Spaces:
Running
Running
Merge pull request #28 from patil-suraj/tpu-demo
Browse files- demo/tpu-demo.ipynb +391 -0
demo/tpu-demo.ipynb
ADDED
@@ -0,0 +1,391 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "code",
|
5 |
+
"execution_count": null,
|
6 |
+
"id": "6eb74941-bb4d-4d7e-97f1-d5a3a07672bf",
|
7 |
+
"metadata": {},
|
8 |
+
"outputs": [],
|
9 |
+
"source": [
|
10 |
+
"# !pip install flax transformers\n",
|
11 |
+
"# !git clone https://github.com/patil-suraj/vqgan-jax.git"
|
12 |
+
]
|
13 |
+
},
|
14 |
+
{
|
15 |
+
"cell_type": "code",
|
16 |
+
"execution_count": 305,
|
17 |
+
"id": "41db7534-f589-4b63-9165-9c9799e1b06e",
|
18 |
+
"metadata": {},
|
19 |
+
"outputs": [
|
20 |
+
{
|
21 |
+
"name": "stdout",
|
22 |
+
"output_type": "stream",
|
23 |
+
"text": [
|
24 |
+
"/home/surajpatil/vqgan-jax\n"
|
25 |
+
]
|
26 |
+
},
|
27 |
+
{
|
28 |
+
"data": {
|
29 |
+
"text/plain": [
|
30 |
+
"[TpuDevice(id=0, process_index=0, coords=(0,0,0), core_on_chip=0),\n",
|
31 |
+
" TpuDevice(id=1, process_index=0, coords=(0,0,0), core_on_chip=1),\n",
|
32 |
+
" TpuDevice(id=2, process_index=0, coords=(1,0,0), core_on_chip=0),\n",
|
33 |
+
" TpuDevice(id=3, process_index=0, coords=(1,0,0), core_on_chip=1),\n",
|
34 |
+
" TpuDevice(id=4, process_index=0, coords=(0,1,0), core_on_chip=0),\n",
|
35 |
+
" TpuDevice(id=5, process_index=0, coords=(0,1,0), core_on_chip=1),\n",
|
36 |
+
" TpuDevice(id=6, process_index=0, coords=(1,1,0), core_on_chip=0),\n",
|
37 |
+
" TpuDevice(id=7, process_index=0, coords=(1,1,0), core_on_chip=1)]"
|
38 |
+
]
|
39 |
+
},
|
40 |
+
"execution_count": 305,
|
41 |
+
"metadata": {},
|
42 |
+
"output_type": "execute_result"
|
43 |
+
}
|
44 |
+
],
|
45 |
+
"source": [
|
46 |
+
"%cd ~/vqgan-jax\n",
|
47 |
+
"\n",
|
48 |
+
"import random\n",
|
49 |
+
"\n",
|
50 |
+
"\n",
|
51 |
+
"import jax\n",
|
52 |
+
"import flax.linen as nn\n",
|
53 |
+
"from flax.training.common_utils import shard\n",
|
54 |
+
"from flax.jax_utils import replicate, unreplicate\n",
|
55 |
+
"\n",
|
56 |
+
"from transformers.models.bart.modeling_flax_bart import *\n",
|
57 |
+
"from transformers import BartTokenizer, FlaxBartForConditionalGeneration\n",
|
58 |
+
"\n",
|
59 |
+
"import io\n",
|
60 |
+
"\n",
|
61 |
+
"import requests\n",
|
62 |
+
"from PIL import Image\n",
|
63 |
+
"import numpy as np\n",
|
64 |
+
"import matplotlib.pyplot as plt\n",
|
65 |
+
"\n",
|
66 |
+
"import torch\n",
|
67 |
+
"import torchvision.transforms as T\n",
|
68 |
+
"import torchvision.transforms.functional as TF\n",
|
69 |
+
"from torchvision.transforms import InterpolationMode\n",
|
70 |
+
"\n",
|
71 |
+
"\n",
|
72 |
+
"from modeling_flax_vqgan import VQModel\n",
|
73 |
+
"\n",
|
74 |
+
"jax.devices()"
|
75 |
+
]
|
76 |
+
},
|
77 |
+
{
|
78 |
+
"cell_type": "code",
|
79 |
+
"execution_count": 2,
|
80 |
+
