change input size error

[ren__nannan]

Hi: 

use "Intel\computer_vision_sdk_2018.3.343\deployment_tools\intel_models\head-pose-estimation-adas-0001\FP32\head-pose-estimation-adas-0001.xml"  or " Intel\computer_vision_sdk_2018.3.343\deployment_tools\intel_models\license-plate-recognition-barrier-0001\FP32\license-plate-recognition-barrier-0001.xml", if I  change input size like this：

    SizeVector input_dims = inputInfoFirst->getInputData()->getTensorDesc().getDims();
    input_dims[2] = config_.input_h;
    input_dims[3] = config_.input_w;
    std::map<std::string, SizeVector> input_shapes;
    input_shapes[inputInfo.begin()->first] = input_dims;
    net_reader.getNetwork().reshape(input_shapes);

run "net_ = config_.plugin.LoadNetwork(net_reader.getNetwork(), {});" will report error in the "extract_exception".

[Nikolay_L_Intel1]

Hi!

These models have Fully Connected layer that by definition can't be resized: weights are fixed and rigidly correspond to the input size.

Have you tried to catch exception and to print error message? In your particular case, you should get the following explanation of the failure:

Failed to infer shapes for FullyConnected layer with error: New shapes [1,71,120,494] make Kernels(120x494), Channels(71), Output depth(128), Groups(1) not matching weights size: 799744

Best regards,
Nikolai