Upload 15 files

Embedding.mlmodelc/analytics/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:28c3ba1fd51947d404655ab2322788f821e6851540444829082ad54915bb3363
+oid sha256:4b2449667626ef30bbcf9aad719b8ad5380aab470e3c3bddf5ece68c101cfd5c
 size 243

Embedding.mlmodelc/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:eda47173407bfdadb8af3b6c18f9f2fa3146188ff42f3cb6ae7d9c6657e1fdc1
+oid sha256:8af46b5ba390f4e1898446f30c318d19f114410e91bd4e2a30839ad3900e3f12
 size 646

Embedding.mlmodelc/metadata.json

@@ -20,7 +20,7 @@
     ],
     "author" : "Fluid Inference",
     "specificationVersion" : 8,
-    "storagePrecision" : "Float32",
+    "storagePrecision" : "Mixed (Float32, Int8)",
     "license" : "CC-BY-4.0",
     "mlProgramOperationTypeHistogram" : {
       "Ios16.reduceL2Norm" : 1,
@@ -35,14 +35,15 @@
       "Ios17.gather" : 2,
       "Pad" : 2,
       "Ios17.log" : 1,
+      "Ios16.constexprAffineDequantize" : 40,
       "Ios17.sqrt" : 1,
       "Ios17.sub" : 5,
       "Ios17.conv" : 41,
       "Ios17.clip" : 3,
       "Ios16.relu" : 33,
-      "Ios17.concat" : 1,
       "Ios17.pow" : 2,
       "Ios17.realDiv" : 4,
+      "Ios17.concat" : 1,
       "Ios17.mul" : 9
     },
     "computePrecision" : "Mixed (Float32, Int32)",
@@ -81,7 +82,7 @@
       "com.github.apple.coremltools.version" : "9.0b1",
       "com.github.apple.coremltools.source_dialect" : "TorchScript"
     },
-    "generatedClassName" : "Embedding",
+    "generatedClassName" : "embedding_community_1_int8_per_channel",
     "method" : "predict"
   }
 ]

Embedding.mlmodelc/weights/weight.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ccbaea148fafa13a1741620e1cf41cb4027160da2a56c3b9a9acdd711252b057
-size 28303104
+oid sha256:9ffc18432c23faf3d6b233c85e859876b3aaf0f1557cd4432bb24b8881eddecc
+size 7124224

FBank.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e8bd3a8b82ac123580989f490e4d9245127c535857630b543311268accc3f0a
+size 243

FBank.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57ac436bb0671cbb5527a339134d695f752eb77f7a18966b93c6835335595759
+size 853

FBank.mlmodelc/metadata.json

@@ -0,0 +1,80 @@
+[
+  {
+    "shortDescription" : "pyannote community-1 FBANK frontend (10 s audio preprocessing to 80×998 features, batch 1-32, CPU preferred)",
+    "metadataOutputVersion" : "3.0",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32)",
+        "shortDescription" : "",
+        "shape" : "[]",
+        "name" : "fbank_features",
+        "type" : "MultiArray"
+      }
+    ],
+    "version" : "pyannote-speaker-diarization-community-1",
+    "modelParameters" : [
+
+    ],
+    "author" : "Fluid Inference",
+    "specificationVersion" : 8,
+    "storagePrecision" : "Float32",
+    "license" : "CC-BY-4.0",
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.mul" : 3,
+      "Ios17.transpose" : 2,
+      "Ios17.sub" : 3,
+      "Ios17.conv" : 4,
+      "Ios17.log" : 1,
+      "Ios17.sliceByIndex" : 1,
+      "Ios16.reduceMean" : 2,
+      "Ios17.add" : 2,
+      "Ios17.clip" : 1,
+      "Ios17.pow" : 2,
+      "Ios17.expandDims" : 4,
+      "Ios17.squeeze" : 4,
+      "Ios17.reshape" : 2,
+      "Pad" : 2
+    },
+    "computePrecision" : "Mixed (Float32, Int32)",
+    "stateSchema" : [
+
+    ],
+    "isUpdatable" : "0",
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "shortDescription" : "",
+        "dataType" : "Float32",
+        "hasShapeFlexibility" : "1",
+        "isOptional" : "0",
+        "shapeFlexibility" : "1 × 1 × 160000 | 2 × 1 × 160000 | 3 × 1 × 160000 | 4 × 1 × 160000 | 5 × 1 × 160000 | 6 × 1 × 160000 | 7 × 1 × 160000 | 8 × 1 × 160000 | 9 × 1 × 160000 | 10 × 1 × 160000 | 11 × 1 × 160000 | 12 × 1 × 160000 | 13 × 1 × 160000 | 14 × 1 × 160000 | 15 × 1 × 160000 | 16 × 1 × 160000 | 17 × 1 × 160000 | 18 × 1 × 160000 | 19 × 1 × 160000 | 20 × 1 × 160000 | 21 × 1 × 160000 | 22 × 1 × 160000 | 23 × 1 × 160000 | 24 × 1 × 160000 | 25 × 1 × 160000 | 26 × 1 × 160000 | 27 × 1 × 160000 | 28 × 1 × 160000 | 29 × 1 × 160000 | 30 × 1 × 160000 | 31 × 1 × 160000 | 32 × 1 × 160000",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 160000)",
+        "type" : "MultiArray",
+        "shape" : "[1, 1, 160000]",
+        "name" : "audio",
+        "enumeratedShapes" : "[[1, 1, 160000], [2, 1, 160000], [3, 1, 160000], [4, 1, 160000], [5, 1, 160000], [6, 1, 160000], [7, 1, 160000], [8, 1, 160000], [9, 1, 160000], [10, 1, 160000], [11, 1, 160000], [12, 1, 160000], [13, 1, 160000], [14, 1, 160000], [15, 1, 160000], [16, 1, 160000], [17, 1, 160000], [18, 1, 160000], [19, 1, 160000], [20, 1, 160000], [21, 1, 160000], [22, 1, 160000], [23, 1, 160000], [24, 1, 160000], [25, 1, 160000], [26, 1, 160000], [27, 1, 160000], [28, 1, 160000], [29, 1, 160000], [30, 1, 160000], [31, 1, 160000], [32, 1, 160000]]"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2025-10-15",
+      "com.github.apple.coremltools.source" : "torch==2.8.0",
+      "com.github.apple.coremltools.version" : "9.0b1",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "fbank_community_1",
+    "method" : "predict"
+  }
+]

FBank.mlmodelc/model.mil

@@ -0,0 +1,96 @@
+program(1.0)
+[buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.8.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0b1"}})]
+{
+    func main<ios17>(tensor<fp32, [?, 1, 160000]> audio) [FlexibleShapeInformation = tuple<tuple<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>, tuple<tensor<string, []>, dict<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>>>((("DefaultShapes", {{"audio", [1, 1, 160000]}}), ("EnumeratedShapes", {{"audio_1_1_10_1_160000_", {{"audio", [10, 1, 160000]}}}, {"audio_1_1_11_1_160000_", {{"audio", [11, 1, 160000]}}}, {"audio_1_1_12_1_160000_", {{"audio", [12, 1, 160000]}}}, {"audio_1_1_13_1_160000_", {{"audio", [13, 1, 160000]}}}, {"audio_1_1_14_1_160000_", {{"audio", [14, 1, 160000]}}}, {"audio_1_1_15_1_160000_", {{"audio", [15, 1, 160000]}}}, {"audio_1_1_16_1_160000_", {{"audio", [16, 1, 160000]}}}, {"audio_1_1_17_1_160000_", {{"audio", [17, 1, 160000]}}}, {"audio_1_1_18_1_160000_", {{"audio", [18, 1, 160000]}}}, {"audio_1_1_19_1_160000_", {{"audio", [19, 1, 160000]}}}, {"audio_1_1_1_1_160000_", {{"audio", [1, 1, 160000]}}}, {"audio_1_1_20_1_160000_", {{"audio", [20, 1, 160000]}}}, {"audio_1_1_21_1_160000_", {{"audio", [21, 1, 160000]}}}, {"audio_1_1_22_1_160000_", {{"audio", [22, 1, 160000]}}}, {"audio_1_1_23_1_160000_", {{"audio", [23, 1, 160000]}}}, {"audio_1_1_24_1_160000_", {{"audio", [24, 1, 160000]}}}, {"audio_1_1_25_1_160000_", {{"audio", [25, 1, 160000]}}}, {"audio_1_1_26_1_160000_", {{"audio", [26, 1, 160000]}}}, {"audio_1_1_27_1_160000_", {{"audio", [27, 1, 160000]}}}, {"audio_1_1_28_1_160000_", {{"audio", [28, 1, 160000]}}}, {"audio_1_1_29_1_160000_", {{"audio", [29, 1, 160000]}}}, {"audio_1_1_2_1_160000_", {{"audio", [2, 1, 160000]}}}, {"audio_1_1_30_1_160000_", {{"audio", [30, 1, 160000]}}}, {"audio_1_1_31_1_160000_", {{"audio", [31, 1, 160000]}}}, {"audio_1_1_32_1_160000_", {{"audio", [32, 1, 160000]}}}, {"audio_1_1_3_1_160000_", {{"audio", [3, 1, 160000]}}}, {"audio_1_1_4_1_160000_", {{"audio", [4, 1, 160000]}}}, {"audio_1_1_5_1_160000_", {{"audio", [5, 1, 160000]}}}, {"audio_1_1_6_1_160000_", {{"audio", [6, 1, 160000]}}}, {"audio_1_1_7_1_160000_", {{"audio", [7, 1, 160000]}}}, {"audio_1_1_8_1_160000_", {{"audio", [8, 1, 160000]}}}, {"audio_1_1_9_1_160000_", {{"audio", [9, 1, 160000]}}}})))] {
+            tensor<fp32, [80, 257, 1]> _fbank_mel_weight = const()[name = tensor<string, []>("_fbank_mel_weight"), val = tensor<fp32, [80, 257, 1]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp32, [257, 1, 512]> _fbank_dft_imag_weight = const()[name = tensor<string, []>("_fbank_dft_imag_weight"), val = tensor<fp32, [257, 1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(82368)))];
+            tensor<fp32, [257, 1, 512]> _fbank_dft_real_weight = const()[name = tensor<string, []>("_fbank_dft_real_weight"), val = tensor<fp32, [257, 1, 512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(608768)))];
+            tensor<fp32, [1, 400]> _fbank_window = const()[name = tensor<string, []>("_fbank_window"), val = tensor<fp32, [1, 400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1135168)))];
+            tensor<fp32, []> _fbank_eps = const()[name = tensor<string, []>("_fbank_eps"), val = tensor<fp32, []>(0x1.0c6f7ap-20)];
+            tensor<fp32, [400, 1, 400]> _fbank_frame_kernel = const()[name = tensor<string, []>("_fbank_frame_kernel"), val = tensor<fp32, [400, 1, 400]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1136832)))];
+            tensor<fp32, []> var_3_promoted = const()[name = tensor<string, []>("op_3_promoted"), val = tensor<fp32, []>(0x1p+15)];
+            tensor<fp32, [?, 1, 160000]> waveforms_3 = mul(x = audio, y = var_3_promoted)[name = tensor<string, []>("waveforms_3")];
+            tensor<string, []> frames_1_pad_type_0 = const()[name = tensor<string, []>("frames_1_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> frames_1_strides_0 = const()[name = tensor<string, []>("frames_1_strides_0"), val = tensor<int32, [1]>([160])];
+            tensor<int32, [2]> frames_1_pad_0 = const()[name = tensor<string, []>("frames_1_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> frames_1_dilations_0 = const()[name = tensor<string, []>("frames_1_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> frames_1_groups_0 = const()[name = tensor<string, []>("frames_1_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [?, 400, 998]> frames_1 = conv(dilations = frames_1_dilations_0, groups = frames_1_groups_0, pad = frames_1_pad_0, pad_type = frames_1_pad_type_0, strides = frames_1_strides_0, weight = _fbank_frame_kernel, x = waveforms_3)[name = tensor<string, []>("frames_1")];
+            tensor<int32, [3]> frames_3_perm_0 = const()[name = tensor<string, []>("frames_3_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [2]> concat_0x = const()[name = tensor<string, []>("concat_0x"), val = tensor<int32, [2]>([-1, 400])];
+            tensor<fp32, [?, 998, 400]> frames_3 = transpose(perm = frames_3_perm_0, x = frames_1)[name = tensor<string, []>("transpose_1")];
+            tensor<fp32, [?, 400]> frames_5 = reshape(shape = concat_0x, x = frames_3)[name = tensor<string, []>("frames_5")];
+            tensor<int32, [1]> var_53_axes_0 = const()[name = tensor<string, []>("op_53_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, []> var_53_keep_dims_0 = const()[name = tensor<string, []>("op_53_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp32, [?, 1]> var_53 = reduce_mean(axes = var_53_axes_0, keep_dims = var_53_keep_dims_0, x = frames_5)[name = tensor<string, []>("op_53")];
+            tensor<fp32, [?, 400]> frames_7 = sub(x = frames_5, y = var_53)[name = tensor<string, []>("frames_7")];
+            tensor<int32, [1]> input_1_axes_0 = const()[name = tensor<string, []>("input_1_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [?, 1, 400]> input_1 = expand_dims(axes = input_1_axes_0, x = frames_7)[name = tensor<string, []>("input_1")];
+            tensor<fp32, []> const_0 = const()[name = tensor<string, []>("const_0"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [6]> var_57_pad_0 = const()[name = tensor<string, []>("op_57_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 1, 0])];
+            tensor<string, []> var_57_mode_0 = const()[name = tensor<string, []>("op_57_mode_0"), val = tensor<string, []>("replicate")];
+            tensor<fp32, [?, 1, 401]> var_57 = pad(constant_val = const_0, mode = var_57_mode_0, pad = var_57_pad_0, x = input_1)[name = tensor<string, []>("op_57")];
+            tensor<int32, [1]> padded_axes_0 = const()[name = tensor<string, []>("padded_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [?, 401]> padded = squeeze(axes = padded_axes_0, x = var_57)[name = tensor<string, []>("padded")];
+            tensor<int32, [2]> var_60_begin_0 = const()[name = tensor<string, []>("op_60_begin_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [2]> var_60_end_0 = const()[name = tensor<string, []>("op_60_end_0"), val = tensor<int32, [2]>([0, 400])];
+            tensor<bool, [2]> var_60_end_mask_0 = const()[name = tensor<string, []>("op_60_end_mask_0"), val = tensor<bool, [2]>([true, false])];
+            tensor<fp32, [?, 400]> var_60 = slice_by_index(begin = var_60_begin_0, end = var_60_end_0, end_mask = var_60_end_mask_0, x = padded)[name = tensor<string, []>("op_60")];
+            tensor<fp32, []> var_61 = const()[name = tensor<string, []>("op_61"), val = tensor<fp32, []>(0x1.f0a3d8p-1)];
+            tensor<fp32, [?, 400]> var_62 = mul(x = var_60, y = var_61)[name = tensor<string, []>("op_62")];
+            tensor<fp32, [?, 400]> frames_9 = sub(x = frames_7, y = var_62)[name = tensor<string, []>("frames_9")];
+            tensor<fp32, [?, 400]> frames_11 = mul(x = frames_9, y = _fbank_window)[name = tensor<string, []>("frames_11")];
+            tensor<int32, [1]> input_axes_0 = const()[name = tensor<string, []>("input_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [?, 1, 400]> input = expand_dims(axes = input_axes_0, x = frames_11)[name = tensor<string, []>("input")];
+            tensor<fp32, []> const_1 = const()[name = tensor<string, []>("const_1"), val = tensor<fp32, []>(0x0p+0)];
+            tensor<int32, [6]> var_67_pad_0 = const()[name = tensor<string, []>("op_67_pad_0"), val = tensor<int32, [6]>([0, 0, 0, 0, 0, 112])];
+            tensor<string, []> var_67_mode_0 = const()[name = tensor<string, []>("op_67_mode_0"), val = tensor<string, []>("constant")];
+            tensor<fp32, [?, 1, 512]> var_67 = pad(constant_val = const_1, mode = var_67_mode_0, pad = var_67_pad_0, x = input)[name = tensor<string, []>("op_67")];
+            tensor<string, []> var_74_pad_type_0 = const()[name = tensor<string, []>("op_74_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> var_74_strides_0 = const()[name = tensor<string, []>("op_74_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_74_pad_0 = const()[name = tensor<string, []>("op_74_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_74_dilations_0 = const()[name = tensor<string, []>("op_74_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> var_74_groups_0 = const()[name = tensor<string, []>("op_74_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [?, 257, 1]> var_74 = conv(dilations = var_74_dilations_0, groups = var_74_groups_0, pad = var_74_pad_0, pad_type = var_74_pad_type_0, strides = var_74_strides_0, weight = _fbank_dft_real_weight, x = var_67)[name = tensor<string, []>("op_74")];
+            tensor<int32, [1]> real_axes_0 = const()[name = tensor<string, []>("real_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [?, 257]> real = squeeze(axes = real_axes_0, x = var_74)[name = tensor<string, []>("real")];
+            tensor<string, []> var_80_pad_type_0 = const()[name = tensor<string, []>("op_80_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> var_80_strides_0 = const()[name = tensor<string, []>("op_80_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_80_pad_0 = const()[name = tensor<string, []>("op_80_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_80_dilations_0 = const()[name = tensor<string, []>("op_80_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> var_80_groups_0 = const()[name = tensor<string, []>("op_80_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [?, 257, 1]> var_80 = conv(dilations = var_80_dilations_0, groups = var_80_groups_0, pad = var_80_pad_0, pad_type = var_80_pad_type_0, strides = var_80_strides_0, weight = _fbank_dft_imag_weight, x = var_67)[name = tensor<string, []>("op_80")];
+            tensor<int32, [1]> imag_axes_0 = const()[name = tensor<string, []>("imag_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [?, 257]> imag = squeeze(axes = imag_axes_0, x = var_80)[name = tensor<string, []>("imag")];
+            tensor<fp32, []> var_22_promoted = const()[name = tensor<string, []>("op_22_promoted"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [?, 257]> var_82 = pow(x = real, y = var_22_promoted)[name = tensor<string, []>("op_82")];
+            tensor<fp32, []> var_22_promoted_1 = const()[name = tensor<string, []>("op_22_promoted_1"), val = tensor<fp32, []>(0x1p+1)];
+            tensor<fp32, [?, 257]> var_83 = pow(x = imag, y = var_22_promoted_1)[name = tensor<string, []>("op_83")];
+            tensor<fp32, [?, 257]> power = add(x = var_82, y = var_83)[name = tensor<string, []>("power")];
+            tensor<int32, [1]> var_85_axes_0 = const()[name = tensor<string, []>("op_85_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [?, 257, 1]> var_85 = expand_dims(axes = var_85_axes_0, x = power)[name = tensor<string, []>("op_85")];
+            tensor<string, []> var_90_pad_type_0 = const()[name = tensor<string, []>("op_90_pad_type_0"), val = tensor<string, []>("valid")];
+            tensor<int32, [1]> var_90_strides_0 = const()[name = tensor<string, []>("op_90_strides_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, [2]> var_90_pad_0 = const()[name = tensor<string, []>("op_90_pad_0"), val = tensor<int32, [2]>([0, 0])];
+            tensor<int32, [1]> var_90_dilations_0 = const()[name = tensor<string, []>("op_90_dilations_0"), val = tensor<int32, [1]>([1])];
+            tensor<int32, []> var_90_groups_0 = const()[name = tensor<string, []>("op_90_groups_0"), val = tensor<int32, []>(1)];
+            tensor<fp32, [?, 80, 1]> var_90 = conv(dilations = var_90_dilations_0, groups = var_90_groups_0, pad = var_90_pad_0, pad_type = var_90_pad_type_0, strides = var_90_strides_0, weight = _fbank_mel_weight, x = var_85)[name = tensor<string, []>("op_90")];
+            tensor<int32, [1]> mel_1_axes_0 = const()[name = tensor<string, []>("mel_1_axes_0"), val = tensor<int32, [1]>([-1])];
+            tensor<fp32, [?, 80]> mel_1 = squeeze(axes = mel_1_axes_0, x = var_90)[name = tensor<string, []>("mel_1")];
+            tensor<fp32, [?, 80]> mel_3 = add(x = mel_1, y = _fbank_eps)[name = tensor<string, []>("mel_3")];
+            tensor<fp32, []> const_2 = const()[name = tensor<string, []>("const_2"), val = tensor<fp32, []>(0x1.fffffep+127)];
+            tensor<fp32, [?, 80]> clip_0 = clip(alpha = _fbank_eps, beta = const_2, x = mel_3)[name = tensor<string, []>("clip_0")];
+            tensor<fp32, []> mel_epsilon_0 = const()[name = tensor<string, []>("mel_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
+            tensor<fp32, [?, 80]> mel = log(epsilon = mel_epsilon_0, x = clip_0)[name = tensor<string, []>("mel")];
+            tensor<int32, [3]> concat_1x = const()[name = tensor<string, []>("concat_1x"), val = tensor<int32, [3]>([-1, 998, 80])];
+            tensor<fp32, [?, 998, 80]> var_96 = reshape(shape = concat_1x, x = mel)[name = tensor<string, []>("op_96")];
+            tensor<int32, [1]> centered_axes_0 = const()[name = tensor<string, []>("centered_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<bool, []> centered_keep_dims_0 = const()[name = tensor<string, []>("centered_keep_dims_0"), val = tensor<bool, []>(true)];
+            tensor<fp32, [?, 1, 80]> centered = reduce_mean(axes = centered_axes_0, keep_dims = centered_keep_dims_0, x = var_96)[name = tensor<string, []>("centered")];
+            tensor<fp32, [?, 998, 80]> features = sub(x = var_96, y = centered)[name = tensor<string, []>("features")];
+            tensor<int32, [3]> var_115 = const()[name = tensor<string, []>("op_115"), val = tensor<int32, [3]>([0, 2, 1])];
+            tensor<int32, [1]> var_118_axes_0 = const()[name = tensor<string, []>("op_118_axes_0"), val = tensor<int32, [1]>([1])];
+            tensor<fp32, [?, 80, 998]> var_116 = transpose(perm = var_115, x = features)[name = tensor<string, []>("transpose_0")];
+            tensor<fp32, [?, 1, 80, 998]> fbank_features = expand_dims(axes = var_118_axes_0, x = var_116)[name = tensor<string, []>("op_118")];
+        } -> (fbank_features);
+}

FBank.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e83fdd3ea78064b078069e4d9141603c61c47a27fd19e7e3142ff7476f8db36
+size 1776896

Segmentation.mlmodelc/analytics/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:64265f8e7ad41a5f68d630c15288c2499cca5892ad49e20096819cdeac004cdb
+oid sha256:518731a13df50d0924c86c919d8b587280dfc40b2f4cd7b7ef3aac75a037305e
 size 243

Segmentation.mlmodelc/coremldata.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5b18b494dc985ac5dabaad0c8ed283044439e4d045dc16eaf5b2aa4bfae4555d
+oid sha256:59adec00fd697d87742db7052380da0ac91bde91acc37ac48a843f0ff07517bb
 size 812

Segmentation.mlmodelc/metadata.json

@@ -20,7 +20,7 @@
     ],
     "author" : "Fluid Inference",
     "specificationVersion" : 8,
-    "storagePrecision" : "Float32",
+    "storagePrecision" : "Float16",
     "license" : "CC-BY-4.0",
     "mlProgramOperationTypeHistogram" : {
       "Ios17.linear" : 3,
@@ -31,16 +31,17 @@
       "Ios17.leakyRelu" : 5,
       "Ios17.gather" : 1,
       "Ios17.concat" : 9,
-      "Fill" : 1,
-      "Ios17.abs" : 1,
       "Ios16.maxPool" : 3,
+      "Ios17.abs" : 1,
+      "Fill" : 1,
       "Ios17.lstm" : 4,
       "Ios16.softmax" : 1,
       "Ios17.instanceNorm" : 4,
+      "Ios17.cast" : 4,
       "Split" : 10,
       "Ios17.squeeze" : 8
     },
-    "computePrecision" : "Mixed (Float32, Int32)",
+    "computePrecision" : "Mixed (Float16, Float32, Int16, Int32, UInt16)",
     "stateSchema" : [
 
     ],
@@ -71,7 +72,7 @@
       }
     ],
     "userDefinedMetadata" : {
-      "com.github.apple.coremltools.conversion_date" : "2025-10-13",
+      "com.github.apple.coremltools.conversion_date" : "2025-10-15",
       "com.github.apple.coremltools.source" : "torch==2.8.0",
       "com.github.apple.coremltools.version" : "9.0b1",
       "com.github.apple.coremltools.source_dialect" : "TorchScript"