"id": "b6a3462a-9004-4121-b365-3ae3aaf94dd2",
|
81 |
+
"metadata": {},
|
82 |
+
"outputs": [],
|
83 |
+
"source": [
|
84 |
+
"# TODO: set those args in a config file\n",
|
85 |
+
"OUTPUT_VOCAB_SIZE = 16384 + 1 # encoded image token space + 1 for bos\n",
|
86 |
+
"OUTPUT_LENGTH = 256 + 1 # number of encoded tokens + 1 for bos\n",
|
87 |
+
"BOS_TOKEN_ID = 16384\n",
|
88 |
+
"BASE_MODEL = 'facebook/bart-large'"
|
89 |
+
]
|
90 |
+
},
|
91 |
+
{
|
92 |
+
"cell_type": "code",
|
93 |
+
"execution_count": 3,
|
94 |
+
"id": "bbef1afb-0b36-44a5-83f7-643d7e2c0e30",
|
95 |
+
"metadata": {},
|
96 |
+
"outputs": [],
|
97 |
+
"source": [
|
98 |
+
"class CustomFlaxBartModule(FlaxBartModule):\n",
|
99 |
+
" def setup(self):\n",
|
100 |
+
" # we keep shared to easily load pre-trained weights\n",
|
101 |
+
" self.shared = nn.Embed(\n",
|
102 |
+
" self.config.vocab_size,\n",
|
103 |
+
" self.config.d_model,\n",
|
104 |
+
" embedding_init=jax.nn.initializers.normal(self.config.init_std, self.dtype),\n",
|
105 |
+
" dtype=self.dtype,\n",
|
106 |
+
" )\n",
|
107 |
+
" # a separate embedding is used for the decoder\n",
|
108 |
+
" self.decoder_embed = nn.Embed(\n",
|
109 |
+
" OUTPUT_VOCAB_SIZE,\n",
|
110 |
+
" self.config.d_model,\n",
|
111 |
+
" embedding_init=jax.nn.initializers.normal(self.config.init_std, self.dtype),\n",
|
112 |
+
" dtype=self.dtype,\n",
|
113 |
+
" )\n",
|
114 |
+
" self.encoder = FlaxBartEncoder(self.config, dtype=self.dtype, embed_tokens=self.shared)\n",
|
115 |
+
"\n",
|
116 |
+
" # the decoder has a different config\n",
|
117 |
+
" decoder_config = BartConfig(self.config.to_dict())\n",
|
118 |
+
" decoder_config.max_position_embeddings = OUTPUT_LENGTH\n",
|
119 |
+
" decoder_config.vocab_size = OUTPUT_VOCAB_SIZE\n",
|
120 |
+
" self.decoder = FlaxBartDecoder(decoder_config, dtype=self.dtype, embed_tokens=self.decoder_embed)\n",
|
121 |
+
"\n",
|
122 |
+
"class CustomFlaxBartForConditionalGenerationModule(FlaxBartForConditionalGenerationModule):\n",
|
123 |
+
" def setup(self):\n",
|
124 |
+
" self.model = CustomFlaxBartModule(config=self.config, dtype=self.dtype)\n",
|
125 |
+
" self.lm_head = nn.Dense(\n",
|
126 |
+
" OUTPUT_VOCAB_SIZE,\n",
|
127 |
+
" use_bias=False,\n",
|
128 |
+
" dtype=self.dtype,\n",
|
129 |
+
" kernel_init=jax.nn.initializers.normal(self.config.init_std, self.dtype),\n",
|
130 |
+
" )\n",
|
131 |
+
" self.final_logits_bias = self.param(\"final_logits_bias\", self.bias_init, (1, OUTPUT_VOCAB_SIZE))\n",
|
132 |
+
"\n",
|
133 |
+
"class CustomFlaxBartForConditionalGeneration(FlaxBartForConditionalGeneration):\n",
|
134 |
+
" module_class = CustomFlaxBartForConditionalGenerationModule"
|
135 |
+
]
|
136 |
+
},
|
137 |
+
{
|
138 |
+
"cell_type": "code",
|
139 |
+
"execution_count": null,
|
140 |
+
"id": "879320b7-eaa0-4dc9-bbf2-c81efc53301d",
|
141 |
+
"metadata": {},
|
142 |
