/session/quantization/quantize-constituency-model.ipynb

https://github.com/huseinzol05/Malaya · Jupyter · 471 lines · 471 code · 0 blank · 0 comment · 0 complexity · 620ff2d6942eb1c7423a043c77a10a75 MD5 · raw file 

    

  1. {
    2. 

  2.  "cells": [
    3. 

  3.   {
    4. 

  4.    "cell_type": "code",
    5. 

  5.    "execution_count": 1,
    6. 

  6.    "metadata": {},
    7. 

  7.    "outputs": [],
    8. 

  8.    "source": [
    9. 

  9.     "import os\n",
    10. 

  10.     "os.environ['CUDA_VISIBLE_DEVICES'] = ''"
    11. 

  11.    ]
    12. 

  12.   },
    13. 

  13.   {
    14. 

  14.    "cell_type": "code",
    15. 

  15.    "execution_count": 2,
    16. 

  16.    "metadata": {},
    17. 

  17.    "outputs": [],
    18. 

  18.    "source": [
    19. 

  19.     "S3_PATH_CONSTITUENCY = {\n",
    20. 

  20.     "    'bert': {\n",
    21. 

  21.     "        'model': 'v38/constituency/bert-base.pb',\n",
    22. 

  22.     "        'dictionary': 'v38/constituency/vocab-bert-base.json',\n",
    23. 

  23.     "        'vocab': 'tokenizer/sp10m.cased.bert.vocab',\n",
    24. 

  24.     "        'tokenizer': 'tokenizer/sp10m.cased.bert.model',\n",
    25. 

  25.     "    },\n",
    26. 

  26.     "    'tiny-bert': {\n",
    27. 

  27.     "        'model': 'v38/constituency/tiny-bert.pb',\n",
    28. 

  28.     "        'dictionary': 'v38/constituency/vocab-tiny-bert.json',\n",
    29. 

  29.     "        'vocab': 'tokenizer/sp10m.cased.bert.vocab',\n",
    30. 

  30.     "        'tokenizer': 'tokenizer/sp10m.cased.bert.model',\n",
    31. 

  31.     "    },\n",
    32. 

  32.     "#     'albert': {\n",
    33. 

  33.     "#         'model': 'v38/constituency/albert-base.pb',\n",
    34. 

  34.     "#         'dictionary': 'v38/constituency/vocab-albert-base.json',\n",
    35. 

  35.     "#         'vocab': 'tokenizer/sp10m.cased.v10.vocab',\n",
    36. 

  36.     "#         'tokenizer': 'tokenizer/sp10m.cased.v10.model',\n",
    37. 

  37.     "#     },\n",
    38. 

  38.     "#     'tiny-albert': {\n",
    39. 

  39.     "#         'model': 'v38/constituency/albert-tiny.pb',\n",
    40. 

  40.     "#         'dictionary': 'v38/constituency/vocab-albert-tiny.json',\n",
    41. 

  41.     "#         'vocab': 'tokenizer/sp10m.cased.v10.vocab',\n",
    42. 

  42.     "#         'tokenizer': 'tokenizer/sp10m.cased.v10.model',\n",
    43. 

  43.     "#     },\n",
    44. 

  44.     "#     'xlnet': {\n",
    45. 

  45.     "#         'model': 'v40/constituency/xlnet-base.pb',\n",
    46. 

  46.     "#         'quantized': 'v40/constituency/xlnet-base.pb.quantized',\n",
    47. 

  47.     "#         'dictionary': 'v40/constituency/vocab-xlnet-base.json',\n",
    48. 

  48.     "#         'vocab': 'tokenizer/sp10m.cased.v9.vocab',\n",
    49. 

  49.     "#         'tokenizer': 'tokenizer/sp10m.cased.v9.model',\n",
    50. 

  50.     "#     },\n",
    51. 

  51.     "}"
    52. 

  52.    ]
    53. 

  53.   },
    54. 

  54.   {
    55. 

  55.    "cell_type": "code",
    56. 

  56.    "execution_count": 3,
    57. 

  57.    "metadata": {},
    58. 

  58.    "outputs": [
    59. 

  59.     {
    60. 

  60.      "name": "stdout",
    61. 

  61.      "output_type": "stream",
    62. 

  62.      "text": [
    63. 

  63.       "bert\n",
    64. 

  64.       "tiny-bert\n"
    65. 

  65.      ]
    66. 

  66.     }
    67. 

  67.    ],
    68. 

  68.    "source": [
    69. 

  69.     "for k in S3_PATH_CONSTITUENCY.keys():\n",
    70. 

  70.     "    if k not in ['multinomial']:\n",
    71. 

  71.     "        print(k)\n",
    72. 

  72.     "        os.system(f\"wget https://f000.backblazeb2.com/file/malaya-model/{S3_PATH_CONSTITUENCY[k]['model']}\")"
    73. 

