Skip to content
This repository was archived by the owner on Dec 14, 2023. It is now read-only.

Commit 8ed13d7

Browse files
kabachuhakabachuha
committed
Create convert_diffusers_to_original_ms_text_to_video.py
1 parent 51c809b commit 8ed13d7

File tree

1 file changed

+386
-0
lines changed

1 file changed

+386
-0
lines changed

‎utils/convert_diffusers_to_original_ms_text_to_video.py‎

Lines changed: 386 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,386 @@
1+
# Script for converting a HF Diffusers saved pipeline to a Stable Diffusion checkpoint.
2+
# *Only* converts the UNet, and Text Encoder.
3+
# Does not convert optimizer state or any other thing.
4+
5+
import argparse
6+
import os.path as osp
7+
import re
8+
9+
import torch
10+
from safetensors.torch import load_file, save_file
11+
12+
# =================#
13+
# UNet Conversion #
14+
# =================#
15+
16+
print ('Initializing the conversion map')
17+
18+
unet_conversion_map = [
19+
# (ModelScope, HF Diffusers)
20+
21+
# from Vanilla ModelScope/StableDiffusion
22+
("time_embed.0.weight", "time_embedding.linear_1.weight"),
23+
("time_embed.0.bias", "time_embedding.linear_1.bias"),
24+
("time_embed.2.weight", "time_embedding.linear_2.weight"),
25+
("time_embed.2.bias", "time_embedding.linear_2.bias"),
26+
27+
# from Modelscope only
28+
("label_emb.0.0.weight", "class_embedding.linear_1.weight"),
29+
("label_emb.0.0.bias", "class_embedding.linear_1.bias"),
30+
("label_emb.0.2.weight", "class_embedding.linear_2.weight"),
31+
("label_emb.0.2.bias", "class_embedding.linear_2.bias"),
32+
33+
# from Vanilla ModelScope/StableDiffusion
34+
("input_blocks.0.0.weight", "conv_in.weight"),
35+
("input_blocks.0.0.bias", "conv_in.bias"),
36+
37+
# from Modelscope only
38+
("input_blocks.0.1.weight", "transformer_in.weight"),
39+
("input_blocks.0.1.bias", "transformer_in.bias"),
40+
41+
# from Vanilla ModelScope/StableDiffusion
42+
("out.0.weight", "conv_norm_out.weight"),
43+
("out.0.bias", "conv_norm_out.bias"),
44+
("out.2.weight", "conv_out.weight"),
45+
("out.2.bias", "conv_out.bias"),
46+
]
47+
48+
unet_conversion_map_resnet = [
49+
# (ModelScope, HF Diffusers)
50+
("in_layers.0", "norm1"),
51+
("in_layers.2", "conv1"),
52+
("out_layers.0", "norm2"),
53+
("out_layers.3", "conv2"),
54+
("emb_layers.1", "time_emb_proj"),
55+
("skip_connection", "conv_shortcut"),
56+
]
57+
58+
unet_conversion_map_layer = []
59+
60+
# Reference for the default settings
61+
62+
# "model_cfg": {
63+
# "unet_in_dim": 4,
64+
# "unet_dim": 320,
65+
# "unet_y_dim": 768,
66+
# "unet_context_dim": 1024,
67+
# "unet_out_dim": 4,
68+
# "unet_dim_mult": [1, 2, 4, 4],
69+
# "unet_num_heads": 8,
70+
# "unet_head_dim": 64,
71+
# "unet_res_blocks": 2,
72+
# "unet_attn_scales": [1, 0.5, 0.25],
73+
# "unet_dropout": 0.1,
74+
# "temporal_attention": "True",
75+
# "num_timesteps": 1000,
76+
# "mean_type": "eps",
77+
# "var_type": "fixed_small",
78+
# "loss_type": "mse"
79+
# }
80+
81+
# hardcoded number of downblocks and resnets/attentions...
82+
# would need smarter logic for other networks.
83+
for i in range(4):
84+
# loop over downblocks/upblocks
85+
86+
for j in range(2):
87+
# loop over resnets/attentions for downblocks
88+
hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}."
89+
sd_down_res_prefix = f"input_blocks.{3*i + j + 1}.0."
90+
unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix))
91+
92+
if i < 3:
93+
# no attention layers in down_blocks.3
94+
hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}."
95+
sd_down_atn_prefix = f"input_blocks.{3*i + j + 1}.1."
96+
unet_conversion_map_layer.append((sd_down_atn_prefix, hf_down_atn_prefix))
97+
98+
for j in range(3):
99+
# loop over resnets/attentions for upblocks
100+
hf_up_res_prefix = f"up_blocks.{i}.resnets.{j}."
101+
