Search code examples
pythontensorflowtensorboard

How can I select which checkpoint to view in TensorBoard's embeddings tab?

Short question: How can I select which checkpoint to view in TensorBoard's embeddings tab?

Longer version of the question:

I want to visualize word embeddings with TensorBoard. To that end, after reading the official tutorial (mirror) I added following code: 

embedding_writer = tf.summary.FileWriter(model_folder)     
embeddings_projector_config = projector.ProjectorConfig()
embedding = embeddings_projector_config.embeddings.add()
embedding.tensor_name = model.W.name # W corresponds to the embeddings' weights.
projector.visualize_embeddings(embedding_writer, embeddings_projector_config)

# Initialize the model
sess.run(tf.global_variables_initializer())

[...]

# Then, for each training epoch:
model_saver.save(sess, os.path.join(model_folder, 'model_{0:05d}.ckpt'.format(epoch_number)))

Looking at the folder where TensorFlow saves the log, I do have a checkpoint for each epoch:

enter image description here

However, in TensorBoard's embeddings tab it seems that I can only view the latest checkpoint:

enter image description here

I sometimes would like to view the embeddings for previous epochs. How can I select which checkpoint to view in TensorBoard's embeddings tab?

Solution

  • I'm one of the engineers working on the embedding visualizer. Thanks for the feedback. We are planning to add a dropdown menu in the UI that allows you to choose different checkpoints.

    In the meantime, there is a workaround. You can edit the projector_config.pbtxt that lives in the folder where TensorBoard saves the log. I'm assuming the contents of projector_config.pbtxt are:

    embeddings {
...
}

    Append the following line at the end of the file:

    model_checkpoint_path: "path_to_log_dir/model_0000N.ckpt"

    pointing to the exact checkpoint you want to visualize, and remove (if it exists) the line model_checkpoint_dir: "...". Then refresh the page (and potentially re-run TensorBoard).

    For example, if you have launched TensorBoard with tensorboard --logdir=output, and the model checkpoint absolute path is C:\Users\a\output\en_2017-03-08_17-42-09-310106\model\model_00004.ckpt, then you should append to projector_config.pbtxt:

    model_checkpoint_path: "output\en_2017-03-08_17-42-09-310106\model\model_00004.ckpt"

    Example of projector_config.pbtxt:

    embeddings {
  tensor_name: "token_embedding/W:0"
}
model_checkpoint_path: "output\en_2017-03-08_17-42-09-310106\model\model_00004.ckpt"

    If when you click to the embeddings tab in TensorBoard nothing appears, it means that the model_checkpoint_path you have entered is incorrect.

    Hope this helps!