  73.    ]
    74. 

  74.   },
    75. 

  75.   {
    76. 

  76.    "cell_type": "code",
    77. 

  77.    "execution_count": 4,
    78. 

  78.    "metadata": {},
    79. 

  79.    "outputs": [],
    80. 

  80.    "source": [
    81. 

  81.     "import tensorflow as tf\n",
    82. 

  82.     "from tensorflow.tools.graph_transforms import TransformGraph\n",
    83. 

  83.     "from glob import glob\n",
    84. 

  84.     "tf.set_random_seed(0)"
    85. 

  85.    ]
    86. 

  86.   },
    87. 

  87.   {
    88. 

  88.    "cell_type": "code",
    89. 

  89.    "execution_count": 5,
    90. 

  90.    "metadata": {},
    91. 

  91.    "outputs": [
    92. 

  92.     {
    93. 

  93.      "data": {
    94. 

  94.       "text/plain": [
    95. 

  95.        "['tiny-bert.pb',\n",
    96. 

  96.        " 'albert-tiny-similarity.pb',\n",
    97. 

  97.        " 'bert-base.pb',\n",
    98. 

  98.        " 'xlnet-base-similarity.pb',\n",
    99. 

  99.        " 'albert-base-similarity.pb',\n",
    100. 

  100.        " 'bert-base-similarity.pb',\n",
    101. 

  101.        " 'alxlnet-base-similarity.pb',\n",
    102. 

  102.        " 'tiny-bert-similarity.pb']"
    103. 

  103.       ]
    104. 

  104.      },
    105. 

  105.      "execution_count": 5,
    106. 

  106.      "metadata": {},
    107. 

  107.      "output_type": "execute_result"
    108. 

  108.     }
    109. 

  109.    ],
    110. 

  110.    "source": [
    111. 

  111.     "pbs = glob('*.pb')\n",
    112. 

  112.     "pbs"
    113. 

  113.    ]
    114. 

  114.   },
    115. 

  115.   {
    116. 

  116.    "cell_type": "code",
    117. 

  117.    "execution_count": 7,
    118. 

  118.    "metadata": {},
    119. 

  119.    "outputs": [
    120. 

  120.     {
    121. 

  121.      "data": {
    122. 

  122.       "text/plain": [
    123. 

  123.        "[(<tf.Tensor 'import/bert/encoder/layer_8/attention/output/LayerNorm/batchnorm/mul_2:0' shape=(?, ?, 312) dtype=float32>,),\n",
    124. 

  124.        " (<tf.Tensor 'import/bert/encoder/layer_8/attention/output/LayerNorm/batchnorm/sub:0' shape=(?, ?, 312) dtype=float32>,),\n",
    125. 

  125.        " (<tf.Tensor 'import/bert/encoder/layer_8/attention/output/LayerNorm/batchnorm/add_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    126. 

  126.        " (<tf.Tensor 'import/bert/encoder/layer_8/intermediate/einsum/Reshape:0' shape=(?, 312) dtype=float32>,),\n",
    127. 

  127.        " (<tf.Tensor 'import/bert/encoder/layer_8/output/einsum/MatMul:0' shape=(?, 312) dtype=float32>,),\n",
    128. 

  128.        " (<tf.Tensor 'import/bert/encoder/layer_8/output/einsum/Reshape_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    129. 

  129.        " (<tf.Tensor 'import/bert/encoder/layer_8/output/add:0' shape=(?, ?, 312) dtype=float32>,),\n",
    130. 

  130.        " (<tf.Tensor 'import/bert/encoder/layer_8/output/add_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    131. 

  131.        " (<tf.Tensor 'import/bert/encoder/layer_8/output/LayerNorm/moments/SquaredDifference:0' shape=(?, ?, 312) dtype=float32>,),\n",
    132. 

  132.        " (<tf.Tensor 'import/bert/encoder/layer_8/output/LayerNorm/batchnorm/mul:0' shape=(?, ?, 312) dtype=float32>,),\n",
    133. 

  133.        " (<tf.Tensor 'import/bert/encoder/layer_8/output/LayerNorm/batchnorm/mul_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    134. 

  134.        " (<tf.Tensor 'import/bert/encoder/layer_8/output/LayerNorm/batchnorm/mul_2:0' shape=(?, ?, 312) dtype=float32>,),\n",
    135. 

  135.        " (<tf.Tensor 'import/bert/encoder/layer_8/output/LayerNorm/batchnorm/sub:0' shape=(?, ?, 312) dtype=float32>,),\n",
    136. 

  136.        " (<tf.Tensor 'import/bert/encoder/layer_8/output/LayerNorm/batchnorm/add_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    137. 

  137.        " (<tf.Tensor 'import/bert/encoder/layer_9/attention/self/query/einsum/Reshape:0' shape=(?, 312) dtype=float32>,),\n",
    138. 

