Sylvain Lejamble, PhD comments

Results 12 comments of


                                            Sylvain Lejamble, PhD

intermediate data files

You can download the dataset on the website of the contest : https://biendata.com/competition/zhihu/ Here is the dropbox link, I don't know how long it will work : https://www.dropbox.com/s/3sk2yojptodkmb2/ieee_zhihu_cup.rar?dl=0

Question about learning rate

LR decay is all done in keras https://keras.io/api/optimizers/learning_rate_schedules/

Output of Gan Images is just Black

Try de-normalize your image before plotting it !

data.py: why "x_train_s" and "x_ref" are identical?

Should be different, yes ! You can use any image dataset ! It is used to compute descriptiveness loss ! There is an error in this code "An external multi-class...

ValueError: Dimensions must be equal

Are you using custom data ?

Training exception

Exception caused by some images, i don't know why, i'm investigating. I fixed it by casting in loss function ```python def class_loss_regr_fixed_num(y_true, y_pred): num_classes_float = K.cast(num_classes, 'float32') indice_slice = K.cast(4...

AP=1.0 for all but one classes and no boxes drawn on test set

try lowering the bbox_threshold and debug using the initial list of probabilities of predicted bbxs

AP=1.0 for all but one classes and no boxes drawn on test set

`probs` is the list of predicted classes, you should see here some class and their probabilities. `if np.max(P_cls[0, ii, :]) < 0.7 or np.argmax(P_cls[0, ii, :]) == (P_cls.shape[2] - 1):...

convert_from_path returns a blank page to some pdf

Same here !

Use the discriminator of the model after training

Hi, I have a great interest on this feature also !