github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-01-09 15:08:02 -05:00
Code Issues Packages Projects Releases Wiki Activity

switch quantization to unsigned/unsigned + add Ops.REDUCE (#9527)

Browse Source 
* switch quantization to unsigned/unsigned + add Ops.REDUCE

* tests

* nhwc + replay pkl
This commit is contained in:
George Hotz
2025-03-21 17:02:37 +08:00
committed by GitHub
parent a35b0a88bf
commit 8e555c586c
6 changed files with 238 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
examples/test_onnx_imagenet.py
View File

@@ -58,7 +58,7 @@ if __name__ == "__main__":
return None
return {"input": img.numpy()}
quantize_static(model_fp32, fn, ImagenetReader(), quant_format=QuantFormat.QDQ, per_channel=False,
activation_type=QuantType.QUInt8, weight_type=QuantType.QInt8,
activation_type=QuantType.QUInt8, weight_type=QuantType.QUInt8,
extra_options={"ActivationSymmetric": False})
run_onnx_jit, input_specs = load_onnx_model(fetch(fn))
@@ -74,5 +74,8 @@ if __name__ == "__main__":
hit += y==t
print(f"target: {y:3d} pred: {t:3d} acc: {hit/(i+1)*100:.2f}%")
MS_TARGET = 13.4
print(f"need {GlobalCounters.global_ops/1e9*(1000/MS_TARGET):.2f} GFLOPS for {MS_TARGET:.2f} ms")
import pickle
with open("/tmp/im.pkl", "wb") as f: pickle.dump(run_onnx_jit, f)