huolongguo10 commited on
Commit
67b476c
·
verified ·
1 Parent(s): 7e67fb4

Upload 5 files

Browse files
Files changed (5) hide show
  1. Dockerfile +18 -12
  2. ag4masses-public.ipynb +473 -0
  3. alphageometry.py +778 -0
  4. download.sh +31 -0
  5. models.py +178 -0
Dockerfile CHANGED
@@ -1,7 +1,7 @@
1
  FROM nvidia/cuda:12.5.1-cudnn-devel-ubuntu20.04
2
 
3
  ENV DEBIAN_FRONTEND=noninteractive \
4
- TZ=Europe/Paris
5
 
6
  # Remove any third-party apt sources to avoid issues with expiring keys.
7
  # Install some basic utilities
@@ -51,7 +51,7 @@ RUN mkdir $HOME/.cache $HOME/.config \
51
  # Set up the Conda environment
52
  ENV CONDA_AUTO_UPDATE_CONDA=false \
53
  PATH=$HOME/miniconda/bin:$PATH
54
- RUN curl -sLo ~/miniconda.sh https://repo.continuum.io/miniconda/Miniconda3-py39_4.10.3-Linux-x86_64.sh \
55
  && chmod +x ~/miniconda.sh \
56
  && ~/miniconda.sh -b -p ~/miniconda \
57
  && rm ~/miniconda.sh \
@@ -73,7 +73,7 @@ RUN --mount=target=/root/packages.txt,source=packages.txt \
73
  && rm -rf /var/lib/apt/lists/*
74
 
75
  RUN --mount=target=/root/on_startup.sh,source=on_startup.sh,readwrite \
76
- bash /root/on_startup.sh
77
 
78
  RUN mkdir /data && chown user:user /data
79
 
@@ -92,14 +92,20 @@ COPY --chown=user . $HOME/app
92
 
93
  RUN chmod +x start_server.sh
94
 
95
- COPY --chown=user login.html /home/user/miniconda/lib/python3.9/site-packages/jupyter_server/templates/login.html
 
 
 
 
 
 
96
 
97
  ENV PYTHONUNBUFFERED=1 \
98
- GRADIO_ALLOW_FLAGGING=never \
99
- GRADIO_NUM_PORTS=1 \
100
- GRADIO_SERVER_NAME=0.0.0.0 \
101
- GRADIO_THEME=huggingface \
102
- SYSTEM=spaces \
103
- SHELL=/bin/bash
104
-
105
- CMD ["./start_server.sh"]
 
1
  FROM nvidia/cuda:12.5.1-cudnn-devel-ubuntu20.04
2
 
3
  ENV DEBIAN_FRONTEND=noninteractive \
4
+ TZ=Asia/Shanghai
5
 
6
  # Remove any third-party apt sources to avoid issues with expiring keys.
7
  # Install some basic utilities
 
51
  # Set up the Conda environment
52
  ENV CONDA_AUTO_UPDATE_CONDA=false \
53
  PATH=$HOME/miniconda/bin:$PATH
54
+ RUN curl -sLo ~/miniconda.sh https://repo.continuum.io/miniconda/Miniconda3-py310_4.10.3-Linux-x86_64.sh \
55
  && chmod +x ~/miniconda.sh \
56
  && ~/miniconda.sh -b -p ~/miniconda \
57
  && rm ~/miniconda.sh \
 
73
  && rm -rf /var/lib/apt/lists/*
74
 
75
  RUN --mount=target=/root/on_startup.sh,source=on_startup.sh,readwrite \
76
+ bash /root/on_startup.sh
77
 
78
  RUN mkdir /data && chown user:user /data
79
 
 
92
 
93
  RUN chmod +x start_server.sh
94
 
95
+ RUN download.sh
96
+
97
+ COPY --chown=user login.html /home/user/miniconda/lib/python3.10/site-packages/jupyter_server/templates/login.html
98
+
99
+ COPY --chown=user public.ipynb /data/public.ipynb
100
+
101
+ # COPY --chown=user models.py /data/public.ipynb
102
 
103
  ENV PYTHONUNBUFFERED=1 \
104
+ GRADIO_ALLOW_FLAGGING=never \
105
+ GRADIO_NUM_PORTS=1 \
106
+ GRADIO_SERVER_NAME=0.0.0.0 \
107
+ GRADIO_THEME=huggingface \
108
+ SYSTEM=spaces \
109
+ SHELL=/bin/bash
110
+
111
+ CMD ["./start_server.sh"]
ag4masses-public.ipynb ADDED
@@ -0,0 +1,473 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "cells": [
3
+ {
4
+ "cell_type": "code",
5
+ "execution_count": null,
6
+ "metadata": {
7
+ "executionInfo": {
8
+ "elapsed": 611,
9
+ "status": "ok",
10
+ "timestamp": 1733595497864,
11
+ "user": {
12
+ "displayName": "Tong Peng",
13
+ "userId": "14680520704856526492"
14
+ },
15
+ "user_tz": 300
16
+ },
17
+ "id": "-IHoHd-t5sLP",
18
+ "trusted": true
19
+ },
20
+ "outputs": [],
21
+ "source": [
22
+ "import sys, os\n",
23
+ "\n",
24
+ "AG4MDIR='/home/user/ag4masses'\n",
25
+ "AGLIB=f'{AG4MDIR}/aglib'\n",
26
+ "AGDIR=f\"{AGLIB}/alphageometry\"\n",
27
+ "MELIAD_PATH=f\"{AGDIR}/meliad\"\n",
28
+ "DATA=f\"{AGLIB}/ag_ckpt_vocab\"\n",
29
+ "TESTDIR=f\"/data/ag4mtest\""
30
+ ]
31
+ },
32
+ {
33
+ "cell_type": "markdown",
34
+ "metadata": {
35
+ "id": "ASgGFu0NYHUH"
36
+ },
37
+ "source": [
38
+ "# Download Files"
39
+ ]
40
+ },
41
+ {
42
+ "cell_type": "code",
43
+ "execution_count": null,
44
+ "metadata": {
45
+ "trusted": true
46
+ },
47
+ "outputs": [],
48
+ "source": [
49
+ "# Run this cell to refresh code and get the latest versions\n",
50
+ "# AG4MDIR and MELIAD_PATH are in /kaggle/working and will be saved as outputs of the Notebook\n",
51
+ "# !rm -fr {AG4MDIR} {MELIAD_PATH}"
52
+ ]
53
+ },
54
+ {
55
+ "cell_type": "code",
56
+ "execution_count": null,
57
+ "metadata": {
58
+ "executionInfo": {
59
+ "elapsed": 5,
60
+ "status": "ok",
61
+ "timestamp": 1733594216384,
62
+ "user": {
63
+ "displayName": "Tong Peng",
64
+ "userId": "14680520704856526492"
65
+ },
66
+ "user_tz": 300
67
+ },
68
+ "id": "GgR_vO8XX9Vr",
69
+ "trusted": true
70
+ },
71
+ "outputs": [],
72
+ "source": [
73
+ "if not os.path.exists(MELIAD_PATH):\n",
74
+ " !git clone https://github.com/google-research/meliad.git {MELIAD_PATH}\n",
75
+ "\n",
76
+ "if not os.path.exists(AG4MDIR):\n",
77
+ " !git clone https://github.com/tpgh24/ag4masses.git {AG4MDIR}\n",
78
+ "\n",
79
+ "# Temporarily modified files, upload into dataset tmpfiles\n",
80
+ "# !cp /kaggle/input/tmpfiles/numericals.py {AGDIR}\n",
81
+ "# !cp /kaggle/input/tmpfiles/alphageometry.py {AGDIR}\n",
82
+ "\n",
83
+ "if not os.path.exists(TESTDIR):\n",
84
+ " !mkdir {TESTDIR}"
85
+ ]
86
+ },
87
+ {
88
+ "cell_type": "code",
89
+ "execution_count": null,
90
+ "metadata": {
91
+ "executionInfo": {
92
+ "elapsed": 40247,
93
+ "status": "ok",
94
+ "timestamp": 1733594312243,
95
+ "user": {
96
+ "displayName": "Tong Peng",
97
+ "userId": "14680520704856526492"
98
+ },
99
+ "user_tz": 300
100
+ },
101
+ "id": "gP4zAZh2MHcv",
102
+ "outputId": "4796397b-8952-411e-bd33-8fd813865735",
103
+ "trusted": true
104
+ },
105
+ "outputs": [],
106
+ "source": [
107
+ "if not os.path.exists(DATA):\n",
108
+ " # download data: vocabulary, trained model\n",
109
+ " !mkdir {DATA}\n",
110
+ "\n",
111
+ " ### Convoluted process. This does no work, seems due to Google Drive restrictions\n",
112
+ " # !gdown --folder https://bit.ly/alphageometry/\n",
113
+ " #\n",
114
+ " ### First got file links from Google Drive web UI, under Share menu. It will download HTML files with download button. Extract URL from the file.\n",
115
+ " ### For checkpoint_10999999, because the file is big, there is an additional step asking for confirmation, got the final URL from the 2nd HTML,\n",
116
+ " ### Link constructed using AI from HTML form.\n",
117
+ " # !gdown https://drive.google.com/file/d/1mRd6J0UkeWoFUjeVB7BQi5lVNLvPBe31/view?usp=drive_link -O {AGLIB}/ag_ckpt_vocab/geometry.757.vocab\n",
118
+ " # !gdown https://drive.google.com/file/d/1t-r3KfU8aDbS1UHpdyM3LH21rwSCIXTz/view?usp=drive_link -O {AGLIB}/ag_ckpt_vocab/geometry.757.model\n",
119
+ " # !gdown https://drive.google.com/file/d/1qXkmmgoJ8oTYJdFV1xw0xGPpQj6SyOYA/view?usp=drive_link -O {AGLIB}/ag_ckpt_vocab/checkpoint_10999999\n",
120
+ "\n",
121
+ " !wget -O {DATA}/geometry.757.vocab \"https://drive.usercontent.google.com/uc?id=1mRd6J0UkeWoFUjeVB7BQi5lVNLvPBe31&export=download\"\n",
122
+ " !wget -O {DATA}/geometry.757.model \"https://drive.usercontent.google.com/uc?id=1t-r3KfU8aDbS1UHpdyM3LH21rwSCIXTz&export=download\"\n",
123
+ " !wget -O {DATA}/checkpoint_10999999 \"https://drive.usercontent.google.com/download?id=1qXkmmgoJ8oTYJdFV1xw0xGPpQj6SyOYA&export=download&confirm=t&uuid=ae22f4de-cb77-4145-af5f-8cfbb59e867e\"\n",
124
+ "\n",
125
+ "!ls -l {DATA}"
126
+ ]
127
+ },
128
+ {
129
+ "cell_type": "markdown",
130
+ "metadata": {},
131
+ "source": [
132
+ "# Setup Env"
133
+ ]
134
+ },
135
+ {
136
+ "cell_type": "code",
137
+ "execution_count": null,
138
+ "metadata": {
139
+ "executionInfo": {
140
+ "elapsed": 14190,
141
+ "status": "ok",
142
+ "timestamp": 1733594230570,
143
+ "user": {
144
+ "displayName": "Tong Peng",
145
+ "userId": "14680520704856526492"
146
+ },
147
+ "user_tz": 300
148
+ },
149
+ "id": "X8Aj3G0neT6K",
150
+ "outputId": "9538ceba-8065-44d6-a32f-35127e5f2575",
151
+ "trusted": true
152
+ },
153
+ "outputs": [],
154
+ "source": [
155
+ "# Python packages for AlphaGeometry\n",
156
+ "!pip cache purge\n",
157
+ "!pip install --upgrade pip\n",
158
+ "!pip install --upgrade packaging setuptools setuptools_scm wheel \n",
159
+ "!pip install --require-hashes --no-deps -r {AG4MDIR}/alphageometry/requirements.txt"
160
+ ]
161
+ },
162
+ {
163
+ "cell_type": "code",
164
+ "execution_count": null,
165
+ "metadata": {
166
+ "executionInfo": {
167
+ "elapsed": 15694,
168
+ "status": "ok",
169
+ "timestamp": 1733594246256,
170
+ "user": {
171
+ "displayName": "Tong Peng",
172
+ "userId": "14680520704856526492"
173
+ },
174
+ "user_tz": 300
175
+ },
176
+ "id": "u9fuBSr2qEwN",
177
+ "outputId": "97bbce78-8b49-4d3b-a831-d188a4a9e536",
178
+ "trusted": true
179
+ },
180
+ "outputs": [],
181
+ "source": [
182
+ "# Python packages for Nvidia gpu.\n",
183
+ "# The versions of Python packages used by AlphaGeometry seem to only work with Cuda 11, not 12\n",
184
+ "!pip install -U \"jax==0.4.6\" \"jaxlib[cuda11_cudnn86]==0.4.6\" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html\n",
185
+ "!pip install nvidia-cuda-runtime-cu11\n",
186
+ "!pip install nvidia-pyindex\n",
187
+ "# !pip list"
188
+ ]
189
+ },
190
+ {
191
+ "cell_type": "code",
192
+ "execution_count": null,
193
+ "metadata": {
194
+ "trusted": true
195
+ },
196
+ "outputs": [],
197
+ "source": [
198
+ "# Linux packages for Nvidia gpu.\n",
199
+ "# The versions of Python packages used by AlphaGeometry seem to only work with Cuda 11, not 12\n",
200
+ "!apt-get update\n",
201
+ "!DEBIAN_FRONTEND=noninteractive apt install -y cuda-11-8\n",
202
+ "# !apt list|grep cuda"
203
+ ]
204
+ },
205
+ {
