search

gcucurull / visual-compatibility

Context-Aware Visual Compatibility Prediction (https://arxiv.org/abs/1902.03646)
115 stars 37 forks source link

Migrate to tf v2 compatible #3

Closed zyyang closed 4 years ago

zyyang commented 4 years ago

Migrate to tf v2 compatible.

Tested: Original V1.* training result: { "dataset": "polyvore", "learning_rate": 0.001, "weight_decay": 0.0, "epochs": 4000, "hidden": [350, 350, 350, 350], "dropout": 0.5, "degree": 1, "summaries_dir": "logs/", "support_dropout": 0.15, "write_summary": false, "batch_norm": true, "amz_data": "Men_bought_together", "best_val_score": 0.9276251196861267, "best_epoch": 3954, "best_epoch_train_score": 0.8384166955947876, "best_train_score": 0.840186357498169, "seed": 1572082204 }

V2 comp training result: { "dataset": "polyvore", "learning_rate": 0.001, "weight_decay": 0.0, "epochs": 4000, "hidden": [350, 350, 350, 350], "dropout": 0.5, "degree": 1, "summaries_dir": "logs/", "support_dropout": 0.15, "write_summary": false, "batch_norm": true, "amz_data": "Men_bought_together", "best_val_score": 0.9242312908172607, "best_epoch": 3982, "best_epoch_train_score": 0.8369424939155579, "best_train_score": 0.8399352431297302, "seed": 1572506422 }

gcucurull commented 4 years ago

Thank you so much for this improvement, I am wondering whether to merge this into master or create a new branch for TF 2, this way people can still use the 1.X version if they haven't migrated to TF 2 yet.

What do you think?

zyyang commented 4 years ago

That's what I am planning to do. I need, also, to make sure the test scripts still work, and make it running on real tf 2 api. Let's close this pull request.