  138.        " (<tf.Tensor 'import/bert/encoder/layer_9/attention/self/query/einsum/MatMul:0' shape=(?, 312) dtype=float32>,),\n",
    139. 

  139.        " (<tf.Tensor 'import/bert/encoder/layer_9/attention/self/key/einsum/Reshape:0' shape=(?, 312) dtype=float32>,),\n",
    140. 

  140.        " (<tf.Tensor 'import/bert/encoder/layer_9/attention/self/key/einsum/MatMul:0' shape=(?, 312) dtype=float32>,),\n",
    141. 

  141.        " (<tf.Tensor 'import/bert/encoder/layer_9/attention/self/value/einsum/Reshape:0' shape=(?, 312) dtype=float32>,),\n",
    142. 

  142.        " (<tf.Tensor 'import/bert/encoder/layer_9/attention/self/value/einsum/MatMul:0' shape=(?, 312) dtype=float32>,),\n",
    143. 

  143.        " (<tf.Tensor 'import/bert/encoder/layer_9/attention/output/einsum/Reshape:0' shape=(?, 312) dtype=float32>,),\n",
    144. 

  144.        " (<tf.Tensor 'import/bert/encoder/layer_9/attention/output/einsum/MatMul:0' shape=(?, 312) dtype=float32>,),\n",
    145. 

  145.        " (<tf.Tensor 'import/bert/encoder/layer_9/attention/output/einsum/Reshape_2:0' shape=(?, ?, 312) dtype=float32>,),\n",
    146. 

  146.        " (<tf.Tensor 'import/bert/encoder/layer_9/attention/output/add:0' shape=(?, ?, 312) dtype=float32>,),\n",
    147. 

  147.        " (<tf.Tensor 'import/bert/encoder/layer_9/attention/output/add_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    148. 

  148.        " (<tf.Tensor 'import/bert/encoder/layer_9/attention/output/LayerNorm/moments/SquaredDifference:0' shape=(?, ?, 312) dtype=float32>,),\n",
    149. 

  149.        " (<tf.Tensor 'import/bert/encoder/layer_9/attention/output/LayerNorm/batchnorm/mul:0' shape=(?, ?, 312) dtype=float32>,),\n",
    150. 

  150.        " (<tf.Tensor 'import/bert/encoder/layer_9/attention/output/LayerNorm/batchnorm/mul_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    151. 

  151.        " (<tf.Tensor 'import/bert/encoder/layer_9/attention/output/LayerNorm/batchnorm/mul_2:0' shape=(?, ?, 312) dtype=float32>,),\n",
    152. 

  152.        " (<tf.Tensor 'import/bert/encoder/layer_9/attention/output/LayerNorm/batchnorm/sub:0' shape=(?, ?, 312) dtype=float32>,),\n",
    153. 

  153.        " (<tf.Tensor 'import/bert/encoder/layer_9/attention/output/LayerNorm/batchnorm/add_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    154. 

  154.        " (<tf.Tensor 'import/bert/encoder/layer_9/intermediate/einsum/Reshape:0' shape=(?, 312) dtype=float32>,),\n",
    155. 

  155.        " (<tf.Tensor 'import/bert/encoder/layer_9/output/einsum/MatMul:0' shape=(?, 312) dtype=float32>,),\n",
    156. 

  156.        " (<tf.Tensor 'import/bert/encoder/layer_9/output/einsum/Reshape_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    157. 

  157.        " (<tf.Tensor 'import/bert/encoder/layer_9/output/add:0' shape=(?, ?, 312) dtype=float32>,),\n",
    158. 

  158.        " (<tf.Tensor 'import/bert/encoder/layer_9/output/add_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    159. 

  159.        " (<tf.Tensor 'import/bert/encoder/layer_9/output/LayerNorm/moments/SquaredDifference:0' shape=(?, ?, 312) dtype=float32>,),\n",
    160. 

  160.        " (<tf.Tensor 'import/bert/encoder/layer_9/output/LayerNorm/batchnorm/mul:0' shape=(?, ?, 312) dtype=float32>,),\n",
    161. 

  161.        " (<tf.Tensor 'import/bert/encoder/layer_9/output/LayerNorm/batchnorm/mul_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    162. 

  162.        " (<tf.Tensor 'import/bert/encoder/layer_9/output/LayerNorm/batchnorm/mul_2:0' shape=(?, ?, 312) dtype=float32>,),\n",
    163. 

  163.        " (<tf.Tensor 'import/bert/encoder/layer_9/output/LayerNorm/batchnorm/sub:0' shape=(?, ?, 312) dtype=float32>,),\n",
    164. 

  164.        " (<tf.Tensor 'import/bert/encoder/layer_9/output/LayerNorm/batchnorm/add_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    165. 

  165.        " (<tf.Tensor 'import/bert/encoder/layer_10/attention/self/query/einsum/Reshape:0' shape=(?, 312) dtype=float32>,),\n",
    166. 

