Bug fix of norm minimization for dev mode (#31462)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/31462
Fix the divide by zero issue in norm minimization in dev mode
Test Plan: buck run mode/dev vision/video_modeling/classification/tools:test_octGloRe_quantization -- --test_data=/mnt/vol/gfsadslearner-frc3c01/fblearner_flow/users/summerdeng/xray_video/deep_vision_video_yufei_test_data_fcc_v4p2_10.csv --output_dir /mnt/vol/gfsadslearner-frc3c01/fblearner_flow/users/summerdeng/xray_video/octGloRe --load_model_path=/mnt/vol/gfsfblearner-oregon/flow/data/2019-10-15/e2681db8-e4f5-4b70-ae18-45bf0b8fbfbc/train_model_epoch0_inputcount0_final.mdl --dataset_name="FCC V4P2" --num_labels=1099 --column_handle="handle" --clip_per_video=1 --num_groups=24 --width_per_group=2 --batch_size=32 --histogram_file=/mnt/vol/gfsadslearner-frc3c01/fblearner_flow/users/summerdeng/xray_video/octGloRe/hist_octGloRe_final_24x2_fcc_v4p2_1clip_f144586257_nullfix_100k_compiled.hist --int8_model_type="pb" --int8_predict_net_path="reproduce_octGloRe_final_24x2_predict_net_int8_l2approx_wminmax_from_mdl.pb" --int8_init_net_path="reproduce_octGloRe_final_24x2_init_net_int8_l2approx_wminmax_from_mdl.pb" --weight_quant="l2_approx" --activation_quant="l2_approx" --print_model --int8_model_saved --num_iter 10
Reviewed By: jspark1105
Differential Revision: D19172591
fbshipit-source-id: 994a20e3364b0dc33623a11281e0bdbc2e06159d
