SW: need Tensorflow/Python integration with TIDL

[jadonk]

Tensorflow Lite support is on the roadmap for 2020Q1.

[pra-dan]

With the release of TIDL 3.15.1, TensorFlow and TensorFlow Lite has been added to the supported frameworks among Caffe, TensorFlow Lite and ONNX although, not yet a verified model like the Jacinto11, MobileNetV2 ,etc.

The model import feature of TIDL there is provides constraint on the supported operators and layers. Irrespective of the framework, the supported layers are

Pooling Layer (Average and Max Pooling)
ReLU Layer
Element Wise Layer (Add, Max, Product)
Inner Product Layer (Fully Connected Layer)
Soft Max Layer
Bias Layer
Deconvolution Layer
Concatenate layer
ArgMax Layer
Scale Layer
PReLU Layer
Batch Normalization layer
ReLU6 Layer
Crop layer
Slice layer
Flatten layer
Split Layer
Detection Output Layer

But for the supported operators, during the import process, some operators/layers are combined or converted into TIDL layers. For TensorFlow specifically, the supported operators are:

TensorFlow Operator	TIDL Layer
Placeholder	TIDL_DataLayer
Conv2D	TIDL_ConvolutionLayer
DepthwiseConv2dNative	TIDL_ConvolutionLayer
BiasAdd	TIDL_BiasLayer
Add	TIDL_EltWiseLayer
Mul	TIDL_ScaleLayer
FusedBatchNorm	TIDL_BatchNormLayer
Relu	TIDL_ReLULayer
Relu6	TIDL_ReLULayer
MaxPool	TIDL_PoolingLayer
AvgPool	TIDL_PoolingLayer
ConcatV2	TIDL_ConcatLayer
Slice	TIDL_SliceLayer
Squeeze	See note below
Reshape	See note below
Softmax	TIDL_SoftMaxLayer
Pad	TIDL_PadLayer
Mean	TIDL_PoolingLayer

Whereas for TensorFlow Lite, TensorFlow Operator	TIDL Layer
Placeholder	TIDL_DataLayer
CONV_2D	TIDL_ConvolutionLayer
TRANSPOSE_CONV	TIDL_Deconv2DLayer
DEPTHWISE_CONV_2D	TIDL_ConvolutionLayer
ADD	TIDL_EltWiseLayer
MUL	TIDL_ScaleLayer
RELU	TIDL_ReLULayer
RELU6	TIDL_ReLULayer
MAX_POOL_2D	TIDL_PoolingLayer
AVERAGE_POOL_2D	TIDL_PoolingLayer
CONCATENATION	TIDL_ConcatLayer
RESHAPE	See note below
SOFTMAX	TIDL_SoftMaxLayer
ARG_MAX	TIDL_ArgMaxLayer
PAD	TIDL_PadLayer
MEAN	TIDL_PoolingLayer
FULLY_CONNECTED	TIDL_InnerProductLayer

Note:

“Reshape” and “Squeeze” are supported by being coalesced into other layers: If “Reshape” immediately follows “Squeeze”, they both are coalesced into TIDL_FlattenLayer. If “Reshape” immediately follows “AvgPool”, “Reshape” is coalesced into TIDL_PoolingLayer. If “Reshape” immediately follows TIDL_InnerProductLayer, it is coalesced into TIDL_InnerProductLayer.

Source: Official docs for TIDL: Foundational Components

[matthewanderson23]

Any updates on whether TF Lite support is imminent?