  166.        " (<tf.Tensor 'import/bert/encoder/layer_10/attention/self/query/einsum/MatMul:0' shape=(?, 312) dtype=float32>,),\n",
    167. 

  167.        " (<tf.Tensor 'import/bert/encoder/layer_10/attention/self/key/einsum/Reshape:0' shape=(?, 312) dtype=float32>,),\n",
    168. 

  168.        " (<tf.Tensor 'import/bert/encoder/layer_10/attention/self/key/einsum/MatMul:0' shape=(?, 312) dtype=float32>,),\n",
    169. 

  169.        " (<tf.Tensor 'import/bert/encoder/layer_10/attention/self/value/einsum/Reshape:0' shape=(?, 312) dtype=float32>,),\n",
    170. 

  170.        " (<tf.Tensor 'import/bert/encoder/layer_10/attention/self/value/einsum/MatMul:0' shape=(?, 312) dtype=float32>,),\n",
    171. 

  171.        " (<tf.Tensor 'import/bert/encoder/layer_10/attention/output/einsum/Reshape:0' shape=(?, 312) dtype=float32>,),\n",
    172. 

  172.        " (<tf.Tensor 'import/bert/encoder/layer_10/attention/output/einsum/MatMul:0' shape=(?, 312) dtype=float32>,),\n",
    173. 

  173.        " (<tf.Tensor 'import/bert/encoder/layer_10/attention/output/einsum/Reshape_2:0' shape=(?, ?, 312) dtype=float32>,),\n",
    174. 

  174.        " (<tf.Tensor 'import/bert/encoder/layer_10/attention/output/add:0' shape=(?, ?, 312) dtype=float32>,),\n",
    175. 

  175.        " (<tf.Tensor 'import/bert/encoder/layer_10/attention/output/add_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    176. 

  176.        " (<tf.Tensor 'import/bert/encoder/layer_10/attention/output/LayerNorm/moments/SquaredDifference:0' shape=(?, ?, 312) dtype=float32>,),\n",
    177. 

  177.        " (<tf.Tensor 'import/bert/encoder/layer_10/attention/output/LayerNorm/batchnorm/mul:0' shape=(?, ?, 312) dtype=float32>,),\n",
    178. 

  178.        " (<tf.Tensor 'import/bert/encoder/layer_10/attention/output/LayerNorm/batchnorm/mul_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    179. 

  179.        " (<tf.Tensor 'import/bert/encoder/layer_10/attention/output/LayerNorm/batchnorm/mul_2:0' shape=(?, ?, 312) dtype=float32>,),\n",
    180. 

  180.        " (<tf.Tensor 'import/bert/encoder/layer_10/attention/output/LayerNorm/batchnorm/sub:0' shape=(?, ?, 312) dtype=float32>,),\n",
    181. 

  181.        " (<tf.Tensor 'import/bert/encoder/layer_10/attention/output/LayerNorm/batchnorm/add_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    182. 

  182.        " (<tf.Tensor 'import/bert/encoder/layer_10/intermediate/einsum/Reshape:0' shape=(?, 312) dtype=float32>,),\n",
    183. 

  183.        " (<tf.Tensor 'import/bert/encoder/layer_10/output/einsum/MatMul:0' shape=(?, 312) dtype=float32>,),\n",
    184. 

  184.        " (<tf.Tensor 'import/bert/encoder/layer_10/output/einsum/Reshape_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    185. 

  185.        " (<tf.Tensor 'import/bert/encoder/layer_10/output/add:0' shape=(?, ?, 312) dtype=float32>,),\n",
    186. 

  186.        " (<tf.Tensor 'import/bert/encoder/layer_10/output/add_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    187. 

  187.        " (<tf.Tensor 'import/bert/encoder/layer_10/output/LayerNorm/moments/SquaredDifference:0' shape=(?, ?, 312) dtype=float32>,),\n",
    188. 

  188.        " (<tf.Tensor 'import/bert/encoder/layer_10/output/LayerNorm/batchnorm/mul:0' shape=(?, ?, 312) dtype=float32>,),\n",
    189. 

  189.        " (<tf.Tensor 'import/bert/encoder/layer_10/output/LayerNorm/batchnorm/mul_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    190. 

  190.        " (<tf.Tensor 'import/bert/encoder/layer_10/output/LayerNorm/batchnorm/mul_2:0' shape=(?, ?, 312) dtype=float32>,),\n",
    191. 

  191.        " (<tf.Tensor 'import/bert/encoder/layer_10/output/LayerNorm/batchnorm/sub:0' shape=(?, ?, 312) dtype=float32>,),\n",
    192. 

  192.        " (<tf.Tensor 'import/bert/encoder/layer_10/output/LayerNorm/batchnorm/add_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    193. 