+
"outputs": [],
|
143 |
+
"source": [
|
144 |
+
"import wandb\n",
|
145 |
+
"run = wandb.init()\n",
|
146 |
+
"artifact = run.use_artifact('wandb/hf-flax-dalle-mini/model-3h3x3565:v7', type='bart_model')\n",
|
147 |
+
"artifact_dir = artifact.download()"
|
148 |
+
]
|
149 |
+
},
|
150 |
+
{
|
151 |
+
"cell_type": "code",
|
152 |
+
"execution_count": 164,
|
153 |
+
"id": "e8bcff33-e95b-4c01-b162-ee857a55c3e6",
|
154 |
+
"metadata": {},
|
155 |
+
"outputs": [
|
156 |
+
{
|
157 |
+
"name": "stderr",
|
158 |
+
"output_type": "stream",
|
159 |
+
"text": [
|
160 |
+
"/home/surajpatil/transformers/src/transformers/models/bart/configuration_bart.py:177: UserWarning: Please make sure the config includes `forced_bos_token_id=16384` in future versions.The config can simply be saved and uploaded again to be fixed.\n",
|
161 |
+
" warnings.warn(\n"
|
162 |
+
]
|
163 |
+
},
|
164 |
+
{
|
165 |
+
"data": {
|
166 |
+
"text/plain": [
|
167 |
+
"(1, 16385)"
|
168 |
+
]
|
169 |
+
},
|
170 |
+
"execution_count": 164,
|
171 |
+
"metadata": {},
|
172 |
+
"output_type": "execute_result"
|
173 |
+
}
|
174 |
+
],
|
175 |
+
"source": [
|
176 |
+
"# create our model and initialize it randomly\n",
|
177 |
+
"tokenizer = BartTokenizer.from_pretrained(BASE_MODEL)\n",
|
178 |
+
"model = CustomFlaxBartForConditionalGeneration.from_pretrained(artifact_dir)\n",
|
179 |
+
"model.config.force_bos_token_to_be_generated = False\n",
|
180 |
+
"model.config.forced_bos_token_id = None\n",
|
181 |
+
"model.config.forced_eos_token_id = None\n",
|
182 |
+
"\n",
|
183 |
+
"# we verify that the shape has not been modified\n",
|
184 |
+
"model.params['final_logits_bias'].shape"
|
185 |
+
]
|
186 |
+
},
|
187 |
+
{
|
188 |
+
"cell_type": "code",
|
189 |
+
"execution_count": 6,
|
190 |
+
"id": "8d5e0f14-2502-470e-9553-daee6748601f",
|
191 |
+
"metadata": {},
|
192 |
+
"outputs": [
|
193 |
+
{
|
194 |
+
"data": {
|
195 |
+
"application/vnd.jupyter.widget-view+json": {
|
196 |
+
"model_id": "9b979a72ab9e449387a89bf9b3012af5",
|
197 |
+
"version_major": 2,
|
198 |
+
"version_minor": 0
|
199 |
+
},
|
200 |
+
"text/plain": [
|
201 |
+
"HBox(children=(FloatProgress(value=0.0, description='Downloading', max=433.0, style=ProgressStyle(description_…"
|
202 |
+
]
|
203 |
+
},
|
204 |
+
"metadata": {},
|
205 |
+
"output_type": "display_data"
|
206 |
+
},
|
207 |
+
{
|
208 |
+
"name": "stdout",
|
209 |
+
"output_type": "stream",
|
210 |
+
"text": [
|
211 |
+
"\n"
|
212 |
+
]
|
213 |
+
},
|
214 |
+
{
|
215 |
+
"data": {
|
216 |
+
"application/vnd.jupyter.widget-view+json": {
|
217 |
+
"model_id": "01730e0e9d02428ca9dad680f9fdda42",
|
218 |
+
"version_major": 2,
|
219 |
+
"version_minor": 0
|
220 |
+
},
|
221 |
+
"text/plain": [
|
222 |
+
"HBox(children=(FloatProgress(value=0.0, description='Downloading', max=304307206.0, style=ProgressStyle(descri…"
