-
Notifications
You must be signed in to change notification settings - Fork 0
/
SDToONNX.py
210 lines (144 loc) · 6.94 KB
/
SDToONNX.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
import sys
import gc
sys.path.append('./stable-diffusion')
sys.path.append('./k-diffusion')
sys.path.append('./src')
import k_diffusion as K
import torch
import torch.onnx
from torch import nn
from torch import autocast
from onnxruntime.quantization import quantize_dynamic, QuantType, quantize, QuantizationMode
from contextlib import nullcontext
import onnx
import onnxruntime
import modelWrap
import CompVisSDModel
import denoisers
#
# Edit paths below in 'main' to point to models / output paths
#
## Note:
# .\stable-diffusion\ldm\modules\diffusionmodules\util.py
# will give an error when converting to ONNX due to being unable to handle 'CheckpointFunction'
# this call is not needed for inference, so just hard code the if statement to always be false:
#
#def checkpoint(func, inputs, params, flag):
# ...
# flag = False <---- add this
# if flag:
# args = tuple(inputs) + tuple(params)
# return CheckpointFunction.apply(func, len(inputs), *args)
# else:
# return func(*inputs)
#
# main executing function
def Main():
#fastest on GPU appears to be autocast = true, dtype = float32
#decoder for FP16 no autocast fails to load as a session for some reason...
torchDevice = torch.device('cuda:0') #'cpu'
modelPrefix = "sd-v1-4"
model_path = "E:/MLModels/stableDiffusion/sd-v1-4.ckpt"
CreateONNXModels(model_path, torch.float32, True, modelPrefix + "-fp32-cuda-auto", torchDevice)
CreateONNXModels(model_path, torch.float32, False, modelPrefix + "-fp32-cuda", torchDevice)
CreateONNXModels(model_path, torch.float16, True, modelPrefix + "-fp16-cuda-auto", torchDevice)
#cant process this one
#CreateONNXModels(model_path, torch.float16, False, modelPrefix + "-fp16-cuda", torchDevice)
def CreateONNXModels(model_path, dtype, autocast, modelName, device):
modelwrapper:modelWrap.ModelContext = CompVisSDModel.CompVisSDModel(dtype)
modelwrapper.model_path = model_path
modelwrapper.modelName = modelName
modelwrapper.ModelLoadSettings()
modelwrapper.LoadModel(device)
outFile = "E:/onnxOut/" + modelwrapper.modelName + "/model.onnx"
ConvertToONNXCuda(modelwrapper, outFile, dtype, autocast, device)
outFile = "E:/onnxOut/" + modelwrapper.modelName + "-decode/model.onnx"
ConvertDecodeToONNXCuda(modelwrapper, outFile, dtype, autocast, device)
modelwrapper.model = None
modelwrapper = None
gc.collect()
torch.cuda.empty_cache()
def CheckONNX(modelPath:str):
onnx_model = onnx.load(modelPath)
print('The model is:\n{}'.format(onnx_model))
# Check the model
try:
onnx.checker.check_model(onnx_model)
except onnx.checker.ValidationError as e:
print('The model is invalid: %s' % e)
else:
print('The model is valid!')
def ConvertToONNXCuda(modelWrapper:modelWrap.ModelWrap, outFilePath:str, dtype, autocast_enable:bool, device):
sigmaTensor = torch.FloatTensor([0.1]).to(device).to(dtype)
condTensor = modelWrapper.frozenClip.encode(['']).to(device).to(dtype)
uncondTensor = modelWrapper.frozenClip.encode(['sample stupid prompt hellooo']).to(device).to(dtype)
condscaleTensor = torch.FloatTensor([0.1]).to(device).to(dtype)
image_input = torch.rand((1, 4, 64, 64)).to(device).to(dtype) #1 512x512 image
#put this through the KDiff wrapper, it seems to work well
modelCtx:modelWrap.ModelContext = modelWrap.ModelContext()
modelCtx.modelWrap = modelWrapper
#modelCtx.extra_args = {'cond': c, 'uncond': uc, 'cond_scale': condScale}
modelCtx.kdiffModelWrap = denoisers.CFGDenoiser(modelWrapper.kdiffExternalModelWrap)
if str(device) == 'cpu':
prec_dev = 'cpu'
else:
prec_dev = 'cuda'
precision_scope = autocast if autocast_enable else nullcontext
with torch.no_grad():
with precision_scope(prec_dev):
torch.onnx.export(modelCtx.kdiffModelWrap, # model being run
(image_input, sigmaTensor, condTensor, uncondTensor, condscaleTensor), # model input (or a tuple for multiple inputs)
outFilePath, # where to save the model
export_params=True, # store the trained parameter weights inside the model file
opset_version=16, # the ONNX version to export the model to
do_constant_folding=True, # whether to execute constant folding for optimization
input_names = ['modelInput', 'sigma', 'uncond', 'cond', 'cond_scale'], # the model's input names
output_names = ['modelOutput'], # the model's output names
dynamic_axes={'modelInput' : {0 : 'batch_size'},
'uncond' : {0 : 'batch_size'},
'cond' : {0 : 'batch_size'},
'sigma' : {0 : 'batch_size'},
'modelOutput' : {0 : 'batch_size'}})
print(" ")
print('Model has been converted to ONNX')
#CheckONNX(outFilePath)
# this doesnt work yet
def OptimizeONNX(onnxModelInPath:str, onnxModelOutPath:str):
#self.model = onnx.load(self.model_path)
model_in = onnxModelInPath
model_out = onnxModelOutPath
quantize.quantize_dynamic(model_in, model_out)
#onnx.save(qm, model_out)
#hacky stuff to treat encode / decode as models
class decodeWrap(nn.Module):
def __init__(self, model):
super().__init__()
self.model = model
def forward(self, inputImageTensor):
return self.model.decode_first_stage(inputImageTensor)
def ConvertDecodeToONNXCuda(modelWrapper:modelWrap.ModelWrap, outFilePath:str, dtype, autocast_enable:bool, device):
#input_tensor = torch.rand((1, 4, 64, 64)).to(device).to(dtype) #1 512x512 image
input_tensor = torch.randn([1, 4, 64, 64], device=device).to(dtype) + 0.01 * 10
dw = decodeWrap(modelWrapper.model)
if str(device) == 'cpu':
prec_dev = 'cpu'
else:
prec_dev = 'cuda'
precision_scope = autocast if autocast_enable else nullcontext
with torch.no_grad():
with precision_scope(prec_dev):
torch.onnx.export(dw, # model being run
(input_tensor), # model input (or a tuple for multiple inputs)
outFilePath, # where to save the model
export_params=True, # store the trained parameter weights inside the model file
opset_version=16, # the ONNX version to export the model to
do_constant_folding=True, # whether to execute constant folding for optimization
input_names = ['modelInput'], # the model's input names
output_names = ['modelOutput'], # the model's output names
dynamic_axes={'modelInput' : {0 : 'batch_size'},
'modelOutput' : {0 : 'batch_size'}})
print(" ")
print('Decode has been converted to ONNX')
#CheckONNX(outFilePath)
return
Main()