20161119-meganeco

「TensorFlowで趣味の画像収集サーバーを作る 11月号」
大量の画像からデータセットを作成するために構築したワークフローについて解説します。

Transcript

1. C-LIS CO., LTD. 5FOTPS'MPXͰझຯͷը૾ऩूαʔόʔΛ࡞Δ
   ݄߸ BCDB
2. C-LIS CO., LTD.  ༗ࢁ
   ܓೋ
   ʢ,FJKJ
   "3*:"."ʣ $-*4
   $0
   -5%
   "OESPJEΞϓϦ։ൃνϣοτσΩϧ
   ػցֶशॳ৺ऀ Photo : Koji MORIGUCHI (AUN CREATIVE FIRM) ΍ͬͯ·ͤΜ
3.  ݄೔
   ʙ https://devfestkansai.connpass.com/event/42864/
4. C-LIS CO., LTD.  (%(
   ,PCFʢʣ
   (%(
   ,ZPUPʢʣ
   ೔ຊ"OESPJEͷձ
   "OESPJE
   #B[BBS
   BOE
   $POGFSFODFʢ
   ʣ
   (%(
   ,PCFʢ
   ڞ࠵
   ($16#
   0TBLBʣ
   5FOTPS'MPXษڧձʢ̏ʣʢ
   
