import torch import utils.logging import os import torchvision from torchvision.transforms.functional import crop # This script is adapted from the following repository: https://github.com/ermongroup/ddim def compute_alpha(beta, t): beta = torch.cat([torch.zeros(1).to(beta.device), beta], dim=0) a = (1 - beta).cumprod(dim=0).index_select(0, t + 1).view(-1, 1, 1, 1) return a def data_transform(X): return 2 * X - 1.0 def inverse_data_transform(X): return torch.clamp((X + 1.0) / 2.0, 0.0, 1.0) def generalized_steps(x, x_cond,seq, model, b,ii,jj,osize,eta=0.): with torch.no_grad(): n = x.size(0) seq_next = [-1] + list(seq[:-1]) x0_preds = [] xs = [x] for i, j in zip(reversed(seq), reversed(seq_next)): t = (torch.ones(n) * i).to(x.device) next_t = (torch.ones(n) * j).to(x.device) at = compute_alpha(b, t.long()) at_next = compute_alpha(b, next_t.long()) xt = xs[-1].to(x.device) ii = ii.to(x.device) jj = jj.to(x.device) osize = osize.to(x.device) # print(xt.shape) # print(x_cond.shape) et = model(torch.cat([x_cond, xt], dim=1), t,ii,jj,osize) x0_t = (xt - et * (1 - at).sqrt()) / at.sqrt() x0_preds.append(x0_t.to('cpu')) c1 = eta * ((1 - at / at_next) * (1 - at_next) / (1 - at)).sqrt() c2 = ((1 - at_next) - c1 ** 2).sqrt() xt_next = at_next.sqrt() * x0_t + c1 * torch.randn_like(x) + c2 * et xs.append(xt_next.to('cpu')) return xs, x0_preds def generalized_steps_overlapping(x, x_cond, seq, model, b,ii,jj,osize, eta=0., corners=None, p_size=None,corners1=None,corners2=None,manual_batching=True): with torch.no_grad(): n = x.size(0) seq_next = [-1] + list(seq[:-1]) x0_preds = [] xs = [x] x_grid_mask = torch.zeros(x_cond.size(0),3,x_cond.size(2),x_cond.size(3), device=x.device) for (hi, wi) in corners: x_grid_mask[:, :, hi:hi + p_size, wi:wi + p_size] += 1 if corners1!=None: p_size1=96 x_grid_mask1 = torch.zeros(x_cond.size(0),3,x_cond.size(2),x_cond.size(3), device=x.device) for (hi, wi) in corners1: x_grid_mask1[:, :, hi:hi + p_size1, wi:wi + p_size1] += 1 if corners2 != None: p_size2=128 x_grid_mask2 = torch.zeros(x_cond.size(0),3,x_cond.size(2),x_cond.size(3), device=x.device) for (hi, wi) in corners2: x_grid_mask2[:, :, hi:hi + p_size2, wi:wi + p_size2] += 1 for i, j in zip(reversed(seq), reversed(seq_next)): t = (torch.ones(n) * i).to(x.device) next_t = (torch.ones(n) * j).to(x.device) at = compute_alpha(b, t.long()) at_next = compute_alpha(b, next_t.long()) xt = xs[-1].to(x.device) et_output = torch.zeros(x_cond.size(0),3,x_cond.size(2),x_cond.size(3), device=x.device) if manual_batching==True: manual_batching_size = p_size xt_patch = torch.cat([crop(xt, hi, wi, p_size, p_size) for (hi, wi) in corners], dim=0) x_cond_patch = torch.cat([data_transform(crop(x_cond, hi, wi, p_size, p_size)) for (hi, wi) in corners], dim=0) for i in range(0, len(corners), manual_batching_size):#以16为步长的块即为corners ii_input=torch.unsqueeze(ii[i],dim=0) jj_input=torch.unsqueeze(jj[i],dim=0) osize_input=torch.unsqueeze(osize[i],dim=0) x_input=torch.cat([x_cond_patch[i:i + manual_batching_size], xt_patch[i:i + manual_batching_size]], dim=1) # print("x_input", x_input.shape) outputs = model(x_input, t,ii_input,jj_input,osize_input) # print("outputs", outputs.shape) for idx, (hi, wi) in enumerate(corners[i:i+manual_batching_size]): # print("idx", idx) et_output[0, :, hi:hi + p_size, wi:wi + p_size] += outputs[idx] if corners1 != None: et_output1 = torch.zeros(x_cond.size(0), 3, x_cond.size(2), x_cond.size(3), device=x.device) manual_batching_size = p_size1 xt_patch = torch.cat([crop(xt, hi, wi, p_size1, p_size1) for (hi, wi) in corners1], dim=0) x_cond_patch = torch.cat( [data_transform(crop(x_cond, hi, wi, p_size1, p_size1)) for (hi, wi) in corners1], dim=0) for i in range(0, len(corners1), manual_batching_size): # 以16为步长的块即为corners ii_input = torch.unsqueeze(ii[i], dim=0) jj_input = torch.unsqueeze(jj[i], dim=0) osize_input = torch.unsqueeze(osize[i], dim=0) x_input = torch.cat( [x_cond_patch[i:i + manual_batching_size], xt_patch[i:i + manual_batching_size]], dim=1) # print("x_input", x_input.shape) outputs1 = model(x_input, t, ii_input, jj_input, osize_input) # print("outputs", outputs.shape) for idx, (hi, wi) in enumerate(corners1[i:i + manual_batching_size]): # print("idx", idx) et_output1[0, :, hi:hi + p_size1, wi:wi + p_size1] += outputs1[idx] if corners2 != None: et_output2= torch.zeros(x_cond.size(0), 3, x_cond.size(2), x_cond.size(3), device=x.device) manual_batching_size = p_size2 xt_patch = torch.cat([crop(xt, hi, wi, p_size2, p_size2) for (hi, wi) in corners2], dim=0) x_cond_patch = torch.cat( [data_transform(crop(x_cond, hi, wi, p_size2, p_size2)) for (hi, wi) in corners2], dim=0) for i in range(0, len(corners2), manual_batching_size): # 以16为步长的块即为corners ii_input = torch.unsqueeze(ii[i], dim=0) jj_input = torch.unsqueeze(jj[i], dim=0) osize_input = torch.unsqueeze(osize[i], dim=0) x_input = torch.cat( [x_cond_patch[i:i + manual_batching_size], xt_patch[i:i + manual_batching_size]], dim=1) # print("x_input", x_input.shape) outputs2 = model(x_input, t, ii_input, jj_input, osize_input) # print("outputs", outputs.shape) for idx, (hi, wi) in enumerate(corners2[i:i + manual_batching_size]): # print("idx", idx) et_output2[0, :, hi:hi + p_size2, wi:wi + p_size2] += outputs2[idx] else: for (hi, wi) in corners: xt_patch = crop(xt, hi, wi, p_size, p_size) x_cond_patch = crop(x_cond, hi, wi, p_size, p_size) x_cond_patch = data_transform(x_cond_patch) et_output[:, :, hi:hi + p_size, wi:wi + p_size] += model(torch.cat([x_cond_patch, xt_patch], dim=1), t) et0 = torch.div(et_output, x_grid_mask) if corners1 != None: et1 = torch.div(et_output1, x_grid_mask1) et=(et0+et1)/2.0 if corners2 != None: et2 = torch.div(et_output2, x_grid_mask2) et = (et0 +et1+ et2) / 3.0 else: et=et0 x0_t = (xt - et * (1 - at).sqrt()) / at.sqrt() x0_preds.append(x0_t.to('cpu')) c1 = eta * ((1 - at / at_next) * (1 - at_next) / (1 - at)).sqrt() c2 = ((1 - at_next) - c1 ** 2).sqrt() xt_next = at_next.sqrt() * x0_t + c1 * torch.randn_like(x) + c2 * et # if i>500: # xt_next=xt_next/torch.mean(xt_next)*torch.mean(x_cond[:,3:6,:,]) xs.append(xt_next.to('cpu')) return xs, x0_preds