What I tried so far:
pre-train a model using unsupervised method in PyTorch, and save off the checkpoint file (using torch.save(state, filename))
convert the checkpoint file to onnx format (using torch.onnx.export)
convert the onnx to tensorflow saved model (using onnx-tf)
trying to load the variables in saved_model folder as checkpoint in my tensorflow training code (using tf.train.init_from_checkpoint) for fine-tuning
But now I am getting stuck at step 4 because I notice that variables.index and variables.data#1 files are basically empty (probably because of this: https://github.com/onnx/onnx-tensorflow/issues/994)
Also, specifically, if I try to use tf.train.NewCheckpointReader to load the files and call ckpt_reader.get_variable_to_shape_map(), _CHECKPOINTABLE_OBJECT_GRAPH is empty
Any suggestions/experience are appreciated :-)
Related
I am aiming to inference tensorflow slim model with Intel OpenVINO optimizer. Using open vino docs and slides for inference and tf slim docs for training model.
It's a multi-class classification problem. I have trained tf slim mobilnet_v2 model from scratch (using sript train_image_classifier.py). Evaluation of trained model on test set gives relatively good results to begin with (using script eval_image_classifier.py):
eval/Accuracy[0.8017]eval/Recall_5[0.9993]
However, single .ckpt file is not saved (even though at the end of train_image_classifier.py run there is a message like "model.ckpt is saved to checkpoint_dir"), there are 3 files (.ckpt-180000.data-00000-of-00001, .ckpt-180000.index, .ckpt-180000.meta) instead.
OpenVINO model optimizer requires a single checkpoint file.
According to docs I call mo_tf.py with following params:
python mo_tf.py --input_model D:/model/mobilenet_v2_224.pb --input_checkpoint D:/model/model.ckpt-180000 -b 1
It gives the error (same if pass --input_checkpoint D:/model/model.ckpt):
[ ERROR ] The value for command line parameter "input_checkpoint" must be existing file/directory, but "D:/model/model.ckpt-180000" does not exist.
Error message is clear, there are not such files on disk. But as I know most tf utilities convert .ckpt-????.meta to .ckpt under the hood.
Trying to call:
python mo_tf.py --input_model D:/model/mobilenet_v2_224.pb --input_meta_graph D:/model/model.ckpt-180000.meta -b 1
Causes:
[ ERROR ] Unknown configuration of input model parameters
It doesn't matter for me in which way I will transfer graph to OpenVINO intermediate representation, just need to reach that result.
Thanks a lot.
EDIT
I managed to run OpenVINO model optimizer on frozen graph of tf slim model. However I still have no idea why had my previous attempts (based on docs) failed.
you can try converting the model to frozen format (.pb) and then convert the model using OpenVINO.
.ckpt-meta has the metagraph. The computation graph structure without variable values.
the one you can observe in tensorboard.
.ckpt-data has the variable values,without the skeleton or structure. to restore a model we need both meta and data files.
.pb file saves the whole graph (meta+data)
As per the documentation of OpenVINO:
When a network is defined in Python* code, you have to create an inference graph file. Usually, graphs are built in a form that allows model training. That means that all trainable parameters are represented as variables in the graph. To use the graph with the Model Optimizer, it should be frozen.
https://software.intel.com/en-us/articles/OpenVINO-Using-TensorFlow
the OpenVINO optimizes the model by converting the weighted graph passed in frozen form.
Anyone knows how to save tensorflow keras checkpoint file into .pb binary file for serving?
I am not 100% sure how to implement this.
Do I need to declare tf.keras.backend.get_session() to save graph and pb file?
In sum,
1. Method to save pb file via tf.keras
or
2. Any approaches using tf.keras.estimator.model_to_estimator(keras_model=model)
Thanks
Here is a link on how to use keras with tensorflow. Basically you should get the tensorflow model from keras (K.get_session() and K.get_graph().graph) and use the tensorflow methotd to save it.
I used Keras to build a model and trained it. Then I saved the model as an h5 file, i.e. model.save('name.h5'). Now I want to reload the model in tensorflow such that I have access to .meta file, for example I want to import the computational graph from the .meta file, i.e., tf.train.import_meta_graph('name_of_the_file.meta').
So, the question is how to convert .h5 file of Keras to the following four files of TensorFlow:
.meta
checkpoint
.data-00000-of-00001
.index
You can use 3rd party packages, for example keras_to_tensorflow
keras_to_tensorflow: General code to convert a trained keras model into an inference tensorflow model
The conversion can be done by
python3 keras_to_tensorflow.py -input_model_file model.h5
Tensorflow 2.x will do that automatically. The function you are using to save (see also) is:
save(
filepath,
overwrite=True,
include_optimizer=True,
save_format=None
)
The save format let's you choose either 'h5' or 'tf'. However, for tensorflow 1.x is not implemented yet (and probably never will).
save_format: Either 'tf' or 'h5', indicating whether to save the model
to Tensorflow SavedModel or HDF5. The default is currently 'h5', but
will switch to 'tf' in TensorFlow 2.0. The 'tf' option is currently
disabled (use tf.keras.experimental.export_saved_model instead).
You can do as it says and use the tf.keras.experimental.export_saved_model but it will still not create the .meta file.
I am new to keras and tensorflow.
I have trained a model using keras and saved it using saved_model_builder.SavedModelBuilder
I have .pb file and variables folder with two files variables.index and variables.data-00000-of-00001
Now I am trying to freeze the graph and variables but getting following error
google.protobuf.text_format.ParseError: 1:1 : Expected identifier or number, got.
please help
I want to train my own Word2Vec model for my text corpus. I can get the code from TensorFlow's tutorial. What I don't know is how to save this model to use for CNN text classification later? Should I use pickle to save it and then read it later?
No pickling is not the way of saving the model in case of tensorflow.
Tensorflow provides with tensorflow serving for saving the models as proto bufs(for exporting the model). The way to save model would be to save the tensorflow session as:
saver.save(sess, 'my_test_model',global_step=1000)
Heres the link for complete answer:
Tensorflow: how to save/restore a model?
You can use pickle to save it to disk. Then when you are creating the CNN model, load the saved word embedding table and use it to initialize the TensorFlow variable that holds the word embeddings for your CNN classifier.