   
   ڞ࠵೔ຊ"OESPJEͷձʣ
   5FOTPS'MPXษڧձʢ̐ʣʢ
   
   ʣ
   ͳʹΘ5&$)ಓʢ
   
   ʣ
   5FOTPS'MPXษڧձʢʣʢ
   
   ʣ
   ೔ຊ"OESPJEͷձ
   "OESPJE
   #B[BBS
   BOE
   $POGFSFODFʢ
   ʣ ˡ
   /FXʂ
5.  10% 1SJOU
   0O
   %FNBOE
   
   ిࢠॻ੶ Impress R&D͔Βൃചத http://amzn.to/2axRog
6. C-LIS CO., LTD. 
7.  ޷Έͷʰ؟ڸ່ͬʱը૾Λ
   ࣗಈͰऩू͍ͨ͠
8. 4PGUXBSF
   %FTJHO
   ೥݄߸
   ʰ"OESPJE
   4UVEJPͷελΠϧͰޮ཰Ξοϓʱ
   ΑΓ © ࠜઇΕ͍ ؟ ڸ ͬ ່
9.  ̎Ϋϥε෼ྨ  
10.  ؟ڸ່ͬσʔληοτ
11. ؟ڸ່ͬը૾
    ຕ
    ඇ؟ڸ່ͬ
    ຕ
    ܭɿ ຕ  ೥݄ ؟ڸ່ͬ ඇ؟ڸ່ͬ ޡݕग़ ؟ڸ່ͬ ඇ؟ڸ່ͬ
12. ؟ڸ່ͬը૾
     ຕ
    ඇ؟ڸ່ͬ
     ຕ
    ܭɿ ຕ  ೥݄ ؟ڸ່ͬ ඇ؟ڸ່ͬ
13. ؟ڸ່ͬը૾
     ຕ
    ඇ؟ڸ່ͬ
     ຕ
    ܭɿ ຕ  ೥݄ ؟ڸ່ͬ ඇ؟ڸ່ͬ
14.  ݱࡏͷߏ੒ Downloader Face Detection Megane Detection API TensorFlow ֬ೝɾमਖ਼ ೝࣝ݁Ռ JSON ܇࿅ λΠϜϥΠϯ ϝσΟΞ σʔληοτ ܇࿅
15.  %PXOMPBEFS Downloader Face Detection Megane Detection API TensorFlow ֬ೝɾमਖ਼ ೝࣝ݁Ռ JSON ܇࿅ λΠϜϥΠϯ ϝσΟΞ σʔληοτ ܇࿅
16. w͋Β͔͡Ίొ࿥ͨ͠5XJUUFS
    *%ͷ౤ߘ͢ΔϝσΟΞΛμ΢ϯϩʔυ
    w5XJUUFS
    *%ͱը૾63-ΛσʔλϕʔεͰ؅ཧ͢Δ͜ͱͰɺ
 ը૾ͷݖརऀΛ֬ೝՄೳʹ  %PXOMPBEFS
17. wը૾ͷॏෳ͕සൟʹൃੜͯ͠͠·͏ɻ
    ఆظతͳ޿ࠂπΠʔτ
    ιʔγϟϧήʔϜͷϓϨΠը໘
    େચɹͳͲ  ՝୊
18.  JNBHF)BTI 67d85827b2396276 = = =
19.  'BDF
    %FUFDUJPO Downloader Face Detection Megane Detection API TensorFlow ֬ೝɾमਖ਼ ೝࣝ݁Ռ JSON ܇࿅ λΠϜϥΠϯ ϝσΟΞ σʔληοτ ܇࿅
20.  © ࠜઇΕ͍ Sliding Window Selective Search
21.  إೝࣝ༻ͷϞσϧ DPO
    YY DPO
    YY GD
     QPPM
    Y DPO
    YY DPO
    YY QPPM
    Y GD
     PVUQVU
     MSO MSO Y
22.  ܇࿅ 12ສεςοϓֶशޙͷςετσʔλʹΑΔਖ਼౴཰ 96%
23.  .FHBOF
    %FUFDUJPO Downloader Face Detection Megane Detection API TensorFlow ֬ೝɾमਖ਼ ೝࣝ݁Ռ JSON ܇࿅ λΠϜϥΠϯ ϝσΟΞ σʔληοτ ܇࿅
24.  ؟ڸೝࣝ༻ͷϞσϧ DPO
    YY DPO
    YY GD
     QPPM
    Y DPO
    YY DPO
    YY QPPM
    Y GD
     PVUQVU
     MSO MSO Y
25.  ܇࿅ 8ສεςοϓֶशޙͷςετσʔλʹΑΔਖ਼౴཰ 96%
26. C-LIS CO., LTD. 
27.  ೝࣝ݁Ռ Downloader Face Detection Megane Detection API TensorFlow ֬ೝɾमਖ਼ ೝࣝ݁Ռ JSON ܇࿅ λΠϜϥΠϯ ϝσΟΞ σʔληοτ ܇࿅
28. { "generator": "Megane Co", "file_name": "neyuki_rei-CZY6tnUUYAAJVe6.jpg", "created_at": "2016-11-18T10:19:04.879281", "detected_faces": { "mode": "selective_search", "model_version": "20161118003546" "configurations": { }, "regions": [ { "rect": { "top": 59.0, "bottom": 175.0, "left": 85.0, "right": 176.0 }, "label": 2, "probability": 0.9962707757949829 },  ೝࣝ݁Ռͷ+40/ { "rect": { "top": 220.0, "bottom": 274.0, "left": 455.0, "right": 524.0 }, "label": 1, "probability": 0.9997575879096985 }, { "rect": { "top": 196.0, "bottom": 282.0, "left": 244.0, "right": 330.0 }, "label": 1, "probability": 1.0 }
