-
Notifications
You must be signed in to change notification settings - Fork 228
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Makes the image retrieval part work without too much effort. (should address issues #64 and #109) #156
Makes the image retrieval part work without too much effort. (should address issues #64 and #109) #156
Conversation
under vg_coco_id_to_capgraphs ). As the solution relies on Spacy Parser, and as we don t know which version was used, the results might differ.
Own metrics
I added a script that shows how the sentence scene graphs were created (can''t recreate exactly the same text graph, as this is based on Spacy Parsing and it has changed since the original version). I also added an evaluation script for image retrieval with another logic. PS: I believe, that could have achieve better performance, if you had trained for only 9 / 10 epochs. I am using your code for a project and some results can be found here: |
I have made some modifications to the code to be able to run the image retrieval part with the pre-trained SGDet Model you uploaded. With proper pre-configuration, the only hard coded path are in tools/image_retrieval_main.py.
The main changes:
The evaluator uses normalized cosine similarity for the scoring.
There s some handling for bad training samples (samples with no relationships)
Some input Tensors needed to be transposed to enable the network to train properly.