  193.        " (<tf.Tensor 'import/bert/encoder/layer_11/attention/self/query/einsum/Reshape:0' shape=(?, 312) dtype=float32>,),\n",
    194. 

  194.        " (<tf.Tensor 'import/bert/encoder/layer_11/attention/self/query/einsum/MatMul:0' shape=(?, 312) dtype=float32>,),\n",
    195. 

  195.        " (<tf.Tensor 'import/bert/encoder/layer_11/attention/self/key/einsum/Reshape:0' shape=(?, 312) dtype=float32>,),\n",
    196. 

  196.        " (<tf.Tensor 'import/bert/encoder/layer_11/attention/self/key/einsum/MatMul:0' shape=(?, 312) dtype=float32>,),\n",
    197. 

  197.        " (<tf.Tensor 'import/bert/encoder/layer_11/attention/self/value/einsum/Reshape:0' shape=(?, 312) dtype=float32>,),\n",
    198. 

  198.        " (<tf.Tensor 'import/bert/encoder/layer_11/attention/self/value/einsum/MatMul:0' shape=(?, 312) dtype=float32>,),\n",
    199. 

  199.        " (<tf.Tensor 'import/bert/encoder/layer_11/attention/output/einsum/Reshape:0' shape=(?, 312) dtype=float32>,),\n",
    200. 

  200.        " (<tf.Tensor 'import/bert/encoder/layer_11/attention/output/einsum/MatMul:0' shape=(?, 312) dtype=float32>,),\n",
    201. 

  201.        " (<tf.Tensor 'import/bert/encoder/layer_11/attention/output/einsum/Reshape_2:0' shape=(?, ?, 312) dtype=float32>,),\n",
    202. 

  202.        " (<tf.Tensor 'import/bert/encoder/layer_11/attention/output/add:0' shape=(?, ?, 312) dtype=float32>,),\n",
    203. 

  203.        " (<tf.Tensor 'import/bert/encoder/layer_11/attention/output/add_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    204. 

  204.        " (<tf.Tensor 'import/bert/encoder/layer_11/attention/output/LayerNorm/moments/SquaredDifference:0' shape=(?, ?, 312) dtype=float32>,),\n",
    205. 

  205.        " (<tf.Tensor 'import/bert/encoder/layer_11/attention/output/LayerNorm/batchnorm/mul:0' shape=(?, ?, 312) dtype=float32>,),\n",
    206. 

  206.        " (<tf.Tensor 'import/bert/encoder/layer_11/attention/output/LayerNorm/batchnorm/mul_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    207. 

  207.        " (<tf.Tensor 'import/bert/encoder/layer_11/attention/output/LayerNorm/batchnorm/mul_2:0' shape=(?, ?, 312) dtype=float32>,),\n",
    208. 

  208.        " (<tf.Tensor 'import/bert/encoder/layer_11/attention/output/LayerNorm/batchnorm/sub:0' shape=(?, ?, 312) dtype=float32>,),\n",
    209. 

  209.        " (<tf.Tensor 'import/bert/encoder/layer_11/attention/output/LayerNorm/batchnorm/add_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    210. 

  210.        " (<tf.Tensor 'import/bert/encoder/layer_11/intermediate/einsum/Reshape:0' shape=(?, 312) dtype=float32>,),\n",
    211. 

  211.        " (<tf.Tensor 'import/bert/encoder/layer_11/output/einsum/MatMul:0' shape=(?, 312) dtype=float32>,),\n",
    212. 

  212.        " (<tf.Tensor 'import/bert/encoder/layer_11/output/einsum/Reshape_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    213. 

  213.        " (<tf.Tensor 'import/bert/encoder/layer_11/output/add:0' shape=(?, ?, 312) dtype=float32>,),\n",
    214. 

  214.        " (<tf.Tensor 'import/bert/encoder/layer_11/output/add_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    215. 

  215.        " (<tf.Tensor 'import/bert/encoder/layer_11/output/LayerNorm/moments/SquaredDifference:0' shape=(?, ?, 312) dtype=float32>,),\n",
    216. 

  216.        " (<tf.Tensor 'import/bert/encoder/layer_11/output/LayerNorm/batchnorm/mul:0' shape=(?, ?, 312) dtype=float32>,),\n",
    217. 

  217.        " (<tf.Tensor 'import/bert/encoder/layer_11/output/LayerNorm/batchnorm/mul_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    218. 

  218.        " (<tf.Tensor 'import/bert/encoder/layer_11/output/LayerNorm/batchnorm/mul_2:0' shape=(?, ?, 312) dtype=float32>,),\n",
    219. 

  219.        " (<tf.Tensor 'import/bert/encoder/layer_11/output/LayerNorm/batchnorm/sub:0' shape=(?, ?, 312) dtype=float32>,),\n",
    220. 

