diff options
author | Roman Grundkiewicz <rgrundkiewicz@gmail.com> | 2020-11-13 21:20:49 +0300 |
---|---|---|
committer | Roman Grundkiewicz <rgrundkiewicz@gmail.com> | 2020-11-13 21:20:49 +0300 |
commit | 525fc8697e4436ab9f1b58abd3e5de409ba28e3b (patch) | |
tree | 58eb71c060769f1f47a25119ecc036cb3135943b | |
parent | a05a7f0099b17f798ddd15ae36d51796b89219d6 (diff) |
Use clip-norm 0 in training/restarting
4 files changed, 31 insertions, 17 deletions
diff --git a/tests/training/restarting/sgd_2e.expected b/tests/training/restarting/sgd_2e.expected index 2a801f2..15bbe18 100644 --- a/tests/training/restarting/sgd_2e.expected +++ b/tests/training/restarting/sgd_2e.expected @@ -1,16 +1,16 @@ Ep. 1 : Up. 4 : Sen. 128 : Cost 257.99652100 Ep. 1 : Up. 8 : Sen. 256 : Cost 267.93783569 -Ep. 1 : Up. 12 : Sen. 384 : Cost 243.39039612 -Ep. 1 : Up. 16 : Sen. 512 : Cost 235.87208557 +Ep. 1 : Up. 12 : Sen. 384 : Cost 243.39041138 +Ep. 1 : Up. 16 : Sen. 512 : Cost 235.87210083 Ep. 1 : Up. 20 : Sen. 640 : Cost 204.79017639 -Ep. 1 : Up. 24 : Sen. 768 : Cost 240.11624146 -Ep. 1 : Up. 28 : Sen. 896 : Cost 208.47099304 -Ep. 1 : Up. 32 : Sen. 1,024 : Cost 199.36221313 -Ep. 2 : Up. 36 : Sen. 128 : Cost 213.58728027 -Ep. 2 : Up. 40 : Sen. 256 : Cost 220.51084900 -Ep. 2 : Up. 44 : Sen. 384 : Cost 199.10847473 -Ep. 2 : Up. 48 : Sen. 512 : Cost 194.56945801 -Ep. 2 : Up. 52 : Sen. 640 : Cost 171.77461243 -Ep. 2 : Up. 56 : Sen. 768 : Cost 208.75405884 -Ep. 2 : Up. 60 : Sen. 896 : Cost 186.17434692 -Ep. 2 : Up. 64 : Sen. 1,024 : Cost 179.63316345 +Ep. 1 : Up. 24 : Sen. 768 : Cost 240.11631775 +Ep. 1 : Up. 28 : Sen. 896 : Cost 208.47109985 +Ep. 1 : Up. 32 : Sen. 1,024 : Cost 199.36233521 +Ep. 2 : Up. 36 : Sen. 128 : Cost 213.58744812 +Ep. 2 : Up. 40 : Sen. 256 : Cost 220.51107788 +Ep. 2 : Up. 44 : Sen. 384 : Cost 199.10870361 +Ep. 2 : Up. 48 : Sen. 512 : Cost 194.56954956 +Ep. 2 : Up. 52 : Sen. 640 : Cost 171.77453613 +Ep. 2 : Up. 56 : Sen. 768 : Cost 208.75399780 +Ep. 2 : Up. 60 : Sen. 896 : Cost 186.17416382 +Ep. 2 : Up. 64 : Sen. 1,024 : Cost 179.63301086 diff --git a/tests/training/restarting/sgd_sync_2e.expected b/tests/training/restarting/sgd_sync_2e.expected index 07061db..a83990d 100644 --- a/tests/training/restarting/sgd_sync_2e.expected +++ b/tests/training/restarting/sgd_sync_2e.expected @@ -7,10 +7,10 @@ Ep. 1 : Up. 24 : Sen. 768 : Cost 240.11631775 Ep. 1 : Up. 28 : Sen. 896 : Cost 208.47109985 Ep. 1 : Up. 32 : Sen. 1,024 : Cost 199.36233521 Ep. 2 : Up. 36 : Sen. 128 : Cost 213.58744812 -Ep. 2 : Up. 40 : Sen. 256 : Cost 220.51107788 +Ep. 2 : Up. 40 : Sen. 256 : Cost 220.51104736 Ep. 2 : Up. 44 : Sen. 384 : Cost 199.10870361 Ep. 2 : Up. 48 : Sen. 512 : Cost 194.56954956 Ep. 2 : Up. 52 : Sen. 640 : Cost 171.77453613 -Ep. 2 : Up. 56 : Sen. 768 : Cost 208.75396729 +Ep. 2 : Up. 56 : Sen. 768 : Cost 208.75399780 Ep. 2 : Up. 60 : Sen. 896 : Cost 186.17416382 Ep. 2 : Up. 64 : Sen. 1,024 : Cost 179.63301086 diff --git a/tests/training/restarting/test_sgd_for_two_epochs.sh b/tests/training/restarting/test_sgd_for_two_epochs.sh index ad92b5f..11bf76e 100644 --- a/tests/training/restarting/test_sgd_for_two_epochs.sh +++ b/tests/training/restarting/test_sgd_for_two_epochs.sh @@ -1,5 +1,11 @@ #!/bin/bash -x +##################################################################### +# SUMMARY: Restaring training after the 1st epoch (async) +# AUTHOR: snukky +# TAGS: optimizer clip-norm +##################################################################### + # Exit on error set -e @@ -7,12 +13,13 @@ set -e rm -rf sgd_2e sgd_1st_epoch.log sgd_2nd_epoch.log mkdir -p sgd_2e -extra_opts="--no-shuffle --seed 1111 --maxi-batch 1 --maxi-batch-sort none --mini-batch 32 --optimizer sgd" +extra_opts="--no-shuffle --clip-norm 1 --seed 1111 --maxi-batch 1 --maxi-batch-sort none --mini-batch 32 --optimizer sgd" # Added because default options has changes extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" # Uncomment to prepare the expected output +#rm -f sgd_two_epochs.log #$MRT_MARIAN/marian \ #-m sgd_2e/model_2e.npz -t $MRT_DATA/train.max50.{en,de} -v vocab.en.yml vocab.de.yml \ #--disp-freq 4 --save-freq 32 --after-epochs 2 -l 0.1 $extra_opts \ diff --git a/tests/training/restarting/test_sgd_for_two_epochs_sync.sh b/tests/training/restarting/test_sgd_for_two_epochs_sync.sh index d3ee295..8615e80 100644 --- a/tests/training/restarting/test_sgd_for_two_epochs_sync.sh +++ b/tests/training/restarting/test_sgd_for_two_epochs_sync.sh @@ -1,5 +1,11 @@ #!/bin/bash -x +##################################################################### +# SUMMARY: Restaring training after the 1st epoch (sync-sgd) +# AUTHOR: snukky +# TAGS: optimizer clip-norm +##################################################################### + # Exit on error set -e @@ -7,12 +13,13 @@ set -e rm -rf sgd_sync_2e sgd_sync_*_epoch.log mkdir -p sgd_sync_2e -extra_opts="--no-shuffle --seed 1111 --maxi-batch 1 --maxi-batch-sort none --mini-batch 32 --optimizer sgd --sync-sgd" +extra_opts="--no-shuffle --clip-norm 1 --seed 1111 --maxi-batch 1 --maxi-batch-sort none --mini-batch 32 --optimizer sgd --sync-sgd" # Added because default options has changes extra_opts="$extra_opts --cost-type ce-mean --disp-label-counts false" # Uncomment to prepare the expected output +#rm -f sgd_sync_two_epochs.log #$MRT_MARIAN/marian \ #-m sgd_sync_2e/model_2e.npz -t $MRT_DATA/train.max50.{en,de} -v vocab.en.yml vocab.de.yml \ #--disp-freq 4 --save-freq 32 --after-epochs 2 -l 0.1 $extra_opts \ |