ColossalAI/examples/images/diffusion/configs/train_ddp.yaml

120 lines
3.1 KiB
YAML
Raw Normal View History

2022-11-08 08:14:45 +00:00
model:
2022-12-12 09:35:23 +00:00
base_learning_rate: 1.0e-4
#target: ldm.models.diffusion.ddpm.LatentDiffusion
2022-11-08 08:14:45 +00:00
params:
2022-12-12 09:35:23 +00:00
parameterization: "v"
2022-11-08 08:14:45 +00:00
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: image
2022-12-12 09:35:23 +00:00
cond_stage_key: txt
image_size: 64
2022-11-08 08:14:45 +00:00
channels: 4
2022-12-12 09:35:23 +00:00
cond_stage_trainable: false
2022-11-08 08:14:45 +00:00
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
2022-12-12 09:35:23 +00:00
use_ema: False # we set this to false because this is an inference only config
2022-11-08 08:14:45 +00:00
scheduler_config: # 10000 warmup steps
target: ldm.lr_scheduler.LambdaLinearScheduler
params:
2022-12-12 09:35:23 +00:00
warm_up_steps: [ 1 ] # NOTE for resuming. use 10000 if starting from scratch
2022-11-08 08:14:45 +00:00
cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
f_start: [ 1.e-6 ]
f_max: [ 1.e-4 ]
2022-12-12 09:35:23 +00:00
f_min: [ 1.e-10 ]
2022-11-08 08:14:45 +00:00
unet_config:
#target: ldm.modules.diffusionmodules.openaimodel.UNetModel
2022-11-08 08:14:45 +00:00
params:
2022-12-12 09:35:23 +00:00
use_checkpoint: True
use_fp16: True
2022-11-08 08:14:45 +00:00
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
2022-12-12 09:35:23 +00:00
num_head_channels: 64 # need to fix for flash-attn
2022-11-08 08:14:45 +00:00
use_spatial_transformer: True
2022-12-12 09:35:23 +00:00
use_linear_in_transformer: True
2022-11-08 08:14:45 +00:00
transformer_depth: 1
2022-12-12 09:35:23 +00:00
context_dim: 1024
2022-11-08 08:14:45 +00:00
legacy: False
first_stage_config:
#target: ldm.models.autoencoder.AutoencoderKL
2022-11-08 08:14:45 +00:00
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
2022-12-12 09:35:23 +00:00
#attn_type: "vanilla-xformers"
2022-11-08 08:14:45 +00:00
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
#target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder
2022-11-08 08:14:45 +00:00
params:
2022-12-12 09:35:23 +00:00
freeze: True
layer: "penultimate"
2022-11-08 08:14:45 +00:00
data:
#target: main.DataModuleFromConfig
2022-11-08 08:14:45 +00:00
params:
batch_size: 128
# num_workwers should be 2 * batch_size, and the total num less than 1024
# e.g. if use 8 devices, no more than 128
num_workers: 128
2022-11-08 08:14:45 +00:00
train:
target: ldm.data.base.Txt2ImgIterableBaseDataset
2022-11-08 08:14:45 +00:00
params:
file_path: # YOUR DATAPATH
world_size: 1
rank: 0
2022-11-08 08:14:45 +00:00
lightning:
trainer:
accelerator: 'gpu'
devices: 8
2022-11-08 08:14:45 +00:00
log_gpu_memory: all
max_epochs: 2
precision: 16
auto_select_gpus: False
strategy:
#target: strategies.DDPStrategy
2022-11-08 08:14:45 +00:00
params:
find_unused_parameters: False
log_every_n_steps: 2
# max_steps: 6o
logger: True
default_root_dir: "/tmp/diff_log/"
# profiler: pytorch
logger_config:
wandb:
#target: loggers.WandbLogger
2022-11-08 08:14:45 +00:00
params:
name: nowname
2022-12-12 09:35:23 +00:00
save_dir: "/data2/tmp/diff_log/"
2022-11-08 08:14:45 +00:00
offline: opt.debug
2022-12-12 09:35:23 +00:00
id: nowname