|
223 |
+
]
|
224 |
+
},
|
225 |
+
"metadata": {},
|
226 |
+
"output_type": "display_data"
|
227 |
+
},
|
228 |
+
{
|
229 |
+
"name": "stdout",
|
230 |
+
"output_type": "stream",
|
231 |
+
"text": [
|
232 |
+
"\n",
|
233 |
+
"Working with z of shape (1, 256, 16, 16) = 65536 dimensions.\n"
|
234 |
+
]
|
235 |
+
}
|
236 |
+
],
|
237 |
+
"source": [
|
238 |
+
"vqgan = VQModel.from_pretrained(\"flax-community/vqgan_f16_16384\")"
|
239 |
+
]
|
240 |
+
},
|
241 |
+
{
|
242 |
+
"cell_type": "code",
|
243 |
+
"execution_count": 295,
|
244 |
+
"id": "6cca395a-93c2-49bc-a3be-98287e4403d4",
|
245 |
+
"metadata": {},
|
246 |
+
"outputs": [],
|
247 |
+
"source": [
|
248 |
+
"def custom_to_pil(x):\n",
|
249 |
+
" x = np.clip(x, 0., 1.)\n",
|
250 |
+
" x = (255*x).astype(np.uint8)\n",
|
251 |
+
" x = Image.fromarray(x)\n",
|
252 |
+
" if not x.mode == \"RGB\":\n",
|
253 |
+
" x = x.convert(\"RGB\")\n",
|
254 |
+
" return x\n",
|
255 |
+
"\n",
|
256 |
+
"def generate(input, rng, params):\n",
|
257 |
+
" return model.generate(\n",
|
258 |
+
" **input,\n",
|
259 |
+
" max_length=257,\n",
|
260 |
+
" num_beams=1,\n",
|
261 |
+
" do_sample=True,\n",
|
262 |
+
" prng_key=rng,\n",
|
263 |
+
" eos_token_id=50000,\n",
|
264 |
+
" pad_token_id=50000,\n",
|
265 |
+
" params=params\n",
|
266 |
+
" )\n",
|
267 |
+
"\n",
|
268 |
+
"def get_images(indices, params):\n",
|
269 |
+
" return vqgan.decode_code(indices, params=params)\n",
|
270 |
+
"\n",
|
271 |
+
"\n",
|
272 |
+
"def plot_images(images):\n",
|
273 |
+
" fig = plt.figure(figsize=(40, 20))\n",
|
274 |
+
" columns = 4\n",
|
275 |
+
" rows = 2\n",
|
276 |
+
" plt.subplots_adjust(hspace=0, wspace=0)\n",
|
277 |
+
"\n",
|
278 |
+
" for i in range(1, columns*rows +1):\n",
|
279 |
+
" fig.add_subplot(rows, columns, i)\n",
|
280 |
+
" plt.imshow(images[i-1])\n",
|
281 |
+
" plt.gca().axes.get_yaxis().set_visible(False)\n",
|
282 |
+
" plt.show()\n",
|
283 |
+
" \n",
|
284 |
+
"def stack_reconstructions(images):\n",
|
285 |
+
" w, h = images[0].size[0], images[0].size[1]\n",
|
286 |
+
" img = Image.new(\"RGB\", (len(images)*w, h))\n",
|
287 |
+
" for i, img_ in enumerate(images):\n",
|
288 |
+
" img.paste(img_, (i*w,0))\n",
|
289 |
+
" return img"
|
290 |
+
]
|
291 |
+
},
|
292 |
+
{
|
293 |
+
"cell_type": "code",
|
294 |
+
"execution_count": 166,
|
295 |
+
"id": "b1bec3d2-ef17-4feb-aa0d-b51ed2fdcd3e",
|
296 |
+
"metadata": {},
|
297 |
+
"outputs": [],
|
298 |
+
"source": [
|
299 |
+
"p_generate = jax.pmap(generate, \"batch\")\n",
|
300 |
+
"p_get_images = jax.pmap(get_images, \"batch\")"
|
301 |
+
]
|
302 |
+
},
|
303 |
+
{
|
304 |
+
"cell_type": "code",
|
305 |
+
"execution_count": null,
|
306 |
+
"id": "a539823a-a775-4d92-96a5-dc8b1eef69c5",
|
307 |
+
"metadata": {},
|
308 |
+
"outputs": [],
|
309 |
+
"source": [
|
310 |
+
"bart_params = replicate(model.params)\n",
