Pytorch implementation to reproduce experiments from "Alleviating Sequence Information Loss with Data Overlapping and Prime Batch Sizes" - (poster).
If you use this code or our results in your research, please cite as appropriate:
@inproceedings{kocher-etal-2019-alleviating,
title = "Alleviating Sequence Information Loss with Data Overlapping and Prime Batch Sizes",
author = "Kocher, No{\'e}mien and
Scuito, Christian and
Tarantino, Lorenzo and
Lazaridis, Alexandros and
Fischer, Andreas and
Musat, Claudiu",
booktitle = "Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL)",
month = nov,
year = "2019",
address = "Hong Kong, China",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/K19-1083",
doi = "10.18653/v1/K19-1083",
pages = "890--899",
}
This repo holds experiments on 4 models using the "overlapping" method:
- awd, AWD ASGD Weight-Dropped LSTM, (
/awd
) - text simple, a very basic lstm for language modelling, (
/simple
) - mos, MOS Mixture of Softmaxes, (
/mos
) - voice simple, a very basic LSTM for emotion detection on voice, (
/emotions
)
To specify which model to run, use --main-model {simple-lstm | awd-lstm |Β mos-lstm | emotions-simple-lstm}
argument. There are additional common paramaters, as well as specific parameters for each model. Those can be found in main_run.py
.
The taxonomy in the code may differe a bit from the paper, especially regarding the type of experiments. Here is the corresponding terms:
In the code | In the paper |
---|---|
No order | Extreme TOI |
Local order | Inter-batch TOI |
Standard order | Standard TOI |
Total order (P) | Alleviated TOI (P) |
Experiments were run on a Tesla P100 GPU. Results are very likely to differ based on the GPU used.
Download the data (PTB, WT2, WT103):
chmod +x get_data.sh
./get_data.sh
For emotions, add in data/IEMOCAP/
the all_features_cv
files.
We use python 3.6
with Pytorch 0.4.1
. To create a new python environement and install dependencies, run:
python3 -m virtualenv venv
source venv/bin/activate
pip3 install -r requirements.txt
You can check your setup by launching a quick training over one epoch with the following command:
python3 main_run.py --main-model awd-lstm --batch-size 20 --data data/penn --epochs 1 --nhid 5 --emsize 5 --nlayers 1 --bptt 5
The program should exit without error and write the logs in the logs/
folder. You can watch the logs with tensorboard by launching the following command:
tensorboard --logdir logs/
main_run.py
is the main entry point that parses arguments, does the global initialization and runs the corresponding model and task.
awd/
, emotions/
, mos/
and simple/
are the different models directories. common/
holds the common initilization and utilities, such as the different data iterators, which are in the DataSelector
class in common/excavator.py
.
The main_run.py
file, after performing the common initilizations, imports the main.py
file corresponding to the choosen model.
Note: Those results do not use prime batch size, but the default parameters. To have better results, adapt the --batch-size
param to the closest prime number.
Quick anchors navigation:
Model | Dataset | Experiments |
---|---|---|
AWD | PTB | Extreme / Inter-batch / Original / Alleviated TOI |
WT2 | Extreme / Inter-batch / Original / Alleviated TOI | |
WT103 | Extreme / Inter-batch / Original / Alleviated TOI | |
Text simple LSTM | PTB | Extreme / Inter-batch / Original / Alleviated TOI |
WT2 | Extreme / Inter-batch / Original / Alleviated TOI | |
MOS | PTB | Original / Alleviated TOI |
Voice simple LSTM | IEMOCAP | Extreme / Inter-batch / Original / Alleviated TOI |
Extreme TOI:
Expected results: 66.38
/ 63.49
(validation / testing)
