,0) loss=abs(v1[0,0]-V1).pow(2).sum()+0.000*l1_reg+10.0*Smoothness(z1[0,0,:,:]) if ep1%10==0: print(ep1,loss.item()) optimizer.zero_grad() loss.backward() optimizer.step() scheduler.step(loss.detach()) #pp=np.append(pp,loss.item()) del kdaT,smap,adjnufft_ob #%% def Smoothness(zvec): zsmoothness = zvec[:,1:]-zvec[:,:-1] zsmoothness = torch.sum(zsmoothness*zsmoothness,axis=1).squeeze() zsmoothness = torch.sum(zsmoothness,axis=0) return(zsmoothness) z=recT1#+torch.normal(0,0.03,(atb.shape[0],atb.shape[1],atb.shape[2],atb.shape[3])).cuda()#.permute(1,0,2,3) z=torch.reshape(z,(1,2*nbasis,nx,nx)) #z1=torch.cuda.FloatTensor(1,1,nbasis,nf1).fill_(0) #z1=torch.normal(0,1,(1,1,nbasis,nf1)) z1=torch.reshape(torch.tensor(V[:,0:nf1]),(1,1,nbasis,nf1)) #z1 = Variable(z1,requires_grad=True) #%% torch.cuda.empty_cache() #from dn_modelU2 import SmallModel1 # G=SmallModel1().to(gpu) # GV=SmallModel().to(gpu) # G.load_state_dict(torch.load('tempp_199.pt')) ## G.load_state_dict(torch.load('wtsN-U3.pt')) ## GV.load_state_dict(torch.load('wtsN-V3.pt')) # #optimizer=torch.optim.AdamW([{'params':G.parameters(),'lr':1e-4},{'params':GV.parameters(),'lr':1e-4},{'params':z,'lr':1e-4}]) #v1=(z1.cuda()).unsqueeze(-1) #v1 = Variable(v1,requires_grad=True) optimizer=torch.optim.Adam([{'params':G.parameters(),'lr':1e-5},{'params':GV.parameters(),'lr':1e-4},{'params':z1,'lr':1e-6}]) # #optimizer=torch.optim.AdamW([{'params':G.parameters(),'lr':1e-4},{'params':GV.parameters(),'lr':1e-4}]) scheduler = ReduceLROnPlateau(optimizer, mode='min', factor=0.7, patience=6, verbose=True, min_lr=1e-5) #optimizer=torch.optim.SGD([{'params':v1,'lr':1e-4}])#,{'params':z,'lr':5e-3,'momentum':0.9}]) smapT=smapT.cpu() ktrajT=ktrajT.to(cpu) dcf=torch.reshape(dcf,(nbasis,3,2,NF,nintl*nx)) plt.figure(2) dd=(np.reshape(V[:,26],(nbasis,1,1,1))*recT.detach().cpu().numpy()).sum(axis=0) plt.imshow(np.abs((dd[0,100:400,100:400]+dd[1,100:400,100:400]*1j)),cmap='gray') pp=np.array([]) plt.pause(1) plt.figure(1) sf.tic() for it in range(500): # for bat in range(1): v1=GV(z1.cuda()).unsqueeze(-1) u1=G(z.cuda())#+z.cuda() u1=torch.reshape(u1,(nbasis,2,nx,nx)) u1=u1.unsqueeze(1) l1_regU=0. for param in G.parameters(): l1_regU += param.abs().sum() l1_regV=0. for param in GV.parameters(): l1_regV += param.abs().sum() err=abs(AUV1(u1,v1)-kdataT1[:,:,0:nf1].cuda()).pow(2) loss=((dcf[0,:,:,0:nf1])*(err)).sum()+0.000005*l1_regU+0.00001*l1_regV+10.0*Smoothness(z1[0,0,:,:]) # err=abs(yy1.squeeze(1)-atb).pow(2) # loss=(err).sum()#+10*l1_reg#+0.1*(z.cpu()-u2.cpu()).pow(2).sum()+0.1*(z1[0,0].cpu()-v2.cpu()).pow(2).sum() if it%10==0: plt.figure(it) dd=(v1[0,0,:,20].unsqueeze(1).unsqueeze(1)*u1.squeeze(1)).sum(dim=0).detach().cpu().numpy() plt.imshow(np.abs((dd[0,100:400,100:400]+dd[1,100:400,100:400]*1j)),cmap='gray') plt.show() plt.pause(1) # if it%10==0: # torch.save(u1,'tmpU'+str(it)+'.pt') optimizer.zero_grad() loss.backward() optimizer.step() scheduler.step(loss.detach()) dd=(v2[0,0,:,10:50].unsqueeze(-1).unsqueeze(-1).detach().cpu()*u1[:,:,:,100:400,100:400].detach().cpu()).sum(dim=0).detach().cpu() org1=(v2[0,0,:,10:50].unsqueeze(-1).unsqueeze(-1).detach().cpu()*recT[:,:,:,100:400,100:400].detach().cpu()).sum(dim=0) psnr=torch.mean(myPSNR(org1,dd)) print(it,loss.item(),psnr.item()) pp=np.append(pp,psnr) sf.toc()
# nch=csm.shape[1] # nbas=x.shape[0] # tmp1=torch.zeros((nch,nbas,2,NF*nx*nintl)).cuda() # for i in range(nch): # tmp=sf.complex_mult(x,csm[:,i].repeat(nbas,1,1,1,1),dim=2) # tmp=torch.reshape(tmp,(nbas,1,2,nx,nx)) # #xx=ktrajT[0].unsqueeze(0) # tmp1[i]=nufft_ob(tmp,traj.repeat(nbas,1,1).cuda()).squeeze(1) # tmp1=tmp1.permute(1,0,2,3) # return tmp1 #%% G = UnetClass().to(gpu) G = SmallModel1().to(gpu) GV = SmallModel().to(gpu) #G.load_state_dict(torch.load('wts-10U2.pt')) #GV.load_state_dict(torch.load('wts-10V2.pt')) #GV.load_state_dict(torch.load('./PTmodels/27Oct_112451am_500ep_27oct/wts-500.pt')) #optimizer=torch.optim.SGD([{'params':G.parameters(),'lr':5e-3,'momentum':0.9}]) optimizer = torch.optim.AdamW([{'params': G.parameters(), 'lr': 1e-3}]) #optimizer=torch.optim.AdamW([{'params':G.parameters(),'lr':1e-3},{'params':GV.parameters(),'lr':1e-3}]) scheduler = ReduceLROnPlateau(optimizer, mode='min', factor=0.7, patience=6, verbose=True, min_lr=5e-5)
n_select = 30 nbasis = 30 #%%Generate a meaningful filename to save the trainined models for testing print('*************************************************') start_time = time.time() saveDir = 'PTmodels/' cwd = os.getcwd() directory=saveDir+datetime.now().strftime("%d%b_%I%M%S%P_")+ \ str(epochs)+'ep_' +'27oct' if not os.path.exists(directory): os.makedirs(directory) sessFileName = directory + '/modelVb' #%% creating the training model unet = SmallModel() unet = unet.to(gpu) optimizer = torch.optim.Adam(unet.parameters(), lr=1e-3) def lossFun(pred, org): loss = torch.mean(torch.abs(pred - org)) return loss #%% training code #print ('training started on', datetime.now().strftime("%d-%b-%Y at %I:%M %P")) #start_time=time.time() torch.save(unet.state_dict(), directory + '/wts-0.pt') #writer = SummaryWriter(log_dir=directory+'/') #writer.add_graph(unet,torch.randn(1,2,512,512,30).to(gpu))
res = res.to(gpu) for k in range(bsz): tmp3 = msk[k].repeat(nbasis, 1) #*tmp2 res[k] = torch.diag(Vv[k]) @ tmp3.to(torch.float32) return res #%% Loading trained networks parameters wts1 = dir2 + 'wts-20.pt' wts2 = dir3 + 'wts-500.pt' gnu = SmallModel1().to(gpu) gnu.load_state_dict(torch.load(wts1)) gnv = SmallModel().to(gpu) gnv.load_state_dict(torch.load(wts2)) wtsFname1 = 'wts-10U' + str(2) + '.pt' #wtsFname1='wts-2U'+str(60+1)+'.pt' G = UnetClass().to(gpu) G.load_state_dict(torch.load(wtsFname1)) #atbT=atbT*W z = recT.permute(3, 0, 1, 2) z = torch.reshape(z, (1, 2 * nbasis, 512, 512)).to(gpu) #z = torch.randn((1,60,512,512),device=gpu, dtype=dtype) z = Variable(z, requires_grad=False) wtsFname2 = 'wts-10V' + str(2) + '.pt'
self.nx * self.nx * 2).fill_(0) x = torch.reshape(x, (nbas, self.nx, self.nx, 2)) for i in range(nch): tmp2 = sf.pt_fft2c( sf.pt_cpx_multipy(x, self.csmT[i].repeat(nbas, 1, 1, 1))) tmp2 = torch.reshape(tmp2, (nbas, self.NX)) tmp2 = tmp2.repeat(self.NF, 1, 1) * mask tmp5[i] = tmp2.sum(axis=1) del tmp2, x return tmp5.cpu() #%% G = UnetClass().to(gpu) GV = SmallModel().to(gpu) G.load_state_dict(torch.load('wts-5aU22.pt')) GV.load_state_dict(torch.load('wts-5aV22.pt')) #optimizer=torch.optim.SGD([{'params':G.parameters(),'lr':5e-3,'momentum':0.9}]) optimizer = torch.optim.AdamW([{ 'params': G.parameters(), 'lr': 1e-4 }, { 'params': GV.parameters(), 'lr': 1e-4 }]) scheduler = ReduceLROnPlateau(optimizer, mode='min', factor=0.7, patience=6, verbose=True,
# nch=csm.shape[1] # nbas=x.shape[0] # tmp1=torch.zeros((nch,nbas,2,NF*nx*nintl)).cuda() # for i in range(nch): # tmp=sf.complex_mult(x,csm[:,i].repeat(nbas,1,1,1,1),dim=2) # tmp=torch.reshape(tmp,(nbas,1,2,nx,nx)) # #xx=ktrajT[0].unsqueeze(0) # tmp1[i]=nufft_ob(tmp,traj.repeat(nbas,1,1).cuda()).squeeze(1) # tmp1=tmp1.permute(1,0,2,3) # return tmp1 #%% #G=UnetClass().to(gpu) G=SmallModel1().to(gpu) GV=SmallModel().to(gpu) #G.load_state_dict(torch.load('wtsUB-131.pt')) #GV.load_state_dict(torch.load('wtsVB-400.pt')) #GV.load_state_dict(torch.load('./PTmodels/27Oct_112451am_500ep_27oct/wts-500.pt')) #optimizer=torch.optim.SGD([{'params':G.parameters(),'lr':5e-3,'momentum':0.9}]) #optimizer=torch.optim.AdamW([{'params':G.parameters(),'lr':1e-4}]) optimizer=torch.optim.AdamW([{'params':G.parameters(),'lr':1e-4},{'params':GV.parameters(),'lr':1e-4}]) scheduler = ReduceLROnPlateau(optimizer, mode='min', factor=0.7, patience=6, verbose=True, min_lr=1e-5) trnFiles=os.listdir('/Shared/lss_jcb/abdul/prashant_cardiac_data/Data/d2/') sz=len(trnFiles) #data2=np.zeros((sz,1,n_select,N,N)).astype(np.complex64) rndm=random.sample(range(sz),sz)