facenet
facenet copied to clipboard
davidsandberg
A new model trained on VGGFace2
I trained a model on VGGFace2 using center loss. The embedding is powerful than the subset of MS-Celeb. I can make the model public with the two available mode. @davidsandberg
@Shahnawazgrewal can you please share with us how you've trained the checkpoint? and can you also share with us the checkpoint?
I trained with VGGFace2 as well, the model is not as good as MS Cele. Although the accuracy might be equal or better than MS, the TPR at 0.001 FPR is much lower (98.X compared to 99.X)
I eventually combined this two datasets and reached 99.73% accuracy and 99.63% TPR w/ 0.001 FPR.
@Shahnawazgrewal could U please kindly share with us how do U train the VGGFace2 model? When we try to train the data, we always got error as below. Thanks! OutOfRangeError (see above for traceback): FIFOQueue '_1_batch_join/fifo_queue' is closed and has insufficient elements (requested 9, current size 0) [[Node: batch_join = QueueDequeueUpToV2[component_types=[DT_FLOAT, DT_INT64], timeout_ms=-1, _device="/job:localhost/replica:0/task:0/device:CPU:0"](batch_join/fifo_queue, _arg_batch_size_0_0)]]
@Shahnawazgrewal For sure, if U can share your model via Google Drive and etc., that is also much appreciated. Thanks!
@Shahnawazgrewal I add center loss to caffe, I met the error "math_functions.cu:155] Check failed: error == cudaSuccess (11 vs. 0) invalid argument" when I train,have you met this problem or you know how to fix it? thx
can you please decrease maximum number of epochs. Please read issue #105 I have similar error for MS-Celeb-1M dataset. @syy6
@Shahnawazgrewal my subset of MS-Celeb-1M is 70k identities, 4.5 million images. I can achieve 99.5% accuracy and 99.3% TPR with it
@Shahnawazgrewal , actually I even tried to reduce the number of epochs, but the issue still exists......
@Shahnawazgrewal it would be very nice if you could share the hyperparameters you used in training. I've recently tried to use the VGGFace2 to train by triplet loss but with no luck. The LFW accuracy and validation rate just levelled off at around 0.96 and 0.7.
@yipsang @Shahnawazgrewal , I just find the issue, one of the input png is broken in my computer, so I got this error. After removing the png, it seems to be fine now.
@JianbangZ, could U please share with us how U take the duplicate between MS & VGG2 dataset? If U look at the namelist of two datasets, certain names are duplicate in both dataset.
Here is the link to download a pre-trained model trained with inception-ResNet-v1 with center loss function on VGGFace2 dataset. Please give your general feedback.
@Shahnawazgrewal , Did you perform the pre-training on MS-Celeb-1M and then fine-tune on VGGFace2 dataset?
No. I didn't I trained with inception-ResNet-v1 with center loss function on VGGFace2 dataset from scratch. More specifically I downloaded loosely cropped faces dataset from the VG-GFace2 (http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/ . The dataset is aligned with 160×160 image size and 32 pixels margin based on Multi-task CNN. I trained the model on aligned dataset for 100 epochs with an RMSProp optimizer. @Yeongjae
@Shahnawazgrewal could you make proper comments for checkpoint files which has been uploaded in dropbox!
@Shahnawazgrewal based on our evaluation, your model is truly powerful than both provided pretrain-model. I wonder the reason behind this wonderful improvement? Is the dataset used for training the root cause?
P.S. Thank your for uploading this wonderful pretrained checkpoint
@tenggyut , VGGFace2 dataset is considered to be a deep dataset (higher number of image per identity). In my opinion, this could be the reason. In addition, I observed that the model trained on VGGFace2 produced better representation of previously unseen faces.
@Shahnawazgrewal did you train the model as a classifier or using triple loss?
@Shahnawazgrewal
I have few questions about your implement detail:
- Do you do 2D alignment or just crop 160x160 bounding box after MTCNN.
- How is the learning rate you use in RMSProp? Do you decrease learning rate? Thanks for your sharing!
- just crop 160x160 bounding box after MTCNN.
- I used default settings available.
Did you train the model with softmax loss combined with center loss? or you just train it with center loss?
Validation on LFW dataset with the model trained on VGGFace2 Runnning forward pass on LFW images Model directory: /home/super/datasets/lfw/vggface2-cl Metagraph file: model-20171216-232945.meta Checkpoint file: model-20171216-232945.ckpt-100000 Runnning forward pass on LFW images Accuracy: 0.992+-0.004 Validation rate: 0.96000+-0.01880 @ FAR=0.00067 Area Under Curve (AUC): 0.999 Equal Error Rate (EER): 0.008
I trained with cosine Face algorithms . accuracy is 0.995, validation rate = 0.985
@JianbangZ what do you mean cosine face algorithms? did you replace the L2 norm in center loss with cosine similarity? Or you mean the paper the author released last year, SphereFace?
