Merge pull request espnet#4149 from brianyan918/fix_st_bug

fix bug in mt/st templates for having separate token lists

egs2/TEMPLATE/asr1/scripts/utils/show_translation_result.sh

@@ -0,0 +1,67 @@
+#!/usr/bin/env bash
+mindepth=0
+maxdepth=3
+case=tc
+
+. utils/parse_options.sh
+
+if [ $# -gt 1 ]; then
+    echo "Usage: $0 --mindepth 0 --maxdepth 1 [exp]" 1>&2
+    echo ""
+    echo "Show the system environments and the evaluation results in Markdown format."
+    echo 'The default of <exp> is "exp/".'
+    exit 1
+fi
+
+[ -f ./path.sh ] && . ./path.sh
+set -euo pipefail
+if [ $# -eq 1 ]; then
+    exp=$1
+else
+    exp=exp
+fi
+
+
+cat << EOF
+<!-- Generated by $0 -->
+# RESULTS
+## Environments
+- date: \`$(LC_ALL=C date)\`
+EOF
+
+python3 << EOF
+import sys, espnet, torch
+pyversion = sys.version.replace('\n', ' ')
+
+print(f"""- python version: \`{pyversion}\`
+- espnet version: \`espnet {espnet.__version__}\`
+- pytorch version: \`pytorch {torch.__version__}\`""")
+EOF
+
+cat << EOF
+- Git hash: \`$(git rev-parse HEAD)\`
+  - Commit date: \`$(git log -1 --format='%cd')\`
+
+EOF
+
+metrics="bleu"
+
+while IFS= read -r expdir; do
+    if ls "${expdir}"/*/*/score_*/result.${case}.txt &> /dev/null; then
+        echo "## $(basename ${expdir})"
+        for type in $metrics; do
+                	cat << EOF
+### ${type^^}
+
+|dataset|bleu_score|verbose_score|
+|---|---|---|
+EOF
+    data=$(echo "${expdir}"/*/*/score_*/result.${case}.txt | cut -d '/' -f4)
+    bleu=$(sed -n '5p' "${expdir}"/*/*/score_*/result.${case}.txt | cut -d ' ' -f 3 | tr -d ',')
+    verbose=$(sed -n '7p' "${expdir}"/*/*/score_*/result.${case}.txt | cut -d ' ' -f 3- | tr -d '",')
+    echo "${data}|${bleu}|${verbose}"
+
+        done
+    fi
+
+done < <(find ${exp} -mindepth ${mindepth} -maxdepth ${maxdepth} -type d)

egs2/TEMPLATE/mt1/mt.sh

@@ -299,7 +299,7 @@ if "${token_joint}"; then
     src_bpetoken_list="${tgt_bpetoken_list}"
     src_chartoken_list="${tgt_chartoken_list}"
 else
-    src_bpedir="${token_listdir}/src_bpe_${tgt_bpemode}${tgt_nbpe}"
+    src_bpedir="${token_listdir}/src_bpe_${src_bpemode}${src_nbpe}"
     src_bpeprefix="${src_bpedir}"/bpe
     src_bpemodel="${src_bpeprefix}".model
     src_bpetoken_list="${src_bpedir}"/tokens.txt

egs2/TEMPLATE/st1/st.sh

@@ -325,7 +325,7 @@ if "${token_joint}"; then
     src_bpetoken_list="${tgt_bpetoken_list}"
     src_chartoken_list="${tgt_chartoken_list}"
 else
-    src_bpedir="${token_listdir}/src_bpe_${tgt_bpemode}${tgt_nbpe}"
+    src_bpedir="${token_listdir}/src_bpe_${src_bpemode}${src_nbpe}"
     src_bpeprefix="${src_bpedir}"/bpe
     src_bpemodel="${src_bpeprefix}".model
     src_bpetoken_list="${src_bpedir}"/tokens.txt

espnet2/st/espnet_model.py

@@ -78,6 +78,8 @@ def __init__(
         # note that eos is the same as sos (equivalent ID)
         self.sos = vocab_size - 1
         self.eos = vocab_size - 1
+        self.src_sos = src_vocab_size - 1
+        self.src_eos = src_vocab_size - 1
         self.vocab_size = vocab_size
         self.src_vocab_size = src_vocab_size
         self.ignore_id = ignore_id
@@ -409,7 +411,9 @@ def _calc_asr_att_loss(
         ys_pad: torch.Tensor,
         ys_pad_lens: torch.Tensor,
     ):
-        ys_in_pad, ys_out_pad = add_sos_eos(ys_pad, self.sos, self.eos, self.ignore_id)
+        ys_in_pad, ys_out_pad = add_sos_eos(
+            ys_pad, self.src_sos, self.src_eos, self.ignore_id
+        )
         ys_in_lens = ys_pad_lens + 1
 
         # 1. Forward decoder
@@ -420,7 +424,7 @@ def _calc_asr_att_loss(
         # 2. Compute attention loss
         loss_att = self.criterion_asr(decoder_out, ys_out_pad)
         acc_att = th_accuracy(
-            decoder_out.view(-1, self.vocab_size),
+            decoder_out.view(-1, self.src_vocab_size),
             ys_out_pad,
             ignore_label=self.ignore_id,
         )