-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathrun_additional_data.sh
226 lines (183 loc) · 9.26 KB
/
run_additional_data.sh
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
# this script assumes there are 2 GPU cards available in this machine (at least)
# please edit the script accordingly in case there are less GPU cards
DATA_PATH="/path/to/data/multi30k"
MODEL_PATH="/path/to/variational-multimodal-nmt-model-snapshots"
MODEL_FILE_NAME="MMT_VI_Model_AdditionalTrainingData"
# multi30k validation set
VAL_SRC="${DATA_PATH}/val.lc.norm.tok.bpe-en-de-30000.en"
VAL_TGT="${DATA_PATH}/val.lc.norm.tok.bpe-en-de-30000.de"
VAL_IMGS="${DATA_PATH}/flickr30k_valid_resnet50_cnn_features.hdf5"
# multi30k training set
TRAIN_SRC="${DATA_PATH}/train.lc.norm.tok.bpe-en-de-30000.en"
TRAIN_TGT="${DATA_PATH}/train.lc.norm.tok.bpe-en-de-30000.de"
TRAIN_IMGS="${DATA_PATH}/flickr30k_train_resnet50_cnn_features.hdf5"
# multi30k training set
PRETRAIN_SRC="${DATA_PATH}/flickr30k_train_translated-5x-comparable-1x.lc.norm.tok.bpe-en-de-30000.not-numbered.shuffled.en"
PRETRAIN_TGT="${DATA_PATH}/flickr30k_train_translated-5x-comparable-1x.lc.norm.tok.bpe-en-de-30000.not-numbered.shuffled.de"
PRETRAIN_IMGS="${DATA_PATH}/flickr30k_train_translated-5x-comparable-1x_resnet50_cnn_features.shuffled.hdf5"
# multi30k test set (2016)
TEST_2016_SRC="${DATA_PATH}/test_2016_flickr.lc.norm.tok.bpe-en-de-30000.en"
TEST_2016_TGT="${DATA_PATH}/test_2016_flickr.lc.norm.tok.bpe-en-de-30000.de"
TEST_2016_IMGS="${DATA_PATH}/flickr30k_test_resnet50_cnn_features.hdf5"
# multi30k test set (2017)
TEST_2017_SRC="${DATA_PATH}/test_2017_flickr.lc.norm.tok.bpe-en-de-30000.en"
TEST_2017_TGT="${DATA_PATH}/test_2017_flickr.lc.norm.tok.bpe-en-de-30000.de"
TEST_2017_IMGS="${DATA_PATH}/flickr30k_test_2017_flickr_resnet50_cnn_features.hdf5"
# ambiguous MSCOCO test set (2017)
TEST_2017_MSCOCO_SRC="${DATA_PATH}/test_2017_mscoco.lc.norm.tok.bpe-en-de-30000.en"
TEST_2017_MSCOCO_TGT="${DATA_PATH}/test_2017_mscoco.lc.norm.tok.bpe-en-de-30000.de"
TEST_2017_MSCOCO_IMGS="${DATA_PATH}/flickr30k_test_2017_mscoco_resnet50_cnn_features.hdf5"
# number of pre-training epochs (1 epochs == 290K examples)
PRETRAIN_EPOCHS=3
# number of fine-tuning epochs (1 epochs == 29K translated Multi30k examples)
EPOCHS=33
# LSTM hidden state size
RNN_SIZE=500
# latent variable size for fixed-prior model
Z_FIXED_SIZE=50
# latent variable size for conditional prior model
Z_COND_SIZE=500
#PRETRAIN_EPOCHS=3
#EPOCHS=2
#RNN_SIZE=10
#Z_FIXED_SIZE=10
#Z_COND_SIZE=10
##########
# train
##########
# pretrain the model on the concatenation of the translated Multi30k (~29K src/tgt/img instances)
# and the back-translated comparable Multi30k (additional ~145K src/tgt/img instances)
PRETRAIN_DATASET=${DATA_PATH}/concat-multi30k-translational-5times-comparable-1time-shuffled_correct
# fine-tune the pre-trained models on the translated Multi30k only
FINETUNE_DATASET=${DATA_PATH}/m30k
# pre-train one conditional prior and one fixed-prior model
# one model on gpu 0, another one on gpu 1 (both spawn validation set translations on gpu 1)
python train_mm_vi_model1.py \
