search

facebookresearch / deepcluster

Deep Clustering for Unsupervised Learning of Visual Features
Other
1.69k stars 325 forks source link

Time for validating #18

Closed pvskand closed 5 years ago

pvskand commented 5 years ago

Hi there. Could you mention the time taken for the validation step (the one where you train mlp on imagenet data) with only conv1 freezed features on Pascal P100?

pvskand commented 5 years ago

@mathildecaron31 could you answer the above query?

mathildecaron31 commented 5 years ago

Hi, During the validation step, no gradient is computed. Hence, this step is the same regardless of which parameters are freezed.

I don't get what experiment precisely you are referring to. Is it: 1- Re-training the MLP on ImageNet classification task, all conv parameters freezed (experiment of the supplementary material); 2- Training a linear classifier on ImageNet classification task on top of conv1 features ?

pvskand commented 5 years ago

Hi, I'll rephrase my question. By validation I meant the validation of the model trained in unsupervised fashion (via clustering). So I wanted the time taken for convergence of 2 ( Training of linear classifier on conv1 features).

pvskand commented 5 years ago

@mathildecaron31 any updates on the aforementioned question?

mathildecaron31 commented 5 years ago

Hi, I ran 90 epochs for convergence; I don't have the log specifying the time taken precisely on hand so I'll need to re-run the experiment to be able to reply.

mathildecaron31 commented 5 years ago

Hi, I have finally reproduced the experiment: it takes 130 sec per training epoch and it takes 50 epochs to converge to a final precision (central crop) of 12.9 for conv1. Thus this experiment on a single P100 gpu takes a total time of roughly 1h50.