Egor Beliaev

Built histogram of predictions (red, on 0.3 subset of entire set  ) similar to the one in the article:  It seems this model does not predict mean scores...

Btw, how did you get such good variance? I've trained for 50 epochs in tensorflow with the same settings (lr=10-4), imagenet mean and stddev - but it's still a spike:...

I did not succeed in training with SGD with lr=10-7 for base net and 10-6 for last layer, as authors. It just does not converge. I try to use Adam...

They seem to use Adam with following parameters: ``` "batch_size": 96, "epochs_train_dense": 5, "learning_rate_dense": 0.001, "decay_dense": 0, "epochs_train_all": 9, "learning_rate_all": 0.00003, "decay_all": 0.000023, ``` Did not study closely though. Will...

This keras implementation: https://github.com/truskovskiyk/nima.pytorch They do adjusting images with ImageNet mean and variance, and use Adam with lr=1e-4.  Histogram is built on 0.3 subset of entire set.

UPD: I found that the change of image size happens in this place of `dataset_mapper.py`: ``` def __call__(self, dataset_dict): dataset_dict = copy.deepcopy(dataset_dict) image = utils.read_image(dataset_dict["file_name"], format=self.img_format) utils.check_image_size(dataset_dict, image) sh1 =...

This dataset [https://drive.google.com/file/d/1vgUp88Co0Rzoxk4aNwJwTa44zFlIGRRu/view?usp=sharing](https://drive.google.com/file/d/1vgUp88Co0Rzoxk4aNwJwTa44zFlIGRRu/view?usp=sharing) LIP dataset seems to be found here: [https://github.com/Engineering-Course/LIP_SSL](https://github.com/Engineering-Course/LIP_SSL) [https://drive.google.com/drive/folders/0BzvH3bSnp3E9QjVYZlhWSjltSWM?resourcekey=0-nkS8bDVjPs3bEw3UZW-omA&usp=sharing](https://drive.google.com/drive/folders/0BzvH3bSnp3E9QjVYZlhWSjltSWM?resourcekey=0-nkS8bDVjPs3bEw3UZW-omA&usp=sharing)

This is celeba model which you disabled somewhy: