Replace with portable 4096-d speaker head

README.md

@@ -1,3 +1,3 @@
----
-license: cc-by-4.0
----
+# Portable Speaker Head
+
+This directory was exported from an Ascend/NPU checkpoint. It contains inference-only, CPU tensors and does not require `torch_npu` to load. Prefer `speaker_head.safetensors` in CUDA/runtime integrations. `speaker_head_cpu.pt` is an equivalent CPU-only PyTorch checkpoint with payload key `speaker_head`. See `export_manifest.json` for hashes and exact validation results.

export_manifest.json

@@ -0,0 +1,131 @@
+{
+  "artifacts": {
+    "config.json": {
+      "bytes": 914,
+      "sha256": "21d52e5f9df53913e1175f8201b43ff15ec25be0618ad4416d97a963b80ae004"
+    },
+    "speaker_head.safetensors": {
+      "bytes": 39912720,
+      "sha256": "8c1e662886a3470c56777b6695b3a24b1b3ec44c6a2f91dbcda3aef4708c81cd"
+    },
+    "speaker_head_cpu.pt": {
+      "bytes": 39916519,
+      "sha256": "4ecea48baa199e61f1eeb48f2bbd6995a6d76c68c0c7ce957ad1c6387d45ded0"
+    }
+  },
+  "format_version": "speaker_head_portable_cpu_v1",
+  "source": {
+    "checkpoint": "/home/zhangchao/project/salmonn_omni_transform_v2_siyuan/siyuanTest/train_speaker_embedding/outputs/mix_5_7_8_ddp_4096_proj1024_min8/best/speaker_head.pt",
+    "checkpoint_sha256": "92351f854ca9ba634af8703ca24a0486f67000b7c024cf133d2e9e63468acf66",
+    "config": "/home/zhangchao/project/salmonn_omni_transform_v2_siyuan/siyuanTest/train_speaker_embedding/outputs/mix_5_7_8_ddp_4096_proj1024_min8/best/config.json",
+    "config_sha256": "21d52e5f9df53913e1175f8201b43ff15ec25be0618ad4416d97a963b80ae004"
+  },
+  "speaker_head": {
+    "parameter_count": 9977860,
+    "tensor_count": 13,
+    "tensors": {
+      "frame_projection.0.bias": {
+        "dtype": "torch.float32",
+        "numel": 1024,
+        "shape": [
+          1024
+        ]
+      },
+      "frame_projection.0.weight": {
+        "dtype": "torch.float32",
+        "numel": 1310720,
+        "shape": [
+          1024,
+          1280
+        ]
+      },
+      "input_norm.bias": {
+        "dtype": "torch.float32",
+        "numel": 1280,
+        "shape": [
+          1280
+        ]
+      },
+      "input_norm.weight": {
+        "dtype": "torch.float32",
+        "numel": 1280,
+        "shape": [
+          1280
+        ]
+      },
+      "layer_logits": {
+        "dtype": "torch.float32",
+        "numel": 3,
+        "shape": [
+          3
+        ]
+      },
+      "output_norm.bias": {
+        "dtype": "torch.float32",
+        "numel": 4096,
+        "shape": [
+          4096
+        ]
+      },
+      "output_norm.weight": {
+        "dtype": "torch.float32",
+        "numel": 4096,
+        "shape": [
+          4096
+        ]
+      },
+      "output_projection.bias": {
+        "dtype": "torch.float32",
+        "numel": 4096,
+        "shape": [
+          4096
+        ]
+      },
+      "output_projection.weight": {
+        "dtype": "torch.float32",
+        "numel": 8388608,
+        "shape": [
+          4096,
+          2048
+        ]
+      },
+      "pooling.attention.0.bias": {
+        "dtype": "torch.float32",
+        "numel": 256,
+        "shape": [
+          256
+        ]
+      },
+      "pooling.attention.0.weight": {
+        "dtype": "torch.float32",
+        "numel": 262144,
+        "shape": [
+          256,
+          1024
+        ]
+      },
+      "pooling.attention.2.bias": {
+        "dtype": "torch.float32",
+        "numel": 1,
+        "shape": [
+          1
+        ]
+      },
+      "pooling.attention.2.weight": {
+        "dtype": "torch.float32",
+        "numel": 256,
+        "shape": [
+          1,
+          256
+        ]
+      }
+    }
+  },
+  "validation": {
+    "cpu_pt_contains_torch_npu_reference": false,
+    "cpu_pt_exact_tensor_match": true,
+    "forward_max_abs_error": 0.0,
+    "safetensors_exact_tensor_match": true,
+    "strict_speaker_head_load": true
+  }
+}

metrics.json

@@ -1,16 +0,0 @@
-{
-  "speaker_count": 1151,
-  "embedding_count": 12000,
-  "positive_pairs": 50000,
-  "negative_pairs": 50000,
-  "eer": 0.026460000000000018,
-  "eer_threshold": 0.21380455791950226,
-  "auc": 0.9962478252,
-  "same_mean": 0.6784662008285522,
-  "diff_mean": 0.0021326725836843252,
-  "top1": 0.970284563082347,
-  "open_eer": 0.045812369269037256,
-  "known_best_mean": 0.775897741317749,
-  "unknown_best_mean": 0.37951692938804626,
-  "step": 12000
-}

speaker_head.pt → speaker_head.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:92351f854ca9ba634af8703ca24a0486f67000b7c024cf133d2e9e63468acf66
-size 1266370097
+oid sha256:8c1e662886a3470c56777b6695b3a24b1b3ec44c6a2f91dbcda3aef4708c81cd
+size 39912720

speaker_head_cpu.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ecea48baa199e61f1eeb48f2bbd6995a6d76c68c0c7ce957ad1c6387d45ded0
+size 39916519