29.  ೝࣝ݁Ռͷ֬ೝͱमਖ਼ Downloader Face Detection Megane Detection API TensorFlow ֬ೝɾमਖ਼ ೝࣝ݁Ռ JSON ܇࿅ λΠϜϥΠϯ ϝσΟΞ σʔληοτ ܇࿅
30. C-LIS CO., LTD. https://github.com/keiji/region_cropper © ࠜઇΕ͍
31.  ೝࣝ݁Ռͷ֬ೝͱमਖ਼ Downloader Face Detection Megane Detection API TensorFlow ֬ೝɾमਖ਼ ೝࣝ݁Ռ JSON ܇࿅ λΠϜϥΠϯ ϝσΟΞ σʔληοτ ܇࿅
32.  ཷ·Δະॲཧը૾ ॲཧࡁΈɿ20,000ຕ ະॲཧɿ 24,000ຕ
33.  ສ݅Λ௒͑ͨ͋ͨΓͰ
 ํ਑Λมߋ
34.  ೝࣝʹࣦഊͨ͠ը૾͚ͩ
    σʔληοτ΁௥Ճ
35.  Ұ൪ࠨͷ؟ڸ່ͬͷإΛೝࣝͰ͖͍ͯͳ͍
    © ࠜઇΕ͍
36.  ؟ڸΛ͔͚͍ͯͳ͍ͷʹ؟ڸ່ͬͱೝࣝ
    © ࠜઇΕ͍
37.  إͱೝࣝ͢Δ৔ॴ͕ؒҧ͍ͬͯΔ
    ؟ڸ່ͬͱೝ͍ࣝͯ͠Δ © ࠜઇΕ͍
38.  σʔληοτ
    ສຕʹ౸ୡ
39. C-LIS CO., LTD. 
40. C-LIS CO., LTD. 
41. C-LIS CO., LTD.  ࢲͷ
    σʔληοτ͸
    ສઍͰ͢ɻ
42.  ͗͢ΌʔΜϝϞ http://memo.sugyan.com/
43. C-LIS CO., LTD. 
44. C-LIS CO., LTD.  ࢲͷ
    σʔληοτ͸
    ສઍͰ͢ɻ
45.  ΋͏ͪΐͬͱ͚ͩ
    ଓ͘Μ͡Ό
46. w؟ڸ່ͬը૾Λൃݟͨ͠Βϝϯγϣϯ͢ΔॲཧΛ࣮૷
    wإೝࣝͷਫ਼౓޲্
    ϞσϧͷܰྔԽ
    4FMFDUJWF4FBSDIͷύϥϝʔλʔௐ੔
    إͷύλʔϯʢਖ਼໘ɾԣإɾԼ޲͖ͳͲʣΛΫϥεʹ෼཭͢Δ
    wΠϥετإೝࣝͷ"1*Խ
    w؟ڸೝࣝͷΫϥεΛ૿΍͢  ࠓޙͷ՝୊
47.  ؟ڸೝࣝͷΫϥεΛ૿΍͢ʁ ԑͳ͠ ্ԑɾηϧϑϨʔϜ
48.  ҧ͍·͢
49.  ޷Έͷʰ؟ڸ່ͬʱը૾Λ
    ࣗಈͰऩू͍ͨ͠
50.  ๻ʹʰ؟ڸ່ͬʱͷ࿩ͤͨ͞Β
    ௕͘ͳΓ·͢Αʁ
51.  ؟ڸೝࣝͷΫϥεΛ૿΍͢ 4PGUXBSF
    %FTJHO
    ೥݄߸
    ʰ"OESPJE
    4UVEJPͷελΠϧͰޮ཰Ξοϓʱ
    ΑΓ © ࠜઇΕ͍
52.  ؟ڸೝࣝͷΫϥεΛ૿΍͢ © ࠜઇΕ͍ ϥϑஈ֊
53.  ؟ڸೝࣝͷΫϥεΛ૿΍͢ 4PGUXBSF
    %FTJHO
    ೥݄߸
    ʰ"OESPJE
    4UVEJPͷελΠϧͰޮ཰Ξοϓʱ
    ΑΓ © ࠜઇΕ͍
54.  ؟ڸೝࣝͷΫϥεΛ૿΍͢ © ࠜઇΕ͍
55.  ग़ྗ૚Λ௥Ճ͢Δ DPO
    YY DPO
    YY GD
     QPPM
    Y DPO
    YY DPO
    YY QPPM
    Y GD
     PVUQVU
     MSO MSO Y φΠε؟ڸ່ͬʂ ؟ڸΛ͔͚ͨঁࢠ Human
56.  άϥϑͷΤΫεϙʔτ
57. 5FOTPS'MPXͰߏஙͨ͠ʮάϥϑʯͱʮύϥϝʔλʔʯΛͦͷ··ॻ͖ग़͢ɻ
    γϦΞϥΠβʔʹ1SPUPDPM#V⒎FSTΛ࢖༻͍ͯ͠Δɻ  άϥϑͷΤΫεϙʔτͱ͸ʁ
58. "OESPJEͳͲͰಡΈࠐΉ৔߹͸άϥϑͱύϥϝʔλʔΛΤΫεϙʔτ͢Δඞཁ ͕͋ΔʢΒ͍͠ʣɻ
    ࣥචͨ͠ͱ͖͸ɺ܇࿅݁ՌΛ֤ΤϙοΫ͝ͱʹΤΫεϙʔτͯ͠ɺ͋ͱͰҰؾ ʹਖ਼౴཰ΛݟΔͱ͍͏͜ͱΛͨ͠ɻ  4BWFSͱͷҧ͍
59. ܇࿅ࡁΈͷϞσϧʢάϥϑʴύϥϝʔλʔʣ͕͋Δɻ
    άϥϑʹυϩοϓΞ΢τͳͲͷ܇࿅༻ͷॲཧؚ͕·Ε͍ͯΔɻ  લఏ
60. def inference(image_node, keep_prob, batch_size=1):
 conv1_1 = _conv('conv1_1', image_node, [3, 3, 3, 64], 64)
 conv1_2 = _conv('conv1_2', conv1_1, [3, 3, 64, 64], 64)
 pool1 = _pool('pool1', conv1_2, [1, 2, 2, 1])
 