  220.        " (<tf.Tensor 'import/bert/encoder/layer_11/output/LayerNorm/batchnorm/add_1:0' shape=(?, ?, 312) dtype=float32>,),\n",
    221. 

  221.        " (<tf.Tensor 'import/Reshape:0' shape=(?, 312) dtype=float32>,),\n",
    222. 

  222.        " (<tf.Tensor 'import/GatherV2:0' shape=(?, 312) dtype=float32>,)]"
    223. 

  223.       ]
    224. 

  224.      },
    225. 

  225.      "execution_count": 7,
    226. 

  226.      "metadata": {},
    227. 

  227.      "output_type": "execute_result"
    228. 

  228.     }
    229. 

  229.    ],
    230. 

  230.    "source": [
    231. 

  231.     "with tf.gfile.GFile('tiny-bert.pb', \"rb\") as f:\n",
    232. 

  232.     "    graph_def = tf.GraphDef()\n",
    233. 

  233.     "    graph_def.ParseFromString(f.read())\n",
    234. 

  234.     "\n",
    235. 

  235.     "with tf.Graph().as_default() as graph:\n",
    236. 

  236.     "    tf.import_graph_def(graph_def)\n",
    237. 

  237.     "\n",
    238. 

  238.     "op = graph.get_operations()\n",
    239. 

  239.     "x = []\n",
    240. 

  240.     "for i in op:\n",
    241. 

  241.     "    try:\n",
    242. 

  242.     "        if i.values()[0].shape[-1] == 312:\n",
    243. 

  243.     "        #if 'batchnorm/add_1' in i.values()[0].name:\n",
    244. 

  244.     "            x.append(i.values())\n",
    245. 

  245.     "    except:\n",
    246. 

  246.     "        pass\n",
    247. 

  247.     "    \n",
    248. 

  248.     "x[-100:]"
    249. 

  249.    ]
    250. 

  250.   },
    251. 

  251.   {
    252. 

  252.    "cell_type": "code",
    253. 

  253.    "execution_count": 33,
    254. 

  254.    "metadata": {},
    255. 

  255.    "outputs": [],
    256. 

  256.    "source": [
    257. 

  257.     "mapping = {'albert-base.pb': 'import/bert/encoder/transformer/group_0_11/layer_11/inner_group_0/LayerNorm_1/batchnorm/add_1:0',\n",
    258. 

  258.     "          'albert-tiny.pb': 'import/bert/encoder/transformer/group_0_3/layer_3/inner_group_0/LayerNorm_1/batchnorm/add_1:0',\n",
    259. 

  259.     "          'bert-base.pb': 'import/bert/encoder/layer_11/output/LayerNorm/batchnorm/add_1:0',\n",
    260. 

  260.     "          'tiny-bert.pb': 'import/bert/encoder/layer_11/output/LayerNorm/batchnorm/add_1:0',\n",
    261. 

  261.     "          'xlnet-base.pb': 'import/model/transformer/layer_11/ff/LayerNorm/batchnorm/add_1:0'}"
    262. 

  262.    ]
    263. 

  263.   },
    264. 

  264.   {
    265. 

  265.    "cell_type": "code",
    266. 

  266.    "execution_count": 35,
    267. 

  267.    "metadata": {},
    268. 

  268.    "outputs": [
    269. 

  269.     {
    270. 

  270.      "data": {
    271. 

  271.       "text/plain": [
    272. 

  272.        "'bert/encoder/transformer/group_0_11/layer_11/inner_group_0/LayerNorm_1/batchnorm/add_1'"
    273. 

  273.       ]
    274. 

  274.      },
    275. 

  275.      "execution_count": 35,
    276. 

  276.      "metadata": {},
    277. 

  277.      "output_type": "execute_result"
    278. 

  278.     }
    279. 

  279.    ],
    280. 

  280.    "source": [
    281. 

  281.     "mapping[pbs[0]].replace('import/','').replace(':0','')"
    282. 

  282.    ]
    283. 

  283.   },
    284. 

  284.   {
    285. 

  285.    "cell_type": "code",
    286. 

  286.    "execution_count": 37,
    287. 

  287.    "metadata": {},
    288. 

  288.    "outputs": [],
    289. 

  289.    "source": [
    290. 

  290.     "transforms = ['add_default_attributes',\n",
    291. 

  291.     "             'remove_nodes(op=Identity, op=CheckNumerics, op=Dropout)',\n",
    292. 

  292.     "             'fold_batch_norms',\n",
    293. 

  293.     "             'fold_old_batch_norms',\n",
    294. 

  294.     "             'quantize_weights(fallback_min=-10, fallback_max=10)',\n",
    295. 

  295.     "             'strip_unused_nodes',\n",
    296. 

  296.     "             'sort_by_execution_order']\n",
    297. 

  297.     "\n",
    298. 

  298.     "for pb in pbs:\n",
    299. 

  299.     "    input_graph_def = tf.GraphDef()\n",
    300. 

