run.sh

CUDA_VISIBLE_DEVICES=7
DATA_PATH=data/raw.both.back

# python preprocess.py -train_src ${DATA_PATH}/train.en \
# -train_tgt ${DATA_PATH}/train.de \
# -save_data ${DATA_PATH}/demo
# -valid_src ${DATA_PATH}/val.en.tok \
# -valid_tgt ${DATA_PATH}/val.de.tok \

# ./tools/embeddings_to_torch.py -emb_file "glove_dir/glove.6B.300d.txt" \
# -dict_file ${DATA_PATH}/demo.vocab.pt \
# -output_file ${DATA_PATH}/embeddings

MODEL_PATH=model.query.back.b128
MODEL_SNAPSHOT=ADAM_acc_87.16_ppl_1.74_e22.pt

echo ${MODEL_PATH}/${MODEL_SNAPSHOT}

# python train_mm.py -data ${DATA_PATH}/demo -save_model ${MODEL_PATH}/ADAM \
# -path_to_train_img_feats ${DATA_PATH}/flickr30k_train_resnet50_cnn_features.hdf5 \
# -path_to_valid_img_feats data/flickr30k/features/flickr30k_valid_resnet50_cnn_features.hdf5 \
# -gpuid 0 -epochs 200 -layers 6 -rnn_size 300 -word_vec_size 512 \
# -encoder_type transformer -decoder_type transformer -position_encoding \
# -max_generator_batches 2 -dropout 0.1 \
# -batch_size 128 -accum_count 1 -use_both \
# -optim adam -adam_beta2 0.998 -decay_method noam -warmup_steps 8000 -learning_rate 2 \
# -max_grad_norm 0 -param_init 0 \
# -label_smoothing 0.1 --multimodal_model_type graphtransformer \
# -pre_word_vecs_enc ${DATA_PATH}/embeddings.enc.pt \
# -pre_word_vecs_dec ${DATA_PATH}/embeddings.dec.pt


# -train_from ${MODEL_PATH}/${MODEL_SNAPSHOT} \

# -train_from model_snapshots.stanford.graphtransformer.query/IMGD_ADAM_acc_68.29_ppl_10.79_e94.pt \
 
python translate_mm.py -src ${DATA_PATH}/test_2016_flickr.lc.norm.tok.en -model ${MODEL_PATH}/${MODEL_SNAPSHOT} -path_to_test_img_feats data/flickr30k/features/flickr30k_test_resnet50_cnn_features.hdf5 -output ${MODEL_PATH}/${MODEL_SNAPSHOT}.translations-test2016

perl tools/multi-bleu.perl ${DATA_PATH}/test_2016_flickr.lc.norm.tok.de < ${MODEL_PATH}/${MODEL_SNAPSHOT}.translations-test2016 #> result.query.back.b128.86ep

echo ${MODEL_PATH}/${MODEL_SNAPSHOT}