sd_up_res_prefix = f"output_blocks.{3*i + j}.0."
102+
unet_conversion_map_layer.append((sd_up_res_prefix, hf_up_res_prefix))
103+
104+
if i > 0:
105+
# no attention layers in up_blocks.0
106+
hf_up_atn_prefix = f"up_blocks.{i}.attentions.{j}."
107+
sd_up_atn_prefix = f"output_blocks.{3*i + j}.1."
108+
unet_conversion_map_layer.append((sd_up_atn_prefix, hf_up_atn_prefix))
109+
110+
if i < 3:
111+
# no downsample in down_blocks.3
112+
hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv."
113+
sd_downsample_prefix = f"input_blocks.{3*(i+1)}.0.op."
114+
unet_conversion_map_layer.append((sd_downsample_prefix, hf_downsample_prefix))
115+
116+
# no upsample in up_blocks.3
117+
hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
118+
sd_upsample_prefix = f"output_blocks.{3*i + 2}.{1 if i == 0 else 2}."
119+
unet_conversion_map_layer.append((sd_upsample_prefix, hf_upsample_prefix))
120+
121+
122+
# Handle the middle block
123+
124+
hf_mid_atn_prefix = "mid_block.attentions.0."
125+
sd_mid_atn_prefix = "middle_block.1."
126+
unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix))
127+
128+
for j in range(2):
129+
hf_mid_res_prefix = f"mid_block.resnets.{j}."
130+
sd_mid_res_prefix = f"middle_block.{2*j}."
131+
unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix))
132+
133+
134+
135+
def convert_unet_state_dict(unet_state_dict):
136+
print ('Converting the UNET')
137+
# buyer beware: this is a *brittle* function,
138+
# and correct output requires that all of these pieces interact in
139+
# the exact order in which I have arranged them.
140+
mapping = {k: k for k in unet_state_dict.keys()}
141+
142+
for sd_name, hf_name in unet_conversion_map:
143+
mapping[hf_name] = sd_name
144+
for k, v in mapping.items():
145+
if "resnets" in k:
146+
for sd_part, hf_part in unet_conversion_map_resnet:
147+
v = v.replace(hf_part, sd_part)
148+
mapping[k] = v
149+
for k, v in mapping.items():
150+
for sd_part, hf_part in unet_conversion_map_layer:
151+
v = v.replace(hf_part, sd_part)
152+
mapping[k] = v
153+
new_state_dict = {v: unet_state_dict[k] for k, v in mapping.items()}
154+
return new_state_dict
155+
156+
# TODO: VAE conversion. We doesn't train it in the most cases, but may be handy for the future --kabachuha
157+
158+
# =========================#
159+
# Text Encoder Conversion #
160+
# =========================#
161+
162+
# IT IS THE SAME CLIP ENCODER, SO JUST COPYPASTING IT --kabachuha
163+
164+
# =========================#
165+
# Text Encoder Conversion #
166+
# =========================#
167+
168+
169+
textenc_conversion_lst = [
170+
# (stable-diffusion, HF Diffusers)
171+
("resblocks.", "text_model.encoder.layers."),
172+
("ln_1", "layer_norm1"),
173+
("ln_2", "layer_norm2"),
174+
(".c_fc.", ".fc1."),
175+
(".c_proj.", ".fc2."),
176+
(".attn", ".self_attn"),
177+
("ln_final.", "transformer.text_model.final_layer_norm."),
178+
("token_embedding.weight", "transformer.text_model.embeddings.token_embedding.weight"),
179+
("positional_embedding", "transformer.text_model.embeddings.position_embedding.weight"),
180+
]
181+
protected = {re.escape(x[1]): x[0] for x in textenc_conversion_lst}
182+
textenc_pattern = re.compile("|".join(protected.keys()))
183+
184+
# Ordering is from https://github.com/pytorch/pytorch/blob/master/test/cpp/api/modules.cpp
185+
code2idx = {"q": 0, "k": 1, "v": 2}
186+
187+
188+
def convert_text_enc_state_dict_v20(text_enc_dict):
189+
print ('Converting the text encoder')
190+
new_state_dict = {}
191+
capture_qkv_weight = {}
192+
capture_qkv_bias = {}
193+
for k, v in text_enc_dict.items():
194+
if (
195+
k.endswith(".self_attn.q_proj.weight")
196+
or k.endswith(".self_attn.k_proj.weight")
197+
or k.endswith(".self_attn.v_proj.weight")
198+
):
199+
k_pre = k[: -len(".q_proj.weight")]
200+
