Eason
Eason
I ran into similar problem. Before using 8.4 GA, TRT never reported the issue, so I never realized how different FP16 model performs, because it gives very similar results in...
> > In the only example provided in the toolkit, it loaded the PTQ calibrated weights and did the QAT based on it. There isn't a standalone QAT example without...
> Hi @deephog , we recommend to do PTQ first, then doing the QAT to fine tune the weights using the fixed quant scale. This helps converge. > > In...
> > > > In the only example provided in the toolkit, it loaded the PTQ calibrated weights and did the QAT based on it. There isn't a standalone QAT...
I previously closed the issue because I thought I figured it out by removing --noTF32 flag. It did successfully compile, but the entire engine fell back to TF32, so basically...
> try to reproduce with TRT 8.4. > > ``` > [07/08/2022-06:16:07] [V] [TRT] Swap the layer type of QuantizeLinear_7751_clone_1_clone_2 from QUANTIZE to kQDQ > [07/08/2022-06:16:07] [V] [TRT] Running: QDQToCopy...
> The `fnet.conv1.weight` is shared by multiple conv, and currently TRT cannot constant fold the weights that shared. Dynamic weights input support could fix this, this is already in TRT...
> The `fnet.conv1.weight` is shared by multiple conv, and currently TRT cannot constant fold the weights that shared. Dynamic weights input support could fix this, this is already in TRT...
> The `fnet.conv1.weight` is shared by multiple conv, and currently TRT cannot constant fold the weights that shared. Dynamic weights input support could fix this, this is already in TRT...
Here is an updated version of ONNX [file](https://drive.google.com/file/d/1GVmxOIIGN5btciDzAsa2N2n9TFKNl6xP/view?usp=sharing) The previous "dual weights no input" errors are cleaned or evaded. However, a new error occurred which is difficult to trace. Error[2]:...