export-for-ncnn-bilingual-small.sh

#!/usr/bin/env bash

set -e

# Please download the torchscript model from
# https://huggingface.co/pfluo/k2fsa-zipformer-bilingual-zh-en-t

if [ ! -d bilingual-small ]; then
  mkdir -p bilingual-small
  pushd bilingual-small
  ln -s ~/open-source/icefall-models/k2fsa-zipformer-bilingual-zh-en-t/exp/pretrained.pt epoch-99.pt
  ln -s ~/open-source/icefall-models/k2fsa-zipformer-bilingual-zh-en-t/data .
  popd
fi

./pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py \
  --lang-dir ./bilingual-small/data/lang_char_bpe \
  --exp-dir ./bilingual-small \
  --use-averaged-model 0 \
  --epoch 99 \
  --avg 1 \
  --decode-chunk-len 32 \
  --num-encoder-layers "2,2,2,2,2" \
  --feedforward-dims "768,768,768,768,768" \
  --nhead "4,4,4,4,4" \
  --encoder-dims "256,256,256,256,256" \
  --attention-dims "192,192,192,192,192" \
  --encoder-unmasked-dims "192,192,192,192,192" \
  --zipformer-downsampling-factors "1,2,4,8,2" \
  --cnn-module-kernels "31,31,31,31,31" \
  --decoder-dim 512 \
  --joiner-dim 512

cd bilingual-small

pnnx encoder_jit_trace-pnnx.pt
pnnx decoder_jit_trace-pnnx.pt
pnnx joiner_jit_trace-pnnx.pt
# modify encoder_jit_trace-pnnx.ncnn.param to support sherpa-ncnn
# The following is the diff
# $ diff -uN ./encoder_jit_trace-pnnx.ncnn.param-before encoder_jit_trace-pnnx.ncnn.param
# --- ./encoder_jit_trace-pnnx.ncnn.param-before  2023-02-16 10:40:18.000000000 +0800
# +++ encoder_jit_trace-pnnx.ncnn.param   2023-02-16 10:43:21.000000000 +0800
# @@ -1,5 +1,6 @@
#  7767517
# -1423 1762
# +1424 1762
# +SherpaMetaData           sherpa_meta_data1        0 0 0=2 1=32 2=4 3=7 -23316=5,2,2,2,2,2 -23317=5,256,256,256,256,256 -23318=5,192,192,192,192,192 -23319=5,1,2,4,8,2 -23320=5,31,31,31,31,31
#  Input                    in0                      0 1 in0
#  Input                    in1                      0 1 in1
#  Split                    splitncnn_0              1 2 in1 2 3
#
#------
# Explanation:
#
# (1) 1423 is changed to 1424 as an extra layer SherpaMetaData is added
# (2) SherpaMetaData is the layer type
# (3) sherpa_meta_data1 is the name of this layer. Must be sherpa_meta_data1
# (4) 0 0 means this layer has no input or output
# (5) 1=32, attribute 1, 32 is the value of --decode-chunk-len
# (6) 2=4, attribute 2, 4 is the value of --num-left-chunks
# (7) 3=7, attribute 3, 7 is the pad length. The first subsampling layer is using (x_len - 7) // 2, so we use 7 here
# (8) -23316=5,2,2,2,2,2, attribute 16, this is an array attribute. It is attribute 16 since -23300 - (-23316) = 16
#       the first element of the array is the length of the array, which is 5 in our case.
#       2,2,2,2,2 is the value of --num-encoder-layers
# (9) -23317=5,256,256,256,256,256, attribute 17. 256,256,256,256,256 is the value of --encoder-dims
# (10) -23318=5,192,192,192,192,192, attribute 18, 192,192,192,192,192 is the value of --attention-dims
# (11) -23319=5,1,2,4,8,2, attribute 19, 1,2,4,8,2 is the value of --zipformer-downsampling-factors
# (12) -23320=5,31,31,31,31,31, attribute 20, 31,31,31,31,31 is the value of --cnn-module-kernels