测试 about rddm

[TonFard]

original_ddim_ddpm = False
num_unet = 2
objective = 'pred_res_noise'
test_res_or_noise = "res_noise"
if original_ddim_ddpm:
model = Unet(
dim=64,
dim_mults=(1, 2, 4, 8)
)
diffusion = GaussianDiffusion(
model,
image_size=image_size,
timesteps=1000, # number of steps
sampling_timesteps=sampling_timesteps_original_ddim_ddpm,
loss_type='l1', # L1 or L2
)
else:
model = UnetRes(
dim=64,
dim_mults=(1, 2, 4, 8),
num_unet=num_unet,
condition=condition,
input_condition=input_condition,
objective=objective,
test_res_or_noise=test_res_or_noise
)
diffusion = ResidualDiffusion(
model,
image_size=image_size,
timesteps=1000, # number of steps
# number of sampling timesteps (using ddim for faster inference [see citation for ddim paper])
sampling_timesteps=sampling_timesteps,
objective=objective,
loss_type='l2', # L1 or L2
condition=condition,
sum_scale=sum_scale,
input_condition=input_condition,
input_condition_mask=input_condition_mask,
test_res_or_noise=test_res_or_noise
)

这是我的配置

[Z0Victor]

请问您是用多少张卡训练的

[nachifur]

感谢大家关注RDDM。去噪可以参考一下另一个issue，https://github.com/nachifur/RDDM/issues/2。

单卡3090就可以跑论文里绝大多数的实验。

抱歉，这几天事情比较多，过段时间会更新代码。

[TonFard]

8g显存的卡一个batchsize能跑起来

[Z0Victor]

8g显存的卡一个batchsize能跑起来

好的，感谢您