Input & output node names about dh3d

[JuanDuGit]

Hi, what's your output of the node list specifically?

[cwlroda]

The link to the text file containing the node list is here (not posting directly as it's quite long)

This is the code I'm running to generate the node list:

with tf.Session(graph=tf.Graph()) as sess:
        # Restore local model
        loader = tf.compat.v1.train.import_meta_graph(
            trained_checkpoint_prefix + ".meta"
        )
        loader.restore(sess, trained_checkpoint_prefix)

        graph_def = tf.get_default_graph().as_graph_def()

        # All nodes to list
        node_list = [n.name for n in graph_def.node]

[JuanDuGit]

Then "pointcloud" exists as a tensor instead of a node in the graph. That's why it does not appear in your node list.
It is not obvious to me what you are going to do with tf.Session, but it's not a proper way to run inference with a model restored from the metagraph as said in this link.
In our evaluation scripts, we use OfflinePreditor from tensorpack to do offline inference:

DH3D/evaluate/local_eval/localdesc_extract.py

Line 81 in b1ec580

predictor = OfflinePredictor(pred_config)

You can refer to its docs for details on how it builds a graph and creates a session, etc .

[cwlroda]

I have read through the tensorpack docs and figured out how to export to SavedModel format, thank you for your help!

Here's the skeleton code if anyone reading this in the future would like to perform the same task.

pred_config = PredictConfig(
    model=DH3D(config),
    session_init=SmartInit(trained_checkpoint_prefix),
    input_names=[
        "pointclouds",
    ],
    output_names=["xyz_feat", "xyz_feat_att"],
)

ModelExporter(pred_config).export_serving(export_dir)