tf pb 轉 tfjs 將固定大小的輸入 改爲任意輸入

訓練時使用固定大小, 方便編程實現和速度優化

部署時使用任意大小, 提高體驗

但是採用固定大小輸入做訓練, 部署時採用任意大小, 可能效果有點差別吧....

 

pb文件的網絡結構

1. import tensorflow as tf

2.  

3. output_graph_def = tf.GraphDef()

4. PB_PATH = r"./pb/feathers.pb"

5.  

6. with open(PB_PATH, "rb") as f:

7. output_graph_def.ParseFromString(f.read())

8. tf.import_graph_def(

9. output_graph_def,

10. name='', # 默認name爲import,類似scope

11. )

12.  

13. with tf.Session() as sess:

14. sess.run(tf.global_variables_initializer())

15. tf.summary.FileWriter('./log/', sess.graph)

tensorboard.exe --logdir .

將固定輸入大小的pb文件,轉化爲任意輸入大小的tfjs格式 

1. import tensorflow as tf

2. from tensorflow.python.framework import graph_util

3. import tensorflowjs as tfjs

4.  

5. sess = tf.Session()

6. output_graph_def = tf.GraphDef()

7. # feathers starry candy

8. PB_PATH = r"./pb/candy.pb"

9. TFJS_PATH = r'./tfjs/candy'

10.  

11. in_image = tf.placeholder(tf.float32, (None, None, None, 3), name='in_x')

12. with open(PB_PATH, "rb") as f:

13. output_graph_def.ParseFromString(f.read())

14. tf.import_graph_def(

15. output_graph_def,

16. input_map={

17. 'in_x:0': in_image

18. },

19. name='', # 默認name爲import,類似scope

20. # return_elements=['generator/mul:0']

21. )

22. sess.run(tf.global_variables_initializer())

23. output = sess.graph.get_tensor_by_name("generator/output:0")

24.  

25. with tf.Session() as sess:

26. sess.run(tf.global_variables_initializer())

27. constant_graph = graph_util.convert_variables_to_constants(sess, sess.graph_def, ['generator/output'])

28. with tf.gfile.FastGFile("./pb/tmp.pb", mode='wb') as f:

29. f.write(constant_graph.SerializeToString())

30.  

31. tfjs.converters.tf_saved_model_conversion_pb.convert_tf_frozen_model(

32. "./pb/tmp.pb",

33. 'generator/output',

34. TFJS_PATH

35. )