Segmentation.mlmodelc/model.mil

@@ -2,217 +2,225 @@ program(1.0)
 [buildInfo = dict<tensor<string, []>, tensor<string, []>>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"}, {"coremltools-component-torch", "2.8.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "9.0b1"}})]
 {
     func main<ios17>(tensor<fp32, [?, 1, 160000]> audio) [FlexibleShapeInformation = tuple<tuple<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>, tuple<tensor<string, []>, dict<tensor<string, []>, dict<tensor<string, []>, tensor<int32, [?]>>>>>((("DefaultShapes", {{"audio", [32, 1, 160000]}}), ("EnumeratedShapes", {{"audio_1_1_10_1_160000_", {{"audio", [10, 1, 160000]}}}, {"audio_1_1_11_1_160000_", {{"audio", [11, 1, 160000]}}}, {"audio_1_1_12_1_160000_", {{"audio", [12, 1, 160000]}}}, {"audio_1_1_13_1_160000_", {{"audio", [13, 1, 160000]}}}, {"audio_1_1_14_1_160000_", {{"audio", [14, 1, 160000]}}}, {"audio_1_1_15_1_160000_", {{"audio", [15, 1, 160000]}}}, {"audio_1_1_16_1_160000_", {{"audio", [16, 1, 160000]}}}, {"audio_1_1_17_1_160000_", {{"audio", [17, 1, 160000]}}}, {"audio_1_1_18_1_160000_", {{"audio", [18, 1, 160000]}}}, {"audio_1_1_19_1_160000_", {{"audio", [19, 1, 160000]}}}, {"audio_1_1_1_1_160000_", {{"audio", [1, 1, 160000]}}}, {"audio_1_1_20_1_160000_", {{"audio", [20, 1, 160000]}}}, {"audio_1_1_21_1_160000_", {{"audio", [21, 1, 160000]}}}, {"audio_1_1_22_1_160000_", {{"audio", [22, 1, 160000]}}}, {"audio_1_1_23_1_160000_", {{"audio", [23, 1, 160000]}}}, {"audio_1_1_24_1_160000_", {{"audio", [24, 1, 160000]}}}, {"audio_1_1_25_1_160000_", {{"audio", [25, 1, 160000]}}}, {"audio_1_1_26_1_160000_", {{"audio", [26, 1, 160000]}}}, {"audio_1_1_27_1_160000_", {{"audio", [27, 1, 160000]}}}, {"audio_1_1_28_1_160000_", {{"audio", [28, 1, 160000]}}}, {"audio_1_1_29_1_160000_", {{"audio", [29, 1, 160000]}}}, {"audio_1_1_2_1_160000_", {{"audio", [2, 1, 160000]}}}, {"audio_1_1_30_1_160000_", {{"audio", [30, 1, 160000]}}}, {"audio_1_1_31_1_160000_", {{"audio", [31, 1, 160000]}}}, {"audio_1_1_32_1_160000_", {{"audio", [32, 1, 160000]}}}, {"audio_1_1_3_1_160000_", {{"audio", [3, 1, 160000]}}}, {"audio_1_1_4_1_160000_", {{"audio", [4, 1, 160000]}}}, {"audio_1_1_5_1_160000_", {{"audio", [5, 1, 160000]}}}, {"audio_1_1_6_1_160000_", {{"audio", [6, 1, 160000]}}}, {"audio_1_1_7_1_160000_", {{"audio", [7, 1, 160000]}}}, {"audio_1_1_8_1_160000_", {{"audio", [8, 1, 160000]}}}, {"audio_1_1_9_1_160000_", {{"audio", [9, 1, 160000]}}}})))] {
-            tensor<fp32, [1]> sincnet_wav_norm1d_bias = const()[name = tensor<string, []>("sincnet_wav_norm1d_bias"), val = tensor<fp32, [1]>([0x1.73505ep-5])];
-            tensor<fp32, [1]> sincnet_wav_norm1d_weight = const()[name = tensor<string, []>("sincnet_wav_norm1d_weight"), val = tensor<fp32, [1]>([0x1.43f862p-7])];
-            tensor<fp32, [80]> sincnet_norm1d_0_bias = const()[name = tensor<string, []>("sincnet_norm1d_0_bias"), val = tensor<fp32, [80]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
-            tensor<fp32, [80]> sincnet_norm1d_0_weight = const()[name = tensor<string, []>("sincnet_norm1d_0_weight"), val = tensor<fp32, [80]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(448)))];
-            tensor<fp32, [60]> sincnet_conv1d_1_bias = const()[name = tensor<string, []>("sincnet_conv1d_1_bias"), val = tensor<fp32, [60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(832)))];
-            tensor<fp32, [60, 80, 5]> sincnet_conv1d_1_weight = const()[name = tensor<string, []>("sincnet_conv1d_1_weight"), val = tensor<fp32, [60, 80, 5]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1152)))];
-            tensor<fp32, [60]> sincnet_norm1d_1_bias = const()[name = tensor<string, []>("sincnet_norm1d_1_bias"), val = tensor<fp32, [60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97216)))];
-            tensor<fp32, [60]> sincnet_norm1d_1_weight = const()[name = tensor<string, []>("sincnet_norm1d_1_weight"), val = tensor<fp32, [60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97536)))];
-            tensor<fp32, [60]> sincnet_conv1d_2_bias = const()[name = tensor<string, []>("sincnet_conv1d_2_bias"), val = tensor<fp32, [60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(97856)))];
-            tensor<fp32, [60, 60, 5]> sincnet_conv1d_2_weight = const()[name = tensor<string, []>("sincnet_conv1d_2_weight"), val = tensor<fp32, [60, 60, 5]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(98176)))];
-            tensor<fp32, [60]> sincnet_norm1d_2_bias = const()[name = tensor<string, []>("sincnet_norm1d_2_bias"), val = tensor<fp32, [60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170240)))];
-            tensor<fp32, [60]> sincnet_norm1d_2_weight = const()[name = tensor<string, []>("sincnet_norm1d_2_weight"), val = tensor<fp32, [60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170560)))];
-            tensor<fp32, [128]> linear_0_bias = const()[name = tensor<string, []>("linear_0_bias"), val = tensor<fp32, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(170880)))];
-            tensor<fp32, [128, 256]> linear_0_weight = const()[name = tensor<string, []>("linear_0_weight"), val = tensor<fp32, [128, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(171456)))];
-            tensor<fp32, [128]> linear_1_bias = const()[name = tensor<string, []>("linear_1_bias"), val = tensor<fp32, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(302592)))];
-            tensor<fp32, [128, 128]> linear_1_weight = const()[name = tensor<string, []>("linear_1_weight"), val = tensor<fp32, [128, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(303168)))];
-            tensor<fp32, [7]> classifier_bias = const()[name = tensor<string, []>("classifier_bias"), val = tensor<fp32, [7]>([-0x1.00e888p+0, 0x1.67cb52p-2, 0x1.3d87fp-1, 0x1.c8aa8p-2, -0x1.445f5ep-2, -0x1.591274p-1, -0x1.8fb70ep-2])];
-            tensor<fp32, [7, 128]> classifier_weight = const()[name = tensor<string, []>("classifier_weight"), val = tensor<fp32, [7, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(368768)))];
             tensor<fp32, []> var_9 = const()[name = tensor<string, []>("op_9"), val = tensor<fp32, []>(0x1.47ae14p-7)];
-            tensor<fp32, []> var_24 = const()[name = tensor<string, []>("op_24"), val = tensor<fp32, []>(0x1.4f8b58p-17)];
-            tensor<fp32, [?, 1, 160000]> waveform = instance_norm(beta = sincnet_wav_norm1d_bias, epsilon = var_24, gamma = sincnet_wav_norm1d_weight, x = audio)[name = tensor<string, []>("waveform")];
-            tensor<fp32, [80, 1, 251]> filters = const()[name = tensor<string, []>("filters"), val = tensor<fp32, [80, 1, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(372416)))];
+            tensor<string, []> audio_to_fp16_dtype_0 = const()[name = tensor<string, []>("audio_to_fp16_dtype_0"), val = tensor<string, []>("fp16")];
+            tensor<fp16, [1]> sincnet_wav_norm1d_weight_to_fp16 = const()[name = tensor<string, []>("sincnet_wav_norm1d_weight_to_fp16"), val = tensor<fp16, [1]>([0x1.44p-7])];
+            tensor<fp16, [1]> sincnet_wav_norm1d_bias_to_fp16 = const()[name = tensor<string, []>("sincnet_wav_norm1d_bias_to_fp16"), val = tensor<fp16, [1]>([0x1.734p-5])];
+            tensor<fp16, []> var_24_to_fp16 = const()[name = tensor<string, []>("op_24_to_fp16"), val = tensor<fp16, []>(0x1.5p-17)];
+            tensor<fp16, [?, 1, 160000]> audio_to_fp16 = cast(dtype = audio_to_fp16_dtype_0, x = audio)[name = tensor<string, []>("cast_19")];
+            tensor<fp16, [?, 1, 160000]> waveform_cast_fp16 = instance_norm(beta = sincnet_wav_norm1d_bias_to_fp16, epsilon = var_24_to_fp16, gamma = sincnet_wav_norm1d_weight_to_fp16, x = audio_to_fp16)[name = tensor<string, []>("waveform_cast_fp16")];
             tensor<string, []> outputs_pad_type_0 = const()[name = tensor<string, []>("outputs_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [1]> outputs_strides_0 = const()[name = tensor<string, []>("outputs_strides_0"), val = tensor<int32, [1]>([10])];
             tensor<int32, [2]> outputs_pad_0 = const()[name = tensor<string, []>("outputs_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> outputs_dilations_0 = const()[name = tensor<string, []>("outputs_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> outputs_groups_0 = const()[name = tensor<string, []>("outputs_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [?, 80, 15975]> outputs = conv(dilations = outputs_dilations_0, groups = outputs_groups_0, pad = outputs_pad_0, pad_type = outputs_pad_type_0, strides = outputs_strides_0, weight = filters, x = waveform)[name = tensor<string, []>("outputs")];
-            tensor<fp32, [?, 80, 15975]> input_1 = abs(x = outputs)[name = tensor<string, []>("input_1")];
+            tensor<fp16, [80, 1, 251]> filters_to_fp16 = const()[name = tensor<string, []>("filters_to_fp16"), val = tensor<fp16, [80, 1, 251]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(64)))];
+            tensor<fp16, [?, 80, 15975]> outputs_cast_fp16 = conv(dilations = outputs_dilations_0, groups = outputs_groups_0, pad = outputs_pad_0, pad_type = outputs_pad_type_0, strides = outputs_strides_0, weight = filters_to_fp16, x = waveform_cast_fp16)[name = tensor<string, []>("outputs_cast_fp16")];
+            tensor<fp16, [?, 80, 15975]> input_1_cast_fp16 = abs(x = outputs_cast_fp16)[name = tensor<string, []>("input_1_cast_fp16")];
             tensor<int32, [1]> var_119 = const()[name = tensor<string, []>("op_119"), val = tensor<int32, [1]>([3])];
             tensor<int32, [1]> var_120 = const()[name = tensor<string, []>("op_120"), val = tensor<int32, [1]>([3])];
             tensor<string, []> input_3_pad_type_0 = const()[name = tensor<string, []>("input_3_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [2]> input_3_pad_0 = const()[name = tensor<string, []>("input_3_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<bool, []> input_3_ceil_mode_0 = const()[name = tensor<string, []>("input_3_ceil_mode_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [?, 80, 5325]> input_3 = max_pool(ceil_mode = input_3_ceil_mode_0, kernel_sizes = var_119, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = var_120, x = input_1)[name = tensor<string, []>("input_3")];
-            tensor<fp32, [?, 80, 5325]> input_5 = instance_norm(beta = sincnet_norm1d_0_bias, epsilon = var_24, gamma = sincnet_norm1d_0_weight, x = input_3)[name = tensor<string, []>("input_5")];
-            tensor<fp32, [?, 80, 5325]> input_7 = leaky_relu(alpha = var_9, x = input_5)[name = tensor<string, []>("input_7")];
+            tensor<fp16, [?, 80, 5325]> input_3_cast_fp16 = max_pool(ceil_mode = input_3_ceil_mode_0, kernel_sizes = var_119, pad = input_3_pad_0, pad_type = input_3_pad_type_0, strides = var_120, x = input_1_cast_fp16)[name = tensor<string, []>("input_3_cast_fp16")];
+            tensor<fp16, [80]> sincnet_norm1d_0_weight_to_fp16 = const()[name = tensor<string, []>("sincnet_norm1d_0_weight_to_fp16"), val = tensor<fp16, [80]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40320)))];
