6

I've retrained my model using tensorflow and now want to use keras to avoid session stuffs. How can I convert .pb file to .h5

Sagar P. Ghagare
  • 542
  • 2
  • 12
  • 25
Avinash Kumar
  • 298
  • 3
  • 13

2 Answers2

4
import tensorflow as tf
from tensorflow.keras.models import save_model, Sequential

model_path = r"c:\temp\model.pb"

model = tf.keras.models.load_model(model_path)

save_model(model,model_path + r"\new_model.h5", save_format='h5')
nivpeled
  • 1,810
  • 5
  • 17
  • 1
    Although this code might solve the problem, a good answer should explain **what** the code does and **how** it helps. – BDL Jun 23 '20 at 13:53
  • @BDL generally I tend to agree, but in this specific case I am pretty sure the code is self explanatory. i.e. "tf.keras.models.load_model" will load a tensor flow model, while save_model with argument save_format='h5' will, hmmm, save it as an h5 keras file. – nivpeled Jun 24 '20 at 06:22
  • @nivpeled I tried same code and got AttributeError: '_UserObject' object has no attribute '_is_graph_network' – H.H Aug 09 '20 at 14:55
  • I get OSError: SavedModel file does not exist at: saved_model_dir/{saved_model.pbtxt|saved_model.pb} error – AnupamChugh Aug 29 '20 at 08:50
  • AttributeError: 'AutoTrackable' object has no attribute '_is_graph_network' .I am also getting this error, have you solved? – maryam mehboob Jul 01 '21 at 14:29
  • model_path should be the path of directory not path of file so should use like "c:\temp" – Ali Aug 28 '21 at 21:25
1

I found this link useful: https://backstreetcoder.com/convert-tensorflow-pb-model-to-keras-h5-model/

Apparently, you need to use the enclosing directory name and not the .pb filename when loading the model initially. This solves the error:

SavedModel file does not exist at: saved_model_dir/{saved_model.pbtxt|saved_model.pb}
Nnamdi Affia
  • 11
  • 5