av-savchenko / hse_facerec_tf Goto Github PK

@HSE-asavchenko Can a different face detector be used instead of mtcnn?

Hello, I read your article which led me here.
Because at the repo of efficientNet they have graphic which shows their B6, B7 models as far superior to the ResNets in both inference performance and precision, I begun wondering whether this is true. Since you have FaceRec, based on both models, it is interesting how do they compare in your implementation.
I saw that ResNet50 gave you best results for Face identification accuracy, but no results in the "fast-image-recognition" repo are published. Also while I haven't read the code yet, I wonder whether the embedding feature vectors are of the same size and what size did you choose.
Thanks !

Hi,

I have been working on trying to understand the steps for training and obtaining my model for - age_gender_tf2_224_deep-03-0.13-0.97.pb
I followed a previously asked question - #2 (comment)
mentioning the two scripts - facerec_keras_train.py and age_gender_train.py. As per my understanding, the following two step training process is required :

1. Training the baseline model with facerec_keras_train.py:

• Download the VGGFace2 dataset.
• Run the Training Script - Execute facerec_keras_train.py to start the training on VGGFace2.
• (I presume this will give me - vgg2_mobilenet.pb

2. Training the age and gender model with age_gender_train.py:

• Data Preparation: Concatenate IMDB-Wiki and Adience datasets
• Organize data into two main folders: age and gender. Each of them should contain train and val subfolders that include images for training and validation respectively.
• Correct age mistakes manually in IMDB-Wiki and utilize the midpoint of age ranges in Adience as ground truth labels (not sure about this)
• Adjust Paths and Labels : Ensure paths in age_gender_train.py are correctly pointing to the data.
• Run the Training Script: Execute age_gender_train.py

Is my understanding correct?
Also another issue I am facing is - I haven't been able to download VGGFace2 Dataset. While I came across VGGFace2 HQ repo so I have downloaded these - https://github.com/NNNNAI/VGGFace2-HQ but would training on these give a different model backbone model?
Thank you so much in advance!!

How do I train vggface2 and get my own pre model?