Upload 2 files
Browse files- facescape.yaml +67 -0
- thuman.yaml +67 -0
facescape.yaml
ADDED
|
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
model:
|
| 2 |
+
base_learning_rate: 5e-5
|
| 3 |
+
target: ldm.models.diffusion.morphable_diffusion.SyncMultiviewDiffusion
|
| 4 |
+
params:
|
| 5 |
+
view_num: 16
|
| 6 |
+
image_size: 256
|
| 7 |
+
cfg_scale: 2.0
|
| 8 |
+
output_num: 8
|
| 9 |
+
batch_view_num: 4
|
| 10 |
+
finetune_unet: True
|
| 11 |
+
drop_conditions: false
|
| 12 |
+
projection: 'perspective'
|
| 13 |
+
use_spatial_volume: False
|
| 14 |
+
clip_image_encoder_path: ./ckpt/ViT-L-14.pt
|
| 15 |
+
target_elevation: 0
|
| 16 |
+
|
| 17 |
+
scheduler_config: # 10000 warmup steps
|
| 18 |
+
target: ldm.lr_scheduler.LambdaLinearScheduler
|
| 19 |
+
params:
|
| 20 |
+
warm_up_steps: [ 100 ]
|
| 21 |
+
cycle_lengths: [ 100000 ]
|
| 22 |
+
f_start: [ 0.02 ]
|
| 23 |
+
f_max: [ 1.0 ]
|
| 24 |
+
f_min: [ 1.0 ]
|
| 25 |
+
|
| 26 |
+
unet_config:
|
| 27 |
+
target: ldm.models.diffusion.attention.DepthWiseAttention
|
| 28 |
+
params:
|
| 29 |
+
volume_dims: [64, 128, 256, 512]
|
| 30 |
+
image_size: 32
|
| 31 |
+
in_channels: 8
|
| 32 |
+
out_channels: 4
|
| 33 |
+
model_channels: 320
|
| 34 |
+
attention_resolutions: [ 4, 2, 1 ]
|
| 35 |
+
num_res_blocks: 2
|
| 36 |
+
channel_mult: [ 1, 2, 4, 4 ]
|
| 37 |
+
num_heads: 8
|
| 38 |
+
use_spatial_transformer: True
|
| 39 |
+
transformer_depth: 1
|
| 40 |
+
context_dim: 768
|
| 41 |
+
use_checkpoint: True
|
| 42 |
+
legacy: False
|
| 43 |
+
|
| 44 |
+
data:
|
| 45 |
+
target: ldm.data.facescape.FaceScapeDataset
|
| 46 |
+
params:
|
| 47 |
+
data_dir: /cluster/scratch/xiychen/data/facescape_color_calibrated
|
| 48 |
+
mesh_topology: 'flame'
|
| 49 |
+
shuffled_expression: True
|
| 50 |
+
batch_size: 70 # batch size for a single gpu
|
| 51 |
+
num_workers: 1
|
| 52 |
+
|
| 53 |
+
lightning:
|
| 54 |
+
modelcheckpoint:
|
| 55 |
+
params:
|
| 56 |
+
every_n_train_steps: 2000
|
| 57 |
+
callbacks:
|
| 58 |
+
{}
|
| 59 |
+
|
| 60 |
+
trainer:
|
| 61 |
+
benchmark: True
|
| 62 |
+
max_steps: 6000
|
| 63 |
+
val_check_interval: 250 # we will run validation every 1k steps, the validation will output images to <log_dir>/<images>/val
|
| 64 |
+
num_sanity_val_steps: 0
|
| 65 |
+
precision: 32
|
| 66 |
+
check_val_every_n_epoch: null
|
| 67 |
+
accumulate_grad_batches: 1
|
thuman.yaml
ADDED
|
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
model:
|
| 2 |
+
base_learning_rate: 5e-5
|
| 3 |
+
target: ldm.models.diffusion.morphable_diffusion.SyncMultiviewDiffusion
|
| 4 |
+
params:
|
| 5 |
+
view_num: 16
|
| 6 |
+
image_size: 256
|
| 7 |
+
cfg_scale: 2.0
|
| 8 |
+
output_num: 8
|
| 9 |
+
batch_view_num: 4
|
| 10 |
+
finetune_unet: True
|
| 11 |
+
drop_conditions: false
|
| 12 |
+
projection: 'orthographic'
|
| 13 |
+
use_spatial_volume: False
|
| 14 |
+
clip_image_encoder_path: ./ckpt/ViT-L-14.pt
|
| 15 |
+
target_elevation: 0
|
| 16 |
+
|
| 17 |
+
scheduler_config: # 10000 warmup steps
|
| 18 |
+
target: ldm.lr_scheduler.LambdaLinearScheduler
|
| 19 |
+
params:
|
| 20 |
+
warm_up_steps: [ 100 ]
|
| 21 |
+
cycle_lengths: [ 100000 ]
|
| 22 |
+
f_start: [ 0.02 ]
|
| 23 |
+
f_max: [ 1.0 ]
|
| 24 |
+
f_min: [ 1.0 ]
|
| 25 |
+
|
| 26 |
+
unet_config:
|
| 27 |
+
target: ldm.models.diffusion.attention.DepthWiseAttention
|
| 28 |
+
params:
|
| 29 |
+
volume_dims: [64, 128, 256, 512]
|
| 30 |
+
image_size: 32
|
| 31 |
+
in_channels: 8
|
| 32 |
+
out_channels: 4
|
| 33 |
+
model_channels: 320
|
| 34 |
+
attention_resolutions: [ 4, 2, 1 ]
|
| 35 |
+
num_res_blocks: 2
|
| 36 |
+
channel_mult: [ 1, 2, 4, 4 ]
|
| 37 |
+
num_heads: 8
|
| 38 |
+
use_spatial_transformer: True
|
| 39 |
+
transformer_depth: 1
|
| 40 |
+
context_dim: 768
|
| 41 |
+
use_checkpoint: True
|
| 42 |
+
legacy: False
|
| 43 |
+
|
| 44 |
+
data:
|
| 45 |
+
target: ldm.data.thuman.THumanDataset
|
| 46 |
+
params:
|
| 47 |
+
data_dir: /cluster/scratch/xiychen/data/thuman_2.1_preprocessed
|
| 48 |
+
smplx_dir: /cluster/scratch/xiychen/data/thuman_smplx # a list of uids
|
| 49 |
+
batch_size: 70 # batch size for a single gpu
|
| 50 |
+
num_workers: 1
|
| 51 |
+
|
| 52 |
+
lightning:
|
| 53 |
+
modelcheckpoint:
|
| 54 |
+
params:
|
| 55 |
+
every_n_train_steps: 2000 # we will save models every 1k steps
|
| 56 |
+
callbacks:
|
| 57 |
+
{}
|
| 58 |
+
|
| 59 |
+
trainer:
|
| 60 |
+
benchmark: True
|
| 61 |
+
max_steps: 6000
|
| 62 |
+
val_check_interval: 250 # we will run validation every 1k steps, the validation will output images to <log_dir>/<images>/val
|
| 63 |
+
num_sanity_val_steps: 0
|
| 64 |
+
precision: 32
|
| 65 |
+
check_val_every_n_epoch: null
|
| 66 |
+
accumulate_grad_batches: 1
|
| 67 |
+
|