206
+ "cell_type": "code",
207
+ "execution_count": null,
208
+ "metadata": {
209
+ "id": "fChy49CNhf01",
210
+ "trusted": true
211
+ },
212
+ "outputs": [],
213
+ "source": [
214
+ "# Information about Nvidia drivers\n",
215
+ "!nvcc --version\n",
216
+ "!nvidia-smi"
217
+ ]
218
+ },
219
+ {
220
+ "cell_type": "markdown",
221
+ "metadata": {
222
+ "id": "jUWvch7kYhxt"
223
+ },
224
+ "source": [
225
+ "# Execution"
226
+ ]
227
+ },
228
+ {
229
+ "cell_type": "code",
230
+ "execution_count": null,
231
+ "metadata": {
232
+ "trusted": true
233
+ },
234
+ "outputs": [],
235
+ "source": [
236
+ "#!! cannot have ' in the script, including in comments\n",
237
+ "jobScript = '''\n",
238
+ "# !/bin/bash\n",
239
+ "set -e\n",
240
+ "set -x\n",
241
+ "\n",
242
+ "# stdout, solution is written here\n",
243
+ "OUTFILE=$TESTDIR/${PROB}.out\n",
244
+ "# stderr, a lot of information, error message, log etc.\n",
245
+ "ERRFILE=$TESTDIR/${PROB}.log\n",
246
+ "\n",
247
+ "# stdout and stderr are written to both ERRFILF and console\n",
248
+ "exec >$ERRFILE 2>&1\n",
249
+ "\n",
250
+ "echo PROB=$PROB\n",
251
+ "echo PROB_FILE=$PROBFILE\n",
252
+ "echo MODEL=$MODEL\n",
253
+ "\n",
254
+ "# Directory where output files go\n",
255
+ "echo TESTDIR=$TESTDIR\n",
256
+ "# Directory containing AG4Masses source files\n",
257
+ "echo AG4MDIR=$AG4MDIR\n",
258
+ "# Directory containing external libraries including ag_ckpt_vocab and meliad\n",
259
+ "echo AGLIB=$AGLIB\n",
260
+ "\n",
261
+ "AGDIR=$AG4MDIR/alphageometry\n",
262
+ "export PYTHONPATH=$PYTHONPATH:$AGDIR:$AGLIB\n",
263
+ "\n",
264
+ "echo BATCH_SIZE=$BATCH_SIZE\n",
265
+ "echo BEAM_SIZE=$BEAM_SIZE\n",
266
+ "echo DEPTH=$DEPTH\n",
267
+ "echo NWORKERS=$NWORKERS\n",
268
+ "\n",
269
+ "echo ERRFILE=$ERRFILE\n",
270
+ "echo OUTFILE=$OUTFILE\n",
271
+ "\n",
272
+ "DATA=$AGLIB/ag_ckpt_vocab\n",
273
+ "MELIAD_PATH=$AGLIB/meliad\n",
274
+ "export PYTHONPATH=$PYTHONPATH:$MELIAD_PATH\n",
275
+ "\n",
276
+ "DDAR_ARGS=( \\\n",
277
+ " --defs_file=$AGDIR/defs.txt \\\n",
278
+ " --rules_file=$AGDIR/rules.txt \\\n",
279
+ ")\n",
280
+ "\n",
281
+ "SEARCH_ARGS=(\n",
282
+ " --beam_size=$BEAM_SIZE\n",
283
+ " --search_depth=$DEPTH\n",
284
+ ")\n",
285
+ "\n",
286
+ "LM_ARGS=(\n",
287
+ " --ckpt_path=$DATA \\\n",
288
+ " --vocab_path=$DATA/geometry.757.model \\\n",
289
+ " --gin_search_paths=$MELIAD_PATH/transformer/configs,$AGDIR \\\n",
290
+ " --gin_file=base_htrans.gin \\\n",
291
+ " --gin_file=size/medium_150M.gin \\\n",
292
+ " --gin_file=options/positions_t5.gin \\\n",
293
+ " --gin_file=options/lr_cosine_decay.gin \\\n",
294
+ " --gin_file=options/seq_1024_nocache.gin \\\n",
295
+ " --gin_file=geometry_150M_generate.gin \\\n",
296
+ " --gin_param=DecoderOnlyLanguageModelGenerate.output_token_losses=True \\\n",
297
+ " --gin_param=TransformerTaskConfig.batch_size=$BATCH_SIZE \\\n",
298
+ " --gin_param=TransformerTaskConfig.sequence_length=128 \\\n",
299
+ " --gin_param=Trainer.restore_state_variables=False\n",
300
+ ");\n",
301
+ "\n",
302
+ "true \"==========================================\"\n",
303
+ "\n",
304
+ "python -m alphageometry \\\n",
305
+ "--alsologtostderr \\\n",
306
+ "--problems_file=$PROB_FILE \\\n",
307
+ "--problem_name=$PROB \\\n",
308
+ "--mode=$MODEL \\\n",
309
+ "\"${DDAR_ARGS[@]}\" \\\n",
310
+ "\"${SEARCH_ARGS[@]}\" \\\n",
311
+ "\"${LM_ARGS[@]}\" \\\n",
312
+ "--out_file=$OUTFILE \\\n",
313
+ "--n_workers=$NWORKERS 2>&1\n",
314
+ "\n",
315
+ "'''"
316
+ ]
317
+ },
318
+ {
319
+ "cell_type": "code",
320
+ "execution_count": null,
321
+ "metadata": {
322
+ "trusted": true
323
+ },
324
+ "outputs": [],
325
+ "source": [
326
+ "os.environ[\"TESTDIR\"]=TESTDIR\n",
327
+ "os.environ[\"AG4MDIR\"]=AG4MDIR\n",
328
+ "os.environ[\"AGLIB\"]=AGLIB\n",
329
+ "\n",
330
+ "# BATCH_SIZE: number of outputs for each LM query\n",
331
+ "# BEAM_SIZE: size of the breadth-first search queue\n",
332
+ "# DEPTH: search depth (number of auxilary points to add)\n",
333
+ "# NWORKERS: number of parallel run worker processes.\n",
334
+ "# \n",
335
+ "# Memory usage is affected by BATCH_SIZE, NWORKER and complexity of the problem.\n",
336
+ "# Larger NWORKER and BATCH_SIZE tends to cause out of memory issue\n",
337
+ "#\n",
338
+ "# The results in Google paper can be obtained by setting BATCH_SIZE=32, BEAM_SIZE=512, DEPTH=16\n",
339
+ "#\n",
340
+ "# 1/2025: Kaggle free version provides GPU T4x2, 4 virtual CPUs, 29G RAM. Can set \n",
341
+ "# NWORKERS=2\n",
342
+ "# CUDA_VISIBLE_DEVICES=0,1\n",
343
+ "\n",
344
+ "os.environ[\"BATCH_SIZE\"]=\"16\"\n",
345
+ "os.environ[\"BEAM_SIZE\"]=\"64\"\n",
346
+ "os.environ[\"DEPTH\"]=\"8\"\n",
347
+ "os.environ[\"NWORKERS\"]=\"2\"\n",
348
+ "\n",
349
+ "os.environ[\"CUDA_VISIBLE_DEVICES\"]=\"0,1\"\n",
350
+ "\n",
351
+ "# test problems can be uploaded into a dataset, e.g. for dataset \"tmpfiles\", \"/kaggle/input/tmpfiles/test-problems.txt\"\n",
352
+ "os.environ[\"PROB_FILE\"]=f\"{AG4MDIR}/data/ag4m_problems.txt\"\n",
353
+ "PROB=\"imo-2024-q4\"\n",
354
+ "os.environ[\"PROB\"]=PROB\n",
355
+ "# alphageometry|ddar\n",
356
+ "os.environ[\"MODEL\"]=\"alphageometry\"\n",
357
+ "\n",
358
+ "# In an interactive Kaggle session, run the job in background, so we can do other things in the Notebook.\n",
359
+ "# For long jobs, commit the Notebook and run in Batch mode.\n",
360
+ "# An interactive session will be terminated after about 20 minutes of idle time.\n",
361
+ "if os.environ[\"KAGGLE_KERNEL_RUN_TYPE\"]==\"Batch\":\n",
362
+ " os.system(f\"echo '{jobScript}'|bash\")\n",
363
+ "else:\n",
364
+ " os.system(f\"echo '{jobScript}'|bash &\")\n"
365
+ ]
366
+ },
367
+ {
368
+ "cell_type": "code",
369
+ "execution_count": null,
370
+ "metadata": {
371
+ "trusted": true
372
+ },
373
+ "outputs": [],
374
+ "source": [
375
+ "#!cat /kaggle/input/tmpfiles/test-problems.txt"
376
+ ]
377
+ },
378
+ {
379
+ "cell_type": "code",
380
+ "execution_count": null,
381
+ "metadata": {
382
+ "trusted": true
383
+ },
384
+ "outputs": [],
385
+ "source": [
386
+ "# In an interactive Kaggle session, run this to see the log file. We can cancel this cell's execution\n",
387
+ "# to do other things in the Notebook\n",
388
+ "if os.environ[\"KAGGLE_KERNEL_RUN_TYPE\"] != \"Batch\":\n",
389
+ " !tail -f {TESTDIR}/{PROB}.log"
390
+ ]
391
+ },
392
+ {
393
+ "cell_type": "code",
394
+ "execution_count": null,
395
+ "metadata": {
396
+ "trusted": true
397
+ },
398
+ "outputs": [],
399
+ "source": [
400
+ "# Command to kill the background job in an interactive session\n",
401
+ "# !pkill -P `ps -ef|grep -- '-m alphageometry'|grep -v grep|awk '{printf \"%d,%d\", $2, $3}'`\n",
402
+ "\n",
403
+ "# Command to check processes\n",
404
+ "### |cat to show full commandline\n",
405
+ "# !ps -eo pid,ppid,pgid,%cpu,cmd |cat\n",
406
+ "# !ps -ef|cat"
407
+ ]
408
+ },
409
+ {
410
+ "cell_type": "code",
411
+ "execution_count": null,