+            tensor<fp16, [80]> sincnet_norm1d_0_bias_to_fp16 = const()[name = tensor<string, []>("sincnet_norm1d_0_bias_to_fp16"), val = tensor<fp16, [80]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40576)))];
+            tensor<fp16, [?, 80, 5325]> input_5_cast_fp16 = instance_norm(beta = sincnet_norm1d_0_bias_to_fp16, epsilon = var_24_to_fp16, gamma = sincnet_norm1d_0_weight_to_fp16, x = input_3_cast_fp16)[name = tensor<string, []>("input_5_cast_fp16")];
+            tensor<fp16, [?, 80, 5325]> input_7_cast_fp16 = leaky_relu(alpha = var_9, x = input_5_cast_fp16)[name = tensor<string, []>("input_7_cast_fp16")];
             tensor<string, []> input_9_pad_type_0 = const()[name = tensor<string, []>("input_9_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [1]> input_9_strides_0 = const()[name = tensor<string, []>("input_9_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_9_pad_0 = const()[name = tensor<string, []>("input_9_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_9_dilations_0 = const()[name = tensor<string, []>("input_9_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_9_groups_0 = const()[name = tensor<string, []>("input_9_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [?, 60, 5321]> input_9 = conv(bias = sincnet_conv1d_1_bias, dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = sincnet_conv1d_1_weight, x = input_7)[name = tensor<string, []>("input_9")];
+            tensor<fp16, [60, 80, 5]> sincnet_conv1d_1_weight_to_fp16 = const()[name = tensor<string, []>("sincnet_conv1d_1_weight_to_fp16"), val = tensor<fp16, [60, 80, 5]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(40832)))];
+            tensor<fp16, [60]> sincnet_conv1d_1_bias_to_fp16 = const()[name = tensor<string, []>("sincnet_conv1d_1_bias_to_fp16"), val = tensor<fp16, [60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(88896)))];
+            tensor<fp16, [?, 60, 5321]> input_9_cast_fp16 = conv(bias = sincnet_conv1d_1_bias_to_fp16, dilations = input_9_dilations_0, groups = input_9_groups_0, pad = input_9_pad_0, pad_type = input_9_pad_type_0, strides = input_9_strides_0, weight = sincnet_conv1d_1_weight_to_fp16, x = input_7_cast_fp16)[name = tensor<string, []>("input_9_cast_fp16")];
             tensor<int32, [1]> var_135 = const()[name = tensor<string, []>("op_135"), val = tensor<int32, [1]>([3])];
             tensor<int32, [1]> var_136 = const()[name = tensor<string, []>("op_136"), val = tensor<int32, [1]>([3])];
             tensor<string, []> input_11_pad_type_0 = const()[name = tensor<string, []>("input_11_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [2]> input_11_pad_0 = const()[name = tensor<string, []>("input_11_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<bool, []> input_11_ceil_mode_0 = const()[name = tensor<string, []>("input_11_ceil_mode_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [?, 60, 1773]> input_11 = max_pool(ceil_mode = input_11_ceil_mode_0, kernel_sizes = var_135, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = var_136, x = input_9)[name = tensor<string, []>("input_11")];
-            tensor<fp32, [?, 60, 1773]> input_13 = instance_norm(beta = sincnet_norm1d_1_bias, epsilon = var_24, gamma = sincnet_norm1d_1_weight, x = input_11)[name = tensor<string, []>("input_13")];
-            tensor<fp32, [?, 60, 1773]> input_15 = leaky_relu(alpha = var_9, x = input_13)[name = tensor<string, []>("input_15")];
+            tensor<fp16, [?, 60, 1773]> input_11_cast_fp16 = max_pool(ceil_mode = input_11_ceil_mode_0, kernel_sizes = var_135, pad = input_11_pad_0, pad_type = input_11_pad_type_0, strides = var_136, x = input_9_cast_fp16)[name = tensor<string, []>("input_11_cast_fp16")];
+            tensor<fp16, [60]> sincnet_norm1d_1_weight_to_fp16 = const()[name = tensor<string, []>("sincnet_norm1d_1_weight_to_fp16"), val = tensor<fp16, [60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89088)))];
+            tensor<fp16, [60]> sincnet_norm1d_1_bias_to_fp16 = const()[name = tensor<string, []>("sincnet_norm1d_1_bias_to_fp16"), val = tensor<fp16, [60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89280)))];
+            tensor<fp16, [?, 60, 1773]> input_13_cast_fp16 = instance_norm(beta = sincnet_norm1d_1_bias_to_fp16, epsilon = var_24_to_fp16, gamma = sincnet_norm1d_1_weight_to_fp16, x = input_11_cast_fp16)[name = tensor<string, []>("input_13_cast_fp16")];
+            tensor<fp16, [?, 60, 1773]> input_15_cast_fp16 = leaky_relu(alpha = var_9, x = input_13_cast_fp16)[name = tensor<string, []>("input_15_cast_fp16")];
             tensor<string, []> input_17_pad_type_0 = const()[name = tensor<string, []>("input_17_pad_type_0"), val = tensor<string, []>("valid")];
             tensor<int32, [1]> input_17_strides_0 = const()[name = tensor<string, []>("input_17_strides_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, [2]> input_17_pad_0 = const()[name = tensor<string, []>("input_17_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<int32, [1]> input_17_dilations_0 = const()[name = tensor<string, []>("input_17_dilations_0"), val = tensor<int32, [1]>([1])];
             tensor<int32, []> input_17_groups_0 = const()[name = tensor<string, []>("input_17_groups_0"), val = tensor<int32, []>(1)];
-            tensor<fp32, [?, 60, 1769]> input_17 = conv(bias = sincnet_conv1d_2_bias, dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = sincnet_conv1d_2_weight, x = input_15)[name = tensor<string, []>("input_17")];
+            tensor<fp16, [60, 60, 5]> sincnet_conv1d_2_weight_to_fp16 = const()[name = tensor<string, []>("sincnet_conv1d_2_weight_to_fp16"), val = tensor<fp16, [60, 60, 5]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(89472)))];
+            tensor<fp16, [60]> sincnet_conv1d_2_bias_to_fp16 = const()[name = tensor<string, []>("sincnet_conv1d_2_bias_to_fp16"), val = tensor<fp16, [60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125568)))];
+            tensor<fp16, [?, 60, 1769]> input_17_cast_fp16 = conv(bias = sincnet_conv1d_2_bias_to_fp16, dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = sincnet_conv1d_2_weight_to_fp16, x = input_15_cast_fp16)[name = tensor<string, []>("input_17_cast_fp16")];
             tensor<int32, [1]> var_151 = const()[name = tensor<string, []>("op_151"), val = tensor<int32, [1]>([3])];
             tensor<int32, [1]> var_152 = const()[name = tensor<string, []>("op_152"), val = tensor<int32, [1]>([3])];
             tensor<string, []> input_19_pad_type_0 = const()[name = tensor<string, []>("input_19_pad_type_0"), val = tensor<string, []>("custom")];
             tensor<int32, [2]> input_19_pad_0 = const()[name = tensor<string, []>("input_19_pad_0"), val = tensor<int32, [2]>([0, 0])];
             tensor<bool, []> input_19_ceil_mode_0 = const()[name = tensor<string, []>("input_19_ceil_mode_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [?, 60, 589]> input_19 = max_pool(ceil_mode = input_19_ceil_mode_0, kernel_sizes = var_151, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = var_152, x = input_17)[name = tensor<string, []>("input_19")];
-            tensor<fp32, [?, 60, 589]> input_21 = instance_norm(beta = sincnet_norm1d_2_bias, epsilon = var_24, gamma = sincnet_norm1d_2_weight, x = input_19)[name = tensor<string, []>("input_21")];
-            tensor<fp32, [?, 60, 589]> x = leaky_relu(alpha = var_9, x = input_21)[name = tensor<string, []>("x")];
+            tensor<fp16, [?, 60, 589]> input_19_cast_fp16 = max_pool(ceil_mode = input_19_ceil_mode_0, kernel_sizes = var_151, pad = input_19_pad_0, pad_type = input_19_pad_type_0, strides = var_152, x = input_17_cast_fp16)[name = tensor<string, []>("input_19_cast_fp16")];
+            tensor<fp16, [60]> sincnet_norm1d_2_weight_to_fp16 = const()[name = tensor<string, []>("sincnet_norm1d_2_weight_to_fp16"), val = tensor<fp16, [60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125760)))];
+            tensor<fp16, [60]> sincnet_norm1d_2_bias_to_fp16 = const()[name = tensor<string, []>("sincnet_norm1d_2_bias_to_fp16"), val = tensor<fp16, [60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(125952)))];
+            tensor<fp16, [?, 60, 589]> input_21_cast_fp16 = instance_norm(beta = sincnet_norm1d_2_bias_to_fp16, epsilon = var_24_to_fp16, gamma = sincnet_norm1d_2_weight_to_fp16, x = input_19_cast_fp16)[name = tensor<string, []>("input_21_cast_fp16")];
+            tensor<fp16, [?, 60, 589]> x_cast_fp16 = leaky_relu(alpha = var_9, x = input_21_cast_fp16)[name = tensor<string, []>("x_cast_fp16")];
             tensor<int32, [3]> var_163 = const()[name = tensor<string, []>("op_163"), val = tensor<int32, [3]>([0, 2, 1])];
             tensor<int32, []> var_172 = const()[name = tensor<string, []>("op_172"), val = tensor<int32, []>(128)];
             tensor<int32, []> var_173 = const()[name = tensor<string, []>("op_173"), val = tensor<int32, []>(8)];
-            tensor<fp32, [?, 589, 60]> input_23 = transpose(perm = var_163, x = x)[name = tensor<string, []>("transpose_6")];
-            tensor<int32, [3]> var_207_shape = shape(x = input_23)[name = tensor<string, []>("op_207_shape")];
+            tensor<fp16, [?, 589, 60]> input_23_cast_fp16 = transpose(perm = var_163, x = x_cast_fp16)[name = tensor<string, []>("transpose_6")];
+            tensor<int32, [3]> var_207_shape_cast_fp16 = shape(x = input_23_cast_fp16)[name = tensor<string, []>("op_207_shape_cast_fp16")];
+            tensor<int32, []> gather_0_axis_0 = const()[name = tensor<string, []>("gather_0_axis_0"), val = tensor<int32, []>(0)];
             tensor<int32, []> gather_0_batch_dims_0 = const()[name = tensor<string, []>("gather_0_batch_dims_0"), val = tensor<int32, []>(0)];
             tensor<bool, []> gather_0_validate_indices_0 = const()[name = tensor<string, []>("gather_0_validate_indices_0"), val = tensor<bool, []>(false)];
-            tensor<int32, []> select_0 = const()[name = tensor<string, []>("select_0"), val = tensor<int32, []>(0)];
-            tensor<int32, []> gather_0_axis_1 = const()[name = tensor<string, []>("gather_0_axis_1"), val = tensor<int32, []>(0)];
-            tensor<int32, []> gather_0 = gather(axis = gather_0_axis_1, batch_dims = gather_0_batch_dims_0, indices = select_0, validate_indices = gather_0_validate_indices_0, x = var_207_shape)[name = tensor<string, []>("gather_0")];
+            tensor<string, []> var_207_shape_cast_fp16_to_int16_dtype_0 = const()[name = tensor<string, []>("op_207_shape_cast_fp16_to_int16_dtype_0"), val = tensor<string, []>("int16")];
+            tensor<uint16, []> gather_0_indices_0_to_uint16 = const()[name = tensor<string, []>("gather_0_indices_0_to_uint16"), val = tensor<uint16, []>(0)];
+            tensor<int16, [3]> var_207_shape_cast_fp16_to_int16 = cast(dtype = var_207_shape_cast_fp16_to_int16_dtype_0, x = var_207_shape_cast_fp16)[name = tensor<string, []>("cast_18")];
+            tensor<int16, []> gather_0_cast_uint16 = gather(axis = gather_0_axis_0, batch_dims = gather_0_batch_dims_0, indices = gather_0_indices_0_to_uint16, validate_indices = gather_0_validate_indices_0, x = var_207_shape_cast_fp16_to_int16)[name = tensor<string, []>("gather_0_cast_uint16")];
