Lukas Geiger
Lukas Geiger
It would be excellent if `netron` would be able to show the shape of the activations between two layers. Similar to the way it shows the shape of the input...
> Does TFLite support `TRANSPOSE_CONV` with per-channel quantization? Yes it does. `TRANSPOSE_CONV` kernels [supports both](https://github.com/tensorflow/tensorflow/blob/57cf4714fc1ee757b6f5226238035a981802b7d8/tensorflow/lite/kernels/transpose_conv.cc#L379-L380) but internally it will actually [convert it to per-channel](https://github.com/tensorflow/tensorflow/blob/57cf4714fc1ee757b6f5226238035a981802b7d8/tensorflow/lite/kernels/transpose_conv.cc#L384-L389) in both cases using [this code...
> Does TFLite support TRANSPOSE_CONV with per-channel quantization? @Maratyszcza I double check this again, and indeed TFLite supports per-channel quantization and the current converter will always generate per-channel quantized transposed...
This could probably done using something like https://github.com/aszinovyev/bazel-compilation-db
It looks like dilated convolutions weren't properly converted in LCE 0.6.2 which was based on the TensorFlow 2.5 converter, this seems to fixed in LCE 0.7. With respect to XNOR-Net:...
That's really cool, I completely forgot about this 🎉 In larq we already have the unipolar [`SteHeaviside`](https://docs.larq.dev/larq/api/quantizers/#steheaviside) quantizer (although I think the backward pass should actually be a [0, 1]...
Looks like we are running into https://github.com/tensorflow/tensorflow/issues/40055 here since the `bias` tensors are shared which prevents quantization of the 8bit BConv. E.g. the following network with a non-zero bias results...
> @lgeiger do you have any idea why a Conv2D combined with a Conv2D works though, wouldn't that run into the same issue of shared bias tensors? There it is...
Sorry for the late reply. > I noticed some issues with the latest version only (0.7.0) but not the one before (0.6.2). Grouped convolutions (FP or binary) are converted as...
> I noticed some issues with the latest version only (0.7.0) but not the one before (0.6.2). Grouped convolutions (FP or binary) are converted as custom ops in the latest...