Parameter settings for `20170511-185253`

Hi –

I’m trying to reproduce the performance of the 20170511-185253 model, training on the CASIA dataset. Using the following parameter settings from the wiki, I get LFW accuracy of 0.979667 for the final model.

Are these the parameters that were used to train 20170511-185253? If not, are those parameters available somewhere?

Thanks!

export PYTHONPATH="./src"

# Align
for N in {1..4}; do
    python src/align/align_dataset_mtcnn.py \
        ./data/CASIA-WebFace-clean/ \
        ./data/casia_maxpy_mtcnnpy_182 \
        --image_size 182 --margin 44 --random_order --gpu-id 0 --gpu_memory_fraction 0.2 &
done

# Train

python src/train_softmax.py \
    --logs_base_dir ./logs \
    --models_base_dir ./models \
    --data_dir ./data/casia_maxpy_mtcnnpy_182 \
    --image_size 160 \
    --model_def models.inception_resnet_v1 \
    --optimizer RMSPROP \
    --learning_rate -1 \
    --max_nrof_epochs 80 \
    --keep_probability 0.8 \
    --random_crop \
    --random_flip \
    --learning_rate_schedule_file ./data/learning_rate_schedule_classifier_casia.txt \
    --weight_decay 5e-5 \
    --center_loss_factor 1e-2 \
    --center_loss_alfa 0.9

(Also, FWIW, in the last epoch of training loss is approx. 1.35 and regloss is approx. 1.06)

[bkj]

OK – as far as I can tell, the issue is related to using the “fused batch norm” option in Tensorflow:

https://www.tensorflow.org/performance/performance_guide#use_fused_batch_norm

That page makes it sound like fused=True uses a newer, faster but identical batch norm implementation. Before I ran training for the first time, I was seeing if there were any trivial ways to speed up training, so I set fused=True. However, I trained ~ 3 models w/ fused batched norm and never broke acc=0.979 whereas the one model I trained w/o fused batched norm got to 0.983.

I recognize there’s variability between runs, so I guess take this as preliminary evidence. I haven’t looked at the batch norm vs. fused batch norm implementations to see if it’s even possible that they give different results, but will post here with more info if I figure anything out.

EDIT: FWIW, training w/ fused batchnorm gives lower training loss (w/ higher variance?) for the duration:

[AlexeyAL1e]

Okay, this repository has a huge number of problems, but I’m gradually trying to solve them. Firstly, I formed my dataset of 30 personalities, each person has 34 face photos. after that i used these commands to align and get the loss with using script softmax.py Second, I had to run this script to get the .pb file python src/freeze_graph.py .model_dir output_file.pb Thirdly I got a .pkl file using this python src/classifier.py TRAIN ~/datasets/my_dataset/train/ ~/models/model-20170216-091149.pb ~/models/my_classifier.pkl --batch_size 34 Now, I use my model to identify faces in my test dataset, with run script: python src/classifier.py CLASSIFY ~/datasets/my_dataset/test/ ~/models/model-20170216-091149.pb ~/models/my_classifier.pkl --batch_size 34 but I get an error Found array with 0 sample(s) (shape=(0, 512)) while a minimum of 1 is required, it looks like the file .pkl, there are no entries for embedded images Tell me, if these are all points, then what am I doing wrong? @bkj @MaartenBloemen @jerryhouuu