+            tensor<string, []> gather_0_cast_uint16_to_int32_dtype_0 = const()[name = tensor<string, []>("gather_0_cast_uint16_to_int32_dtype_0"), val = tensor<string, []>("int32")];
             tensor<int32, []> concat_0_axis_0 = const()[name = tensor<string, []>("concat_0_axis_0"), val = tensor<int32, []>(0)];
             tensor<bool, []> concat_0_interleave_0 = const()[name = tensor<string, []>("concat_0_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<int32, [3]> concat_0 = concat(axis = concat_0_axis_0, interleave = concat_0_interleave_0, values = (var_173, gather_0, var_172))[name = tensor<string, []>("concat_0")];
-            tensor<fp32, []> hx_1_value_0 = const()[name = tensor<string, []>("hx_1_value_0"), val = tensor<fp32, []>(0x0p+0)];
-            tensor<fp32, [8, ?, 128]> hx_1 = fill(shape = concat_0, value = hx_1_value_0)[name = tensor<string, []>("hx_1")];
+            tensor<int32, []> gather_0_cast_uint16_to_int32 = cast(dtype = gather_0_cast_uint16_to_int32_dtype_0, x = gather_0_cast_uint16)[name = tensor<string, []>("cast_17")];
+            tensor<int32, [3]> concat_0 = concat(axis = concat_0_axis_0, interleave = concat_0_interleave_0, values = (var_173, gather_0_cast_uint16_to_int32, var_172))[name = tensor<string, []>("concat_0")];
+            tensor<fp16, []> hx_1_value_0_to_fp16 = const()[name = tensor<string, []>("hx_1_value_0_to_fp16"), val = tensor<fp16, []>(0x0p+0)];
+            tensor<fp16, [8, ?, 128]> hx_1_cast_fp16 = fill(shape = concat_0, value = hx_1_value_0_to_fp16)[name = tensor<string, []>("hx_1_cast_fp16")];
             tensor<int32, [3]> input_23_batch_first_transpose_perm_0 = const()[name = tensor<string, []>("input_23_batch_first_transpose_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
             tensor<int32, []> split_0_num_splits_0 = const()[name = tensor<string, []>("split_0_num_splits_0"), val = tensor<int32, []>(4)];
             tensor<int32, []> split_0_axis_0 = const()[name = tensor<string, []>("split_0_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [2, ?, 128]> split_0_0, tensor<fp32, [2, ?, 128]> split_0_1, tensor<fp32, [2, ?, 128]> split_0_2, tensor<fp32, [2, ?, 128]> split_0_3 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = hx_1)[name = tensor<string, []>("split_0")];
+            tensor<fp16, [2, ?, 128]> split_0_cast_fp16_0, tensor<fp16, [2, ?, 128]> split_0_cast_fp16_1, tensor<fp16, [2, ?, 128]> split_0_cast_fp16_2, tensor<fp16, [2, ?, 128]> split_0_cast_fp16_3 = split(axis = split_0_axis_0, num_splits = split_0_num_splits_0, x = hx_1_cast_fp16)[name = tensor<string, []>("split_0_cast_fp16")];
             tensor<int32, []> split_1_num_splits_0 = const()[name = tensor<string, []>("split_1_num_splits_0"), val = tensor<int32, []>(4)];
             tensor<int32, []> split_1_axis_0 = const()[name = tensor<string, []>("split_1_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [2, ?, 128]> split_1_0, tensor<fp32, [2, ?, 128]> split_1_1, tensor<fp32, [2, ?, 128]> split_1_2, tensor<fp32, [2, ?, 128]> split_1_3 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = hx_1)[name = tensor<string, []>("split_1")];
-            tensor<fp32, [512]> add_0 = const()[name = tensor<string, []>("add_0"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(452800)))];
-            tensor<fp32, [512]> add_1 = const()[name = tensor<string, []>("add_1"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(454912)))];
-            tensor<fp32, [512, 60]> concat_6 = const()[name = tensor<string, []>("concat_6"), val = tensor<fp32, [512, 60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(457024)))];
-            tensor<fp32, [512, 128]> concat_7 = const()[name = tensor<string, []>("concat_7"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(579968)))];
-            tensor<fp32, [512, 60]> concat_8 = const()[name = tensor<string, []>("concat_8"), val = tensor<fp32, [512, 60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(842176)))];
-            tensor<fp32, [512, 128]> concat_9 = const()[name = tensor<string, []>("concat_9"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(965120)))];
+            tensor<fp16, [2, ?, 128]> split_1_cast_fp16_0, tensor<fp16, [2, ?, 128]> split_1_cast_fp16_1, tensor<fp16, [2, ?, 128]> split_1_cast_fp16_2, tensor<fp16, [2, ?, 128]> split_1_cast_fp16_3 = split(axis = split_1_axis_0, num_splits = split_1_num_splits_0, x = hx_1_cast_fp16)[name = tensor<string, []>("split_1_cast_fp16")];
             tensor<int32, [2]> split_10_split_sizes_0 = const()[name = tensor<string, []>("split_10_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> split_10_axis_0 = const()[name = tensor<string, []>("split_10_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [1, ?, 128]> split_10_0, tensor<fp32, [1, ?, 128]> split_10_1 = split(axis = split_10_axis_0, split_sizes = split_10_split_sizes_0, x = split_0_0)[name = tensor<string, []>("split_10")];
+            tensor<fp16, [1, ?, 128]> split_10_cast_fp16_0, tensor<fp16, [1, ?, 128]> split_10_cast_fp16_1 = split(axis = split_10_axis_0, split_sizes = split_10_split_sizes_0, x = split_0_cast_fp16_0)[name = tensor<string, []>("split_10_cast_fp16")];
             tensor<int32, []> concat_10_axis_0 = const()[name = tensor<string, []>("concat_10_axis_0"), val = tensor<int32, []>(2)];
             tensor<bool, []> concat_10_interleave_0 = const()[name = tensor<string, []>("concat_10_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, ?, 256]> concat_10 = concat(axis = concat_10_axis_0, interleave = concat_10_interleave_0, values = (split_10_0, split_10_1))[name = tensor<string, []>("concat_10")];
+            tensor<fp16, [1, ?, 256]> concat_10_cast_fp16 = concat(axis = concat_10_axis_0, interleave = concat_10_interleave_0, values = (split_10_cast_fp16_0, split_10_cast_fp16_1))[name = tensor<string, []>("concat_10_cast_fp16")];
             tensor<int32, [1]> input_25_lstm_layer_0_lstm_h0_reshaped_axes_0 = const()[name = tensor<string, []>("input_25_lstm_layer_0_lstm_h0_reshaped_axes_0"), val = tensor<int32, [1]>([0])];
-            tensor<fp32, [?, 256]> input_25_lstm_layer_0_lstm_h0_reshaped = squeeze(axes = input_25_lstm_layer_0_lstm_h0_reshaped_axes_0, x = concat_10)[name = tensor<string, []>("input_25_lstm_layer_0_lstm_h0_reshaped")];
+            tensor<fp16, [?, 256]> input_25_lstm_layer_0_lstm_h0_reshaped_cast_fp16 = squeeze(axes = input_25_lstm_layer_0_lstm_h0_reshaped_axes_0, x = concat_10_cast_fp16)[name = tensor<string, []>("input_25_lstm_layer_0_lstm_h0_reshaped_cast_fp16")];
             tensor<int32, [2]> split_11_split_sizes_0 = const()[name = tensor<string, []>("split_11_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> split_11_axis_0 = const()[name = tensor<string, []>("split_11_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [1, ?, 128]> split_11_0, tensor<fp32, [1, ?, 128]> split_11_1 = split(axis = split_11_axis_0, split_sizes = split_11_split_sizes_0, x = split_1_0)[name = tensor<string, []>("split_11")];
+            tensor<fp16, [1, ?, 128]> split_11_cast_fp16_0, tensor<fp16, [1, ?, 128]> split_11_cast_fp16_1 = split(axis = split_11_axis_0, split_sizes = split_11_split_sizes_0, x = split_1_cast_fp16_0)[name = tensor<string, []>("split_11_cast_fp16")];
             tensor<int32, []> concat_11_axis_0 = const()[name = tensor<string, []>("concat_11_axis_0"), val = tensor<int32, []>(2)];
             tensor<bool, []> concat_11_interleave_0 = const()[name = tensor<string, []>("concat_11_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, ?, 256]> concat_11 = concat(axis = concat_11_axis_0, interleave = concat_11_interleave_0, values = (split_11_0, split_11_1))[name = tensor<string, []>("concat_11")];
+            tensor<fp16, [1, ?, 256]> concat_11_cast_fp16 = concat(axis = concat_11_axis_0, interleave = concat_11_interleave_0, values = (split_11_cast_fp16_0, split_11_cast_fp16_1))[name = tensor<string, []>("concat_11_cast_fp16")];
             tensor<int32, [1]> input_25_lstm_layer_0_lstm_c0_reshaped_axes_0 = const()[name = tensor<string, []>("input_25_lstm_layer_0_lstm_c0_reshaped_axes_0"), val = tensor<int32, [1]>([0])];
-            tensor<fp32, [?, 256]> input_25_lstm_layer_0_lstm_c0_reshaped = squeeze(axes = input_25_lstm_layer_0_lstm_c0_reshaped_axes_0, x = concat_11)[name = tensor<string, []>("input_25_lstm_layer_0_lstm_c0_reshaped")];
+            tensor<fp16, [?, 256]> input_25_lstm_layer_0_lstm_c0_reshaped_cast_fp16 = squeeze(axes = input_25_lstm_layer_0_lstm_c0_reshaped_axes_0, x = concat_11_cast_fp16)[name = tensor<string, []>("input_25_lstm_layer_0_lstm_c0_reshaped_cast_fp16")];
             tensor<string, []> input_25_lstm_layer_0_direction_0 = const()[name = tensor<string, []>("input_25_lstm_layer_0_direction_0"), val = tensor<string, []>("bidirectional")];
             tensor<bool, []> input_25_lstm_layer_0_output_sequence_0 = const()[name = tensor<string, []>("input_25_lstm_layer_0_output_sequence_0"), val = tensor<bool, []>(true)];
             tensor<string, []> input_25_lstm_layer_0_recurrent_activation_0 = const()[name = tensor<string, []>("input_25_lstm_layer_0_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
             tensor<string, []> input_25_lstm_layer_0_cell_activation_0 = const()[name = tensor<string, []>("input_25_lstm_layer_0_cell_activation_0"), val = tensor<string, []>("tanh")];
             tensor<string, []> input_25_lstm_layer_0_activation_0 = const()[name = tensor<string, []>("input_25_lstm_layer_0_activation_0"), val = tensor<string, []>("tanh")];
-            tensor<fp32, [589, ?, 60]> input_23_batch_first_transpose = transpose(perm = input_23_batch_first_transpose_perm_0, x = input_23)[name = tensor<string, []>("transpose_5")];
-            tensor<fp32, [589, ?, 256]> input_25_lstm_layer_0_0, tensor<fp32, [?, 256]> input_25_lstm_layer_0_1, tensor<fp32, [?, 256]> input_25_lstm_layer_0_2 = lstm(activation = input_25_lstm_layer_0_activation_0, bias = add_0, bias_back = add_1, cell_activation = input_25_lstm_layer_0_cell_activation_0, direction = input_25_lstm_layer_0_direction_0, initial_c = input_25_lstm_layer_0_lstm_c0_reshaped, initial_h = input_25_lstm_layer_0_lstm_h0_reshaped, output_sequence = input_25_lstm_layer_0_output_sequence_0, recurrent_activation = input_25_lstm_layer_0_recurrent_activation_0, weight_hh = concat_7, weight_hh_back = concat_9, weight_ih = concat_6, weight_ih_back = concat_8, x = input_23_batch_first_transpose)[name = tensor<string, []>("input_25_lstm_layer_0")];
-            tensor<fp32, [512]> add_2 = const()[name = tensor<string, []>("add_2"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1227328)))];
-            tensor<fp32, [512]> add_3 = const()[name = tensor<string, []>("add_3"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1229440)))];
-            tensor<fp32, [512, 256]> concat_16 = const()[name = tensor<string, []>("concat_16"), val = tensor<fp32, [512, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1231552)))];
-            tensor<fp32, [512, 128]> concat_17 = const()[name = tensor<string, []>("concat_17"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1755904)))];
-            tensor<fp32, [512, 256]> concat_18 = const()[name = tensor<string, []>("concat_18"), val = tensor<fp32, [512, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2018112)))];
-            tensor<fp32, [512, 128]> concat_19 = const()[name = tensor<string, []>("concat_19"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2542464)))];
+            tensor<fp16, [512, 60]> concat_6_to_fp16 = const()[name = tensor<string, []>("concat_6_to_fp16"), val = tensor<fp16, [512, 60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(126144)))];
+            tensor<fp16, [512, 128]> concat_7_to_fp16 = const()[name = tensor<string, []>("concat_7_to_fp16"), val = tensor<fp16, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(187648)))];
+            tensor<fp16, [512]> add_0_to_fp16 = const()[name = tensor<string, []>("add_0_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(318784)))];
+            tensor<fp16, [512, 60]> concat_8_to_fp16 = const()[name = tensor<string, []>("concat_8_to_fp16"), val = tensor<fp16, [512, 60]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(319872)))];
+            tensor<fp16, [512, 128]> concat_9_to_fp16 = const()[name = tensor<string, []>("concat_9_to_fp16"), val = tensor<fp16, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(381376)))];
+            tensor<fp16, [512]> add_1_to_fp16 = const()[name = tensor<string, []>("add_1_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(512512)))];
+            tensor<fp16, [589, ?, 60]> input_23_batch_first_transpose_cast_fp16 = transpose(perm = input_23_batch_first_transpose_perm_0, x = input_23_cast_fp16)[name = tensor<string, []>("transpose_5")];
+            tensor<fp16, [589, ?, 256]> input_25_lstm_layer_0_cast_fp16_0, tensor<fp16, [?, 256]> input_25_lstm_layer_0_cast_fp16_1, tensor<fp16, [?, 256]> input_25_lstm_layer_0_cast_fp16_2 = lstm(activation = input_25_lstm_layer_0_activation_0, bias = add_0_to_fp16, bias_back = add_1_to_fp16, cell_activation = input_25_lstm_layer_0_cell_activation_0, direction = input_25_lstm_layer_0_direction_0, initial_c = input_25_lstm_layer_0_lstm_c0_reshaped_cast_fp16, initial_h = input_25_lstm_layer_0_lstm_h0_reshaped_cast_fp16, output_sequence = input_25_lstm_layer_0_output_sequence_0, recurrent_activation = input_25_lstm_layer_0_recurrent_activation_0, weight_hh = concat_7_to_fp16, weight_hh_back = concat_9_to_fp16, weight_ih = concat_6_to_fp16, weight_ih_back = concat_8_to_fp16, x = input_23_batch_first_transpose_cast_fp16)[name = tensor<string, []>("input_25_lstm_layer_0_cast_fp16")];
             tensor<int32, [2]> split_20_split_sizes_0 = const()[name = tensor<string, []>("split_20_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> split_20_axis_0 = const()[name = tensor<string, []>("split_20_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [1, ?, 128]> split_20_0, tensor<fp32, [1, ?, 128]> split_20_1 = split(axis = split_20_axis_0, split_sizes = split_20_split_sizes_0, x = split_0_1)[name = tensor<string, []>("split_20")];
+            tensor<fp16, [1, ?, 128]> split_20_cast_fp16_0, tensor<fp16, [1, ?, 128]> split_20_cast_fp16_1 = split(axis = split_20_axis_0, split_sizes = split_20_split_sizes_0, x = split_0_cast_fp16_1)[name = tensor<string, []>("split_20_cast_fp16")];
             tensor<int32, []> concat_20_axis_0 = const()[name = tensor<string, []>("concat_20_axis_0"), val = tensor<int32, []>(2)];
             tensor<bool, []> concat_20_interleave_0 = const()[name = tensor<string, []>("concat_20_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, ?, 256]> concat_20 = concat(axis = concat_20_axis_0, interleave = concat_20_interleave_0, values = (split_20_0, split_20_1))[name = tensor<string, []>("concat_20")];
+            tensor<fp16, [1, ?, 256]> concat_20_cast_fp16 = concat(axis = concat_20_axis_0, interleave = concat_20_interleave_0, values = (split_20_cast_fp16_0, split_20_cast_fp16_1))[name = tensor<string, []>("concat_20_cast_fp16")];
             tensor<int32, [1]> input_25_lstm_layer_1_lstm_h0_reshaped_axes_0 = const()[name = tensor<string, []>("input_25_lstm_layer_1_lstm_h0_reshaped_axes_0"), val = tensor<int32, [1]>([0])];
-            tensor<fp32, [?, 256]> input_25_lstm_layer_1_lstm_h0_reshaped = squeeze(axes = input_25_lstm_layer_1_lstm_h0_reshaped_axes_0, x = concat_20)[name = tensor<string, []>("input_25_lstm_layer_1_lstm_h0_reshaped")];
+            tensor<fp16, [?, 256]> input_25_lstm_layer_1_lstm_h0_reshaped_cast_fp16 = squeeze(axes = input_25_lstm_layer_1_lstm_h0_reshaped_axes_0, x = concat_20_cast_fp16)[name = tensor<string, []>("input_25_lstm_layer_1_lstm_h0_reshaped_cast_fp16")];
             tensor<int32, [2]> split_21_split_sizes_0 = const()[name = tensor<string, []>("split_21_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> split_21_axis_0 = const()[name = tensor<string, []>("split_21_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [1, ?, 128]> split_21_0, tensor<fp32, [1, ?, 128]> split_21_1 = split(axis = split_21_axis_0, split_sizes = split_21_split_sizes_0, x = split_1_1)[name = tensor<string, []>("split_21")];
+            tensor<fp16, [1, ?, 128]> split_21_cast_fp16_0, tensor<fp16, [1, ?, 128]> split_21_cast_fp16_1 = split(axis = split_21_axis_0, split_sizes = split_21_split_sizes_0, x = split_1_cast_fp16_1)[name = tensor<string, []>("split_21_cast_fp16")];
             tensor<int32, []> concat_21_axis_0 = const()[name = tensor<string, []>("concat_21_axis_0"), val = tensor<int32, []>(2)];
             tensor<bool, []> concat_21_interleave_0 = const()[name = tensor<string, []>("concat_21_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, ?, 256]> concat_21 = concat(axis = concat_21_axis_0, interleave = concat_21_interleave_0, values = (split_21_0, split_21_1))[name = tensor<string, []>("concat_21")];
+            tensor<fp16, [1, ?, 256]> concat_21_cast_fp16 = concat(axis = concat_21_axis_0, interleave = concat_21_interleave_0, values = (split_21_cast_fp16_0, split_21_cast_fp16_1))[name = tensor<string, []>("concat_21_cast_fp16")];
             tensor<int32, [1]> input_25_lstm_layer_1_lstm_c0_reshaped_axes_0 = const()[name = tensor<string, []>("input_25_lstm_layer_1_lstm_c0_reshaped_axes_0"), val = tensor<int32, [1]>([0])];
-            tensor<fp32, [?, 256]> input_25_lstm_layer_1_lstm_c0_reshaped = squeeze(axes = input_25_lstm_layer_1_lstm_c0_reshaped_axes_0, x = concat_21)[name = tensor<string, []>("input_25_lstm_layer_1_lstm_c0_reshaped")];
+            tensor<fp16, [?, 256]> input_25_lstm_layer_1_lstm_c0_reshaped_cast_fp16 = squeeze(axes = input_25_lstm_layer_1_lstm_c0_reshaped_axes_0, x = concat_21_cast_fp16)[name = tensor<string, []>("input_25_lstm_layer_1_lstm_c0_reshaped_cast_fp16")];
             tensor<string, []> input_25_lstm_layer_1_direction_0 = const()[name = tensor<string, []>("input_25_lstm_layer_1_direction_0"), val = tensor<string, []>("bidirectional")];
             tensor<bool, []> input_25_lstm_layer_1_output_sequence_0 = const()[name = tensor<string, []>("input_25_lstm_layer_1_output_sequence_0"), val = tensor<bool, []>(true)];
             tensor<string, []> input_25_lstm_layer_1_recurrent_activation_0 = const()[name = tensor<string, []>("input_25_lstm_layer_1_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
             tensor<string, []> input_25_lstm_layer_1_cell_activation_0 = const()[name = tensor<string, []>("input_25_lstm_layer_1_cell_activation_0"), val = tensor<string, []>("tanh")];
             tensor<string, []> input_25_lstm_layer_1_activation_0 = const()[name = tensor<string, []>("input_25_lstm_layer_1_activation_0"), val = tensor<string, []>("tanh")];
-            tensor<fp32, [589, ?, 256]> input_25_lstm_layer_1_0, tensor<fp32, [?, 256]> input_25_lstm_layer_1_1, tensor<fp32, [?, 256]> input_25_lstm_layer_1_2 = lstm(activation = input_25_lstm_layer_1_activation_0, bias = add_2, bias_back = add_3, cell_activation = input_25_lstm_layer_1_cell_activation_0, direction = input_25_lstm_layer_1_direction_0, initial_c = input_25_lstm_layer_1_lstm_c0_reshaped, initial_h = input_25_lstm_layer_1_lstm_h0_reshaped, output_sequence = input_25_lstm_layer_1_output_sequence_0, recurrent_activation = input_25_lstm_layer_1_recurrent_activation_0, weight_hh = concat_17, weight_hh_back = concat_19, weight_ih = concat_16, weight_ih_back = concat_18, x = input_25_lstm_layer_0_0)[name = tensor<string, []>("input_25_lstm_layer_1")];
-            tensor<fp32, [512]> add_4 = const()[name = tensor<string, []>("add_4"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2804672)))];
-            tensor<fp32, [512]> add_5 = const()[name = tensor<string, []>("add_5"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2806784)))];
-            tensor<fp32, [512, 256]> concat_26 = const()[name = tensor<string, []>("concat_26"), val = tensor<fp32, [512, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2808896)))];
-            tensor<fp32, [512, 128]> concat_27 = const()[name = tensor<string, []>("concat_27"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3333248)))];
-            tensor<fp32, [512, 256]> concat_28 = const()[name = tensor<string, []>("concat_28"), val = tensor<fp32, [512, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(3595456)))];
-            tensor<fp32, [512, 128]> concat_29 = const()[name = tensor<string, []>("concat_29"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4119808)))];
+            tensor<fp16, [512, 256]> concat_16_to_fp16 = const()[name = tensor<string, []>("concat_16_to_fp16"), val = tensor<fp16, [512, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(513600)))];
+            tensor<fp16, [512, 128]> concat_17_to_fp16 = const()[name = tensor<string, []>("concat_17_to_fp16"), val = tensor<fp16, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(775808)))];
+            tensor<fp16, [512]> add_2_to_fp16 = const()[name = tensor<string, []>("add_2_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(906944)))];
+            tensor<fp16, [512, 256]> concat_18_to_fp16 = const()[name = tensor<string, []>("concat_18_to_fp16"), val = tensor<fp16, [512, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(908032)))];
+            tensor<fp16, [512, 128]> concat_19_to_fp16 = const()[name = tensor<string, []>("concat_19_to_fp16"), val = tensor<fp16, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1170240)))];
+            tensor<fp16, [512]> add_3_to_fp16 = const()[name = tensor<string, []>("add_3_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1301376)))];
+            tensor<fp16, [589, ?, 256]> input_25_lstm_layer_1_cast_fp16_0, tensor<fp16, [?, 256]> input_25_lstm_layer_1_cast_fp16_1, tensor<fp16, [?, 256]> input_25_lstm_layer_1_cast_fp16_2 = lstm(activation = input_25_lstm_layer_1_activation_0, bias = add_2_to_fp16, bias_back = add_3_to_fp16, cell_activation = input_25_lstm_layer_1_cell_activation_0, direction = input_25_lstm_layer_1_direction_0, initial_c = input_25_lstm_layer_1_lstm_c0_reshaped_cast_fp16, initial_h = input_25_lstm_layer_1_lstm_h0_reshaped_cast_fp16, output_sequence = input_25_lstm_layer_1_output_sequence_0, recurrent_activation = input_25_lstm_layer_1_recurrent_activation_0, weight_hh = concat_17_to_fp16, weight_hh_back = concat_19_to_fp16, weight_ih = concat_16_to_fp16, weight_ih_back = concat_18_to_fp16, x = input_25_lstm_layer_0_cast_fp16_0)[name = tensor<string, []>("input_25_lstm_layer_1_cast_fp16")];
             tensor<int32, [2]> split_30_split_sizes_0 = const()[name = tensor<string, []>("split_30_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> split_30_axis_0 = const()[name = tensor<string, []>("split_30_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [1, ?, 128]> split_30_0, tensor<fp32, [1, ?, 128]> split_30_1 = split(axis = split_30_axis_0, split_sizes = split_30_split_sizes_0, x = split_0_2)[name = tensor<string, []>("split_30")];
+            tensor<fp16, [1, ?, 128]> split_30_cast_fp16_0, tensor<fp16, [1, ?, 128]> split_30_cast_fp16_1 = split(axis = split_30_axis_0, split_sizes = split_30_split_sizes_0, x = split_0_cast_fp16_2)[name = tensor<string, []>("split_30_cast_fp16")];
             tensor<int32, []> concat_30_axis_0 = const()[name = tensor<string, []>("concat_30_axis_0"), val = tensor<int32, []>(2)];
             tensor<bool, []> concat_30_interleave_0 = const()[name = tensor<string, []>("concat_30_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, ?, 256]> concat_30 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (split_30_0, split_30_1))[name = tensor<string, []>("concat_30")];
+            tensor<fp16, [1, ?, 256]> concat_30_cast_fp16 = concat(axis = concat_30_axis_0, interleave = concat_30_interleave_0, values = (split_30_cast_fp16_0, split_30_cast_fp16_1))[name = tensor<string, []>("concat_30_cast_fp16")];
             tensor<int32, [1]> input_25_lstm_layer_2_lstm_h0_reshaped_axes_0 = const()[name = tensor<string, []>("input_25_lstm_layer_2_lstm_h0_reshaped_axes_0"), val = tensor<int32, [1]>([0])];
-            tensor<fp32, [?, 256]> input_25_lstm_layer_2_lstm_h0_reshaped = squeeze(axes = input_25_lstm_layer_2_lstm_h0_reshaped_axes_0, x = concat_30)[name = tensor<string, []>("input_25_lstm_layer_2_lstm_h0_reshaped")];
+            tensor<fp16, [?, 256]> input_25_lstm_layer_2_lstm_h0_reshaped_cast_fp16 = squeeze(axes = input_25_lstm_layer_2_lstm_h0_reshaped_axes_0, x = concat_30_cast_fp16)[name = tensor<string, []>("input_25_lstm_layer_2_lstm_h0_reshaped_cast_fp16")];
             tensor<int32, [2]> split_31_split_sizes_0 = const()[name = tensor<string, []>("split_31_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> split_31_axis_0 = const()[name = tensor<string, []>("split_31_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [1, ?, 128]> split_31_0, tensor<fp32, [1, ?, 128]> split_31_1 = split(axis = split_31_axis_0, split_sizes = split_31_split_sizes_0, x = split_1_2)[name = tensor<string, []>("split_31")];
+            tensor<fp16, [1, ?, 128]> split_31_cast_fp16_0, tensor<fp16, [1, ?, 128]> split_31_cast_fp16_1 = split(axis = split_31_axis_0, split_sizes = split_31_split_sizes_0, x = split_1_cast_fp16_2)[name = tensor<string, []>("split_31_cast_fp16")];
             tensor<int32, []> concat_31_axis_0 = const()[name = tensor<string, []>("concat_31_axis_0"), val = tensor<int32, []>(2)];
             tensor<bool, []> concat_31_interleave_0 = const()[name = tensor<string, []>("concat_31_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, ?, 256]> concat_31 = concat(axis = concat_31_axis_0, interleave = concat_31_interleave_0, values = (split_31_0, split_31_1))[name = tensor<string, []>("concat_31")];
+            tensor<fp16, [1, ?, 256]> concat_31_cast_fp16 = concat(axis = concat_31_axis_0, interleave = concat_31_interleave_0, values = (split_31_cast_fp16_0, split_31_cast_fp16_1))[name = tensor<string, []>("concat_31_cast_fp16")];
             tensor<int32, [1]> input_25_lstm_layer_2_lstm_c0_reshaped_axes_0 = const()[name = tensor<string, []>("input_25_lstm_layer_2_lstm_c0_reshaped_axes_0"), val = tensor<int32, [1]>([0])];
-            tensor<fp32, [?, 256]> input_25_lstm_layer_2_lstm_c0_reshaped = squeeze(axes = input_25_lstm_layer_2_lstm_c0_reshaped_axes_0, x = concat_31)[name = tensor<string, []>("input_25_lstm_layer_2_lstm_c0_reshaped")];
+            tensor<fp16, [?, 256]> input_25_lstm_layer_2_lstm_c0_reshaped_cast_fp16 = squeeze(axes = input_25_lstm_layer_2_lstm_c0_reshaped_axes_0, x = concat_31_cast_fp16)[name = tensor<string, []>("input_25_lstm_layer_2_lstm_c0_reshaped_cast_fp16")];
             tensor<string, []> input_25_lstm_layer_2_direction_0 = const()[name = tensor<string, []>("input_25_lstm_layer_2_direction_0"), val = tensor<string, []>("bidirectional")];
             tensor<bool, []> input_25_lstm_layer_2_output_sequence_0 = const()[name = tensor<string, []>("input_25_lstm_layer_2_output_sequence_0"), val = tensor<bool, []>(true)];
             tensor<string, []> input_25_lstm_layer_2_recurrent_activation_0 = const()[name = tensor<string, []>("input_25_lstm_layer_2_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
             tensor<string, []> input_25_lstm_layer_2_cell_activation_0 = const()[name = tensor<string, []>("input_25_lstm_layer_2_cell_activation_0"), val = tensor<string, []>("tanh")];
             tensor<string, []> input_25_lstm_layer_2_activation_0 = const()[name = tensor<string, []>("input_25_lstm_layer_2_activation_0"), val = tensor<string, []>("tanh")];
-            tensor<fp32, [589, ?, 256]> input_25_lstm_layer_2_0, tensor<fp32, [?, 256]> input_25_lstm_layer_2_1, tensor<fp32, [?, 256]> input_25_lstm_layer_2_2 = lstm(activation = input_25_lstm_layer_2_activation_0, bias = add_4, bias_back = add_5, cell_activation = input_25_lstm_layer_2_cell_activation_0, direction = input_25_lstm_layer_2_direction_0, initial_c = input_25_lstm_layer_2_lstm_c0_reshaped, initial_h = input_25_lstm_layer_2_lstm_h0_reshaped, output_sequence = input_25_lstm_layer_2_output_sequence_0, recurrent_activation = input_25_lstm_layer_2_recurrent_activation_0, weight_hh = concat_27, weight_hh_back = concat_29, weight_ih = concat_26, weight_ih_back = concat_28, x = input_25_lstm_layer_1_0)[name = tensor<string, []>("input_25_lstm_layer_2")];
-            tensor<fp32, [512]> add_6 = const()[name = tensor<string, []>("add_6"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4382016)))];
-            tensor<fp32, [512]> add_7 = const()[name = tensor<string, []>("add_7"), val = tensor<fp32, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4384128)))];
-            tensor<fp32, [512, 256]> concat_36 = const()[name = tensor<string, []>("concat_36"), val = tensor<fp32, [512, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4386240)))];
-            tensor<fp32, [512, 128]> concat_37 = const()[name = tensor<string, []>("concat_37"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(4910592)))];
-            tensor<fp32, [512, 256]> concat_38 = const()[name = tensor<string, []>("concat_38"), val = tensor<fp32, [512, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5172800)))];
-            tensor<fp32, [512, 128]> concat_39 = const()[name = tensor<string, []>("concat_39"), val = tensor<fp32, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(5697152)))];
+            tensor<fp16, [512, 256]> concat_26_to_fp16 = const()[name = tensor<string, []>("concat_26_to_fp16"), val = tensor<fp16, [512, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1302464)))];
+            tensor<fp16, [512, 128]> concat_27_to_fp16 = const()[name = tensor<string, []>("concat_27_to_fp16"), val = tensor<fp16, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1564672)))];
+            tensor<fp16, [512]> add_4_to_fp16 = const()[name = tensor<string, []>("add_4_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1695808)))];
+            tensor<fp16, [512, 256]> concat_28_to_fp16 = const()[name = tensor<string, []>("concat_28_to_fp16"), val = tensor<fp16, [512, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1696896)))];
+            tensor<fp16, [512, 128]> concat_29_to_fp16 = const()[name = tensor<string, []>("concat_29_to_fp16"), val = tensor<fp16, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(1959104)))];
+            tensor<fp16, [512]> add_5_to_fp16 = const()[name = tensor<string, []>("add_5_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2090240)))];
+            tensor<fp16, [589, ?, 256]> input_25_lstm_layer_2_cast_fp16_0, tensor<fp16, [?, 256]> input_25_lstm_layer_2_cast_fp16_1, tensor<fp16, [?, 256]> input_25_lstm_layer_2_cast_fp16_2 = lstm(activation = input_25_lstm_layer_2_activation_0, bias = add_4_to_fp16, bias_back = add_5_to_fp16, cell_activation = input_25_lstm_layer_2_cell_activation_0, direction = input_25_lstm_layer_2_direction_0, initial_c = input_25_lstm_layer_2_lstm_c0_reshaped_cast_fp16, initial_h = input_25_lstm_layer_2_lstm_h0_reshaped_cast_fp16, output_sequence = input_25_lstm_layer_2_output_sequence_0, recurrent_activation = input_25_lstm_layer_2_recurrent_activation_0, weight_hh = concat_27_to_fp16, weight_hh_back = concat_29_to_fp16, weight_ih = concat_26_to_fp16, weight_ih_back = concat_28_to_fp16, x = input_25_lstm_layer_1_cast_fp16_0)[name = tensor<string, []>("input_25_lstm_layer_2_cast_fp16")];
             tensor<int32, [2]> split_40_split_sizes_0 = const()[name = tensor<string, []>("split_40_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> split_40_axis_0 = const()[name = tensor<string, []>("split_40_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [1, ?, 128]> split_40_0, tensor<fp32, [1, ?, 128]> split_40_1 = split(axis = split_40_axis_0, split_sizes = split_40_split_sizes_0, x = split_0_3)[name = tensor<string, []>("split_40")];
+            tensor<fp16, [1, ?, 128]> split_40_cast_fp16_0, tensor<fp16, [1, ?, 128]> split_40_cast_fp16_1 = split(axis = split_40_axis_0, split_sizes = split_40_split_sizes_0, x = split_0_cast_fp16_3)[name = tensor<string, []>("split_40_cast_fp16")];
             tensor<int32, []> concat_40_axis_0 = const()[name = tensor<string, []>("concat_40_axis_0"), val = tensor<int32, []>(2)];
             tensor<bool, []> concat_40_interleave_0 = const()[name = tensor<string, []>("concat_40_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, ?, 256]> concat_40 = concat(axis = concat_40_axis_0, interleave = concat_40_interleave_0, values = (split_40_0, split_40_1))[name = tensor<string, []>("concat_40")];
+            tensor<fp16, [1, ?, 256]> concat_40_cast_fp16 = concat(axis = concat_40_axis_0, interleave = concat_40_interleave_0, values = (split_40_cast_fp16_0, split_40_cast_fp16_1))[name = tensor<string, []>("concat_40_cast_fp16")];
             tensor<int32, [1]> input_25_batch_first_lstm_h0_reshaped_axes_0 = const()[name = tensor<string, []>("input_25_batch_first_lstm_h0_reshaped_axes_0"), val = tensor<int32, [1]>([0])];
-            tensor<fp32, [?, 256]> input_25_batch_first_lstm_h0_reshaped = squeeze(axes = input_25_batch_first_lstm_h0_reshaped_axes_0, x = concat_40)[name = tensor<string, []>("input_25_batch_first_lstm_h0_reshaped")];
+            tensor<fp16, [?, 256]> input_25_batch_first_lstm_h0_reshaped_cast_fp16 = squeeze(axes = input_25_batch_first_lstm_h0_reshaped_axes_0, x = concat_40_cast_fp16)[name = tensor<string, []>("input_25_batch_first_lstm_h0_reshaped_cast_fp16")];
             tensor<int32, [2]> split_41_split_sizes_0 = const()[name = tensor<string, []>("split_41_split_sizes_0"), val = tensor<int32, [2]>([1, 1])];
             tensor<int32, []> split_41_axis_0 = const()[name = tensor<string, []>("split_41_axis_0"), val = tensor<int32, []>(0)];
-            tensor<fp32, [1, ?, 128]> split_41_0, tensor<fp32, [1, ?, 128]> split_41_1 = split(axis = split_41_axis_0, split_sizes = split_41_split_sizes_0, x = split_1_3)[name = tensor<string, []>("split_41")];
+            tensor<fp16, [1, ?, 128]> split_41_cast_fp16_0, tensor<fp16, [1, ?, 128]> split_41_cast_fp16_1 = split(axis = split_41_axis_0, split_sizes = split_41_split_sizes_0, x = split_1_cast_fp16_3)[name = tensor<string, []>("split_41_cast_fp16")];
             tensor<int32, []> concat_41_axis_0 = const()[name = tensor<string, []>("concat_41_axis_0"), val = tensor<int32, []>(2)];
             tensor<bool, []> concat_41_interleave_0 = const()[name = tensor<string, []>("concat_41_interleave_0"), val = tensor<bool, []>(false)];
-            tensor<fp32, [1, ?, 256]> concat_41 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (split_41_0, split_41_1))[name = tensor<string, []>("concat_41")];
+            tensor<fp16, [1, ?, 256]> concat_41_cast_fp16 = concat(axis = concat_41_axis_0, interleave = concat_41_interleave_0, values = (split_41_cast_fp16_0, split_41_cast_fp16_1))[name = tensor<string, []>("concat_41_cast_fp16")];
             tensor<int32, [1]> input_25_batch_first_lstm_c0_reshaped_axes_0 = const()[name = tensor<string, []>("input_25_batch_first_lstm_c0_reshaped_axes_0"), val = tensor<int32, [1]>([0])];
-            tensor<fp32, [?, 256]> input_25_batch_first_lstm_c0_reshaped = squeeze(axes = input_25_batch_first_lstm_c0_reshaped_axes_0, x = concat_41)[name = tensor<string, []>("input_25_batch_first_lstm_c0_reshaped")];
+            tensor<fp16, [?, 256]> input_25_batch_first_lstm_c0_reshaped_cast_fp16 = squeeze(axes = input_25_batch_first_lstm_c0_reshaped_axes_0, x = concat_41_cast_fp16)[name = tensor<string, []>("input_25_batch_first_lstm_c0_reshaped_cast_fp16")];
             tensor<string, []> input_25_batch_first_direction_0 = const()[name = tensor<string, []>("input_25_batch_first_direction_0"), val = tensor<string, []>("bidirectional")];
             tensor<bool, []> input_25_batch_first_output_sequence_0 = const()[name = tensor<string, []>("input_25_batch_first_output_sequence_0"), val = tensor<bool, []>(true)];
             tensor<string, []> input_25_batch_first_recurrent_activation_0 = const()[name = tensor<string, []>("input_25_batch_first_recurrent_activation_0"), val = tensor<string, []>("sigmoid")];
             tensor<string, []> input_25_batch_first_cell_activation_0 = const()[name = tensor<string, []>("input_25_batch_first_cell_activation_0"), val = tensor<string, []>("tanh")];
             tensor<string, []> input_25_batch_first_activation_0 = const()[name = tensor<string, []>("input_25_batch_first_activation_0"), val = tensor<string, []>("tanh")];
-            tensor<fp32, [589, ?, 256]> input_25_batch_first_0, tensor<fp32, [?, 256]> input_25_batch_first_1, tensor<fp32, [?, 256]> input_25_batch_first_2 = lstm(activation = input_25_batch_first_activation_0, bias = add_6, bias_back = add_7, cell_activation = input_25_batch_first_cell_activation_0, direction = input_25_batch_first_direction_0, initial_c = input_25_batch_first_lstm_c0_reshaped, initial_h = input_25_batch_first_lstm_h0_reshaped, output_sequence = input_25_batch_first_output_sequence_0, recurrent_activation = input_25_batch_first_recurrent_activation_0, weight_hh = concat_37, weight_hh_back = concat_39, weight_ih = concat_36, weight_ih_back = concat_38, x = input_25_lstm_layer_2_0)[name = tensor<string, []>("input_25_batch_first")];
+            tensor<fp16, [512, 256]> concat_36_to_fp16 = const()[name = tensor<string, []>("concat_36_to_fp16"), val = tensor<fp16, [512, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2091328)))];
+            tensor<fp16, [512, 128]> concat_37_to_fp16 = const()[name = tensor<string, []>("concat_37_to_fp16"), val = tensor<fp16, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2353536)))];
+            tensor<fp16, [512]> add_6_to_fp16 = const()[name = tensor<string, []>("add_6_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2484672)))];
+            tensor<fp16, [512, 256]> concat_38_to_fp16 = const()[name = tensor<string, []>("concat_38_to_fp16"), val = tensor<fp16, [512, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2485760)))];
+            tensor<fp16, [512, 128]> concat_39_to_fp16 = const()[name = tensor<string, []>("concat_39_to_fp16"), val = tensor<fp16, [512, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2747968)))];
+            tensor<fp16, [512]> add_7_to_fp16 = const()[name = tensor<string, []>("add_7_to_fp16"), val = tensor<fp16, [512]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2879104)))];
+            tensor<fp16, [589, ?, 256]> input_25_batch_first_cast_fp16_0, tensor<fp16, [?, 256]> input_25_batch_first_cast_fp16_1, tensor<fp16, [?, 256]> input_25_batch_first_cast_fp16_2 = lstm(activation = input_25_batch_first_activation_0, bias = add_6_to_fp16, bias_back = add_7_to_fp16, cell_activation = input_25_batch_first_cell_activation_0, direction = input_25_batch_first_direction_0, initial_c = input_25_batch_first_lstm_c0_reshaped_cast_fp16, initial_h = input_25_batch_first_lstm_h0_reshaped_cast_fp16, output_sequence = input_25_batch_first_output_sequence_0, recurrent_activation = input_25_batch_first_recurrent_activation_0, weight_hh = concat_37_to_fp16, weight_hh_back = concat_39_to_fp16, weight_ih = concat_36_to_fp16, weight_ih_back = concat_38_to_fp16, x = input_25_lstm_layer_2_cast_fp16_0)[name = tensor<string, []>("input_25_batch_first_cast_fp16")];
             tensor<int32, [3]> input_25_perm_0 = const()[name = tensor<string, []>("input_25_perm_0"), val = tensor<int32, [3]>([1, 0, 2])];
-            tensor<fp32, [?, 589, 256]> input_25 = transpose(perm = input_25_perm_0, x = input_25_batch_first_0)[name = tensor<string, []>("transpose_4")];
-            tensor<fp32, [?, 589, 128]> input_27 = linear(bias = linear_0_bias, weight = linear_0_weight, x = input_25)[name = tensor<string, []>("linear_0")];
+            tensor<fp16, [128, 256]> linear_0_weight_to_fp16 = const()[name = tensor<string, []>("linear_0_weight_to_fp16"), val = tensor<fp16, [128, 256]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2880192)))];
+            tensor<fp16, [128]> linear_0_bias_to_fp16 = const()[name = tensor<string, []>("linear_0_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2945792)))];
+            tensor<fp16, [?, 589, 256]> input_25_cast_fp16 = transpose(perm = input_25_perm_0, x = input_25_batch_first_cast_fp16_0)[name = tensor<string, []>("transpose_4")];
+            tensor<fp16, [?, 589, 128]> linear_0_cast_fp16 = linear(bias = linear_0_bias_to_fp16, weight = linear_0_weight_to_fp16, x = input_25_cast_fp16)[name = tensor<string, []>("linear_0_cast_fp16")];
             tensor<fp32, []> var_220 = const()[name = tensor<string, []>("op_220"), val = tensor<fp32, []>(0x1.47ae14p-7)];
-            tensor<fp32, [?, 589, 128]> input_29 = leaky_relu(alpha = var_220, x = input_27)[name = tensor<string, []>("input_29")];
-            tensor<fp32, [?, 589, 128]> input_31 = linear(bias = linear_1_bias, weight = linear_1_weight, x = input_29)[name = tensor<string, []>("linear_1")];
+            tensor<fp16, [?, 589, 128]> input_29_cast_fp16 = leaky_relu(alpha = var_220, x = linear_0_cast_fp16)[name = tensor<string, []>("input_29_cast_fp16")];
+            tensor<fp16, [128, 128]> linear_1_weight_to_fp16 = const()[name = tensor<string, []>("linear_1_weight_to_fp16"), val = tensor<fp16, [128, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2946112)))];
+            tensor<fp16, [128]> linear_1_bias_to_fp16 = const()[name = tensor<string, []>("linear_1_bias_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2978944)))];
+            tensor<fp16, [?, 589, 128]> linear_1_cast_fp16 = linear(bias = linear_1_bias_to_fp16, weight = linear_1_weight_to_fp16, x = input_29_cast_fp16)[name = tensor<string, []>("linear_1_cast_fp16")];
             tensor<fp32, []> var_225 = const()[name = tensor<string, []>("op_225"), val = tensor<fp32, []>(0x1.47ae14p-7)];
-            tensor<fp32, [?, 589, 128]> input_33 = leaky_relu(alpha = var_225, x = input_31)[name = tensor<string, []>("input_33")];
-            tensor<fp32, [?, 589, 7]> input = linear(bias = classifier_bias, weight = classifier_weight, x = input_33)[name = tensor<string, []>("linear_2")];
+            tensor<fp16, [?, 589, 128]> input_33_cast_fp16 = leaky_relu(alpha = var_225, x = linear_1_cast_fp16)[name = tensor<string, []>("input_33_cast_fp16")];
+            tensor<fp16, [7, 128]> classifier_weight_to_fp16 = const()[name = tensor<string, []>("classifier_weight_to_fp16"), val = tensor<fp16, [7, 128]>(BLOBFILE(path = tensor<string, []>("@model_path/weights/weight.bin"), offset = tensor<uint64, []>(2979264)))];
+            tensor<fp16, [7]> classifier_bias_to_fp16 = const()[name = tensor<string, []>("classifier_bias_to_fp16"), val = tensor<fp16, [7]>([-0x1.01p+0, 0x1.67cp-2, 0x1.3d8p-1, 0x1.c8cp-2, -0x1.444p-2, -0x1.59p-1, -0x1.8fcp-2])];
+            tensor<fp16, [?, 589, 7]> linear_2_cast_fp16 = linear(bias = classifier_bias_to_fp16, weight = classifier_weight_to_fp16, x = input_33_cast_fp16)[name = tensor<string, []>("linear_2_cast_fp16")];
+            tensor<string, []> linear_2_cast_fp16_to_fp32_dtype_0 = const()[name = tensor<string, []>("linear_2_cast_fp16_to_fp32_dtype_0"), val = tensor<string, []>("fp32")];
             tensor<int32, []> var_231 = const()[name = tensor<string, []>("op_231"), val = tensor<int32, []>(-1)];
-            tensor<fp32, [?, 589, 7]> var_232_softmax = softmax(axis = var_231, x = input)[name = tensor<string, []>("op_232_softmax")];
+            tensor<fp32, [?, 589, 7]> linear_2_cast_fp16_to_fp32 = cast(dtype = linear_2_cast_fp16_to_fp32_dtype_0, x = linear_2_cast_fp16)[name = tensor<string, []>("cast_16")];
+            tensor<fp32, [?, 589, 7]> var_232_softmax = softmax(axis = var_231, x = linear_2_cast_fp16_to_fp32)[name = tensor<string, []>("op_232_softmax")];
             tensor<fp32, []> var_232_epsilon_0 = const()[name = tensor<string, []>("op_232_epsilon_0"), val = tensor<fp32, []>(0x1p-149)];
             tensor<fp32, [?, 589, 7]> log_probs = log(epsilon = var_232_epsilon_0, x = var_232_softmax)[name = tensor<string, []>("op_232")];
         } -> (log_probs);

Segmentation.mlmodelc/weights/weight.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c3189a64946c75bc24fcb98afe89ad78c52bdbadfdf65e857fb1b81e2cc9fbb2
-size 5959360
+oid sha256:0026d3483c74bc989fdd1649c5765ca5395235a6d140a698a2d87b95cddf56ae
+size 2981120