update model

bar/decoder_jit_trace-v2-iter-468000-avg-16-pnnx.ncnn.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:01787e159327815555f1a8d589c0bd43e5cb609b120de5ec554a2c7d4e1a3583
+oid sha256:d85e49679bb4993e7f87516e3d0a6dde13a833930cafd092aa41e10c1b94d0cf
 size 514056

bar/encoder_jit_trace-v2-iter-468000-avg-16-pnnx.ncnn.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:00a19320bbf51235a0ee7a744533c7d2eaae6d26a57640cdd21c87aeeea8accf
+oid sha256:1e4f42320511992a8e39586286498d3a7950621f5952bee0671aa7fb972236d6
 size 166239840

bar/encoder_jit_trace-v2-iter-468000-avg-16-pnnx.ncnn.param

@@ -1,5 +1,6 @@
 7767517
-267 379
+268 379
+SherpaMetaData           sherpa_meta_data1        0 0 0=3 1=12 2=512 3=1024
 Input                    in0                      0 1 in0
 Input                    in1                      0 1 in1
 Input                    in2                      0 1 in2
@@ -28,7 +29,7 @@ InnerProduct             linear_18                1 1 48 49 0=512 1=1 2=1179648
 Split                    splitncnn_5              1 3 49 50 51 52
 BinaryOp                 mul_6                    2 1 50 51 53 0=2
 Reduction                mean_80                  1 1 53 54 0=3 1=0 -23303=1,-1 4=1 5=1
-BinaryOp                 add_7                    1 1 54 55 0=0 1=1 2=2.499987e-01
+BinaryOp                 add_7                    1 1 54 55 0=0 1=1 2=2.499982e-01
 BinaryOp                 pow_8                    1 1 55 56 0=6 1=1 2=-5.000000e-01
 BinaryOp                 mul_9                    2 1 52 56 57 0=2
 BinaryOp                 sub_10                   1 1 in1 58 0=1 1=1 2=3.000000e+00
@@ -53,7 +54,7 @@ BinaryOp                 add_19                   2 1 72 80 81 0=0
 Split                    splitncnn_9              1 3 81 82 83 84
 BinaryOp                 mul_20                   2 1 82 83 85 0=2
 Reduction                mean_81                  1 1 85 86 0=3 1=0 -23303=1,-1 4=1 5=1
-BinaryOp                 add_21                   1 1 86 87 0=0 1=1 2=1.285161e+00
+BinaryOp                 add_21                   1 1 86 87 0=0 1=1 2=1.285091e+00
 BinaryOp                 pow_22                   1 1 87 88 0=6 1=1 2=-5.000000e-01
 BinaryOp                 mul_23                   2 1 84 88 89 0=2
 Split                    splitncnn_10             1 2 89 90 91
@@ -72,7 +73,7 @@ BinaryOp                 add_27                   2 1 98 106 107 0=0
 Split                    splitncnn_13             1 3 107 108 109 110
 BinaryOp                 mul_28                   2 1 108 109 111 0=2
 Reduction                mean_82                  1 1 111 112 0=3 1=0 -23303=1,-1 4=1 5=1
-BinaryOp                 add_29                   1 1 112 113 0=0 1=1 2=1.770352e+00
+BinaryOp                 add_29                   1 1 112 113 0=0 1=1 2=1.768978e+00
 BinaryOp                 pow_30                   1 1 113 114 0=6 1=1 2=-5.000000e-01
 BinaryOp                 mul_31                   2 1 110 114 115 0=2
 Split                    splitncnn_14             1 2 115 116 117
@@ -91,7 +92,7 @@ BinaryOp                 add_35                   2 1 124 132 133 0=0
 Split                    splitncnn_17             1 3 133 134 135 136
 BinaryOp                 mul_36                   2 1 134 135 137 0=2
 Reduction                mean_83                  1 1 137 138 0=3 1=0 -23303=1,-1 4=1 5=1
-BinaryOp                 add_37                   1 1 138 139 0=0 1=1 2=2.005826e+00
+BinaryOp                 add_37                   1 1 138 139 0=0 1=1 2=2.004356e+00
 BinaryOp                 pow_38                   1 1 139 140 0=6 1=1 2=-5.000000e-01
 BinaryOp                 mul_39                   2 1 136 140 141 0=2
 Split                    splitncnn_18             1 2 141 142 143
@@ -110,7 +111,7 @@ BinaryOp                 add_43                   2 1 150 158 159 0=0
 Split                    splitncnn_21             1 3 159 160 161 162
 BinaryOp                 mul_44                   2 1 160 161 163 0=2
 Reduction                mean_84                  1 1 163 164 0=3 1=0 -23303=1,-1 4=1 5=1
-BinaryOp                 add_45                   1 1 164 165 0=0 1=1 2=1.877999e+00
+BinaryOp                 add_45                   1 1 164 165 0=0 1=1 2=1.876125e+00
 BinaryOp                 pow_46                   1 1 165 166 0=6 1=1 2=-5.000000e-01
 BinaryOp                 mul_47                   2 1 162 166 167 0=2
 Split                    splitncnn_22             1 2 167 168 169
@@ -129,7 +130,7 @@ BinaryOp                 add_51                   2 1 176 184 185 0=0
 Split                    splitncnn_25             1 3 185 186 187 188
 BinaryOp                 mul_52                   2 1 186 187 189 0=2
 Reduction                mean_85                  1 1 189 190 0=3 1=0 -23303=1,-1 4=1 5=1
-BinaryOp                 add_53                   1 1 190 191 0=0 1=1 2=2.357182e+00
+BinaryOp                 add_53                   1 1 190 191 0=0 1=1 2=2.354979e+00
 BinaryOp                 pow_54                   1 1 191 192 0=6 1=1 2=-5.000000e-01
 BinaryOp                 mul_55                   2 1 188 192 193 0=2
 Split                    splitncnn_26             1 2 193 194 195
@@ -148,7 +149,7 @@ BinaryOp                 add_59                   2 1 202 210 211 0=0
 Split                    splitncnn_29             1 3 211 212 213 214
 BinaryOp                 mul_60                   2 1 212 213 215 0=2
 Reduction                mean_86                  1 1 215 216 0=3 1=0 -23303=1,-1 4=1 5=1
-BinaryOp                 add_61                   1 1 216 217 0=0 1=1 2=2.294552e+00
+BinaryOp                 add_61                   1 1 216 217 0=0 1=1 2=2.291315e+00
 BinaryOp                 pow_62                   1 1 217 218 0=6 1=1 2=-5.000000e-01
 BinaryOp                 mul_63                   2 1 214 218 219 0=2
 Split                    splitncnn_30             1 2 219 220 221
@@ -167,7 +168,7 @@ BinaryOp                 add_67                   2 1 228 236 237 0=0
 Split                    splitncnn_33             1 3 237 238 239 240
 BinaryOp                 mul_68                   2 1 238 239 241 0=2
 Reduction                mean_87                  1 1 241 242 0=3 1=0 -23303=1,-1 4=1 5=1
-BinaryOp                 add_69                   1 1 242 243 0=0 1=1 2=3.318981e+00
+BinaryOp                 add_69                   1 1 242 243 0=0 1=1 2=3.315238e+00
 BinaryOp                 pow_70                   1 1 243 244 0=6 1=1 2=-5.000000e-01
 BinaryOp                 mul_71                   2 1 240 244 245 0=2
 Split                    splitncnn_34             1 2 245 246 247
@@ -186,7 +187,7 @@ BinaryOp                 add_75                   2 1 254 262 263 0=0
 Split                    splitncnn_37             1 3 263 264 265 266
 BinaryOp                 mul_76                   2 1 264 265 267 0=2
 Reduction                mean_88                  1 1 267 268 0=3 1=0 -23303=1,-1 4=1 5=1
-BinaryOp                 add_77                   1 1 268 269 0=0 1=1 2=2.019367e+00
+BinaryOp                 add_77                   1 1 268 269 0=0 1=1 2=2.017090e+00
 BinaryOp                 pow_78                   1 1 269 270 0=6 1=1 2=-5.000000e-01
 BinaryOp                 mul_79                   2 1 266 270 271 0=2
 Split                    splitncnn_38             1 2 271 272 273
@@ -205,7 +206,7 @@ BinaryOp                 add_83                   2 1 280 288 289 0=0
 Split                    splitncnn_41             1 3 289 290 291 292
 BinaryOp                 mul_84                   2 1 290 291 293 0=2
 Reduction                mean_89                  1 1 293 294 0=3 1=0 -23303=1,-1 4=1 5=1
-BinaryOp                 add_85                   1 1 294 295 0=0 1=1 2=2.291728e+00
+BinaryOp                 add_85                   1 1 294 295 0=0 1=1 2=2.289504e+00
 BinaryOp                 pow_86                   1 1 295 296 0=6 1=1 2=-5.000000e-01
 BinaryOp                 mul_87                   2 1 292 296 297 0=2
 Split                    splitncnn_42             1 2 297 298 299
@@ -224,7 +225,7 @@ BinaryOp                 add_91                   2 1 306 314 315 0=0
 Split                    splitncnn_45             1 3 315 316 317 318
 BinaryOp                 mul_92                   2 1 316 317 319 0=2
 Reduction                mean_90                  1 1 319 320 0=3 1=0 -23303=1,-1 4=1 5=1
-BinaryOp                 add_93                   1 1 320 321 0=0 1=1 2=1.188224e+00
+BinaryOp                 add_93                   1 1 320 321 0=0 1=1 2=1.186822e+00
 BinaryOp                 pow_94                   1 1 321 322 0=6 1=1 2=-5.000000e-01
 BinaryOp                 mul_95                   2 1 318 322 323 0=2
 Split                    splitncnn_46             1 2 323 324 325
@@ -243,7 +244,7 @@ BinaryOp                 add_99                   2 1 332 340 341 0=0
 Split                    splitncnn_49             1 3 341 342 343 344
 BinaryOp                 mul_100                  2 1 342 343 345 0=2
 Reduction                mean_91                  1 1 345 346 0=3 1=0 -23303=1,-1 4=1 5=1
-BinaryOp                 add_101                  1 1 346 347 0=0 1=1 2=1.185556e+00
+BinaryOp                 add_101                  1 1 346 347 0=0 1=1 2=1.183580e+00
 BinaryOp                 pow_102                  1 1 347 348 0=6 1=1 2=-5.000000e-01
 BinaryOp                 mul_103                  2 1 344 348 349 0=2
 Split                    splitncnn_50             1 2 349 350 351
@@ -262,7 +263,7 @@ BinaryOp                 add_107                  2 1 358 366 367 0=0
 Split                    splitncnn_53             1 3 367 368 369 370
 BinaryOp                 mul_108                  2 1 368 369 371 0=2
 Reduction                mean_92                  1 1 371 372 0=3 1=0 -23303=1,-1 4=1 5=1
-BinaryOp                 add_109                  1 1 372 373 0=0 1=1 2=1.118955e+00
+BinaryOp                 add_109                  1 1 372 373 0=0 1=1 2=1.117711e+00
 BinaryOp                 pow_110                  1 1 373 374 0=6 1=1 2=-5.000000e-01
 BinaryOp                 mul_111                  2 1 370 374 out0 0=2
 Concat                   cat_0                    12 1 70 96 122 148 174 200 226 252 278 304 330 356 out3 0=0

