Why used [kernel_extent_w, kernel_extent_h] as k_size?

[YuqiData-UW]

https://github.com/YuqiData-UW/nn-Meter/blob/4a00b043e7ad5ed37eb1946dc67b55107abdca75/nn_meter/ir_converter/frozenpb_converter/shape_inference.py#L469

{'inbounds': ['model/LAYER_0_pad/Pad', 'model/LAYER_0/Conv2D/ReadVariableOp'], 'attr': {'name': 'model/LAYER_0/Conv2D', 'type': 'Conv2D', 'output_shape': [[-1, 226, 226, 64]], 'attr': {'dilations': [1, 1], 'strides': [1, 1], 'padding': b'VALID', 'data_format': b'NCHW', 'kernel_shape': [3, 3], 'weight_shape': [3, 3, 3, 64], 'pads': [0, 0, 0, 0]}, 'input_shape': [[-1, 226, 226, 3]]}, 'outbounds': ['model/LAYER_0/BiasAdd']}

I got this result from Con2D, shouldn't the output_shape be [[-1, 224, 224, 64]] since the padding is VALID? If I replace [kernel_extent_w, kernel_extent_h] with k_size, the output_shape can be [[-1, 224, 224, 64]].