-gpuid 0 -epochs ${PRETRAIN_EPOCHS} -batch_size 40 -valid_batch_size 40 -optim 'adam' -learning_rate 0.002 -rnn_type LSTM \
-rnn_size ${RNN_SIZE} --z_latent_dim ${Z_FIXED_SIZE} \
-src ${VAL_SRC} \
-tgt ${VAL_TGT} \
-path_to_train_img_feats ${PRETRAIN_IMGS} \
-path_to_valid_img_feats ${VAL_IMGS} \
-data ${PRETRAIN_DATASET} \
--multimodal_model_type vi-model1 --use_global_image_features -dropout 0.5 -dropout_imgs 0.5 \
-save_model ${MODEL_PATH}/${MODEL_FILE_NAME}.fixed-prior \
-overwrite_model_file 2>&1 >> ${MODEL_PATH}/${MODEL_FILE_NAME}.fixed-prior.pretrain.log &
python train_mm_vi_model1.py \
-gpuid 1 -epochs ${PRETRAIN_EPOCHS} -batch_size 40 -valid_batch_size 40 -optim 'adam' -learning_rate 0.002 -rnn_type LSTM \
-rnn_size ${RNN_SIZE} --z_latent_dim ${Z_COND_SIZE} \
-src ${VAL_SRC} \
-tgt ${VAL_TGT} \
-path_to_train_img_feats ${PRETRAIN_IMGS} \
-path_to_valid_img_feats ${VAL_IMGS} \
-data ${PRETRAIN_DATASET} \
--multimodal_model_type vi-model1 --use_global_image_features -dropout 0.5 -dropout_imgs 0.5 \
-save_model ${MODEL_PATH}/${MODEL_FILE_NAME}.conditional-prior \
-overwrite_model_file \
--conditional 2>&1 >> ${MODEL_PATH}/${MODEL_FILE_NAME}.conditional-prior.pretrain.log &
wait;
PRETRAINED_MODEL_FIXED=${MODEL_PATH}/${MODEL_FILE_NAME}.fixed-prior_MostCurrentModel.pt
PRETRAINED_MODEL_COND=${MODEL_PATH}/${MODEL_FILE_NAME}.conditional-prior_MostCurrentModel.pt
FINETUNE_MODEL_FIXED=${MODEL_PATH}/${MODEL_FILE_NAME}.fixed-prior_MostCurrentModel_finetuned.pt
FINETUNE_MODEL_COND=${MODEL_PATH}/${MODEL_FILE_NAME}.conditional-prior_MostCurrentModel_finetuned.pt
# create a copy of the pre-trained models for finetuning,
# so that we can still use pretrained models without fine-tuning if needed
cp ${PRETRAINED_MODEL_FIXED} ${FINETUNE_MODEL_FIXED}
cp ${PRETRAINED_MODEL_COND} ${FINETUNE_MODEL_COND}
# fine-tune each model (conditional- and fixed-prior) on the translated Multi30k only.
# one model on gpu 0, another one on gpu 1 (both spawn validation set translations on gpu 1)
python train_mm_vi_model1.py \
-gpuid 0 -epochs ${EPOCHS} -batch_size 40 -valid_batch_size 40 -optim 'adam' -learning_rate 0.002 -rnn_type LSTM \
-rnn_size ${RNN_SIZE} --z_latent_dim ${Z_FIXED_SIZE} \
-early_stopping_criteria 'bleu' \
-src ${VAL_SRC} \
-tgt ${VAL_TGT} \
-path_to_train_img_feats ${TRAIN_IMGS} \
-path_to_valid_img_feats ${VAL_IMGS} \
-data ${FINETUNE_DATASET} \
--multimodal_model_type vi-model1 --use_global_image_features -dropout 0.5 -dropout_imgs 0.5 \
-save_model ${MODEL_PATH}/${MODEL_FILE_NAME}.fixed-prior_finetune \
-overwrite_model_file \
-train_from ${FINETUNE_MODEL_FIXED} \
-finetune \
2>&1 > ${MODEL_PATH}/${MODEL_FILE_NAME}.fixed-prior.finetune.log &
python train_mm_vi_model1.py \
-gpuid 1 -epochs ${EPOCHS} -batch_size 40 -valid_batch_size 40 -optim 'adam' -learning_rate 0.002 -rnn_type LSTM \
-rnn_size ${RNN_SIZE} --z_latent_dim ${Z_COND_SIZE} \
-early_stopping_criteria 'bleu' \
-src ${VAL_SRC} \
-tgt ${VAL_TGT} \
-path_to_train_img_feats ${TRAIN_IMGS} \
-path_to_valid_img_feats ${VAL_IMGS} \
-data ${FINETUNE_DATASET} \