  300.     "    with tf.gfile.FastGFile(pb, 'rb') as f:\n",
    301. 

  301.     "        input_graph_def.ParseFromString(f.read())\n",
    302. 

  302.     "        \n",
    303. 

  303.     "    a = [mapping[pb].replace('import/','').replace(':0','')]\n",
    304. 

  304.     "    \n",
    305. 

  305.     "    transformed_graph_def = TransformGraph(input_graph_def, \n",
    306. 

  306.     "                                           ['input_ids', 'word_end_mask'],\n",
    307. 

  307.     "                                           ['charts', 'tags'] + a, transforms)\n",
    308. 

  308.     "    \n",
    309. 

  309.     "    with tf.gfile.GFile(f'{pb}.quantized', 'wb') as f:\n",
    310. 

  310.     "        f.write(transformed_graph_def.SerializeToString())"
    311. 

  311.    ]
    312. 

  312.   },
    313. 

  313.   {
    314. 

  314.    "cell_type": "code",
    315. 

  315.    "execution_count": 38,
    316. 

  316.    "metadata": {},
    317. 

  317.    "outputs": [],
    318. 

  318.    "source": [
    319. 

  319.     "def load_graph(frozen_graph_filename, **kwargs):\n",
    320. 

  320.     "    with tf.gfile.GFile(frozen_graph_filename, 'rb') as f:\n",
    321. 

  321.     "        graph_def = tf.GraphDef()\n",
    322. 

  322.     "        graph_def.ParseFromString(f.read())\n",
    323. 

  323.     "\n",
    324. 

  324.     "    # https://github.com/onnx/tensorflow-onnx/issues/77#issuecomment-445066091\n",
    325. 

  325.     "    # to fix import T5\n",
    326. 

  326.     "    for node in graph_def.node:\n",
    327. 

  327.     "        if node.op == 'RefSwitch':\n",
    328. 

  328.     "            node.op = 'Switch'\n",
    329. 

  329.     "            for index in xrange(len(node.input)):\n",
    330. 

  330.     "                if 'moving_' in node.input[index]:\n",
    331. 

  331.     "                    node.input[index] = node.input[index] + '/read'\n",
    332. 

  332.     "        elif node.op == 'AssignSub':\n",
    333. 

  333.     "            node.op = 'Sub'\n",
    334. 

  334.     "            if 'use_locking' in node.attr:\n",
    335. 

  335.     "                del node.attr['use_locking']\n",
    336. 

  336.     "        elif node.op == 'AssignAdd':\n",
    337. 

  337.     "            node.op = 'Add'\n",
    338. 

  338.     "            if 'use_locking' in node.attr:\n",
    339. 

  339.     "                del node.attr['use_locking']\n",
    340. 

  340.     "        elif node.op == 'Assign':\n",
    341. 

  341.     "            node.op = 'Identity'\n",
    342. 

  342.     "            if 'use_locking' in node.attr:\n",
    343. 

  343.     "                del node.attr['use_locking']\n",
    344. 

  344.     "            if 'validate_shape' in node.attr:\n",
    345. 

  345.     "                del node.attr['validate_shape']\n",
    346. 

  346.     "            if len(node.input) == 2:\n",
    347. 

  347.     "                node.input[0] = node.input[1]\n",
    348. 

  348.     "                del node.input[1]\n",
    349. 

  349.     "\n",
    350. 

  350.     "    with tf.Graph().as_default() as graph:\n",
    351. 

  351.     "        tf.import_graph_def(graph_def)\n",
    352. 

  352.     "    return graph"
    353. 

  353.    ]
    354. 

  354.   },
    355. 

  355.   {
    356. 

  356.    "cell_type": "code",
    357. 

  357.    "execution_count": 39,
    358. 

  358.    "metadata": {},
    359. 

  359.    "outputs": [
    360. 

  360.     {
    361. 

  361.      "data": {
    362. 

  362.       "text/plain": [
    363. 

  363.        "['albert-base.pb.quantized',\n",
    364. 

  364.        " 'xlnet-base.pb.quantized',\n",
    365. 

  365.        " 'albert-tiny.pb.quantized',\n",
    366. 

  366.        " 'tiny-bert.pb.quantized',\n",
    367. 

  367.        " 'bert-base.pb.quantized']"
    368. 

  368.       ]
    369. 

  369.      },
    370. 

  370.      "execution_count": 39,
    371. 

  371.      "metadata": {},
    372. 

  372.      "output_type": "execute_result"
    373. 

  373.     }
    374. 

  374.    ],
    375. 

  375.    "source": [
    376. 

  376.     "quantized = glob('*.pb.quantized')\n",
    377. 

  377.     "quantized"
    378. 

  378.    ]
    379. 

  379.   },
    380. 

  380.   {
    381. 

  381.    "cell_type": "code",
    382. 

