Activity

1. keosaly commented on Mar 29, 2024

   

   keosaly

   on Mar 29, 2024

   Author

   More actions

   my code:

   Define callbacks

   earlystopper = EarlyStopping(monitor="val_CER", patience=50, verbose=1, mode="min")
   checkpoint = ModelCheckpoint(f"{configs.model_path}/model.h5", monitor="val_CER", verbose=1, save_best_only=True, mode="min")
   trainLogger = TrainLogger(configs.model_path)
   tb_callback = TensorBoard(f"{configs.model_path}/logs", update_freq=1)
   reduceLROnPlat = ReduceLROnPlateau(monitor="val_CER", factor=0.9, min_delta=1e-10, patience=20, verbose=1, mode="min")
   model2onnx = Model2onnx(f"{configs.model_path}/model.h5")

   Train the model

   model.fit(
   train_data_provider,
   validation_data=val_data_provider,
   epochs=configs.train_epochs,
   callbacks=[earlystopper, checkpoint, trainLogger, reduceLROnPlat, tb_callback, model2onnx],
   workers=configs.train_workers
   )

   Save training and validation datasets as csv files

   train_data_provider.to_csv(os.path.join(configs.model_path, "train.csv"))
   val_data_provider.to_csv(os.path.join(configs.model_path, "val.csv"))
2. mbahadirk commented on May 7, 2024

   

   mbahadirk

   on May 7, 2024

   More actions

   did you find any solutions on that?
   I'am still working on that

   👍React with 👍1Reacted by maxima120
3. pythonlessons commented on May 16, 2024

   

   pythonlessons

   on May 16, 2024

   Owner

   More actions

   It seems you try to use newest tensorflow version (different from tutorial). Recommend to downgrade tensorflow for compatibility. Otherwise you need to do changes manualy. For example change model.h5 to model.keras in all places. But you may face more serious incompatibilities further
4. Mudr0x commented on Aug 8, 2024

   

   Mudr0x

   on Aug 8, 2024

   Last edited by Mudr0x

   More actions

   In Keras 3, for checkpoint filepath, you need to provide in .keras format only, if you're saving only weights file name should end with .weights.h5, so if you want to save your model in .h5 file in any case, set "save_weights_only=True", and change your flie name to "xxx.weights.h5".
   https://keras.io/guides/migrating_to_keras_3/#saving-a-model-in-the-tf-savedmodel-format

   keras-team/keras-io#1844

   Here's changes for your code:

   earlystopper = EarlyStopping(monitor="val_CER", patience=50, verbose=1, mode="min")
   checkpoint = ModelCheckpoint(f"{configs.model_path}/model.h5", monitor="val_CER", verbose=1, save_best_only=True,, save_weights_only=True, mode="min")
   trainLogger = TrainLogger(configs.model_path)
   tb_callback = TensorBoard(f"{configs.model_path}/logs", update_freq=1)
   reduceLROnPlat = ReduceLROnPlateau(monitor="val_CER", factor=0.9, min_delta=1e-10, patience=20, verbose=1, mode="min")
   model2onnx = Model2onnx(f"{configs.model_path}/model.h5")
   
   model.fit(
   train_data_provider,
   validation_data=val_data_provider,
   epochs=configs.train_epochs,
   callbacks=[earlystopper, checkpoint, trainLogger, reduceLROnPlat, tb_callback, model2onnx]
   )
   
   train_data_provider.to_csv(os.path.join(configs.model_path, "train.csv"))
   val_data_provider.to_csv(os.path.join(configs.model_path, "val.csv"))
   
5. maxima120 commented on Feb 2, 2025

   

   maxima120

   on Feb 2, 2025

   More actions

   I have the same problem. I want to save in "tensorflow" format - which is set of files in a directory. I do not want to use either .keras nor .h5

   I want to save everything not just weights. And I want to use it for both - future training and inference.

   If i save it in .keras format - when I load it for further training or inference, sometimes it throws - optimiser variables mismatch, and doesnt use the saved adam vars.

   Which I want to avoid (tell me this is wrong idea).

   Versions:
   TF Version: 2.18.0
   Python Version: 3.11.2 (main, Nov 30 2024, 21:22:50) [GCC 12.2.0]

   I use tf.keras everywhere not keras directly (if it makes any difference).

   What happened to the "tensorflow" checkpoints? have you removed this feature completely? I dont see anything in the docs indicating it and the samples that supposed to use tensorflow format are throwing the same error (want .keras suffix - in either - checkpoint callback and direct model.save)