python3 main_run.py --main-model awd-lstm --batch-size 20 --data data/penn --dropouti 0.4 --dropouth 0.25 --seed 141 --seed-shuffle 141 --epochs 1000 --shuffle-full-seq
Inter-batch TOI:
Expected results: 66.96
/ 64.20
(validation / testing)
python3 main_run.py --main-model awd-lstm --batch-size 20 --data data/penn --dropouti 0.4 --dropouth 0.25 --seed 141 --seed-shuffle 141 --epochs 1000 --shuffle-row-seq
Standard TOI:
Expected results: 61.28
/ 58.94
(validation / testing)
python3 main_run.py --main-model awd-lstm --batch-size 20 --data data/penn --dropouti 0.4 --dropouth 0.25 --seed 141 --epochs 1000
Alleviated TOI {2,5,7,10}:
Expected results (validation / testing):
- 2:
61.73
/59.37
- 5:
63.37
/60.50
- 7:
59.22
/56.7
- 10:
68.09
/65.88
overlaps=(2 5 7 10)
epochs=1000
for k in "${overlaps[@]}"
do
:
python3 main_run.py --main-model awd-lstm --batch-size 20 --data data/penn --dropouti 0.4 --dropouth 0.25 --seed 141 --epochs "$(($epochs/$k))" --init-seq "overlapCN_${k}"
sleep 10
done
π₯ With a prime batch size:
Expected results (validation / testing):
- 2:
60.56
/57.97
- 5:
59.52
/57.14
- 7:
59.43
/57.16
- 10:
58.96
/56.46
overlaps=(2 5 7 10)
epochs=1000
for k in "${overlaps[@]}"
do
:
python3 main_run.py --main-model awd-lstm --batch-size 19 --data data/penn --dropouti 0.4 --dropouth 0.25 --seed 141 --epochs "$(($epochs/$k))" --init-seq "overlapCN_${k}"
sleep 10
done
Extreme TOI
Expected results: 77.14
/ 73.52
(validation / testing)
python3 main_run.py --main-model awd-lstm --epochs 750 --data /data/noemien.kocher/datasets/wikitext-2 --dropouth 0.2 --seed 1882 --batch-size 80 --shuffle-full-seq
Inter-batch TOI
Expected results: 76.08
/ 72.61
(validation / testing)
python main_run.py --main-model awd-lstm --epochs 750 --data /data/noemien.kocher/datasets/wikitext-2 --dropouth 0.2 --seed 1882 --batch-size 80 --shuffle-row-seq
Standard TOI
Expected results: 68.50
/ 65.86
(validation / testing)
python3 main_run.py --main-model awd-lstm --epochs 750 --data /data/noemien.kocher/datasets/wikitext-2 --dropouth 0.2 --seed 1882 --batch-size 80
Alleviated TOI {2,5,7,10}
Expected results (validation / testing):
- 2:
68.56
/65.51
- 5:
69.56
/66.33
- 7:
67.48
/64.87
- 10:
72.95
/69.69
overlaps=(2 5 7 10)
epochs=750
for k in "${overlaps[@]}"
do
:
python3 main_run.py --main-model awd-lstm --data /data/noemien.kocher/datasets/wikitext-2 --dropouth 0.2 --seed 1882 --batch-size 80 --epochs "$(($epochs/$k))" --init-seq "overlapCN_${k}"
sleep 10
done
π₯ With a prime batch size:
Expected results (validation / testing):
- 2:
68.11
/65.14
- 5:
67.74
/65.11
- 7:
67.79
/64.79
- 10:
67.47
/64.73
overlaps=(2 5 7 10)
epochs=750
for k in "${overlaps[@]}"
do
:
python3 main_run.py --main-model awd-lstm --data /data/noemien.kocher/datasets/wikitext-2 --dropouth 0.2 --seed 1882 --batch-size 79 --epochs "$(($epochs/$k))" --init-seq "overlapCN_${k}"
sleep 10
done
Extreme TOI
Expected results: 35.22
/ 36.19
(validation / testing)
python3 -u main_run.py --main-model awd-lstm --epochs 14 --nlayers 4 --emsize 400 --nhid 2500 --alpha 0 --beta 0 --dropoute 0 --dropouth 0.1 --dropouti 0.1 --dropout 0.1 --wdrop 0 --wdecay 0 --bptt 140 --batch-size 60 --optimizer adam --lr 1e-3 --data /data/noemien.kocher/datasets/wikitext-103 --when 12 --model QRNN --shuffle-full-seq
Inter-batch TOI
Expected results: 35.41
/ 36.39
(validation / testing)
python3 -u main_run.py --main-model awd-lstm --epochs 14 --nlayers 4 --emsize 400 --nhid 2500 --alpha 0 --beta 0 --dropoute 0 --dropouth 0.1 --dropouti 0.1 --dropout 0.1 --wdrop 0 --wdecay 0 --bptt 140 --batch-size 60 --optimizer adam --lr 1e-3 --data /data/noemien.kocher/datasets/wikitext-103 --when 12 --model QRNN --shuffle-row-seq
