handle raw types correctly in helper.make_tensor #4262
Conversation
Before this change, the newly added `test_make_int8_tensor` would fail because `TENSOR_TYPE_TO_STORAGE_TENSOR_TYPE[INT8] = FLOAT32`. But raw INT8 values are packed and are not the same size as FLOAT32 values. Signed-off-by: Gary Miguel <garymiguel@microsoft.com>
There was a problem hiding this comment.
Thank you so much for the quick fix! I can confirm this is the right way. Please fix flake8 failure.
I misunderstood so my fix (#4234) was incorrect --
Instead of storage type, actually make_tensor with raw directly uses its original NumPy type (defined in TENSOR_TYPE_TO_NP_TYPE) as bytes. Only TensorProto.FLOAT16 and TensorProto.BFLOAT16 are exceptions. Although now TensorProto.FLOAT16 uses uint16 as raw_data, the itemsize for uint16 and float16 is the same (2) so TensorProto.FLOAT16 is not an issue.
Thus, just like what this PR does, we only need to handle the special case TensorProto.BFLOAT16, because float32's itemsize is 4 and uint16 itemsize is 2 (there is a truncation when processing original float32 NumPy array).
Another finding is for making float16 data with raw, actually there is no need to convert original float16 NumPy array as np_array.view(dtype=np.uint16).flatten().tobytes() for raw_data, because their tobytes() are identical (See here). I will send another PR to improve it. Let's merge this PR first for upcoming ONNX 1.12.
|
@jcwchen please merge |
Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com>
* Use op name rather than hard-coding Hann in window op doc strings. (#4248) And strip trailing spaces. Signed-off-by: Gary Miguel <garymiguel@microsoft.com> Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com> * Fix: Window functions generate incorrect shape for with symmetric attribute (#4256) * fix symmetric function definition, and add shape inference test Signed-off-by: Sheil Kumar <smk2007@gmail.com> * whitespace... Signed-off-by: Sheil Kumar <smk2007@gmail.com> * whitespace... Signed-off-by: Sheil Kumar <smk2007@gmail.com> * flake8 whitespace error Signed-off-by: Sheil Kumar <smk2007@gmail.com> * fix stft test Signed-off-by: Sheil Kumar <smk2007@gmail.com> Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com> * support truncation & fix big-endian support (#4238) Signed-off-by: Ian Bearman <ianb@microsoft.com> Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com> * Fix layer normalization's reference outputs (#4263) Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com> * handle raw types correctly in helper.make_tensor (#4262) Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com> * bump to 1.12.0rc4 for cherry-picks Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com> * Fix sub-graph generation for LN (#4268) Signed-off-by: Wei-Sheng Chin <wschin@outlook.com> Signed-off-by: Chun-Wei Chen <jacky82226@gmail.com> Co-authored-by: Gary Miguel <garymiguel@microsoft.com> Co-authored-by: Sheil Kumar <smk2007@gmail.com> Co-authored-by: Ian Bearman <ianb@microsoft.com> Co-authored-by: Wei-Sheng Chin <wschin@outlook.com>
Before this change, the newly added
test_make_int8_tensorwould failbecause
TENSOR_TYPE_TO_STORAGE_TENSOR_TYPE[INT8] = FLOAT32. But rawINT8 values are packed and are not the same size as FLOAT32 values.
Signed-off-by: Gary Miguel garymiguel@microsoft.com