|
311 |
+
"vqgan_params = replicate(vqgan.params)"
|
312 |
+
]
|
313 |
+
},
|
314 |
+
{
|
315 |
+
"cell_type": "code",
|
316 |
+
"execution_count": 328,
|
317 |
+
"id": "e8b268d8-6992-422a-8373-95651474ae70",
|
318 |
+
"metadata": {},
|
319 |
+
"outputs": [],
|
320 |
+
"source": [
|
321 |
+
"prompts = [\n",
|
322 |
+
" \"man in blue jacket walking on pathway in between trees during daytime\",\n",
|
323 |
+
" 'white snow covered mountain under blue sky during daytime',\n",
|
324 |
+
" 'white snow covered mountain under blue sky during night',\n",
|
325 |
+
" \"orange tabby cat on persons hand\",\n",
|
326 |
+
" \"aerial view of beach during daytime\",\n",
|
327 |
+
" \"chess pieces on chess board\",\n",
|
328 |
+
" \"laptop on brown wooden table\",\n",
|
329 |
+
" \"white bus on road near high rise buildings\",\n",
|
330 |
+
"]\n",
|
331 |
+
"\n",
|
332 |
+
"\n",
|
333 |
+
"prompt = [prompts[-1]] * 8\n",
|
334 |
+
"inputs = tokenizer(prompt, return_tensors='jax', padding=\"max_length\", truncation=True, max_length=128).data\n",
|
335 |
+
"inputs = shard(inputs)"
|
336 |
+
]
|
337 |
+
},
|
338 |
+
{
|
339 |
+
"cell_type": "code",
|
340 |
+
"execution_count": null,
|
341 |
+
"id": "68638cfa-9a4d-4e6a-8630-91aefb627bbd",
|
342 |
+
"metadata": {},
|
343 |
+
"outputs": [],
|
344 |
+
"source": [
|
345 |
+
"%%time\n",
|
346 |
+
"for i in range(8):\n",
|
347 |
+
" key = random.randint(0, 1e7)\n",
|
348 |
+
" rng = jax.random.PRNGKey(key)\n",
|
349 |
+
" rngs = jax.random.split(rng, jax.local_device_count())\n",
|
350 |
+
" indices = p_generate(inputs, rngs, bart_params).sequences\n",
|
351 |
+
" indices = indices[:, :, 1:]\n",
|
352 |
+
"\n",
|
353 |
+
" images = p_get_images(indices, vqgan_params)\n",
|
354 |
+
" images = np.squeeze(np.asarray(images), 1)\n",
|
355 |
+
" imges = [custom_to_pil(image) for image in images]\n",
|
356 |
+
"\n",
|
357 |
+
" plt.figure(figsize=(40, 20))\n",
|
358 |
+
" plt.imshow(stack_reconstructions(imges))"
|
359 |
+
]
|
360 |
+
},
|
361 |
+
{
|
362 |
+
"cell_type": "code",
|
363 |
+
"execution_count": null,
|
364 |
+
"id": "681af54e-da10-4b8e-80d0-ebcbdf23f376",
|
365 |
+
"metadata": {},
|
366 |
+
"outputs": [],
|
367 |
+
"source": []
|
368 |
+
}
|
369 |
+
],
|
370 |
+
"metadata": {
|
371 |
+
"kernelspec": {
|
372 |
+
"display_name": "Python 3",
|
373 |
+
"language": "python",
|
374 |
+
"name": "python3"
|
375 |
+
},
|
376 |
+
"language_info": {
|
377 |
+
"codemirror_mode": {
|
378 |
+
"name": "ipython",
|
379 |
+
"version": 3
|
380 |
+
},
|
381 |
+
"file_extension": ".py",
|
382 |
+
"mimetype": "text/x-python",
|
383 |
+
"name": "python",
|
384 |
+
"nbconvert_exporter": "python",
|
385 |
+
"pygments_lexer": "ipython3",
|
386 |
+
"version": "3.8.10"
|
387 |
+
}
|
388 |
+
},
|
389 |
+
"nbformat": 4,
|
390 |
+
"nbformat_minor": 5
|
391 |
+
}
|