Standard TOI
Expected results: 32.18
/ 32.94
(validation / testing)
python3 -u main_run.py --main-model awd-lstm --epochs 14 --nlayers 4 --emsize 400 --nhid 2500 --alpha 0 --beta 0 --dropoute 0 --dropouth 0.1 --dropouti 0.1 --dropout 0.1 --wdrop 0 --wdecay 0 --bptt 140 --batch-size 60 --optimizer adam --lr 1e-3 --data /data/noemien.kocher/datasets/wikitext-103 --when 12 --model QRNN
Alleviated TOI {2,5,7,10}
Expected results (validation / testing):
- 2:
36.94
/34.31
- 5:
38.50
/40.04
- 7:
31.78
/32.72
- 10:
48.28
/49.49
# base num epochs is 14
overlaps=(2 5 7 10)
when_steps=147456
max_steps=172032
for i in "${!overlaps[@]}"
do
:
python3 -u main_run.py --main-model awd-lstm --epochs 14 --nlayers 4 --emsize 400 --nhid 2500 --alpha 0 --beta 0 --dropoute 0 --dropouth 0.1 --dropouti 0.1 --dropout 0.1 --wdrop 0 --wdecay 0 --bptt 140 --batch-size 60 --optimizer adam --lr 1e-3 --data /data/noemien.kocher/datasets/wikitext-103 --when-steps "$when_steps" --model QRNN --init-seq "overlapCN_${overlaps[$i]}" --log-dir /data/noemien.kocher/logs/ --max-steps "$max_steps"
sleep 10
done
π₯ With a prime batch size:
Expected results (validation / testing):
- 2:
32.00
/32.98
- 5:
31.93
/33.07
- 7:
31.78
/32.89
- 10:
31.92
/32.85
# base num epochs is 14
overlaps=(2 5 7 10)
when_steps=147456
max_steps=172032
for i in "${!overlaps[@]}"
do
:
python3 -u main_run.py --main-model awd-lstm --epochs 14 --nlayers 4 --emsize 400 --nhid 2500 --alpha 0 --beta 0 --dropoute 0 --dropouth 0.1 --dropouti 0.1 --dropout 0.1 --wdrop 0 --wdecay 0 --bptt 140 --batch-size 59 --optimizer adam --lr 1e-3 --data /data/noemien.kocher/datasets/wikitext-103 --when-steps "$when_steps" --model QRNN --init-seq "overlapCN_${overlaps[$i]}" --log-dir /data/noemien.kocher/logs/ --max-steps "$max_steps"
sleep 10
done
Extreme TOI:
Expected results: 81.97
/ 79.08
(validation / testing)
python3 main_run.py --main-model simple-lstm --epochs 100 --batch-size 20 --dropout 0.15 --nlayers 2 --bptt 70 --nhid 1500 --lr-decay 1 --shuffle-full-seq
Inter-batch TOI:
Expected results: 81.67
/ 78.59
(validation / testing)
python3 main_run.py --main-model simple-lstm --epochs 100 --batch-size 20 --dropout 0.15 --nlayers 2 --bptt 70 --nhid 1500 --lr-decay 1 --shuffle-row-seq
Standard TOI:
Expected results: 77.54
/ 75.36
(validation / testing)
python3 main_run.py --main-model simple-lstm --epochs 100 --batch-size 20 --dropout 0.15 --nlayers 2 --bptt 70 --nhid 1500 --lr-decay 1
Alleviated TOI {2,5,7,10}:
Expected results (validation / testing):
- 2:
78.48
/76.55
- 5:
91.95
/89.64
- 7:
77.47
/74.98
- 10:
92.92
/92.07
overlaps=(2 5 7 10)
epochs=100
for k in "${overlaps[@]}"
do
:
python3 main_run.py --main-model simple-lstm --epochs "$(($epochs/$k))" --batch-size 20 --dropout 0.15 --nlayers 2 --bptt 70 --nhid 1500 --lr-decay 1 --init-seq "overlapCN_${k}"
sleep 10
done
Extreme TOI
Expected results: 101.3
/ 96.08
(validation / testing)
python3 main_run.py --main-model simple-lstm --epochs 100 --batch-size 80 --dropout 0.15 --nlayers 2 --bptt 70 --nhid 1150 --lr-decay 1 --data /data/noemien.kocher/datasets/wikitext-2 --shuffle-full-seq
Inter-batch TOI
Expected results: 101.7
/ 96.89
(validation / testing)
python3 main_run.py --main-model simple-lstm --epochs 100 --batch-size 80 --dropout 0.15 --nlayers 2 --bptt 70 --nhid 1150 --lr-decay 1 --data /data/noemien.kocher/datasets/wikitext-2 --shuffle-row-seq
Standard TOI
Expected results: 98.85
/ 93.15
(validation / testing)
python3 main_run.py --main-model simple-lstm --epochs 100 --batch-size 80 --dropout 0.15 --nlayers 2 --bptt 70 --nhid 1150 --lr-decay 1 --data /data/noemien.kocher/datasets/wikitext-2
