-
Notifications
You must be signed in to change notification settings - Fork 56
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
Mobilenet model trained on casia dataset #116
Comments
It could be rather hard reaching any satisfactory result using mobilenet +casia only, even not possible... I'm previously using RTX8000 with 46GB GPU memory.
data_path = "faces_webface_112x112_112x112_folders"
eval_paths = ["faces_webface_112x112/lfw.bin"]
basic_model = models.buildin_models("MobileNet", dropout=0, emb_shape=256, output_layer="E")
tt = train.Train(data_path, save_path='mobilenet_256_adam_E.h5',
eval_paths=eval_paths,
basic_model=basic_model,
batch_size=512, random_status=0,
lr_base=0.001, lr_decay=0.5, lr_decay_steps=16, lr_min=1e-5)
optimizer = keras.optimizers.Adam(learning_rate=0.001, weight_decay=5e-4)
sch = [
{"loss": losses.ArcfaceLoss(scale=16), "epoch": 10, "optimizer": optimizer},
{"loss": losses.ArcfaceLoss(scale=32), "epoch": 10},
{"loss": losses.ArcfaceLoss(scale=64), "epoch": 30},
]
tt.train(sch, 0) |
First of all this repo is really helpful for me and it has helped me a lot , so thank you for this and more over thank you for replying to me . Wow 46 gb vram is impressive probably that is why you were able to use large batch sizes. and I feel the main bottleneck is using casia dataset in my case , as you have got better results with mobilenet when trained on MS1M dataset. |
I was trying to train a mobilenet model using arcloss on casia dataset and I am unable to exceed lfw acc more than ~0.9763.
I wanna know is this the best accuracy I can get using this dataset(casia) ? as training on ms1m is not possible for me as it is really large.
Also as I was checking your training scripts I saw that you are using large batch size and you have done a lot of experiments, didn't those took a lot of this ? what hardware did you used to train them?
I am mainly using google colab and kaggle for training .
training code :
training logs:
The text was updated successfully, but these errors were encountered: