Keras2ncnn

NEW OPS:
24 Add support to SeparableConv2D and BilinearUpsampling 27cc153
27 Add support to Relu6 Activation 943cd5c
33 Add support to Permute ad6ab22
Add support to decoding Functional c6db96e3

BUG FIX:
30 When there is only one layer, the optimizer thought it was a dummy node and removed 87357b0
Remove redundant InputLayer when using a sequential model. fbabcad

BUG FIX:
- 23 Add support to SeparableConv2D. ce4c1c8
- 22 Fix "'str' object has no attribute 'decode'" on certain python version. 9980e9a
- 21 Add limited support to TensorflowOpLayer (Mul with constant). 71dbf08

BUG FIX:
18 Emit default input when no input layer is specific in Keras.
Fix a byte string decoding bug on certain version of h5py.

NEW FEATURES:
MUCH MUCH BETTER Exception Prints. If you meet any issue, attach the exception message will help a lot!

v.0.1.3
BUG FIX:
- Fix a bug in reshape. When using reshape as squeeze, the dim is incorrect.
- Fix a bug in ReLU layer. When ReLU does not have slopt, the converter will throw an error.

NEW FEATURES:
- Better debugging system! Try it out when you converting yout model by ->
bash
python3 -m keras2ncnn -i YOUT_KERAS_FILE.h5 -d


KNOWN BUGS:
- The debugger does not work well with multi out or multi input model.
- The debugger is WIP, so... It will have a lot of bugs.

BUG FIX:
- When emitting fused sigmoid for Dense layer, an extra softmax layer may be inserted after the Dense layer.
- When parsing nested sequential graph, the joint of the the graph may be misconnected.

NEW FEATURES:
- New debugging system allowing for ease comparing accuracy between ncnn and keras graph.

Feature Highlights
- Keras h5df to ncnn param/bin file converter
- Support a variety of models, sequential or not, TF1 or TF2 !
- New weight indexing method, better model compatibility !
- Emended debugger for comparing accuracy with ncnn. (Working on)

Supported Op
- InputLayer
- Conv2D (With fused relu, sigmoid activation)
- Conv2DTranspose (With fused relu, sigmoid activation)
- DepthwiseConv2D
- Add
- Multiply
- ZeroPadding2D
- ReLU
- LeakyReLU
- UpSampling2D
- Concatenate
- GlobalAveragePooling2D
- MaxAveragePooling2D
- AveragePooling2D
- MaxPooling2D
- BatchNormalization
- Dense (With fused relu, sigmoid, and non-fused softmax activation)
- Activation (Support relu, sigmoid)