Update tracing module

[gmalivenko]

Due to new PyTorch updates the only way to proper handle weights and layer names at the same time - to use ONNX export.

From the next version of the converter, tracing module will be replaced. Hopefully it will open new features - converting modules such as ModuleList, converting recurrent networks and so on. Stay tuned.

The new tracing module will be placed to the separated repository https://github.com/nerox8664/onnx2keras and used as a library after implementation.

Progress:

• [x] Model loading / parsing
• [x] Convert convolution layers:
  • [x] Conv2d
  • [x] ConvTranspose2d
• [x] Convert linear layers:
  • [x] Linear (Gemm / MatMul)
• [x] Convert pooling layers:
  • [x] MaxPool
  • [x] AveragePool
  • [x] GlobalAveragePool
• [x] Convert padding layers:
  • [x] Pad*
• [x] Convert normalization layers:
  • [x] BatchNorm2d
  • [x] InstanceNorm2d
• [x] Activations:
  • [x] ReLU
  • [x] LeakyReLU
  • [x] Sigmoid
  • [x] Tanh
  • [x] Selu
• [x] Reshape operations:
  • [x] View / Flatten *
  • [x] Transpose *
  • [x] Unsqueeze *
  • [x] Shape *
• [x] Also element-wise operations:
  • [x] Div
  • [x] Add
  • [x] Mul
  • [x] Sub
• [x] Other operations:
  • [x] Gather *
  • [x] Clip
  • [x] Exp
  • [x] Log
• [x] Implement new tests to cover converter
• [x] Test new available models on torchvision