clh / scripts /convert_weight.py
LiuhanChen's picture
Add files using upload-large-folder tool
f2b657e verified
import torch
import sys
import re
import safetensors
sys.path.append(".")
from causalvideovae.model import CausalVAEModel
origin_ckpt_path = "/remote-home1/clh/models/sd2_1/vae-ft-mse-840000-ema-pruned.ckpt"
config_path = "/remote-home1/clh/models/sd2_1/config.json"
output_path = "/remote-home1/clh/models/norm3d_vae_pretrained_weight"
init_method = "tail"
model = CausalVAEModel.from_config(config_path)
if origin_ckpt_path.endswith('ckpt'):
ckpt = torch.load(origin_ckpt_path, map_location="cpu")['state_dict']
elif origin_ckpt_path.endswith('safetensors'):
ckpt = {}
with safetensors.safe_open(origin_ckpt_path, framework="pt") as file:
for k in file.keys():
ckpt[k] = file.get_tensor(k)
print("key", k)
for name, module in model.named_modules():
if "loss" in name:
continue
if isinstance(module, torch.nn.Conv3d):
in_channels = module.in_channels
out_channels = module.out_channels
kernel_size = module.kernel_size
old_name = re.sub(".conv$", "", name)
if old_name + ".weight" not in ckpt:
print(old_name + ".weight", "not found")
continue
if init_method == "tail":
shape_2d = ckpt[old_name + ".weight"].shape
new_weight = torch.zeros(*shape_2d)
new_weight = new_weight.unsqueeze(2).repeat(1, 1, kernel_size[0], 1, 1)
middle_idx = kernel_size[0] // 2
new_weight[:, :, -1, :, :] = ckpt[old_name + ".weight"]
new_bias = ckpt[old_name + ".bias"]
elif init_method == "avg":
new_weight = ckpt[old_name + ".weight"].unsqueeze(2)
new_weight = new_weight.repeat(1, 1, kernel_size[0], 1, 1) / kernel_size[0]
new_bias = ckpt[old_name + ".bias"]
assert new_weight.shape == module.weight.shape
module.weight.data = new_weight.cpu().float()
module.bias.data = new_bias.cpu().float()
elif isinstance(module, torch.nn.GroupNorm):
old_name = name
if old_name + ".weight" not in ckpt:
print(old_name + ".weight", "not found")
continue
new_weight = ckpt[old_name + ".weight"]
new_bias = ckpt[old_name + ".bias"]
module.weight.data = new_weight.cpu().float()
module.bias.data = new_bias.cpu().float()
elif isinstance(module, torch.nn.Conv2d):
in_channels = module.in_channels
out_channels = module.out_channels
kernel_size = module.kernel_size
old_name = name
if old_name + ".weight" not in ckpt:
print(old_name + ".weight", "not found")
continue
new_weight = ckpt[old_name + ".weight"]
new_bias = ckpt[old_name + ".bias"]
assert new_weight.shape == module.weight.shape
module.weight.data = new_weight.cpu().float()
module.bias.data = new_bias.cpu().float()
model.save_pretrained(output_path)