OSSForks/pytorch
1
0
Fork 0
mirror of https://github.com/zebrajr/pytorch.git synced 2025-12-06 12:20:52 +01:00
Code Issues Actions 38 Packages Projects Releases Wiki Activity

[fbgemm] fix bug handling bias in rowwise quantization of FC (#58022)

Browse Source 
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/58022

Caffe2 Int8FC + rowwise quantization was not handling bias correctly.

Test Plan: The example in D28347336 doesn't show bigger error with rowwise quantization any more

Reviewed By: hx89, janeyx99

Differential Revision: D28347336

fbshipit-source-id: 3ac95fd2f29ef6e52705c3a2361b605813c2bcc5
This commit is contained in:
Jongsoo Park 2021-05-11 08:36:49 -07:00 committed by Facebook GitHub Bot
parent 3e46d6c9e4
commit dc55ab3f77
1 changed files with 4 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
caffe2/quantization/server/fully_connected_dnnlowp_op.cc
View File

@ -815,7 +815,8 @@ bool FullyConnectedDNNLowPOp<T, ReluFused>::GetQuantizationParameters_() {
b_dequantized_.resize(N);
for (int j = 0; j < N; ++j) {
b_dequantized_[j] = fbgemm::Dequantize<int32_t>(
b_quantized_data_[j], in_qparams_[2]);
b_quantized_data_[j],
filter_qparams_[quantize_channelwise_ ? j : 0]);
}
b_dequantized_data_ = b_dequantized_.data();
}
@ -827,7 +828,8 @@ bool FullyConnectedDNNLowPOp<T, ReluFused>::GetQuantizationParameters_() {
(*b_quantized_)[j] = fbgemm::Quantize<int32_t>(
b_dequantized_data_[j],
0,
in_qparams_[0].scale * filter_qparams_[0].scale,
in_qparams_[0].scale *
filter_qparams_[quantize_channelwise_ ? j : 0].scale,
32);
}
b_quantized_data_ = b_quantized_->data();