aljaz/generative-models
1
0
Fork 0
mirror of https://github.com/Stability-AI/generative-models.git synced 2025-12-21 07:14:21 +01:00
Code Issues Packages Projects Releases Wiki Activity

Fixing additional GPU memory on device 0 due to discretization

Browse Source
This commit is contained in:
Tim
2023-07-06 22:25:37 -07:00
parent 061d11d55d
commit ba3e7fed5a
1 changed files with 10 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
sgm/modules/diffusionmodules/discretizer.py
View File

@@ -1,6 +1,7 @@
import torch import torch
import numpy as np import numpy as np
from functools import partial from functools import partial
from abc import abstractmethod
from ...util import append_zero from ...util import append_zero
from ...modules.diffusionmodules.util import make_beta_schedule from ...modules.diffusionmodules.util import make_beta_schedule
@@ -13,11 +14,15 @@ def generate_roughly_equally_spaced_steps(
class Discretization: class Discretization:
def __call__(self, n, do_append_zero=True, device="cuda", flip=False): def __call__(self, n, do_append_zero=True, device="cpu", flip=False):
sigmas = self.get_sigmas(n, device) sigmas = self.get_sigmas(n, device=device)
sigmas = append_zero(sigmas) if do_append_zero else sigmas sigmas = append_zero(sigmas) if do_append_zero else sigmas
return sigmas if not flip else torch.flip(sigmas, (0,)) return sigmas if not flip else torch.flip(sigmas, (0,))
@abstractmethod
def get_sigmas(self, n, device):
raise NotImplementedError("abstract class should not be called")
class EDMDiscretization(Discretization): class EDMDiscretization(Discretization):
def __init__(self, sigma_min=0.02, sigma_max=80.0, rho=7.0): def __init__(self, sigma_min=0.02, sigma_max=80.0, rho=7.0):
@@ -25,7 +30,7 @@ class EDMDiscretization(Discretization):
self.sigma_max = sigma_max self.sigma_max = sigma_max
self.rho = rho self.rho = rho
def get_sigmas(self, n, device): def get_sigmas(self, n, device="cpu"):
ramp = torch.linspace(0, 1, n, device=device) ramp = torch.linspace(0, 1, n, device=device)
min_inv_rho = self.sigma_min ** (1 / self.rho) min_inv_rho = self.sigma_min ** (1 / self.rho)
max_inv_rho = self.sigma_max ** (1 / self.rho) max_inv_rho = self.sigma_max ** (1 / self.rho)
@@ -40,6 +45,7 @@ class LegacyDDPMDiscretization(Discretization):
linear_end=0.0120, linear_end=0.0120,
num_timesteps=1000, num_timesteps=1000,
): ):
super().__init__()
self.num_timesteps = num_timesteps self.num_timesteps = num_timesteps
betas = make_beta_schedule( betas = make_beta_schedule(
"linear", num_timesteps, linear_start=linear_start, linear_end=linear_end "linear", num_timesteps, linear_start=linear_start, linear_end=linear_end
@@ -48,7 +54,7 @@ class LegacyDDPMDiscretization(Discretization):
self.alphas_cumprod = np.cumprod(alphas, axis=0) self.alphas_cumprod = np.cumprod(alphas, axis=0)
self.to_torch = partial(torch.tensor, dtype=torch.float32) self.to_torch = partial(torch.tensor, dtype=torch.float32)
def get_sigmas(self, n, device): def get_sigmas(self, n, device="cpu"):
if n < self.num_timesteps: if n < self.num_timesteps:
timesteps = generate_roughly_equally_spaced_steps(n, self.num_timesteps) timesteps = generate_roughly_equally_spaced_steps(n, self.num_timesteps)
alphas_cumprod = self.alphas_cumprod[timesteps] alphas_cumprod = self.alphas_cumprod[timesteps]