Even after running all notebooks, the metrics one does not display good results

[victoraccete]

Hi!

I ran all the notebooks, but I can't seem to find a way to make the metrics notebook display good results. We are interested in this work, but we are not able to make it work so far. Here's an image of the results:

Would be possible to help us out with this?

Thanks in advance!

[mcartagenah]

Hi, I just pushed a little fix in the training notebook, the cluster_ids for the visual similarity sampling strategies were incorrect.

Let me know if you still have any problems :)

[victoraccete]

Hi, @mcartagenah. I'm still having the same problems. I suppose it doesn't happen to you at all. The metrics do not improve and display very weird results.

I had to do a few changes in the notebooks in order to run here, maybe these changes may be related to this error I'm having? In the training notebook I changed to MODEL_PATH to hold the absolute path.

Before I did this, the training cell (the one with the train_network function) would never load the previously trained model, always training from the start again.
In a similar way, in the precomputation notebook, it would not restore the checkpoint, the function tf.train.latest_checkpoint(MODEL_PATH) was returning None. So in the precomputation notebook I also changed the MODEL_PATH to hold the absolute path.

Then it could successfully restore the parameters. The evaluation notebook ran just fine, but the metrics notebook, even though it ran without errors, did not display good results.

I can send you further information if needed.

Seems like your fix helped me with another problem I was having, so thank you! But I'm still having this other problem. Do you have any ideas on what can be done to fix it?

Again, thanks in advance!