 conv2_1 = _conv('conv2_1', pool1, [3, 3, 64, 128], 128)
 conv2_2 = _conv('conv2_2', conv2_1, [3, 3, 128, 128], 128)
 pool2 = _pool('pool2', conv2_2, [1, 2, 2, 1])
 
 conv3_1 = _conv('conv3_1', pool2, [3, 3, 128, 256], 256)
 conv3_2 = _conv('conv3_2', conv3_1, [3, 3, 256, 256], 256)
 pool3 = _pool('pool3', conv3_2, [1, 2, 2, 1])
  ܇࿅༻ͷάϥϑ reshape = tf.reshape(pool3, [batch_size, -1])
 dim = reshape.get_shape()[1].value
 
 fc4 = _fc('fc4', reshape, [dim, 512], 512, keep_prob)
 fc5 = _fc('fc5', fc4, [512, 256], 256, keep_prob)
 
 # output
 with tf.variable_scope('output') as scope:
 weights = _get_weights(shape=[256, NUM_CLASSES], stddev=1 / 256.0)
 biases = _get_biases([NUM_CLASSES], value=0.0)
 logits = tf.add(tf.matmul(fc5, weights), biases, name='logits')
 
 return logits

61. def _fc(name, input_layer, weights_shape, biases_size, keep_prob):
 with tf.variable_scope(name) as scope:
 weights = _get_weights(shape=weights_shape, stddev=0.05)
 biases = _get_biases([biases_size], value=0.1)
 fc = tf.nn.relu(tf.matmul(input_layer, weights) + biases, name=scope.name)
 fc = tf.nn.dropout(fc, keep_prob)
 return fc
  ܇࿅༻ͷάϥϑ http://goo.gl/forms/2IsX177eJ3CpNKng1 model_vgg_lite_minibatch.py
62. def _fc(name, input_layer, weights_shape, biases_size, keep_prob):
 with tf.variable_scope(name) as scope:
 weights = _get_weights(shape=weights_shape, stddev=0.05)
 biases = _get_biases([biases_size], value=0.1)
 fc = tf.nn.relu(tf.matmul(input_layer, weights) + biases, name=scope.name)
 fc = tf.nn.dropout(fc, keep_prob)
 return fc
  ධՁ༻ͷάϥϑ http://goo.gl/forms/2IsX177eJ3CpNKng1 model_vgg_lite_minibatch_no_dropout.py
63. # coding: UTF-8
 from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
 import tensorflow as tf
 
 from tensorflow.python.framework import graph_util
 
 import model_vgg_lite_minibatch_no_dropout as model
 
 FLAGS = tf.app.flags.FLAGS
 tf.app.flags.DEFINE_string('output_graph_path', None, "ग़ྗ͢ΔάϥϑͷσΟϨΫτϦ")
 tf.app.flags.DEFINE_string('checkpoint_dir', './checkpoints', "νΣοΫϙΠϯτͷσΟϨΫτϦ")
 tf.app.flags.DEFINE_integer('batch_size', 1, "όοναΠζ")
 
