Conversion to tflite fails for a specific network about tinyneuralnetwork

[peterjc123]

It's a known issue that in some legacy versions of PyTorch, the following code won't be traced.

class Model(nn.Module):
    def forward(self, x):
        // ...
        cat_10 = self.float_functional_simple_9.cat([getitem_15, backbone_stage3_5_branch_main_7], 1)
        data_9 = cat_10.data
        size_9 = data_9.size()
        // ...

but it will work when you write like this.

class Model(nn.Module):
    def forward(self, x):
        // ...
        cat_10 = self.float_functional_simple_9.cat([getitem_15, backbone_stage3_5_branch_main_7], 1)
        size_9 = cat_10.size()
        // ...

We will work on our code generator and tracer to get rid of the .data calls.

[peterjc123]

@liamsun2019 We pushed a fix for the issue and the given model can be converted locally. Would you please try to remove the existing model description file and try again?

[liamsun2019]

Big thanks. I will try it out ASAP.

[peterjc123]

I'll close this one for now. If you disagree, please feel free to reopen it.

[liamsun2019]

No problem，thanks.

[liamsun2019]

The converted tflite model cannot pass benchmark test:
./linux_x86-64_benchmark_model --graph=qat_model.tflite
STARTING!
Log parameter values verbosely: [0]
Graph: [qat_model.tflite]
Loaded model qat_model.tflite
ERROR: tensorflow/lite/kernels/concatenation.cc:179 t->params.scale != output->params.scale (3 != 2018333616)
ERROR: Node number 12 (CONCATENATION) failed to prepare.
Failed to allocate tensors!
Benchmarking failed.

Similar errors happen when doing inference with the tflite model if invoking interpreter.allocate_tensors()

[peterjc123]

@liamsun2019 This is because of something else. I've discussed with you through online messages. Next time, please open a new issue for that. Thanks.