Add ONNX exports for in-browser use (Transformers.js): fp32 and per-channel int8
Browse filesChecked on the 763-review held-out split: fp32 identical to PyTorch (F1 0.885); per-channel MatMul int8 F1 0.891 with 1 decision flip.
- onnx/model.onnx +3 -0
- onnx/model_quantized.onnx +3 -0
onnx/model.onnx
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:fd7104b48037c74237c10fc0bbb94639bb390bf097128e2a7ba2a838805fe8f1
|
| 3 |
+
size 409344363
|
onnx/model_quantized.onnx
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3e8c56a42a42c86738429fec8871b539ce983ba39cb2d70aa623515d938bf5a2
|
| 3 |
+
size 153390841
|