Alleviated TOI {2,5,7,10}
Expected results (validation / testing):
- 2:
100.4
/94.49
- 5:
113.5
/106.1
- 7:
98.25
/92.77
- 10:
151.0
/135.1
overlaps=(2 5 7 10)
epochs=100
for k in "${overlaps[@]}"
do
:
python3 main_run.py --main-model simple-lstm --epochs "$(($epochs/$k))" --batch-size 80 --dropout 0.15 --nlayers 2 --bptt 70 --nhid 1150 --lr-decay 1 --data /data/noemien.kocher/datasets/wikitext-2 --init-seq "overlapCN_${k}"
sleep 10
done
Standard TOI:
Expected results: 58.49
/ 56.19
(validation / testing)
python3 main_run.py --main-model mos-lstm --data data/penn --dropouti 0.4 --dropoutl 0.29 --dropouth 0.225 --seed 28 --batch-size 12 --lr 20.0 --epochs 1000 --nhid 960 --nhidlast 620 --emsize 280 --n-experts 15
Alleviated TOI {1..40}:
π₯ With a prime batch size:
epochs=2000
for k in {1..70}
do
:
python3 main_run.py --main-model mos-lstm --data data/penn --dropouti 0.4 --dropoutl 0.29 --dropouth 0.225 --seed 28 --batch-size 13 --lr 20.0 --epochs "$(($epochs/$k))" --nhid 960 --nhidlast 620 --emsize 280 --n-experts 15 --init-seq "overlapCNF_${k}"
sleep 10
done
Expected results (validation / testing):
- 1:
58.36
/56.21
- 2:
58.07
/55.76
- 3:
58.03
/55.79
- 4:
52.82
/55.63
- 5:
57.81
/55.63
- 6:
57.55
/55.32
- 7:
57.47
/55.23
- 8:
57.47
/55.34
- 9:
57.16
/54.93
- 10:
57.34
/54.90
- 11:
57.11
/54.98
- 12:
57.47
/55.44
- 13:
67.77
/66.01
- 14:
56.76
/54.58
(paper's result) - 15:
57.44
/55.20
- 16:
56.95
/54.86
- 17:
57.64
/55.14
- 18:
57.38
/54.93
- 19:
57.55
/55.35
- 20:
57.00
/54.67
- 21:
57.55
/55.22
- 22:
57.54
/55.19
- 23:
57.29
/54.90
- 24:
57.47
/55.11
- 25:
57.12
/54.85
- 26:
66.14
/63.81
- 27:
57.08
/54.85
- 28:
--.--
/--.--
- 29:
--.--
/--.--
- 30:
--.--
/--.--
- 31:
57.74
/55.37
- 32:
57.21
/55.26
- 33:
57.66
/55.40
- 34:
57.48
/55.44
- 35:
56.44
/54.33
(post-result, not in the paper) - 36:
57.10
/55.09
- 37:
57.55
/55.29
- 38:
57.04
/54.87
- 39:
64.37
/62.54
- 40:
57.52
/54.99
Extreme TOI:
Expected result: 0.475
/ 0.377
(WA / UA)
python3 main_run.py --main-model emotions-simple-lstm --cv 5 --data data/IEMOCAP/all_features_cv --test-batch-size 20 --lr 0.05 --log-interval 20 --lr-decay 1 --step-size 0.1 --epochs 60 --order complete_random
Inter-batch TOI:
Expected result: 0.478
/ 0.386
(WA / UA)
python3 main_run.py --main-model emotions-simple-lstm --cv 5 --data data/IEMOCAP/all_features_cv --test-batch-size 20 --lr 0.05 --log-interval 20 --lr-decay 1 --step-size 0.1 --epochs 60 --window-size 300 --order local_order
Standard TOI:
Expected result: 0.486
/ 0.404
(WA / UA)
python3 main_run.py --main-model emotions-simple-lstm --cv 5 --data data/IEMOCAP/all_features_cv --test-batch-size 20 --lr 0.05 --log-interval 20 --lr-decay 1 --step-size 0.1 --epochs 60 --order standard_order
Alleviated TOI 10:
Expected result:
- 15k steps:
0.553
/0.489
(WA / UA) - 60 epochs:
0.591
/0.523
(WA / UA)
python3 main_run.py --main-model emotions-simple-lstm --cv 5 --data data/IEMOCAP/all_features_cv --test-batch-size 20 --lr 0.05 --log-interval 20 --lr-decay 1 --step-size 0.1 --epochs 60 --order total_order
Expected results (validation / testing):
- 1:
61.28
/58.94
- 2:
60.76
/58.55
- 5:
60.10
/57.83
- 7:
60.08
/57.76
- 10:
60.05
/57.78
P=(1 2 5 7 10)
epochs=1000
for k in "${P[@]}"
do
:
python3 main_run.py --main-model awd-lstm-repetitions --batch-size 20 --data data/penn --dropouti 0.4 --dropouth 0.25 --seed 141 --epochs 1000 --use-repetitions "${k}"
sleep 10
done
Code is heavily borrowed from the following sources:
- simple-lstm (
simple/
): https://github.com/deeplearningathome/pytorch-language-model - awd-lstm (
awd/
): https://github.com/salesforce/awd-lstm-lm - mos-lstm: (
mos/
) https://github.com/zihangdai/mos