k_code = k[-len("q_proj.weight")]
201+
if k_pre not in capture_qkv_weight:
202+
capture_qkv_weight[k_pre] = [None, None, None]
203+
capture_qkv_weight[k_pre][code2idx[k_code]] = v
204+
continue
205+
206+
if (
207+
k.endswith(".self_attn.q_proj.bias")
208+
or k.endswith(".self_attn.k_proj.bias")
209+
or k.endswith(".self_attn.v_proj.bias")
210+
):
211+
k_pre = k[: -len(".q_proj.bias")]
212+
k_code = k[-len("q_proj.bias")]
213+
if k_pre not in capture_qkv_bias:
214+
capture_qkv_bias[k_pre] = [None, None, None]
215+
capture_qkv_bias[k_pre][code2idx[k_code]] = v
216+
continue
217+
218+
relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k)
219+
new_state_dict[relabelled_key] = v
220+
221+
for k_pre, tensors in capture_qkv_weight.items():
222+
if None in tensors:
223+
raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")
224+
relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)
225+
new_state_dict[relabelled_key + ".in_proj_weight"] = torch.cat(tensors)
226+
227+
for k_pre, tensors in capture_qkv_bias.items():
228+
if None in tensors:
229+
raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")
230+
relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)
231+
new_state_dict[relabelled_key + ".in_proj_bias"] = torch.cat(tensors)
232+
233+
return new_state_dict
234+
235+
236+
def convert_text_enc_state_dict(text_enc_dict):
237+
return text_enc_dict
238+
239+
textenc_conversion_lst = [
240+
# (stable-diffusion, HF Diffusers)
241+
("resblocks.", "text_model.encoder.layers."),
242+
("ln_1", "layer_norm1"),
243+
("ln_2", "layer_norm2"),
244+
(".c_fc.", ".fc1."),
245+
(".c_proj.", ".fc2."),
246+
(".attn", ".self_attn"),
247+
("ln_final.", "transformer.text_model.final_layer_norm."),
248+
("token_embedding.weight", "transformer.text_model.embeddings.token_embedding.weight"),
249+
("positional_embedding", "transformer.text_model.embeddings.position_embedding.weight"),
250+
]
251+
protected = {re.escape(x[1]): x[0] for x in textenc_conversion_lst}
252+
textenc_pattern = re.compile("|".join(protected.keys()))
253+
254+
# Ordering is from https://github.com/pytorch/pytorch/blob/master/test/cpp/api/modules.cpp
255+
code2idx = {"q": 0, "k": 1, "v": 2}
256+
257+
258+
def convert_text_enc_state_dict_v20(text_enc_dict):
259+
new_state_dict = {}
260+
capture_qkv_weight = {}
261+
capture_qkv_bias = {}
262+
for k, v in text_enc_dict.items():
263+
if (
264+
k.endswith(".self_attn.q_proj.weight")
265+
or k.endswith(".self_attn.k_proj.weight")
266+
or k.endswith(".self_attn.v_proj.weight")
267+
):
268+
k_pre = k[: -len(".q_proj.weight")]
269+
k_code = k[-len("q_proj.weight")]
270+
if k_pre not in capture_qkv_weight:
271+
capture_qkv_weight[k_pre] = [None, None, None]
272+
capture_qkv_weight[k_pre][code2idx[k_code]] = v
273+
continue
274+
275+
if (
276+
k.endswith(".self_attn.q_proj.bias")
277+
or k.endswith(".self_attn.k_proj.bias")
278+
or k.endswith(".self_attn.v_proj.bias")
279+
):
280+
k_pre = k[: -len(".q_proj.bias")]
281+
k_code = k[-len("q_proj.bias")]
282+
if k_pre not in capture_qkv_bias:
283+
capture_qkv_bias[k_pre] = [None, None, None]
284+
capture_qkv_bias[k_pre][code2idx[k_code]] = v
285+
continue
286+
287+
relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k)
288+
new_state_dict[relabelled_key] = v
289+
290+
for k_pre, tensors in capture_qkv_weight.items():
291+
if None in tensors:
292+
raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")
293+
relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)
294+
new_state_dict[relabelled_key + ".in_proj_weight"] = torch.cat(tensors)
295+
296+
for k_pre, tensors in capture_qkv_bias.items():
297+
if None in tensors:
298+
raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")
299+
relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)
300+
