damian0815
commited on
Commit
•
e25b4fd
1
Parent(s):
dc5a3b2
Upload clip-to-coreml.ipynb
Browse files- clip-to-coreml.ipynb +337 -0
clip-to-coreml.ipynb
ADDED
@@ -0,0 +1,337 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "markdown",
|
5 |
+
"id": "1092f43b",
|
6 |
+
"metadata": {},
|
7 |
+
"source": [
|
8 |
+
"# Convert CLIP models to CoreML"
|
9 |
+
]
|
10 |
+
},
|
11 |
+
{
|
12 |
+
"cell_type": "code",
|
13 |
+
"execution_count": null,
|
14 |
+
"id": "e5f63e7a",
|
15 |
+
"metadata": {},
|
16 |
+
"outputs": [],
|
17 |
+
"source": [
|
18 |
+
"!pip install torch transformers coremltools"
|
19 |
+
]
|
20 |
+
},
|
21 |
+
{
|
22 |
+
"cell_type": "code",
|
23 |
+
"execution_count": null,
|
24 |
+
"id": "a7f0ab67",
|
25 |
+
"metadata": {},
|
26 |
+
"outputs": [],
|
27 |
+
"source": [
|
28 |
+
"from transformers import CLIPProcessor, CLIPModel\n",
|
29 |
+
"\n",
|
30 |
+
"model_version = \"laion/CLIP-ViT-H-14-laion2B-s32B-b79K\"\n",
|
31 |
+
"\n",
|
32 |
+
"processor = CLIPProcessor.from_pretrained(model_version)"
|
33 |
+
]
|
34 |
+
},
|
35 |
+
{
|
36 |
+
"cell_type": "markdown",
|
37 |
+
"id": "4bd0aa05",
|
38 |
+
"metadata": {},
|
39 |
+
"source": [
|
40 |
+
"# Text model"
|
41 |
+
]
|
42 |
+
},
|
43 |
+
{
|
44 |
+
"cell_type": "code",
|
45 |
+
"execution_count": null,
|
46 |
+
"id": "19851197",
|
47 |
+
"metadata": {},
|
48 |
+
"outputs": [],
|
49 |
+
"source": [
|
50 |
+
"# wrapped CLIPModel so that forward() function returns get_text_features()\n",
|
51 |
+
"class WrappedCLIPModel_Text(CLIPModel): \n",
|
52 |
+
" def forward(self, *args, **kwargs):\n",
|
53 |
+
" return self.get_text_features(*args, **kwargs)\n",
|
54 |
+
"\n",
|
55 |
+
"model_pt_text = WrappedCLIPModel_Text.from_pretrained(model_version)\n",
|
56 |
+
"model_pt_text.eval()"
|
57 |
+
]
|
58 |
+
},
|
59 |
+
{
|
60 |
+
"cell_type": "code",
|
61 |
+
"execution_count": null,
|
62 |
+
"id": "c8b3a1ca",
|
63 |
+
"metadata": {},
|
64 |
+
"outputs": [],
|
65 |
+
"source": [
|
66 |
+
"import torch\n",
|
67 |
+
"\n",
|
68 |
+
"with torch.no_grad():\n",
|
69 |
+
" text = \"the \" + \" \".join([\"example text\"]*37) # 77 tokens\n",
|
70 |
+
" processed_text = processor(text=text, images=None, return_tensors=\"pt\", padding=True)\n",
|
71 |
+
" print(len(processed_text.input_ids[0]), processed_text.input_ids)\n",
|
72 |
+
" model_traced = torch.jit.trace(model_pt_text, processed_text.input_ids, strict=True)"
|
73 |
+
]
|
74 |
+
},
|
75 |
+
{
|
76 |
+
"cell_type": "code",
|
77 |
+
"execution_count": null,
|
78 |
+
"id": "5066eb03",
|
79 |
+
"metadata": {
|
80 |
+
"scrolled": true
|
81 |
+
},
|
82 |
+
"outputs": [],
|
83 |
+
"source": [
|
84 |
+
"import coremltools as ct\n",
|
85 |
+
"import numpy as np\n",
|
86 |
+
"\n",
|
87 |
+
"# Convert traced model to CoreML\n",
|
88 |
+
"text_input_shape = ct.Shape(shape=(1, 77))\n",
|
89 |
+
"\n",
|
90 |
+
"model_coreml = ct.convert(\n",
|
91 |
+
" model_traced,\n",
|
92 |
+
" inputs=[ct.TensorType(name=\"input_text_token_ids\", shape=text_input_shape, dtype=np.float32)],\n",
|
93 |
+
" outputs=[ct.TensorType(name=\"output_embedding\", dtype=np.float16)],\n",
|
94 |
+
" minimum_deployment_target=ct.target.macOS13,\n",
|
95 |
+
" convert_to='mlprogram'\n",
|
96 |
+
")"
|
97 |
+
]
|
98 |
+
},
|
99 |
+
{
|
100 |
+
"cell_type": "code",
|
101 |
+
"execution_count": null,
|
102 |
+
"id": "a323b1b8",
|
103 |
+
"metadata": {},
|
104 |
+
"outputs": [],
|
105 |
+
"source": [
|
106 |
+
"model_coreml.get_spec().description"
|
107 |
+
]
|
108 |
+
},
|
109 |
+
{
|
110 |
+
"cell_type": "code",
|
111 |
+
"execution_count": null,
|
112 |
+
"id": "04773702",
|
113 |
+
"metadata": {},
|
114 |
+
"outputs": [],
|
115 |
+
"source": [
|
116 |
+
"model_coreml.save(\"CLIP-ViT-H-14-laion2B-s32B-b79K.text-encoder.mlpackage\")"
|
117 |
+
]
|
118 |
+
},
|
119 |
+
{
|
120 |
+
"cell_type": "markdown",
|
121 |
+
"id": "346ade90",
|
122 |
+
"metadata": {},
|
123 |
+
"source": [
|
124 |
+
"## Check correctness\n",
|
125 |
+
"Should see a mean difference on the order of 1e-5 "
|
126 |
+
]
|
127 |
+
},
|
128 |
+
{
|
129 |
+
"cell_type": "code",
|
130 |
+
"execution_count": null,
|
131 |
+
"id": "9fcaef03",
|
132 |
+
"metadata": {},
|
133 |
+
"outputs": [],
|
134 |
+
"source": [
|
135 |
+
"import numpy as np\n",
|
136 |
+
"import torch\n",
|
137 |
+
"with torch.no_grad():\n",
|
138 |
+
" processed_text = processor(text=\"hello there\", images=None, return_tensors=\"pt\", padding=True)\n",
|
139 |
+
" input_ids = processed_text.input_ids\n",
|
140 |
+
" input_ids = torch.cat([input_ids, torch.tensor([[49407] * (77-input_ids.shape[1])])], dim=1)\n",
|
141 |
+
" print(\"input shape:\", input_ids.shape)\n",
|
142 |
+
"\n",
|
143 |
+
" res_pt = model_pt_text(**processed_text)\n",
|
144 |
+
" print(f\"original output: shape {res_pt.shape}, {res_pt}\")\n",
|
145 |
+
" \n",
|
146 |
+
" coreml_out = model_coreml.predict({'input_text_token_ids': input_ids.float()})\n",
|
147 |
+
" res_coreml = torch.tensor(coreml_out['output_embedding'])\n",
|
148 |
+
" print(f\"coreml output: shape {res_coreml.shape}, {res_coreml}, type {type(res_coreml)}\")\n",
|
149 |
+
" \n",
|
150 |
+
" difference = res_pt - res_coreml\n",
|
151 |
+
" print(f\"mean difference: {torch.sum(difference)/difference.shape[1]}, max: {torch.max(difference)}\")\n",
|
152 |
+
"\n"
|
153 |
+
]
|
154 |
+
},
|
155 |
+
{
|
156 |
+
"cell_type": "markdown",
|
157 |
+
"id": "ec415cc5",
|
158 |
+
"metadata": {},
|
159 |
+
"source": [
|
160 |
+
"# Image encoder"
|
161 |
+
]
|
162 |
+
},
|
163 |
+
{
|
164 |
+
"cell_type": "code",
|
165 |
+
"execution_count": null,
|
166 |
+
"id": "9228b9dc",
|
167 |
+
"metadata": {},
|
168 |
+
"outputs": [],
|
169 |
+
"source": [
|
170 |
+
"# wrap CLIPModel so that forward() function returns get_image_features()\n",
|
171 |
+
"class WrappedCLIPModel_Image(CLIPModel): \n",
|
172 |
+
" def forward(self, *args, **kwargs):\n",
|
173 |
+
" return self.get_image_features(*args, **kwargs)\n",
|
174 |
+
"\n",
|
175 |
+
"model_pt_image = WrappedCLIPModel_Image.from_pretrained(model_version)\n",
|
176 |
+
"model_pt_image.eval()"
|
177 |
+
]
|
178 |
+
},
|
179 |
+
{
|
180 |
+
"cell_type": "code",
|
181 |
+
"execution_count": null,
|
182 |
+
"id": "e9560396",
|
183 |
+
"metadata": {},
|
184 |
+
"outputs": [],
|
185 |
+
"source": [
|
186 |
+
"from PIL import Image\n",
|
187 |
+
"import torch\n",
|
188 |
+
"\n",
|
189 |
+
"with torch.no_grad():\n",
|
190 |
+
" image = Image.open(\"example.jpg\") \n",
|
191 |
+
" processed_image = processor(text=None, images=image, return_tensors=\"pt\", padding=True)\n",
|
192 |
+
" trace_input = torch.rand_like(processed_image.pixel_values)\n",
|
193 |
+
" model_traced = torch.jit.trace(model_pt_image, trace_input, strict=True)"
|
194 |
+
]
|
195 |
+
},
|
196 |
+
{
|
197 |
+
"cell_type": "code",
|
198 |
+
"execution_count": null,
|
199 |
+
"id": "37adb85f",
|
200 |
+
"metadata": {
|
201 |
+
"scrolled": true
|
202 |
+
},
|
203 |
+
"outputs": [],
|
204 |
+
"source": [
|
205 |
+
"import coremltools as ct\n",
|
206 |
+
"import numpy as np\n",
|
207 |
+
"\n",
|
208 |
+
"# Convert traced model to CoreML\n",
|
209 |
+
"image_input_shape = ct.Shape(shape=trace_input.shape)\n",
|
210 |
+
"\n",
|
211 |
+
"model_coreml = ct.convert(\n",
|
212 |
+
" model_traced,\n",
|
213 |
+
" inputs=[ct.TensorType(name=\"input_image_preproessed\", shape=image_input_shape, dtype=np.float16)],\n",
|
214 |
+
" outputs=[ct.TensorType(name=\"output_embedding\", dtype=np.float16)],\n",
|
215 |
+
" minimum_deployment_target=ct.target.macOS13,\n",
|
216 |
+
" convert_to='mlprogram'\n",
|
217 |
+
")"
|
218 |
+
]
|
219 |
+
},
|
220 |
+
{
|
221 |
+
"cell_type": "code",
|
222 |
+
"execution_count": null,
|
223 |
+
"id": "9cb1b830",
|
224 |
+
"metadata": {},
|
225 |
+
"outputs": [],
|
226 |
+
"source": [
|
227 |
+
"model_coreml.get_spec().description"
|
228 |
+
]
|
229 |
+
},
|
230 |
+
{
|
231 |
+
"cell_type": "code",
|
232 |
+
"execution_count": null,
|
233 |
+
"id": "281451f8",
|
234 |
+
"metadata": {},
|
235 |
+
"outputs": [],
|
236 |
+
"source": [
|
237 |
+
"model_coreml.save(\"CLIP-ViT-H-14-laion2B-s32B-b79K.image-encoder.mlpackage\")"
|
238 |
+
]
|
239 |
+
},
|
240 |
+
{
|
241 |
+
"cell_type": "markdown",
|
242 |
+
"id": "9f2e43c3",
|
243 |
+
"metadata": {},
|
244 |
+
"source": [
|
245 |
+
"## Check correctness\n",
|
246 |
+
"Should see a mean difference on the order of 1e-5 "
|
247 |
+
]
|
248 |
+
},
|
249 |
+
{
|
250 |
+
"cell_type": "code",
|
251 |
+
"execution_count": null,
|
252 |
+
"id": "7cfe24af",
|
253 |
+
"metadata": {},
|
254 |
+
"outputs": [],
|
255 |
+
"source": [
|
256 |
+
"\n",
|
257 |
+
"with torch.no_grad():\n",
|
258 |
+
" image = Image.open(\"example.jpg\")\n",
|
259 |
+
"\n",
|
260 |
+
" processed_image = processor(text=None, images=image, return_tensors=\"pt\", padding=True)\n",
|
261 |
+
" print(\"input shape:\", processed_image.pixel_values.shape)\n",
|
262 |
+
"\n",
|
263 |
+
" res_pt = model_pt_image.get_image_features(**processed_image)\n",
|
264 |
+
" print(f\"original output: shape {res_pt.shape}, {res_pt}\")\n",
|
265 |
+
"\n",
|
266 |
+
" coreml_out = model_coreml.predict({'input_image_preproessed': processed_image.pixel_values})\n",
|
267 |
+
" res_coreml = torch.tensor(coreml_out['output_embedding'])\n",
|
268 |
+
" print(f\"coreml output: shape {res_coreml.shape}, {res_coreml}, type {type(res_coreml)}\")\n",
|
269 |
+
"\n",
|
270 |
+
" difference = res_pt - res_coreml\n",
|
271 |
+
" print(f\"mean difference: {torch.sum(difference)/difference.shape[1]}, cosine: {torch.nn.functional.cosine_similarity(res_pt, res_coreml)}, max: {torch.max(difference)}\")"
|
272 |
+
]
|
273 |
+
},
|
274 |
+
{
|
275 |
+
"cell_type": "markdown",
|
276 |
+
"id": "154fffa4",
|
277 |
+
"metadata": {},
|
278 |
+
"source": [
|
279 |
+
"# Check performance"
|
280 |
+
]
|
281 |
+
},
|
282 |
+
{
|
283 |
+
"cell_type": "code",
|
284 |
+
"execution_count": null,
|
285 |
+
"id": "55260e23",
|
286 |
+
"metadata": {},
|
287 |
+
"outputs": [],
|
288 |
+
"source": [
|
289 |
+
"import time\n",
|
290 |
+
"from tqdm.auto import tqdm\n",
|
291 |
+
"\n",
|
292 |
+
"model_pt_image = model_pt_image.to('mps', dtype=torch.float16)\n",
|
293 |
+
"\n",
|
294 |
+
"start = time.perf_counter()\n",
|
295 |
+
"for i in tqdm(range(100)):\n",
|
296 |
+
" model_pt_image(pixel_values = torch.rand_like(processed_image.pixel_values, device=model_pt_image.device, dtype=torch.float16))\n",
|
297 |
+
"end = time.perf_counter()\n",
|
298 |
+
"print(\"original (GPU): \", (end-start)/100)\n",
|
299 |
+
"\n",
|
300 |
+
"start = time.perf_counter()\n",
|
301 |
+
"for i in tqdm(range(100)):\n",
|
302 |
+
" model_coreml.predict({'input_image_preproessed': torch.rand_like(processed_image.pixel_values)})\n",
|
303 |
+
"end = time.perf_counter()\n",
|
304 |
+
"print(\"coreml: \", (end-start)/100)\n"
|
305 |
+
]
|
306 |
+
},
|
307 |
+
{
|
308 |
+
"cell_type": "code",
|
309 |
+
"execution_count": null,
|
310 |
+
"id": "41449a3a",
|
311 |
+
"metadata": {},
|
312 |
+
"outputs": [],
|
313 |
+
"source": []
|
314 |
+
}
|
315 |
+
],
|
316 |
+
"metadata": {
|
317 |
+
"kernelspec": {
|
318 |
+
"display_name": "Python 3 (ipykernel)",
|
319 |
+
"language": "python",
|
320 |
+
"name": "python3"
|
321 |
+
},
|
322 |
+
"language_info": {
|
323 |
+
"codemirror_mode": {
|
324 |
+
"name": "ipython",
|
325 |
+
"version": 3
|
326 |
+
},
|
327 |
+
"file_extension": ".py",
|
328 |
+
"mimetype": "text/x-python",
|
329 |
+
"name": "python",
|
330 |
+
"nbconvert_exporter": "python",
|
331 |
+
"pygments_lexer": "ipython3",
|
332 |
+
"version": "3.10.8"
|
333 |
+
}
|
334 |
+
},
|
335 |
+
"nbformat": 4,
|
336 |
+
"nbformat_minor": 5
|
337 |
+
}
|