From c675caf3f9021d257609b03dfd6c36e23c74692e Mon Sep 17 00:00:00 2001 From: cmdr2 Date: Fri, 23 Sep 2022 00:54:35 +0530 Subject: [PATCH] Fix a missing ddim_steps error for the DDIM sampler --- ui/sd_internal/ddim_callback.patch | 45 ++++++++++++++++-------------- 1 file changed, 24 insertions(+), 21 deletions(-) diff --git a/ui/sd_internal/ddim_callback.patch b/ui/sd_internal/ddim_callback.patch index 9852aab8..73dcff22 100644 --- a/ui/sd_internal/ddim_callback.patch +++ b/ui/sd_internal/ddim_callback.patch @@ -1,5 +1,5 @@ diff --git a/optimizedSD/ddpm.py b/optimizedSD/ddpm.py -index b967b55..10a7c32 100644 +index b967b55..e06e653 100644 --- a/optimizedSD/ddpm.py +++ b/optimizedSD/ddpm.py @@ -22,7 +22,7 @@ from ldm.util import exists, default, instantiate_from_config @@ -11,8 +11,11 @@ index b967b55..10a7c32 100644 def disabled_train(self): """Overwrite model.train with this function to make sure train/eval mode -@@ -528,39 +528,46 @@ class UNet(DDPM): +@@ -526,41 +526,49 @@ class UNet(DDPM): + ) + elif sampler == "ddim": ++ self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=False) samples = self.ddim_sampling(x_latent, conditioning, S, unconditional_guidance_scale=unconditional_guidance_scale, unconditional_conditioning=unconditional_conditioning, - mask = mask,init_latent=x_T,use_original_steps=False) @@ -67,7 +70,7 @@ index b967b55..10a7c32 100644 @torch.no_grad() def plms_sampling(self, cond,b, img, ddim_use_original_steps=False, -@@ -599,10 +606,10 @@ class UNet(DDPM): +@@ -599,10 +607,10 @@ class UNet(DDPM): old_eps.append(e_t) if len(old_eps) >= 4: old_eps.pop(0) @@ -81,7 +84,7 @@ index b967b55..10a7c32 100644 @torch.no_grad() def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False, -@@ -706,7 +713,8 @@ class UNet(DDPM): +@@ -706,7 +714,8 @@ class UNet(DDPM): @torch.no_grad() def ddim_sampling(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None, @@ -91,7 +94,7 @@ index b967b55..10a7c32 100644 timesteps = self.ddim_timesteps timesteps = timesteps[:t_start] -@@ -730,10 +738,13 @@ class UNet(DDPM): +@@ -730,10 +739,13 @@ class UNet(DDPM): unconditional_guidance_scale=unconditional_guidance_scale, unconditional_conditioning=unconditional_conditioning) @@ -107,7 +110,7 @@ index b967b55..10a7c32 100644 @torch.no_grad() -@@ -779,13 +790,16 @@ class UNet(DDPM): +@@ -779,13 +791,16 @@ class UNet(DDPM): @torch.no_grad() @@ -125,7 +128,7 @@ index b967b55..10a7c32 100644 s_in = x.new_ones([x.shape[0]]).half() for i in trange(len(sigmas) - 1, disable=disable): gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0. -@@ -807,13 +821,18 @@ class UNet(DDPM): +@@ -807,13 +822,18 @@ class UNet(DDPM): d = to_d(x, sigma_hat, denoised) if callback is not None: callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised}) @@ -146,7 +149,7 @@ index b967b55..10a7c32 100644 """Ancestral sampling with Euler method steps.""" extra_args = {} if extra_args is None else extra_args -@@ -822,6 +841,8 @@ class UNet(DDPM): +@@ -822,6 +842,8 @@ class UNet(DDPM): sigmas = cvd.get_sigmas(S) x = x*sigmas[0] @@ -155,7 +158,7 @@ index b967b55..10a7c32 100644 s_in = x.new_ones([x.shape[0]]).half() for i in trange(len(sigmas) - 1, disable=disable): -@@ -837,17 +858,22 @@ class UNet(DDPM): +@@ -837,17 +859,22 @@ class UNet(DDPM): sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1]) if callback is not None: callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) @@ -180,7 +183,7 @@ index b967b55..10a7c32 100644 """Implements Algorithm 2 (Heun steps) from Karras et al. (2022).""" extra_args = {} if extra_args is None else extra_args -@@ -855,6 +881,8 @@ class UNet(DDPM): +@@ -855,6 +882,8 @@ class UNet(DDPM): sigmas = cvd.get_sigmas(S) x = x*sigmas[0] @@ -189,7 +192,7 @@ index b967b55..10a7c32 100644 s_in = x.new_ones([x.shape[0]]).half() for i in trange(len(sigmas) - 1, disable=disable): -@@ -876,6 +904,9 @@ class UNet(DDPM): +@@ -876,6 +905,9 @@ class UNet(DDPM): d = to_d(x, sigma_hat, denoised) if callback is not None: callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigma_hat, 'denoised': denoised}) @@ -199,7 +202,7 @@ index b967b55..10a7c32 100644 dt = sigmas[i + 1] - sigma_hat if sigmas[i + 1] == 0: # Euler method -@@ -895,11 +926,13 @@ class UNet(DDPM): +@@ -895,11 +927,13 @@ class UNet(DDPM): d_2 = to_d(x_2, sigmas[i + 1], denoised_2) d_prime = (d + d_2) / 2 x = x + d_prime * dt @@ -215,7 +218,7 @@ index b967b55..10a7c32 100644 """A sampler inspired by DPM-Solver-2 and Algorithm 2 from Karras et al. (2022).""" extra_args = {} if extra_args is None else extra_args -@@ -907,6 +940,8 @@ class UNet(DDPM): +@@ -907,6 +941,8 @@ class UNet(DDPM): sigmas = cvd.get_sigmas(S) x = x*sigmas[0] @@ -224,7 +227,7 @@ index b967b55..10a7c32 100644 s_in = x.new_ones([x.shape[0]]).half() for i in trange(len(sigmas) - 1, disable=disable): gamma = min(s_churn / (len(sigmas) - 1), 2 ** 0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0. -@@ -924,7 +959,7 @@ class UNet(DDPM): +@@ -924,7 +960,7 @@ class UNet(DDPM): e_t_uncond, e_t = (x_in + eps * c_out).chunk(2) denoised = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond) @@ -233,7 +236,7 @@ index b967b55..10a7c32 100644 d = to_d(x, sigma_hat, denoised) # Midpoint method, where the midpoint is chosen according to a rho=3 Karras schedule -@@ -945,11 +980,13 @@ class UNet(DDPM): +@@ -945,11 +981,13 @@ class UNet(DDPM): d_2 = to_d(x_2, sigma_mid, denoised_2) x = x + d_2 * dt_2 @@ -249,7 +252,7 @@ index b967b55..10a7c32 100644 """Ancestral sampling with DPM-Solver inspired second-order steps.""" extra_args = {} if extra_args is None else extra_args -@@ -957,6 +994,8 @@ class UNet(DDPM): +@@ -957,6 +995,8 @@ class UNet(DDPM): sigmas = cvd.get_sigmas(S) x = x*sigmas[0] @@ -258,7 +261,7 @@ index b967b55..10a7c32 100644 s_in = x.new_ones([x.shape[0]]).half() for i in trange(len(sigmas) - 1, disable=disable): -@@ -973,6 +1012,9 @@ class UNet(DDPM): +@@ -973,6 +1013,9 @@ class UNet(DDPM): sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1]) if callback is not None: callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised}) @@ -268,7 +271,7 @@ index b967b55..10a7c32 100644 d = to_d(x, sigmas[i], denoised) # Midpoint method, where the midpoint is chosen according to a rho=3 Karras schedule sigma_mid = ((sigmas[i] ** (1 / 3) + sigma_down ** (1 / 3)) / 2) ** 3 -@@ -993,11 +1035,13 @@ class UNet(DDPM): +@@ -993,11 +1036,13 @@ class UNet(DDPM): d_2 = to_d(x_2, sigma_mid, denoised_2) x = x + d_2 * dt_2 x = x + torch.randn_like(x) * sigma_up @@ -284,7 +287,7 @@ index b967b55..10a7c32 100644 extra_args = {} if extra_args is None else extra_args s_in = x.new_ones([x.shape[0]]) -@@ -1005,6 +1049,8 @@ class UNet(DDPM): +@@ -1005,6 +1050,8 @@ class UNet(DDPM): sigmas = cvd.get_sigmas(S) x = x*sigmas[0] @@ -293,7 +296,7 @@ index b967b55..10a7c32 100644 ds = [] for i in trange(len(sigmas) - 1, disable=disable): -@@ -1017,6 +1063,7 @@ class UNet(DDPM): +@@ -1017,6 +1064,7 @@ class UNet(DDPM): e_t_uncond, e_t = (x_in + eps * c_out).chunk(2) denoised = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond) @@ -301,7 +304,7 @@ index b967b55..10a7c32 100644 d = to_d(x, sigmas[i], denoised) ds.append(d) -@@ -1027,4 +1074,5 @@ class UNet(DDPM): +@@ -1027,4 +1075,5 @@ class UNet(DDPM): cur_order = min(i + 1, order) coeffs = [linear_multistep_coeff(cur_order, sigmas.cpu(), i, j) for j in range(cur_order)] x = x + sum(coeff * d for coeff, d in zip(coeffs, reversed(ds)))