ColossalAI/examples/images/diffusion/configs/Inference/x4-upscaling.yaml

76 lines
2.2 KiB
YAML

model:
base_learning_rate: 1.0e-04
target: ldm.models.diffusion.ddpm.LatentUpscaleDiffusion
params:
parameterization: "v"
low_scale_key: "lr"
linear_start: 0.0001
linear_end: 0.02
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
image_size: 128
channels: 4
cond_stage_trainable: false
conditioning_key: "hybrid-adm"
monitor: val/loss_simple_ema
scale_factor: 0.08333
use_ema: False
low_scale_config:
target: ldm.modules.diffusionmodules.upscaling.ImageConcatWithNoiseAugmentation
params:
noise_schedule_config: # image space
linear_start: 0.0001
linear_end: 0.02
max_noise_level: 350
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
use_checkpoint: True
num_classes: 1000 # timesteps for noise conditioning (here constant, just need one)
image_size: 128
in_channels: 7
out_channels: 4
model_channels: 256
attention_resolutions: [ 2,4,8]
num_res_blocks: 2
channel_mult: [ 1, 2, 2, 4]
disable_self_attentions: [True, True, True, False]
disable_middle_self_attn: False
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 1024
legacy: False
use_linear_in_transformer: True
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
ddconfig:
# attn_type: "vanilla-xformers" this model needs efficient attention to be feasible on HR data, also the decoder seems to break in half precision (UNet is fine though)
double_z: True
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult: [ 1,2,4 ] # num_down = len(ch_mult)-1
num_res_blocks: 2
attn_resolutions: [ ]
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder
params:
freeze: True
layer: "penultimate"