412
+ "metadata": {
413
+ "trusted": true
414
+ },
415
+ "outputs": [],
416
+ "source": [
417
+ "# Command to check progress of a running job in an interactive session\n",
418
+ "# !bash {AG4MDIR}/utils/checkprog.sh {TESTDIR}/{PROB}.log"
419
+ ]
420
+ },
421
+ {
422
+ "cell_type": "code",
423
+ "execution_count": null,
424
+ "metadata": {
425
+ "trusted": true
426
+ },
427
+ "outputs": [],
428
+ "source": [
429
+ "# In Batch run, after the job completes, list output files\n",
430
+ "!ls -ltr {TESTDIR}"
431
+ ]
432
+ }
433
+ ],
434
+ "metadata": {
435
+ "accelerator": "GPU",
436
+ "colab": {
437
+ "authorship_tag": "ABX9TyOcsgkfOgCk5oTpUiS6zrgo",
438
+ "collapsed_sections": [
439
+ "pW2KIijZBAdh"
440
+ ],
441
+ "gpuType": "T4",
442
+ "provenance": []
443
+ },
444
+ "kaggle": {
445
+ "accelerator": "nvidiaTeslaT4",
446
+ "dataSources": [],
447
+ "dockerImageVersionId": 30823,
448
+ "isGpuEnabled": true,
449
+ "isInternetEnabled": true,
450
+ "language": "python",
451
+ "sourceType": "notebook"
452
+ },
453
+ "kernelspec": {
454
+ "display_name": "Python 3",
455
+ "language": "python",
456
+ "name": "python3"
457
+ },
458
+ "language_info": {
459
+ "codemirror_mode": {
460
+ "name": "ipython",
461
+ "version": 3
462
+ },
463
+ "file_extension": ".py",
464
+ "mimetype": "text/x-python",
465
+ "name": "python",
466
+ "nbconvert_exporter": "python",
467
+ "pygments_lexer": "ipython3",
468
+ "version": "3.10.12"
469
+ }
470
+ },
471
+ "nbformat": 4,
472
+ "nbformat_minor": 4
473
+ }
alphageometry.py ADDED
@@ -0,0 +1,778 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 DeepMind Technologies Limited
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+ # ==============================================================================
15
+
16
+ """Run DD+AR or AlphaGeometry solver.
17
+
18
+ Please refer to README.md for detailed instructions.
19
+ """
20
+
21
+ import time
22
+ import traceback
23
+
24
+ from absl import app
25
+ from absl import flags
26
+ from absl import logging
27
+ import ddar
28
+ import graph as gh
29
+ import lm_inference as lm
30
+ import pretty as pt
31
+ import problem as pr
32
+
33
+ #=============
34
+ import sys, os, math, re
35
+ import multiprocessing
36
+ model = None # global variable used in multi-processing workers
37
+
38
+ _GIN_SEARCH_PATHS = flags.DEFINE_list(
39
+ 'gin_search_paths',
40
+ ['third_party/py/meliad/transformer/configs'],
41
+ 'List of paths where the Gin config files are located.',
42
+ )
43
+ _GIN_FILE = flags.DEFINE_multi_string(
44
+ 'gin_file', ['base_htrans.gin'], 'List of Gin config files.'
45
+ )
46
+ _GIN_PARAM = flags.DEFINE_multi_string(
47
+ 'gin_param', None, 'Newline separated list of Gin parameter bindings.'
48
+ )
49
+
50
+ _PROBLEMS_FILE = flags.DEFINE_string(
51
+ 'problems_file',
52
+ 'imo_ag_30.txt',
53
+ 'text file contains the problem strings. See imo_ag_30.txt for example.',
54
+ )
55
+ _PROBLEM_NAME = flags.DEFINE_string(
56
+ 'problem_name',
57
+ 'imo_2000_p1',
58
+ 'name of the problem to solve, must be in the problem_file.',
59
+ )
60
+ _MODE = flags.DEFINE_string(
61
+ 'mode', 'ddar', 'either `ddar` (DD+AR) or `alphageometry`')
62
+ _DEFS_FILE = flags.DEFINE_string(
63
+ 'defs_file',
64
+ 'defs.txt',
65
+ 'definitions of available constructions to state a problem.',
66
+ )
67
+ _RULES_FILE = flags.DEFINE_string(
68
+ 'rules_file', 'rules.txt', 'list of deduction rules used by DD.'
69
+ )
70
+ _CKPT_PATH = flags.DEFINE_string('ckpt_path', '', 'checkpoint of the LM model.')
71
+ _VOCAB_PATH = flags.DEFINE_string(
72
+ 'vocab_path', '', 'path to the LM vocab file.'
73
+ )
74
+ _OUT_FILE = flags.DEFINE_string(
75
+ 'out_file', '', 'path to the solution output file.'
76
+ ) # pylint: disable=line-too-long
77
+ _BEAM_SIZE = flags.DEFINE_integer(
78
+ 'beam_size', 1, 'beam size of the proof search.'
79
+ ) # pylint: disable=line-too-long
80
+ _SEARCH_DEPTH = flags.DEFINE_integer(
81
+ 'search_depth', 1, 'search depth of the proof search.'
82
+ ) # pylint: disable=line-too-long
83
+
84
+ #===================================
85
+ _N_WORKSERS = flags.DEFINE_integer(
86
+ 'n_workers', 1, 'number of workers'
87
+ )# pylint: disable=line-too-long
88
+
89
+ DEFINITIONS = None # contains definitions of construction actions
90
+ RULES = None # contains rules of deductions
91
+
92
+
93
+ def natural_language_statement(logical_statement: pr.Dependency) -> str:
94
+ """Convert logical_statement to natural language.
95
+
96
+ Args:
97
+ logical_statement: pr.Dependency with .name and .args
98
+
99
+ Returns:
100
+ a string of (pseudo) natural language of the predicate for human reader.
101
+ """
102
+ names = [a.name.upper() for a in logical_statement.args]
103
+ names = [(n[0] + '_' + n[1:]) if len(n) > 1 else n for n in names]
104
+ return pt.pretty_nl(logical_statement.name, names)
105
+
106
+
107
+ def proof_step_string(
108
+ proof_step: pr.Dependency, refs: dict[tuple[str, ...], int], last_step: bool
109
+ ) -> str:
110
+ """Translate proof to natural language.
111
+
112
+ Args:
113
+ proof_step: pr.Dependency with .name and .args
114
+ refs: dict(hash: int) to keep track of derived predicates
115
+ last_step: boolean to keep track whether this is the last step.
116
+
117
+ Returns:
118
+ a string of (pseudo) natural language of the proof step for human reader.
119
+ """
120
+ premises, [conclusion] = proof_step
121
+
122
+ premises_nl = ' & '.join(
123
+ [
124
+ natural_language_statement(p) + ' [{:02}]'.format(refs[p.hashed()])
125
+ for p in premises
126
+ ]
127
+ )
128
+
129
+ if not premises:
130
+ premises_nl = 'similarly'
131
+
132
+ refs[conclusion.hashed()] = len(refs)
133
+
134
+ conclusion_nl = natural_language_statement(conclusion)
135
+ if not last_step:
136
+ conclusion_nl += ' [{:02}]'.format(refs[conclusion.hashed()])
137
+
138
+ return f'{premises_nl} \u21d2 {conclusion_nl}'
139
+
140
+
141
+ def write_solution(g: gh.Graph, p: pr.Problem, out_file: str) -> None:
142
+ """Output the solution to out_file.
143
+
144
+ Args:
145
+ g: gh.Graph object, containing the proof state.
146
+ p: pr.Problem object, containing the theorem.
147
+ out_file: file to write to, empty string to skip writing to file.
148
+ """
149
+ setup, aux, proof_steps, refs = ddar.get_proof_steps(
150
+ g, p.goal, merge_trivials=False
151
+ )
152
+
153
+ solution = '\n=========================='
154
+ solution += '\n * From theorem premises:\n'
155
+ premises_nl = []
156
+ for premises, [points] in setup:
157
+ solution += ' '.join([p.name.upper() for p in points]) + ' '
158
+ if not premises:
159
+ continue
160
+ premises_nl += [
161
+ natural_language_statement(p) + ' [{:02}]'.format(refs[p.hashed()])
162
+ for p in premises
163
+ ]
164
+ solution += ': Points\n' + '\n'.join(premises_nl)
165
+
166
+ solution += '\n\n * Auxiliary Constructions:\n'
167
+ aux_premises_nl = []
168
+ for premises, [points] in aux:
169
+ solution += ' '.join([p.name.upper() for p in points]) + ' '
170
+ aux_premises_nl += [
171
+ natural_language_statement(p) + ' [{:02}]'.format(refs[p.hashed()])
172
+ for p in premises
173
+ ]
174
+ solution += ': Points\n' + '\n'.join(aux_premises_nl)
175
+
176
+ # some special case where the deduction rule has a well known name.
177
+ r2name = {
178
+ 'r32': '(SSS)',
179
+ 'r33': '(SAS)',
180
+ 'r34': '(Similar Triangles)',
181
+ 'r35': '(Similar Triangles)',
182
+ 'r36': '(ASA)',
183
+ 'r37': '(ASA)',
184
+ 'r38': '(Similar Triangles)',
185
+ 'r39': '(Similar Triangles)',
186
+ 'r40': '(Congruent Triangles)',
187
+ 'a00': '(Distance chase)',
188
+ 'a01': '(Ratio chase)',
189
+ 'a02': '(Angle chase)',
190
+ }
191
+
192
+ solution += '\n\n * Proof steps:\n'
193
+ for i, step in enumerate(proof_steps):
194
+ _, [con] = step
195
+ nl = proof_step_string(step, refs, last_step=i == len(proof_steps) - 1)
196
+ rule_name = r2name.get(con.rule_name, '')
197
+ nl = nl.replace('\u21d2', f'{rule_name}\u21d2 ')
198
+ solution += '{:03}. '.format(i + 1) + nl + '\n'
199
+
200
+ solution += '==========================\n'
201
+ logging.info(solution)
202
+ if out_file:
203
+ with open(out_file, 'w') as f:
204
+ f.write(solution)
205
+ logging.info('Solution written to %s.', out_file)
206
+
207
+
208
+ def get_lm(ckpt_init: str, vocab_path: str) -> lm.LanguageModelInference:
209
+ lm.parse_gin_configuration(
210
+ _GIN_FILE.value, _GIN_PARAM.value, gin_paths=_GIN_SEARCH_PATHS.value
211
+ )
212
+
213
+ return lm.LanguageModelInference(vocab_path, ckpt_init, mode='beam_search')
214
+
215
+
216
+ def run_ddar(g: gh.Graph, p: pr.Problem, out_file: str) -> bool:
217
+ """Run DD+AR.
218
+
219
+ Args:
220
+ g: gh.Graph object, containing the proof state.
221
+ p: pr.Problem object, containing the problem statement.
222
+ out_file: path to output file if solution is found.
223
+
224
+ Returns:
225
+ Boolean, whether DD+AR finishes successfully.
226
+ """
227
+ ddar.solve(g, RULES, p, max_level=1000)
228
+
229
+ goal_args = g.names2nodes(p.goal.args)
230
+ if not g.check(p.goal.name, goal_args):
231
+ logging.info('DD+AR failed to solve the problem.')
232
+ return False
233
+
234
+ write_solution(g, p, out_file)
235
+
236
+ gh.nm.draw(
237
+ g.type2nodes[gh.Point],
238
+ g.type2nodes[gh.Line],
239
+ g.type2nodes[gh.Circle],
240
+ g.type2nodes[gh.Segment])
241
+ return True
242
+
243
+
244
+ def translate_constrained_to_constructive(
245
+ point: str, name: str, args: list[str]
246
+ ) -> tuple[str, list[str]]:
247
+ """Translate a predicate from constraint-based to construction-based.
248
+
249
+ Args:
250
+ point: str: name of the new point
251
+ name: str: name of the predicate, e.g., perp, para, etc.
252
+ args: list[str]: list of predicate args.
253
+
254
+ Returns:
255
+ (name, args): translated to constructive predicate.
256
+ """
257
+ if name in ['T', 'perp']:
258
+ a, b, c, d = args
259
+ if point in [c, d]:
260
+ a, b, c, d = c, d, a, b
261
+ if point == b:
262
+ a, b = b, a
263
+ if point == d:
264
+ c, d = d, c
265
+ if a == c and a == point:
266
+ return 'on_dia', [a, b, d]
267
+ return 'on_tline', [a, b, c, d]
268
+
269
+ elif name in ['P', 'para']:
270
+ a, b, c, d = args
271
+ if point in [c, d]:
272
+ a, b, c, d = c, d, a, b
273
+ if point == b:
274
+ a, b = b, a
275
+ return 'on_pline', [a, b, c, d]
276
+
277
+ elif name in ['D', 'cong']:
278
+ a, b, c, d = args
279
+ if point in [c, d]:
280
+ a, b, c, d = c, d, a, b
281
+ if point == b:
282
+ a, b = b, a
283
+ if point == d:
284
+ c, d = d, c
285
+ if a == c and a == point:
286
+ return 'on_bline', [a, b, d]
287
+ if b in [c, d]:
288
+ if b == d:
289
+ c, d = d, c # pylint: disable=unused-variable
290
+ return 'on_circle', [a, b, d]
291
+ return 'eqdistance', [a, b, c, d]
292
+
293
+ elif name in ['C', 'coll']:
294
+ a, b, c = args
295
+ if point == b:
296
+ a, b = b, a
297
+ if point == c:
298
+ a, b, c = c, a, b
299
+ return 'on_line', [a, b, c]
300
+
301
+ elif name in ['^', 'eqangle']:
302
+ a, b, c, d, e, f = args
303
+
304
+ if point in [d, e, f]:
305
+ a, b, c, d, e, f = d, e, f, a, b, c
306
+
307
+ x, b, y, c, d = b, c, e, d, f
308
+ if point == b:
309
+ a, b, c, d = b, a, d, c
310
+
311
+ if point == d and x == y: # x p x b = x c x p
312
+ return 'angle_bisector', [point, b, x, c]
313
+
314
+ if point == x:
315
+ return 'eqangle3', [x, a, b, y, c, d]
316
+
317
+ return 'on_aline', [a, x, b, c, y, d]
318
+
319
+ elif name in ['cyclic', 'O']:
320
+ a, b, c = [x for x in args if x != point]
321
+ return 'on_circum', [point, a, b, c]
322
+
323
+ return name, args
324
+
325
+
326
+ def check_valid_args(name: str, args: list[str]) -> bool:
327
+ """Check whether a predicate is grammarically correct.
328
+
329
+ Args:
330
+ name: str: name of the predicate
331
+ args: list[str]: args of the predicate
332
+
333
+ Returns:
334
+ bool: whether the predicate arg count is valid.
335
+ """
336
+ if name == 'perp':
337
+ if len(args) != 4:
338
+ return False
339
+ a, b, c, d = args
340
+ if len({a, b}) < 2:
341
+ return False
342
+ if len({c, d}) < 2:
343
+ return False
344
+ elif name == 'para':
345
+ if len(args) != 4:
346
+ return False
347
+ a, b, c, d = args
348
+ if len({a, b, c, d}) < 4:
349
+ return False
350
+ elif name == 'cong':
351
+ if len(args) != 4:
352
+ return False
353
+ a, b, c, d = args
354
+ if len({a, b}) < 2:
355
+ return False
356
+ if len({c, d}) < 2:
357
+ return False
358
+ elif name == 'coll':
359
+ if len(args) != 3:
360
+ return False
361
+ a, b, c = args
362
+ if len({a, b, c}) < 3:
363
+ return False
364
+ elif name == 'cyclic':
365
+ if len(args) != 4:
366
+ return False
367
+ a, b, c, d = args
368
+ if len({a, b, c, d}) < 4:
369
+ return False
370
+ elif name == 'eqangle':
371
+ if len(args) != 8:
372
+ return False
373
+ a, b, c, d, e, f, g, h = args
374
+ if len({a, b, c, d}) < 3:
375
+ return False
376
+ if len({e, f, g, h}) < 3:
377
+ return False
378
+ return True
379
+
380
+
381
+ def try_translate_constrained_to_construct(string: str, g: gh.Graph) -> str:
382
+ """Whether a string of aux construction can be constructed.
383
+
384
+ Args:
385
+ string: str: the string describing aux construction.
386
+ g: gh.Graph: the current proof state.
387
+
388
+ Returns:
389
+ str: whether this construction is valid. If not, starts with "ERROR:".
390
+ """
391
+ if string[-1] != ';':
392
+ return 'ERROR: must end with ;'
393
+
394
+ logging.info(f'PID={os.getpid()}: !! try_translate_constrained_to_construct: string=%s', string)
395
+
396
+ # sometimes the LM may return ill-formed result with multiple colons.
397
+ # example:
398
+ #
399
+ # napoleon2
400
+ # a1 a2 a3 = triangle; c3 = s_angle a1 a2 c3 30, s_angle a2 a1 c3 150; c1 = s_angle a2 a3 c1 30, s_angle a3 a2 c1 150; c2 = s_angle a3 a1 c2 30, s_angle a1 a3 c2 150 ? cong c1 c2 c1 c3
401
+ #
402