--multimodal_model_type vi-model1 --use_global_image_features -dropout 0.5 -dropout_imgs 0.5 \
-save_model ${MODEL_PATH}/${MODEL_FILE_NAME}.conditional-prior_finetune \
-overwrite_model_file \
-train_from ${FINETUNE_MODEL_COND} \
-finetune \
--conditional 2>&1 > ${MODEL_PATH}/${MODEL_FILE_NAME}.conditional-prior.finetune.log &
wait;
#############
# translate
#############
# translate the validation set
SPLIT="validation"
python translate_mm_vi.py \
-model ${MODEL_PATH}/${MODEL_FILE_NAME}.conditional-prior_BestModelBleu.pt \
-src ${VAL_SRC} \
-path_to_test_img_feats ${VAL_IMGS} \
-gpu 0 \
-output ${MODEL_PATH}/${MODEL_FILE_NAME}.conditional-prior_BestModelBleu.pt.${SPLIT}-translations &
python translate_mm_vi.py \
-model ${MODEL_PATH}/${MODEL_FILE_NAME}.fixed-prior_BestModelBleu.pt \
-src ${VAL_SRC} \
-path_to_test_img_feats ${VAL_IMGS} \
-gpu 1 \
-output ${MODEL_PATH}/${MODEL_FILE_NAME}.fixed-prior_BestModelBleu.pt.${SPLIT}-translations &
wait;
# translate the test set (2016)
SPLIT="test2016"
python translate_mm_vi.py \
-model ${MODEL_PATH}/${MODEL_FILE_NAME}.conditional-prior_BestModelBleu.pt \
-src ${TEST_2016_SRC} \
-path_to_test_img_feats ${TEST_2016_IMGS} \
-gpu 0 \
-output ${MODEL_PATH}/${MODEL_FILE_NAME}.conditional-prior_BestModelBleu.pt.${SPLIT}-translations &
python translate_mm_vi.py \
-model ${MODEL_PATH}/${MODEL_FILE_NAME}.fixed-prior_BestModelBleu.pt \
-src ${TEST_2016_SRC} \
-path_to_test_img_feats ${TEST_2016_IMGS} \
-gpu 1 \
-output ${MODEL_PATH}/${MODEL_FILE_NAME}.fixed-prior_BestModelBleu.pt.${SPLIT}-translations &
wait;
# translate the test set (2017)
SPLIT="test2017"
python translate_mm_vi.py \
-model ${MODEL_PATH}/${MODEL_FILE_NAME}.fixed-prior_BestModelBleu.pt \
-src ${TEST_2017_SRC} \
-path_to_test_img_feats ${TEST_2017_IMGS} \
-gpu 0 \
-output ${MODEL_PATH}/${MODEL_FILE_NAME}.fixed-prior_BestModelBleu.pt.${SPLIT}-translations &
python translate_mm_vi.py \
-model ${MODEL_PATH}/${MODEL_FILE_NAME}.conditional-prior_BestModelBleu.pt \
-src ${TEST_2017_SRC} \
-path_to_test_img_feats ${TEST_2017_IMGS} \
-gpu 1 \
-output ${MODEL_PATH}/${MODEL_FILE_NAME}.conditional-prior_BestModelBleu.pt.${SPLIT}-translations &
wait;
# translate the ambiguous MSCOCO test set (2017)
SPLIT="test2017_mscoco"
python translate_mm_vi.py \
-model ${MODEL_PATH}/${MODEL_FILE_NAME}.conditional-prior_BestModelBleu.pt \
-src ${TEST_2017_MSCOCO_SRC} \
-path_to_test_img_feats ${TEST_2017_MSCOCO_IMGS} \
-gpu 0 \
-output ${MODEL_PATH}/${MODEL_FILE_NAME}.conditional-prior_BestModelBleu.pt.${SPLIT}-translations &
python translate_mm_vi.py \
-model ${MODEL_PATH}/${MODEL_FILE_NAME}.fixed-prior_BestModelBleu.pt \
-src ${TEST_2017_MSCOCO_SRC} \
-path_to_test_img_feats ${TEST_2017_MSCOCO_IMGS} \
-gpu 1 \
-output ${MODEL_PATH}/${MODEL_FILE_NAME}.fixed-prior_BestModelBleu.pt.${SPLIT}-translations &
wait;
echo -ne "Finished. Translations of valid/test 2016/test 2017 (Flickr and ambiguous MSCOCO) can be found in:\n${MODEL_PATH}/${MODEL_FILE_NAME}.{fixed,conditional}-prior_BestModelBleu.pt.{validation,test2016,test2017,test2017_mscoco}-translations\n"