Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

support loading stable zero123 in diffuser (fp16 support) #395

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

support loading stable zero123 in diffuser (fp16 support) #395

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
146 changes: 146 additions & 0 deletions configs/stable-zero123-diffusers.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,146 @@
name: "zero123-sai"
tag: "${data.random_camera.height}_${rmspace:${basename:${data.image_path}},_}"
exp_root_dir: "outputs"
seed: 0

data_type: "single-image-datamodule"
data: # threestudio/data/image.py -> SingleImageDataModuleConfig
image_path: ./load/images/hamburger_rgba.png
height: [128, 256, 512]
width: [128, 256, 512]
resolution_milestones: [200, 300]
default_elevation_deg: 5.0
default_azimuth_deg: 0.0
default_camera_distance: 3.8
default_fovy_deg: 20.0
requires_depth: ${cmaxgt0orcmaxgt0:${system.loss.lambda_depth},${system.loss.lambda_depth_rel}}
requires_normal: ${cmaxgt0:${system.loss.lambda_normal}}
random_camera: # threestudio/data/uncond.py -> RandomCameraDataModuleConfig
height: [64, 128, 256]
width: [64, 128, 256]
batch_size: [12, 8, 4]
resolution_milestones: [200, 300]
eval_height: 512
eval_width: 512
eval_batch_size: 1
elevation_range: [-10, 80]
azimuth_range: [-180, 180]
camera_distance_range: [3.8, 3.8]
fovy_range: [20.0, 20.0] # Zero123 has fixed fovy
progressive_until: 0
camera_perturb: 0.0
center_perturb: 0.0
up_perturb: 0.0
light_position_perturb: 1.0
light_distance_range: [7.5, 10.0]
eval_elevation_deg: ${data.default_elevation_deg}
eval_camera_distance: ${data.default_camera_distance}
eval_fovy_deg: ${data.default_fovy_deg}
light_sample_strategy: "dreamfusion"
batch_uniform_azimuth: False
n_val_views: 30
n_test_views: 120

system_type: "zero123-system"
system:
geometry_type: "implicit-volume"
geometry:
radius: 2.0
normal_type: "analytic"

# use Magic3D density initialization instead
density_bias: "blob_magic3d"
density_activation: softplus
density_blob_scale: 10.
density_blob_std: 0.5

# coarse to fine hash grid encoding
# to ensure smooth analytic normals
pos_encoding_config:
otype: HashGrid
n_levels: 16
n_features_per_level: 2
log2_hashmap_size: 19
base_resolution: 16
per_level_scale: 1.447269237440378 # max resolution 4096
mlp_network_config:
otype: "VanillaMLP"
activation: "ReLU"
output_activation: "none"
n_neurons: 64
n_hidden_layers: 2

material_type: "diffuse-with-point-light-material"
material:
ambient_only_steps: 100000
textureless_prob: 0.05
albedo_activation: sigmoid

background_type: "solid-color-background" # unused

renderer_type: "nerf-volume-renderer"
renderer:
radius: ${system.geometry.radius}
num_samples_per_ray: 512
return_comp_normal: ${cmaxgt0:${system.loss.lambda_normal_smooth}}
return_normal_perturb: ${cmaxgt0:${system.loss.lambda_3d_normal_smooth}}

prompt_processor_type: "dummy-prompt-processor" # Zero123 doesn't use prompts
prompt_processor:
pretrained_model_name_or_path: ""
prompt: ""

guidance_type: "zero123-unified-guidance"
guidance:
pretrained_model_name_or_path: "ashawkey/stable-zero123-diffusers"

cond_image_path: ${data.image_path}
cond_elevation_deg: ${data.default_elevation_deg}
cond_azimuth_deg: ${data.default_azimuth_deg}
cond_camera_distance: ${data.default_camera_distance}
guidance_scale: 3.0
min_step_percent: [50, 0.7, 0.3, 200] # (start_iter, start_val, end_val, end_iter)
max_step_percent: [50, 0.98, 0.8, 200]

freq:
ref_only_steps: 0
guidance_eval: 0

loggers:
wandb:
enable: false
project: "threestudio"
name: None

loss:
lambda_sds: 0.1
lambda_rgb: [100, 500., 1000., 400]
lambda_mask: 50.
lambda_depth: 0. # 0.05
lambda_depth_rel: 0. # [0, 0, 0.05, 100]
lambda_normal: 0. # [0, 0, 0.05, 100]
lambda_normal_smooth: [100, 7.0, 5.0, 150, 10.0, 200]
lambda_3d_normal_smooth: [100, 7.0, 5.0, 150, 10.0, 200]
lambda_orient: 1.0
lambda_sparsity: 0.5 # should be tweaked for every model
lambda_opaque: 0.5

optimizer:
name: Adam
args:
lr: 0.01
betas: [0.9, 0.99]
eps: 1.e-8

trainer:
max_steps: 600
log_every_n_steps: 1
num_sanity_val_steps: 0
val_check_interval: 100
enable_progress_bar: true
precision: 32

checkpoint:
save_last: true # save at each validation time
save_top_k: -1
every_n_train_steps: 100 # ${trainer.max_steps}
61 changes: 45 additions & 16 deletions threestudio/models/guidance/zero123_unified_guidance.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -82,17 +82,29 @@ class NonTrainableModules:
torch.float16 if self.cfg.half_precision_weights else torch.float32
)

threestudio.info(f"Loading Zero123 ...")
self.use_stable_zero123 = 'stable' in self.cfg.pretrained_model_name_or_path
if self.use_stable_zero123:
# stable-zero123 has a different camera embedding
threestudio.info(f"Loaded Stable Zero123!")
pipe_kwargs = {
"safety_checker": None,
"requires_safety_checker": False,
"variant": "fp16" if self.cfg.half_precision_weights else None,
"torch_dtype": self.weights_dtype,
}
else:
pipe_kwargs = {
"safety_checker": None,
"requires_safety_checker": False,
"trust_remote_code": True,
# "variant": "fp16" if self.cfg.half_precision_weights else None,
"torch_dtype": self.weights_dtype,
}
threestudio.info(f"Loading Zero123 ...")

# need to make sure the pipeline file is in path
sys.path.append("extern/")

pipe_kwargs = {
"safety_checker": None,
"requires_safety_checker": False,
"variant": "fp16" if self.cfg.half_precision_weights else None,
"torch_dtype": self.weights_dtype,
}
pipe = Zero123Pipeline.from_pretrained(
self.cfg.pretrained_model_name_or_path,
**pipe_kwargs,
Expand Down Expand Up @@ -262,15 +274,31 @@ def get_image_camera_embeddings(
camera_distances: Float[Tensor, "B"],
) -> Float[Tensor, "B 1 D"]:
batch_size = elevation_deg.shape[0]
camera_embeddings: Float[Tensor, "B 1 4"] = torch.stack(
[
torch.deg2rad(self.cfg.cond_elevation_deg - elevation_deg),
torch.sin(torch.deg2rad(azimuth_deg - self.cfg.cond_azimuth_deg)),
torch.cos(torch.deg2rad(azimuth_deg - self.cfg.cond_azimuth_deg)),
camera_distances - self.cfg.cond_camera_distance,
],
dim=-1,
)[:, None, :]
if self.use_stable_zero123:
camera_embeddings: Float[Tensor, "B 1 4"] = torch.stack(
[
torch.deg2rad(
(90 - elevation_deg) - (90 - self.cfg.cond_elevation_deg)
), # Zero123 polar is 90-elevation
torch.sin(torch.deg2rad(azimuth_deg - self.cfg.cond_azimuth_deg)),
torch.cos(torch.deg2rad(azimuth_deg - self.cfg.cond_azimuth_deg)),
torch.deg2rad(
90 - torch.full_like(elevation_deg, self.cfg.cond_elevation_deg)
),
],
dim=-1,
)[:, None, :].to(self.device)
else:
# original zero123 camera embedding
camera_embeddings: Float[Tensor, "B 1 4"] = torch.stack(
[
torch.deg2rad(self.cfg.cond_elevation_deg - elevation_deg),
torch.sin(torch.deg2rad(azimuth_deg - self.cfg.cond_azimuth_deg)),
torch.cos(torch.deg2rad(azimuth_deg - self.cfg.cond_azimuth_deg)),
camera_distances - self.cfg.cond_camera_distance,
],
dim=-1,
)[:, None, :]

image_camera_embeddings = self.pipe.clip_camera_projection(
torch.cat(
Expand Down Expand Up @@ -638,6 +666,7 @@ def forward(
loss_sd = 0.5 * F.mse_loss(latents, target, reduction="sum") / batch_size

guidance_out = {
"loss_sds": loss_sd,
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Magic123 uses loss_sd so it will raise an error. How about changing the config file from loss_sds to loss_sd?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I originally keep/duplicate both the key of both loss_sd and loss_sds, To make it compatible with different models. Yes, i will consider change the cfg. Thanks

"loss_sd": loss_sd,
"grad_norm": grad.norm(),
"timesteps": t,
Expand Down
Loading