+ # in the process,
403
+ # I0210 17:58:01.513668 140016515833856 alphageometry.py:550] Decoding from {S} a : ; b : ; c : ; d : ^ a d a b 5. pi / 6. 00 ^ b d b a 1. pi / 6. 01 ; e : ^ b e b c 5. pi / 6. 02 ^ c e c b 1. pi / 6. 03 ; f : ^ a f a c 1. pi / 6. 04 ^ c f c a 5. pi / 6. 05 ? D e f e d {F1} x00 g : C a b g 06 D a g b g 07 ; x00 h : C c b h 08 D c h b h 09 ; x00
404
+ # I0210 18:01:38.182158 140016515833856 alphageometry.py:384] !! try_translate_constrained_to_construct: string=i : C a c i 10 D a i c i 11 ? V d f {F1} x00 j : D g j h j 12 D h j i j 13 ;
405
+
406
+ #XXX
407
+ # str_parts = string.split(' : ')
408
+ # if len(str_parts) != 2:
409
+ # return f'ERROR: string has multiple colons: |{string}|'
410
+ mch = re.match('(.*?)( \? | \. \{)', string)
411
+ if mch :
412
+ strFixed = mch.group(1) + ';'
413
+ logging.info(f'ID={os.getpid()}: Bad LM output: {string}. Changed to {strFixed}')
414
+ string = strFixed
415
+
416
+ # sometimes the constraint in string is empty:
417
+ # 0407 17:11:35.470240 126383800963072 alphageometry.py:394] !! try_translate_constrained_to_construct: string=j : ;
418
+ hdprem = string.split(' : ')
419
+ if len(hdprem) !=2 or hdprem[1].strip()==';' :
420
+ logging.info(f'ID={os.getpid()}: Bad LM output: {string}. ERROR')
421
+ return f'ERROR: Bad LM output: {string}'
422
+ head, prem_str = hdprem
423
+ point = head.strip()
424
+
425
+ if len(point) != 1 or point == ' ':
426
+ return f'ERROR: invalid point name {point}'
427
+
428
+ existing_points = [p.name for p in g.all_points()]
429
+ if point in existing_points:
430
+ return f'ERROR: point {point} already exists.'
431
+
432
+ prem_toks = prem_str.split()[:-1] # remove the EOS ' ;'
433
+ prems = [[]]
434
+
435
+ for i, tok in enumerate(prem_toks):
436
+ if tok.isdigit():
437
+ if i < len(prem_toks) - 1:
438
+ prems.append([])
439
+ else:
440
+ prems[-1].append(tok)
441
+
442
+ if len(prems) > 2:
443
+ return 'ERROR: there cannot be more than two predicates.'
444
+
445
+ clause_txt = point + ' = '
446
+ constructions = []
447
+
448
+ for prem in prems:
449
+ name, *args = prem
450
+
451
+ if point not in args:
452
+ return f'ERROR: {point} not found in predicate args.'
453
+
454
+ if not check_valid_args(pt.map_symbol(name), args):
455
+ return 'ERROR: Invalid predicate ' + name + ' ' + ' '.join(args)
456
+
457
+ for a in args:
458
+ if a != point and a not in existing_points:
459
+ return f'ERROR: point {a} does not exist.'
460
+
461
+ try:
462
+ name, args = translate_constrained_to_constructive(point, name, args)
463
+ except: # pylint: disable=bare-except
464
+ return 'ERROR: Invalid predicate ' + name + ' ' + ' '.join(args)
465
+
466
+ if name == 'on_aline':
467
+ if args.count(point) > 1:
468
+ return f'ERROR: on_aline involves twice {point}'
469
+
470
+ constructions += [name + ' ' + ' '.join(args)]
471
+
472
+ clause_txt += ', '.join(constructions)
473
+ clause = pr.Clause.from_txt(clause_txt)
474
+
475
+ try:
476
+ g.copy().add_clause(clause, 0, DEFINITIONS)
477
+ except: # pylint: disable=bare-except
478
+ return 'ERROR: ' + traceback.format_exc()
479
+
480
+ return clause_txt
481
+
482
+
483
+ def insert_aux_to_premise(pstring: str, auxstring: str) -> str:
484
+ """Insert auxiliary constructs from proof to premise.
485
+
486
+ Args:
487
+ pstring: str: describing the problem to solve.
488
+ auxstring: str: describing the auxiliar construction.
489
+
490
+ Returns:
491
+ str: new pstring with auxstring inserted before the conclusion.
492
+ """
493
+ setup, goal = pstring.split(' ? ')
494
+ return setup + '; ' + auxstring + ' ? ' + goal
495
+
496
+
497
+ class BeamQueue:
498
+ """Keep only the top k objects according to their values."""
499
+
500
+ def __init__(self, max_size: int = 512):
501
+ self.queue = []
502
+ self.max_size = max_size
503
+
504
+ def add(self, node: object, val: float) -> None:
505
+ """Add a new node to this queue."""
506
+
507
+ if len(self.queue) < self.max_size:
508
+ self.queue.append((val, node))
509
+ return
510
+
511
+ # Find the minimum node:
512
+ min_idx, (min_val, _) = min(enumerate(self.queue), key=lambda x: x[1])
513
+
514
+ # replace it if the new node has higher value.
515
+ if val > min_val:
516
+ self.queue[min_idx] = (val, node)
517
+
518
+ def __iter__(self):
519
+ for val, node in self.queue:
520
+ yield val, node
521
+
522
+ def __len__(self) -> int:
523
+ return len(self.queue)
524
+
525
+ def bqsearch_init(worker_id):
526
+ # When using spawn or forkserver start method for multiprocessing.Pool, need to re-initialize
527
+ flags.FLAGS(sys.argv)
528
+ logging.use_absl_handler()
529
+ logging.set_verbosity(logging.INFO)
530
+ sys.setrecursionlimit(10000)
531
+
532
+ # Global variables initialized in main(). Need to re-initialize
533
+ #
534
+ # definitions of terms used in our domain-specific language.
535
+ global DEFINITIONS, RULES
536
+ DEFINITIONS = pr.Definition.from_txt_file(_DEFS_FILE.value, to_dict=True)
537
+ # load inference rules used in DD.
538
+ RULES = pr.Theorem.from_txt_file(_RULES_FILE.value, to_dict=True)
539
+
540
+ wkrpid = os.getpid()
541
+ logging.info('Worker %d initializing. PID=%d', worker_id, wkrpid)
542
+
543
+ if 'CUDA_VISIBLE_DEVICES' in os.environ and os.environ['CUDA_VISIBLE_DEVICES'].strip():
544
+ os.environ['CUDA_VISIBLE_DEVICES']=f"{worker_id}"
545
+ logging.info('Worker %d: CUDA_VISIBLE_DEVICES=%s', worker_id, os.environ['CUDA_VISIBLE_DEVICES'])
546
+
547
+ global model
548
+ model = get_lm(_CKPT_PATH.value, _VOCAB_PATH.value)
549
+ return wkrpid
550
+
551
+ def bqsearch(i_nd, srch_inputs, out_file) -> tuple[int, bool, list]: # ( iNode, solved, [ (node, score) ] )
552
+ pid = os.getpid()
553
+ logging.info(f'Worker PID={pid} called for beam search node {i_nd}')
554
+
555
+ prev_score, (g, string, pstring) = srch_inputs
556
+ logging.info(f'Worker PID={pid}: Beam-searching and Decoding from {string}')
557
+ outputs = model.beam_decode(string, eos_tokens=[';'])
558
+
559
+ # translate lm output to the constructive language.
560
+ # so that we can update the graph representing proof states:
561
+ translations = [
562
+ try_translate_constrained_to_construct(o, g)
563
+ for o in outputs['seqs_str']
564
+ ]
565
+
566
+ # couple the lm outputs with its translations
567
+ candidates = zip(outputs['seqs_str'], translations, outputs['scores'])
568
+
569
+ # bring the highest scoring candidate first
570
+ candidates = reversed(list(candidates))
571
+
572
+ ret = []
573
+ for lm_out, translation, score in candidates:
574
+ logging.info(f'Worker PID={pid}: LM output (score={score}): "{lm_out}"')
575
+ logging.info(f'Worker PID={pid}: Translation: "{translation}"')
576
+
577
+ if translation.startswith('ERROR:'):
578
+ # the construction is invalid.
579
+ continue
580
+
581
+ # Update the constructive statement of the problem with the aux point:
582
+ candidate_pstring = insert_aux_to_premise(pstring, translation)