bar/export-for-ncnn-lstm.sh

@@ -0,0 +1,58 @@
+#!/usr/bin/env bash
+
+# Please go to
+# https://huggingface.co/csukuangfj/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03
+# to download the pre-trained models
+
+if [ ! -d lstm2 ]; then
+  mkdir lstm2
+  pushd lstm2
+  ln -s ~/open-source/icefall-models/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/exp/pretrained-iter-468000-avg-16.pt epoch-99.pt
+  ln -s ~/open-source/icefall-models/icefall-asr-librispeech-lstm-transducer-stateless2-2022-09-03/data .
+  popd
+fi
+
+export CUDA_VISIBLE_DEVICES=""
+set -ex
+
+./lstm_transducer_stateless2/export-for-ncnn.py \
+  --exp-dir ./lstm2 \
+  --bpe-model ./lstm2/data/lang_bpe_500/bpe.model \
+  --epoch 99 \
+  --avg 1 \
+  --use-averaged-model 0 \
+  --num-encoder-layers 12 \
+  --encoder-dim 512 \
+  --rnn-hidden-size 1024
+
+cd ./lstm2
+
+pnnx encoder_jit_trace-pnnx.pt
+pnnx decoder_jit_trace-pnnx.pt
+pnnx joiner_jit_trace-pnnx.pt
+
+# Now we need to modified encoder_jit_trace-pnnx.ncnn.param for 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  2023-02-15 11:27:14.000000000 +0800
+# +++ ./encoder_jit_trace-pnnx.ncnn.param 2023-02-15 11:35:41.000000000 +0800
+# @@ -1,5 +1,6 @@
+#  7767517
+# -267 379
+# +268 379
+# +SherpaMetaData           sherpa_meta_data1        0 0 0=3 1=12 2=512 3=1024
+#  Input                    in0                      0 1 in0
+#  Input                    in1                      0 1 in1
+#  Input                    in2                      0 1 in2
+
+#
+#
+# Explanation:
+# 1. A new layer is added, so we change the layer number 267 to 268
+# 2. SherpaMetaData, it is the layer type. MUST be SherpaMetaData
+# 3. sherpa_meta_data1, it is the layer name. MUST be sherpa_meta_data1
+# 4. 0=3, attribute 0, MUST be 3
+# 5. 1=12, attribute 1, 12 is the value of --num-encoder-layers
+# 6. 2=512, attribute 2, 512 is the value of --encoder-dim
+# 7. 3=1024, attribute 3, 1024 is the value of --rnn-hidden-size

bar/joiner_jit_trace-v2-iter-468000-avg-16-pnnx.ncnn.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a9e96223c3c1f60d39fca5264a418e561d53a684fc5694c14999b1c29a235ae7
+oid sha256:50a8e255319475d80ac5c0aefd7c91ae557c844700168e536d9af1798f71f67d
 size 1566684