Zulu02

I encountered the same problem and I guess you can fix it by building with an older LLVM and clang version just install llvm-8 with `sudo apt-get install llvm-8` and...

and I realized that Glow expects at least something behind python, on new Ubuntu 20.04 installations, python needs to be redirected with python-is-python3, which is available with `sudo apt-get install...

using -minibatch=1 resulted in the same output. When I try an TFLite Flatbuffer, which has been created from the same keras network, it results in `Model: networks Running 1 thread(s)....

Okay, is it possible, that the problem is related to training the model with a numpy array and the the model input [1, 28, 28, 1] and converting the samples...

Is there a way to feed the image-classifier with binary files instead of PNGs? I found the -input-dataset option in the documentation, but when I tried it, it was not...

Ok, I guess, I've found the issue, for the inital keras training, the input has the size [1,28,28,1], but if I write this numpy array to a file using Pillow...

Almost, it is the other way round: The model, which has been created with TF Keras, expects `[1,28,28,1]` as input and the images are `[28,28]`, I modified the exported ONNX...

Okay, I made some progress by using the precompiled docker instead of compiling Glow myself and using the MNIST model from the ONNX model zoo [MNIST ](https://github.com/onnx/models/tree/master/vision/classification/mnist), which uses NCHW...