583
+
584
+ #XXX
585
+ logging.info(f'Worker PID={pid}: string=|{string}| lm_out=|{lm_out}|')
586
+ logging.info(f'Worker PID={pid}: Solving: "{candidate_pstring}"')
587
+ p_new = pr.Problem.from_txt(candidate_pstring)
588
+
589
+ # This is the new proof state graph representation:
590
+ g_new, _ = gh.Graph.build_problem(p_new, DEFINITIONS)
591
+
592
+ try:
593
+ if run_ddar(g_new, p_new, out_file):
594
+ logging.info(f'Worker PID={pid}: Solved.')
595
+ return (i_nd, True, None)
596
+ except Exception as e:
597
+ logging.info(f'Worker PID={pid}: Error in run_ddar: {e}')
598
+
599
+ # Add the candidate to the beam queue.
600
+ ret.append( [
601
+ # The string for the new node is old_string + lm output +
602
+ # the special token asking for a new auxiliary point ' x00':
603
+ # node
604
+ (g_new, string + ' ' + lm_out + ' x00', candidate_pstring),
605
+ # the score of each node is sum of score of all nodes
606
+ # on the path to itself. For beam search, there is no need to
607
+ # normalize according to path length because all nodes in beam
608
+ # is of the same path length.
609
+ # val
610
+ prev_score + score ]
611
+ )
612
+
613
+ logging.info(f'Worker PID={pid} beam search node {i_nd}: returning')
614
+ return (i_nd, False, ret)
615
+
616
+ def run_alphageometry(
617
+ #XX model: lm.LanguageModelInference,
618
+ p: pr.Problem,
619
+ search_depth: int,
620
+ beam_size: int,
621
+ out_file: str,
622
+ ) -> bool:
623
+ """Simplified code to run AlphaGeometry proof search.
624
+
625
+ We removed all optimizations that are infrastructure-dependent, e.g.
626
+ parallelized model inference on multi GPUs,
627
+ parallelized DD+AR on multiple CPUs,
628
+ parallel execution of LM and DD+AR,
629
+ shared pool of CPU workers across different problems, etc.
630
+
631
+ Many other speed optimizations and abstractions are also removed to
632
+ better present the core structure of the proof search.
633
+
634
+ Args:
635
+ model: Interface with inference-related endpoints to JAX's model.
636
+ p: pr.Problem object describing the problem to solve.
637
+ search_depth: max proof search depth.
638
+ beam_size: beam size of the proof search.
639
+ out_file: path to output file if solution is found.
640
+
641
+ Returns:
642
+ boolean of whether this is solved.
643
+ """
644
+ # translate the problem to a string of grammar that the LM is trained on.
645
+ string = p.setup_str_from_problem(DEFINITIONS)
646
+ # special tokens prompting the LM to generate auxiliary points.
647
+ string += ' {F1} x00'
648
+ # the graph to represent the proof state.
649
+ g, _ = gh.Graph.build_problem(p, DEFINITIONS)
650
+
651
+ # First we run the symbolic engine DD+AR:
652
+ if run_ddar(g, p, out_file):
653
+ return True
654
+
655
+ # ?? when pickling graph for some problems, the default recursion limit 1000 is not enough,
656
+ # got 'maximum recursion depth exceeded while pickling an object' error
657
+ sys.setrecursionlimit(10000)
658
+
659
+ # beam search for the proof
660
+ # each node in the search tree is a 3-tuple:
661
+ # (<graph representation of proof state>,
662
+ # <string for LM to decode from>,
663
+ # <original problem string>)
664
+ beam_queue = BeamQueue(max_size=beam_size)
665
+ # originally the beam search tree starts with a single node (a 3-tuple):
666
+ beam_queue.add(
667
+ node=(g, string, p.txt()), val=0.0 # value of the root node is simply 0.
668
+ )
669
+
670
+ pool = None
671
+ if _N_WORKSERS.value == 1:
672
+ bqsearch_init(0)
673
+ else:
674
+ # Default is 'fork' on Linux, does not work with CUDA. Need to use 'spawn' or 'forkserver'
675
+ multiprocessing.set_start_method('spawn')
676
+ pool = multiprocessing.Pool(_N_WORKSERS.value)
677
+
678
+ logging.info("Initializing workers")
679
+ wkrpids = pool.map(bqsearch_init, range(_N_WORKSERS.value))
680
+ logging.info("Worker PIDs: " + str(wkrpids))
681
+
682
+ for depth in range(search_depth):
683
+ logging.info(
684
+ 'Depth %s. There are %i nodes to expand:', depth, len(beam_queue)
685
+ )
686
+ for _, (_, string, _) in beam_queue:
687
+ logging.info(string)
688
+
689
+ new_queue = BeamQueue(max_size=beam_size) # to replace beam_queue.
690
+ if _N_WORKSERS.value==1:
691
+ for i, srch_inputs in enumerate(beam_queue):
692
+ _, solved, res = bqsearch(i, srch_inputs, out_file)
693
+ if solved:
694
+ return True
695
+ for node, val in res:
696
+ # Add the candidate to the beam queue.
697
+ new_queue.add(node, val)
698
+ # Note that the queue only maintain at most beam_size nodes
699
+ # so this new node might possibly be dropped depending on its value.
700
+ else:
701
+ jobs = [pool.apply_async(bqsearch, (i, srch_inputs, out_file)) for i, srch_inputs in enumerate(beam_queue)]
702
+
703
+ n_done = 0
704
+ while n_done < len(beam_queue):
705
+ for i, jobres in enumerate(jobs):
706
+ if jobres and jobres.ready():
707
+ n_done += 1
708
+ jobs[i] = None
709
+ _, solved, res = jobres.get()
710
+ if solved:
711
+ # Clean up resources
712
+ pool.terminate()
713
+ pool.join()
714
+ return True
715
+ for node, val in res:
716
+ # Add the candidate to the beam queue.
717
+ new_queue.add(node, val)
718
+ # Note that the queue only maintain at most beam_size nodes
719
+ # so this new node might possibly be dropped depending on its value.
720
+ time.sleep(1) # Adjust wait time as needed
721
+
722
+ # replace the old queue with new queue before the new proof search depth.
723
+ beam_queue = new_queue
724
+
725
+ # Clean up resources
726
+ if pool:
727
+ pool.terminate()
728
+ pool.join()
729
+ return False
730
+
731
+ def main(_):
732
+ global DEFINITIONS
733
+ global RULES
734
+
735
+ # definitions of terms used in our domain-specific language.
736
+ DEFINITIONS = pr.Definition.from_txt_file(_DEFS_FILE.value, to_dict=True)
737
+ # load inference rules used in DD.
738
+ RULES = pr.Theorem.from_txt_file(_RULES_FILE.value, to_dict=True)
739
+
740
+ # when using the language model,
741
+ # point names will be renamed to alphabetical a, b, c, d, e, ...
742
+ # instead of staying with their original names,
743
+ # in order to match the synthetic training data generation.
744
+ need_rename = _MODE.value != 'ddar'
745
+
746
+ # load problems from the problems_file,
747
+ problems = pr.Problem.from_txt_file(
748
+ _PROBLEMS_FILE.value, to_dict=True, translate=need_rename
749
+ )
750
+
751
+ if _PROBLEM_NAME.value not in problems:
752
+ raise ValueError(
753
+ f'Problem name `{_PROBLEM_NAME.value}` '
754
+ + f'not found in `{_PROBLEMS_FILE.value}`'
755
+ )
756
+
757
+ this_problem = problems[_PROBLEM_NAME.value]
758
+
759
+ if _MODE.value == 'ddar':
760
+ g, _ = gh.Graph.build_problem(this_problem, DEFINITIONS)
761
+ run_ddar(g, this_problem, _OUT_FILE.value)
762
+
763
+ elif _MODE.value == 'alphageometry':
764
+ #XX model = get_lm(_CKPT_PATH.value, _VOCAB_PATH.value)
765
+ run_alphageometry(
766
+ #XX model,
