I found with original training workflow, the loss is not decling, I am not sure this is because I am using a subset of the training set.
# File modified by authors of InstructDiffusion from original (https://github.com/CompVis/stable-diffusion).
# See more details in LICENSE.
model:
base_learning_rate: 1.0e-04
weight_decay: 0.01
target: ldm.models.diffusion.ddpm_edit.LatentDiffusion
params:
fp16: True
deepspeed: 'deepspeed_1'
ckpt_path: stable_diffusion/models/ldm/stable-diffusion-v1/v1-5-pruned-emaonly-adaption-task.ckpt
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: edited
cond_stage_key: edit
image_size: 32
channels: 4
cond_stage_trainable: false # Note: different from the one we trained before
conditioning_key: hybrid
monitor: val/loss_simple_ema
scale_factor: 0.18215
scheduler_config: # 10000 warmup steps
target: ldm.lr_scheduler.LambdaLinearScheduler
params:
warm_up_steps: [ 0 ]
cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
f_start: [ 1.e-6 ]
f_max: [ 1. ]
f_min: [ 1. ]
unet_config:
target: ldm.modules.diffusionmodules.openaimodel.UNetModel
params:
image_size: 32 # unused
in_channels: 8
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
force_type_convert: True
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
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.FrozenCLIPEmbedder
data:
target: main.DataModuleFromConfig
params:
batch_size: 2
num_workers: 4
train:
- ds1:
target: dataset.editing.edit_zip_dataset.GIERDataset
params:
path: data/GIER_editing_data/
split: train
min_resize_res: 256
max_resize_res: 256
crop_res: 256
flip_prob: 0.0
zip_start_index: 0
zip_end_index: 100
sample_weight: 2.0
validation:
target: dataset.pose.pose.COCODataset
params:
root: data/coco/
image_set: val2017
is_train: False
max_prompt_num: 5
min_prompt_num: 1
radius: 10
trainer:
initial_scale: 13
max_epochs: 200
save_freq: 20
accumulate_grad_batches: 1
clip_grad: 0.0
optimizer: adamw
I found with original training workflow, the loss is not decling, I am not sure this is because I am using a subset of the training set.