initial commit

decoder_jit_trace-pnnx.ncnn.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:26ff4fa3e415f54a7ee6b4360873c25db764f68ca010139864008ab0391e0192
+size 3548808

decoder_jit_trace-pnnx.ncnn.param

@@ -0,0 +1,8 @@
+7767517
+6 6
+Input                    in0                      0 1 in0
+Embed                    embed_1                  1 1 in0 1 0=320 1=5537 2=0 3=1771840
+Permute                  permute_2                1 1 1 2 0=1
+ConvolutionDepthWise1D   convdw1d_4               1 1 2 3 0=320 1=2 2=1 3=1 4=0 5=0 6=2560 7=80
+Permute                  permute_3                1 1 3 4 0=1
+ReLU                     relu_0                   1 1 4 out0

encoder_jit_trace-pnnx.ncnn.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f951dfe1fbee93edaf40ae857b1fb164424316ae702393dd1491efcb2ce2590
+size 17150324

export-ncnn.sh

@@ -0,0 +1,60 @@
+#!/usr/bin/env bash
+
+# Here are the model hyper-parameters required for model exporting
+
+exp_dir=pruned_transducer_stateless7_streaming/exp-14M
+
+python ./pruned_transducer_stateless7_streaming/export-for-ncnn-zh.py \
+    --lang-dir data/lang_char \
+    --exp-dir $exp_dir \
+    --use-averaged-model True \
+    --iter 400000 \
+    --avg 8 \
+    --decode-chunk-len 32 \
+    --num-encoder-layers "2,3,2,2,3" \
+    --feedforward-dims "320,320,640,640,320" \
+    --nhead "4,4,4,4,4" \
+    --encoder-dims "160,160,160,160,160" \
+    --attention-dims "96,96,96,96,96" \
+    --encoder-unmasked-dims "128,128,128,128,128" \
+    --decoder-dim 320 \
+    --joiner-dim 320
+
+pushd $exp_dir
+
+pnnx encoder_jit_trace-pnnx.pt
+pnnx decoder_jit_trace-pnnx.pt
+pnnx joiner_jit_trace-pnnx.pt
+
+popd
+
+# modify encoder_jit_trace-pnnx.ncnn.param to support sherpa-ncnn
+# The following is the diff
+# --- encoder_jit_trace-pnnx.ncnn.param-before    2023-02-14 20:48:52.000000000 +0800
+# +++ encoder_jit_trace-pnnx.ncnn.param   2023-02-14 20:50:15.000000000 +0800
+# @@ -1,5 +1,6 @@
+#  7767517
+# -2028 2547
+# +2029 2547
+# +SherpaMetaData           sherpa_meta_data1        0 0 0=2 1=32 2=4 3=7 -23316=5,2,4,3,2,4 -23317=5,384,384,384,384,384 -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) 2028 is changed to 2029 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=5,2,4,3,2,4, 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.
+#       2,4,3,2,4 is the value of --num-encoder-layers
+# (9) -23317=5,384,384,384,384,384, attribute 17. 384,384,384,384,384 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

joiner_jit_trace-pnnx.ncnn.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5209db16c9308e4b83134829f6de5ce993be5c340fd962b18c140f5e27c5c42
+size 3875600

joiner_jit_trace-pnnx.ncnn.param

@@ -0,0 +1,9 @@
+7767517
+7 7
+Input                    in0                      0 1 in0
+Input                    in1                      0 1 in1
+InnerProduct             linear_2                 1 1 in1 2 0=320 1=1 2=102400
+InnerProduct             linear_1                 1 1 in0 3 0=320 1=1 2=51200
+BinaryOp                 add_0                    2 1 3 2 4 0=0
+TanH                     tanh_0                   1 1 4 5
+InnerProduct             linear_3                 1 1 5 out0 0=5537 1=1 2=1771840