767
+ this_problem,
768
+ _SEARCH_DEPTH.value,
769
+ _BEAM_SIZE.value,
770
+ _OUT_FILE.value,
771
+ )
772
+
773
+ else:
774
+ raise ValueError(f'Unknown FLAGS.mode: {_MODE.value}')
775
+
776
+
777
+ if __name__ == '__main__':
778
+ app.run(main)
download.sh ADDED
@@ -0,0 +1,31 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ HOME=/home/user
2
+ AG4MDIR=$HOME/ag4masses
3
+ # Directory containing external libraries including ag_ckpt_vocab and meliad
4
+ AGLIB=$HOME/aglib
5
+
6
+ AGDIR=$AG4MDIR/alphageometry
7
+
8
+ DATA=$AGLIB/ag_ckpt_vocab
9
+ MELIAD_PATH=$AGLIB/meliad
10
+
11
+ cd /home/user
12
+ mkdir aglib
13
+ cd aglib
14
+ git clone https://github.com/tpgh24/ag4masses.git
15
+ cd alphageometry
16
+ git clone https://github.com/google-research/meliad.git
17
+
18
+ cd ..
19
+ # mkdir aglib
20
+ gdown --folder https://bit.ly/alphageometry
21
+ export DATA=ag_ckpt_vocab
22
+
23
+ pip cache purge
24
+ pip install --upgrade pip
25
+ pip install --upgrade packaging setuptools setuptools_scm wheel
26
+ pip install --require-hashes --no-deps -r /home/user/ag4masses/alphageometry/requirements.txt
27
+
28
+ # some patch for cpu
29
+ cp models.py /home/user/ag4masses/alphageometry/models.py
30
+ cp alphageometry.py /home/user/ag4masses/alphageometry/alphageometry.py
31
+ cp ag4masses-public.ipynb /data/ag4masses-public.ipynb
models.py ADDED
@@ -0,0 +1,178 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2023 DeepMind Technologies Limited
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+ # ==============================================================================
15
+
16
+ """Transformer language model generate mode."""
17
+
18
+ from typing import Any, Tuple
19
+ import beam_search
20
+ import decoder_stack
21
+ import gin
22
+ import jax
23
+ import jax.numpy as jnp
24
+ from transformer import models
25
+
26
+
27
+ @gin.configurable
28
+ class DecoderOnlyLanguageModelGenerate(models.DecoderOnlyLanguageModel):
29
+ """Decoder only language modeling in inference mode."""
30
+
31
+ decoder_factory = decoder_stack.DecoderStackGenerate
32
+
33
+ num_heads: int = gin.REQUIRED
34
+ head_size: int = gin.REQUIRED
35
+
36
+ def get_fake_input(self) -> dict[str, Any]:
37
+ fake_input_dict = super().get_fake_input()
38
+ b = self.task_config.batch_size
39
+ n = self.num_heads
40
+ h = self.head_size
41
+ fake_input_dict.update({
42
+ 'dstate': tuple(
43
+ [{
44
+ 'current_index': jnp.array([0] * b, dtype=jnp.int32),
45
+ 'keys': jnp.zeros((b, 2048, n, h), dtype=jnp.float32),
46
+ 'values': jnp.zeros((b, 2048, n, h), dtype=jnp.float32),
47
+ 'recurrent_kvq': None,
48
+ 'relative_position_bias': jnp.zeros(
49
+ (b, n, 1, 1024), dtype=jnp.float32
50
+ ),
51
+ }]
52
+ * 12
53
+ ),
54
+ 'eos': jnp.zeros([1024], dtype=jnp.float32),
55
+ 'mask': jnp.ones([1024], dtype=jnp.float32),
56
+ 'length': 1,
57
+ 'temperature': 1.0,
58
+ })
59
+ return fake_input_dict
60
+
61
+ def __call__(self, inputs: ...) -> tuple[Any, dict[str, Any]]:
62
+ # Make sure this code is not used on untested cases.
63
+ if self.mode not in ['init', 'beam_search']:
64
+ raise ValueError(f'{type(self)} cannot do mode {self.mode}')
65
+ if self.decoder.supports_generate():
66
+ raise ValueError(f'{type(self)}.decoder cannot supports_generate()')
67
+
68
+ self.decoder(
69
+ input_tokens=inputs['targets'][:, 0:1],
70
+ target_tokens=None,
71
+ start_of_sequence=inputs['start_of_sequence'],
72
+ )
73
+
74
+ b = inputs['targets'].shape[0]
75
+ no_start_of_seq = jnp.array([False] * b, dtype=jnp.bool_)
76
+
77
+ # This fn is used in both beam_search or topk_sampling.
78
+ def tokens_to_logits_fn(
79
+ input_token: jnp.ndarray, dstate: tuple[dict[str, jnp.ndarray], ...]
80
+ ) -> tuple[jnp.ndarray, tuple[dict[str, jnp.ndarray], ...]]:
81
+ (logits, dstate, _) = self.decoder(
82
+ input_tokens=input_token,
83
+ target_tokens=None,
84
+ start_of_sequence=no_start_of_seq,
85
+ decoder_state=dstate,
86
+ )
87
+ return logits[:, -1, :], dstate
88
+
89
+ last_token = jax.lax.dynamic_slice_in_dim(
90
+ inputs['targets'], inputs['length'] - 1, 1, axis=1
91
+ )
92
+
93
+ # last token is used to seed beam_search
94
+ inputs['targets'] = inputs['targets'][:, 0:-1]
95
+ dstate = jax.lax.cond(
96
+ inputs['start_of_sequence'][0],
97
+ lambda: self.generate(inputs)[0],
98
+ lambda: inputs['dstate'],
99
+ )
100
+
101
+ # Then we run beam search, init with last_token & dstate.
102
+ finished_seqs, finished_scores, dstate = beam_search.beam_search_flat(
103
+ last_token,
104
+ dstate,
105
+ tokens_to_logits_fn,
106
+ max_decode_len=512,
107
+ eos=inputs['eos'].reshape((1, 1, -1)),
108
+ mask=inputs['mask'].reshape((1, 1, -1)),
109
+ )
110
+
111
+ return 0.0, {
112
+ 'finished_seqs': finished_seqs,
113
+ 'finished_scores': finished_scores,
114
+ 'dstate': dstate,
115
+ }
116
+
117
+ def generate(
118
+ self, inputs: ...
119
+ ) -> tuple[tuple[dict[str, jnp.ndarray, ...], ...], jnp.ndarray]:
120
+ """Generate an output sequence.
121
+
122
+ Args:
123
+ inputs: the same as argument to _call_.
124
+
125
+ Returns:
126
+ An array of generated tokens of shape (batch_size, sequence_length).
127
+ """
128
+ input_tokens = inputs['targets'] # [b,seq_len]
129
+ start_of_sequence = inputs['start_of_sequence'] # [b]
130
+ target_tokens = jnp.pad(input_tokens[:, 1:], [(0, 0), (0, 1)])
131
+ batch_size = target_tokens.shape[0]
132
+
133
+ # Assuming all sequences start at the same time.
134
+ start0 = inputs['start_of_sequence'][0]
135
+ dstate = jax.lax.cond(
136
+ start0,
137
+ lambda: self.decoder.init_decoder_state_vanilla( # pylint: disable=g-long-lambda
138
+ 1024, start_of_sequence
139
+ ),
140
+ lambda: inputs['dstate'],
141
+ )
142
+
143
+ first_token = input_tokens[:, 0:1]
144
+ no_start_of_seq = jnp.array([False] * batch_size, dtype=jnp.bool_)
145
+ temperature = 1
146
+ if 'temperature' in inputs:
147
+ temperature = inputs['temperature']
148
+
149
+ num_steps = inputs['length']
150
+ if self.mode == 'beam_search':
151
+ num_steps -= 1
152
+
153
+ def cond_fn(scan_state) -> jnp.bool_:
154
+ _, _, i, _ = scan_state
155
+ return i < num_steps
156
+
157
+ def loop_fn(scan_state: Any) -> Tuple[Any, Any, Any, Any]:
158
+ (dstate, input_token, i, _) = scan_state
159
+
160
+ (logits, dstate, _) = self.decoder(
161
+ input_tokens=input_token,
162
+ target_tokens=None,
163
+ start_of_sequence=no_start_of_seq,
164
+ decoder_state=dstate,
165
+ )
166
+
167
+ logits = logits / temperature
168
+ output_token = jax.lax.dynamic_slice_in_dim(target_tokens, i, 1, axis=1)
169
+
170
+ return (dstate, output_token, i + 1, logits)
171
+
172
+ # Scan over the sequence length.
173
+ dummy_logits = jnp.zeros((batch_size, 1, 1024))
174
+ initial_scan_state = (dstate, first_token, 0, dummy_logits)
175
+ dstate, _, _, logits = jax.lax.while_loop(
176
+ cond_fn, loop_fn, initial_scan_state
177
+ )
178
+ return dstate, logits