64. def main(argv=None):
 # eval༻ͷΤϯτϦϙΠϯτ
 input_image = tf.placeholder(tf.float32, shape=[FLAGS.batch_size, 32, 32, 3], name='input_image')
 logits = model.inference(input_image, batch_size=FLAGS.batch_size)
 
 saver = tf.train.Saver()
 
 with tf.Session() as sess: 
 # ύϥϝʔλʔͷಡΈࠐΈ
 checkpoint = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)
 if checkpoint and checkpoint.model_checkpoint_path:
 saver.restore(sees, checkpoint.model_checkpoint_path)
 # άϥϑʴύϥϝʔλʔͷΤΫεϙʔτ
 constant_graph_def = graph_util.convert_variables_to_constants(
 sess, sess.graph_def, ["output/logits"])
 with tf.gfile.FastGFile(FLAGS.output_graph_path, 'wb') as f:
 f.write(constant_graph_def.SerializeToString()) 
65. $ python3 convert_to_graph.py --checkpoint_dir ./checkpoints \ --output_graph_path ./graphs/exported_graph.pb \ --batch_size 1 Converted 18 variables to const ops.  ධՁ༻άϥϑʴύϥϝʔλʔͷΤΫεϙʔτ
66. with tf.gfile.FastGFile(graph_file, 'rb') as f:
 graph_def = tf.GraphDef()
 graph_def.ParseFromString(f.read())
 _ = tf.import_graph_def(graph_def, name='')
  ධՁଆͰάϥϑΛΠϯϙʔτ
67. predictions = sess.run(logits,
 feed_dict={
 input_image: batch_byte_array,
 labels: batch_label,
 })
  MPHJUTͱMBCFM͔Βਖ਼౴൑ఆ
68. labels = tf.placeholder(tf.int32, shape=[FLAGS.batch_size], name='label')
 input_image = tf.get_default_graph().get_tensor_by_name('input_image:0')
 logits = tf.get_default_graph().get_tensor_by_name('output/logits:0')
 