  382.    "execution_count": null,
    383. 

  383.    "metadata": {},
    384. 

  384.    "outputs": [],
    385. 

  385.    "source": [
    386. 

  386.     "# converter = tf.compat.v1.lite.TFLiteConverter.from_frozen_graph(\n",
    387. 

  387.     "#     graph_def_file='test.pb',\n",
    388. 

  388.     "#     input_arrays=['Placeholder', 'Placeholder_1'],\n",
    389. 

  389.     "#     input_shapes={'Placeholder' : [None, 512], 'Placeholder_1': [None, 512]},\n",
    390. 

  390.     "#     output_arrays=['logits'],\n",
    391. 

  391.     "# )\n",
    392. 

  392.     "# # converter.allow_custom_ops=True"
    393. 

  393.    ]
    394. 

  394.   },
    395. 

  395.   {
    396. 

  396.    "cell_type": "code",
    397. 

  397.    "execution_count": null,
    398. 

  398.    "metadata": {},
    399. 

  399.    "outputs": [],
    400. 

  400.    "source": [
    401. 

  401.     "# converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS, tf.lite.OpsSet.SELECT_TF_OPS]\n",
    402. 

  402.     "# converter.target_spec.supported_types = [tf.float16]\n",
    403. 

  403.     "# converter.optimizations = [tf.lite.Optimize.DEFAULT]\n",
    404. 

  404.     "# converter.experimental_new_converter = True\n",
    405. 

  405.     "# tflite_model = converter.convert()"
    406. 

  406.    ]
    407. 

  407.   },
    408. 

  408.   {
    409. 

  409.    "cell_type": "code",
    410. 

  410.    "execution_count": null,
    411. 

  411.    "metadata": {},
    412. 

  412.    "outputs": [],
    413. 

  413.    "source": [
    414. 

  414.     "# converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS, \n",
    415. 

  415.     "#                                        tf.lite.OpsSet.SELECT_TF_OPS]\n",
    416. 

  416.     "# converter.target_spec.supported_types = [tf.float16]\n",
    417. 

  417.     "# converter.optimizations = [tf.lite.Optimize.DEFAULT]\n",
    418. 

  418.     "# tflite_model = converter.convert()\n",
    419. 

  419.     "\n",
    420. 

  420.     "# with open('tiny-bert-sentiment-float16.tflite', 'wb') as f:\n",
    421. 

  421.     "#     f.write(tflite_model)"
    422. 

  422.    ]
    423. 

  423.   },
    424. 

  424.   {
    425. 

  425.    "cell_type": "code",
    426. 

  426.    "execution_count": null,
    427. 

  427.    "metadata": {},
    428. 

  428.    "outputs": [],
    429. 

  429.    "source": [
    430. 

  430.     "# converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS, \n",
    431. 

  431.     "#                                        tf.lite.OpsSet.SELECT_TF_OPS]\n",
    432. 

  432.     "# converter.optimizations = [tf.lite.Optimize.OPTIMIZE_FOR_SIZE]\n",
    433. 

  433.     "# tflite_model = converter.convert()\n",
    434. 

  434.     "\n",
    435. 

  435.     "# with open('tiny-bert-sentiment-hybrid.tflite', 'wb') as f:\n",
    436. 

  436.     "#     f.write(tflite_model)"
    437. 

  437.    ]
    438. 

  438.   },
    439. 

  439.   {
    440. 

  440.    "cell_type": "code",
    441. 

  441.    "execution_count": null,
    442. 

  442.    "metadata": {},
    443. 

  443.    "outputs": [],
    444. 

  444.    "source": [
    445. 

  445.     "# interpreter = tf.lite.Interpreter(model_path='tiny-bert-sentiment-hybrid.tflite')\n",
    446. 

  446.     "# interpreter.allocate_tensors()"
    447. 

  447.    ]
    448. 

  448.   }
    449. 

  449.  ],
    450. 

  450.  "metadata": {
    451. 

  451.   "kernelspec": {
    452. 

  452.    "display_name": "Python 3",
    453. 

  453.    "language": "python",
    454. 

  454.    "name": "python3"
    455. 

  455.   },
    456. 

  456.   "language_info": {
    457. 

  457.    "codemirror_mode": {
    458. 

  458.     "name": "ipython",
    459. 

  459.     "version": 3
    460. 

  460.    },
    461. 

  461.    "file_extension": ".py",
    462. 

  462.    "mimetype": "text/x-python",
    463. 

  463.    "name": "python",
    464. 

  464.    "nbconvert_exporter": "python",
    465. 

  465.    "pygments_lexer": "ipython3",
    466. 

  466.    "version": "3.6.8"
    467. 

  467.   }
    468. 

  468.  },
    469. 

  469.  "nbformat": 4,
    470. 

  470.  "nbformat_minor": 2
    471. 

  471. }
    472.