new_state_dict[relabelled_key + ".in_proj_bias"] = torch.cat(tensors)
301+
302+
return new_state_dict
303+
304+
305+
def convert_text_enc_state_dict(text_enc_dict):
306+
return text_enc_dict
307+
308+
if __name__ == "__main__":
309+
parser = argparse.ArgumentParser()
310+
311+
parser.add_argument("--model_path", default=None, type=str, required=True, help="Path to the model to convert.")
312+
parser.add_argument("--checkpoint_path", default=None, type=str, required=True, help="Path to the output model.")
313+
parser.add_argument("--half", action="store_true", help="Save weights in half precision.")
314+
parser.add_argument(
315+
"--use_safetensors", action="store_true", help="Save weights use safetensors, default is ckpt."
316+
)
317+
318+
args = parser.parse_args()
319+
320+
assert args.model_path is not None, "Must provide a model path!"
321+
322+
assert args.checkpoint_path is not None, "Must provide a checkpoint path!"
323+
324+
# Path for safetensors
325+
unet_path = osp.join(args.model_path, "unet", "diffusion_pytorch_model.safetensors")
326+
vae_path = osp.join(args.model_path, "vae", "diffusion_pytorch_model.safetensors")
327+
text_enc_path = osp.join(args.model_path, "text_encoder", "model.safetensors")
328+
329+
# Load models from safetensors if it exists, if it doesn't pytorch
330+
if osp.exists(unet_path):
331+
unet_state_dict = load_file(unet_path, device="cpu")
332+
else:
333+
unet_path = osp.join(args.model_path, "unet", "diffusion_pytorch_model.bin")
334+
unet_state_dict = torch.load(unet_path, map_location="cpu")
335+
336+
if osp.exists(vae_path):
337+
vae_state_dict = load_file(vae_path, device="cpu")
338+
else:
339+
vae_path = osp.join(args.model_path, "vae", "diffusion_pytorch_model.bin")
340+
vae_state_dict = torch.load(vae_path, map_location="cpu")
341+
342+
if osp.exists(text_enc_path):
343+
text_enc_dict = load_file(text_enc_path, device="cpu")
344+
else:
345+
text_enc_path = osp.join(args.model_path, "text_encoder", "pytorch_model.bin")
346+
text_enc_dict = torch.load(text_enc_path, map_location="cpu")
347+
348+
# Convert the UNet model
349+
unet_state_dict = convert_unet_state_dict(unet_state_dict)
350+
unet_state_dict = {"model.diffusion_model." + k: v for k, v in unet_state_dict.items()}
351+
352+
# Convert the VAE model
353+
# vae_state_dict = convert_vae_state_dict(vae_state_dict)
354+
# vae_state_dict = {"first_stage_model." + k: v for k, v in vae_state_dict.items()}
355+
356+
# Easiest way to identify v2.0 model seems to be that the text encoder (OpenCLIP) is deeper
357+
is_v20_model = "text_model.encoder.layers.22.layer_norm2.bias" in text_enc_dict
358+
359+
if is_v20_model:
360+
361+
# MODELSCOPE always uses the 2.X encoder, btw --kabachuha
362+
363+
# Need to add the tag 'transformer' in advance so we can knock it out from the final layer-norm
364+
text_enc_dict = {"transformer." + k: v for k, v in text_enc_dict.items()}
365+
text_enc_dict = convert_text_enc_state_dict_v20(text_enc_dict)
366+
text_enc_dict = {"cond_stage_model.model." + k: v for k, v in text_enc_dict.items()}
367+
else:
368+
text_enc_dict = convert_text_enc_state_dict(text_enc_dict)
369+
text_enc_dict = {"cond_stage_model.transformer." + k: v for k, v in text_enc_dict.items()}
370+
371+
# DON'T PUT TOGETHER FOR THE NEW CHECKPOINT AS MODELSCOPE USES THEM IN THE SPLITTED FORM --kabachuha
372+
373+
374+
375+
#state_dict = {**unet_state_dict, **vae_state_dict, **text_enc_dict}
376+
print ('Saving UNET')
377+
state_dict = {**unet_state_dict}
378+
379+
if args.half:
380+
state_dict = {k: v.half() for k, v in state_dict.items()}
381+
382+
if args.use_safetensors:
383+
save_file(state_dict, args.checkpoint_path)
384+
else:
385+
state_dict = {"state_dict": state_dict}
386+
torch.save(state_dict, args.checkpoint_path)

0 commit comments

Comments
 (0)