 top_k_op = tf.nn.in_top_k(logits, labels, 1)
  ΦϖϨʔγϣϯͷऔಘͱάϥϑͷߏங
69.  ࣮ߦ
70. E tensorflow/core/common_runtime/executor.cc:334] Executor failed to create kernel. Invalid argument: NodeDef mentions attr 'Tshape' not in Op<name=Reshape; signature=tensor:T, shape:int32 -> output:T; attr=T:type>; NodeDef: Reshape = Reshape[T=DT_FLOAT, Tshape=DT_INT32, _device="/job:localhost/replica:0/task:0/cpu:0"] (pool3, Reshape/shape) [[Node: Reshape = Reshape[T=DT_FLOAT, Tshape=DT_INT32, _device="/job:localhost/ replica:0/task:0/cpu:0"](pool3, Reshape/shape)]] E tensorflow/core/client/tensor_c_api.cc:485] NodeDef mentions attr 'Tshape' not in Op<name=Reshape; signature=tensor:T, shape:int32 -> output:T; attr=T:type>; NodeDef: Reshape = Reshape[T=DT_FLOAT, Tshape=DT_INT32, _device="/job:localhost/replica:0/task: 0/cpu:0"](pool3, Reshape/shape) [[Node: Reshape = Reshape[T=DT_FLOAT, Tshape=DT_INT32, _device="/job:localhost/ replica:0/task:0/cpu:0"](pool3, Reshape/shape)]]  ಈ͖·ͤΜ
71. ['input_image', 'conv1_1/weights', 'conv1_1/weights/read', 'conv1_1/Conv2D', 'conv1_1/biases', 'conv1_1/biases/read', 'conv1_1/BiasAdd', 'conv1_1/conv1_1', 'conv1_2/weights', 'conv1_2/weights/read', 'conv1_2/Conv2D', 'conv1_2/biases', 'conv1_2/biases/read', 'conv1_2/BiasAdd', 'conv1_2/conv1_2', 'pool1',  ߏ੒͢Δϊʔυ 'conv2_1/weights', 'conv2_1/weights/read', 'conv2_1/Conv2D', 'conv2_1/biases', 'conv2_1/biases/read', 'conv2_1/BiasAdd', 'conv2_1/conv2_1', 'conv2_2/weights', 'conv2_2/weights/read', 'conv2_2/Conv2D', 'conv2_2/biases', 'conv2_2/biases/read', 'conv2_2/BiasAdd', 'conv2_2/conv2_2', 'pool2', 'conv3_1/weights', 'conv3_1/weights/read', 'conv3_1/Conv2D', 'conv3_1/biases', 'conv3_1/biases/read', 'conv3_1/BiasAdd', 'conv3_1/conv3_1', 'conv3_2/weights', 'conv3_2/weights/read', 'conv3_2/Conv2D', 'conv3_2/biases', 'conv3_2/biases/read', 'conv3_2/BiasAdd', 'conv3_2/conv3_2', 'pool3', 'Reshape/shape', 'Reshape', 'fc4/weights', 'fc4/weights/read', 'fc4/biases', 'fc4/biases/read', 'fc4/MatMul', 'fc4/add', 'fc4/fc4', 'fc5/weights', 'fc5/weights/read', 'fc5/biases', 'fc5/biases/read', 'fc5/MatMul', 'fc5/add', 'fc5/fc5', 'output/weights', 'output/weights/read', 'output/biases', 'output/biases/read', 'output/MatMul', 'output/logits']
72. ['train_image', 'input_image', 'conv1_1/weights', 'conv1_1/weights/read', 'conv1_1/Conv2D', 'conv1_1/biases', 'conv1_1/biases/read', 'conv1_1/BiasAdd', 'conv1_1/conv1_1', 'conv1_2/weights', 'conv1_2/weights/read', 'conv1_2/Conv2D', 'conv1_2/biases', 'conv1_2/biases/read', 'conv1_2/BiasAdd', 'conv1_2/conv1_2', 'pool1',  ܇࿅தʹΤΫεϙʔτͨ͠΋ͷͱൺֱ 'conv2_1/weights', 'conv2_1/weights/read', 'conv2_1/Conv2D', 'conv2_1/biases', 'conv2_1/biases/read', 'conv2_1/BiasAdd', 'conv2_1/conv2_1', 'conv2_2/weights', 'conv2_2/weights/read', 'conv2_2/Conv2D', 'conv2_2/biases', 'conv2_2/biases/read', 'conv2_2/BiasAdd', 'conv2_2/conv2_2', 'pool2', 'conv3_1/weights', 'conv3_1/weights/read', 'conv3_1/Conv2D', 'conv3_1/biases', 'conv3_1/biases/read', 'conv3_1/BiasAdd', 'conv3_1/conv3_1', 'conv3_2/weights', 'conv3_2/weights/read', 'conv3_2/Conv2D', 'conv3_2/biases', 'conv3_2/biases/read', 'conv3_2/BiasAdd', 'conv3_2/conv3_2', 'pool3', 'Reshape/shape', 'Reshape', 'fc4/weights', 'fc4/weights/read', 'fc4/biases', 'fc4/biases/read', 'fc4/MatMul', 'fc4/add', 'fc4/fc4', 'fc5/weights', 'fc5/weights/read', 'fc5/biases', 'fc5/biases/read', 'fc5/MatMul', 'fc5/add', 'fc5/fc5', 'output/weights', 'output/weights/read', 'output/biases', 'output/biases/read', 'output/MatMul', 'output/logits']
73.  ݁࿦ Α͘Θ͔Βͳ͍
74.  /FYU ʰ5FOTPS'MPXΛ"OESPJEͰ࣮ߦ͠Α͏ʂʱ ݹ઒ɹ৽ ։ൃτϥοΫ
    ϧʔϜ
    5
75.  ػցֶशϋϯζΦϯ
    
     https://gcpug-tokushima.connpass.com/event/43430/
76. C-LIS CO., LTD. ֤੡඼໊ɾϒϥϯυ໊ɺձ໊ࣾͳͲ͸ɺҰൠʹ֤ࣾͷ঎ඪ·ͨ͸ొ࿥঎ඪͰ͢ɻຊࢿྉதͰ͸ɺ˜ɺšɺäΛׂѪ͍ͯ͠·͢ɻ ຊࢿྉ͸ɺ༗ݶձࣾγʔϦεͷஶ࡞෺Ͱ͢ɻܝࡌ͞Ε͍ͯΔΠϥετ͸ࠜઇΕ͍ͷஶ࡞෺Ͱ͢ɻ
 ಺༰ͷશ෦ɺ·ͨ͸Ұ෦ʹ͍ͭͯɺஶ࡞ऀ͔ΒจॻʹΑΔڐ୚Λಘͣʹෳ੡͢Δ͜ͱ͸ې͡ΒΕ͍ͯ·͢ɻ