import paddle import math from x2paddle.op_mapper.onnx2paddle import onnx_custom_layer as x2paddle_nn class ONNXModel(paddle.nn.Layer): def __init__(self): super(ONNXModel, self).__init__() self.x2paddle_fc_weight = self.create_parameter(shape=[8, 64], attr='x2paddle_fc_weight', dtype='float32', default_initializer=paddle.nn.initializer.Constant(value=0.0)) self.x2paddle_fc_bias = self.create_parameter(shape=[8], attr='x2paddle_fc_bias', dtype='float32', default_initializer=paddle.nn.initializer.Constant(value=0.0)) self.conv0 = paddle.nn.Conv2D(in_channels=3, out_channels=16, kernel_size=[3, 3], padding=1) self.relu0 = paddle.nn.ReLU() self.conv1 = paddle.nn.Conv2D(in_channels=16, out_channels=16, kernel_size=[3, 3], padding=1) self.relu1 = paddle.nn.ReLU() self.conv2 = paddle.nn.Conv2D(in_channels=16, out_channels=16, kernel_size=[3, 3], padding=1) self.relu2 = paddle.nn.ReLU() self.conv3 = paddle.nn.Conv2D(in_channels=16, out_channels=16, kernel_size=[3, 3], padding=1) self.relu3 = paddle.nn.ReLU() self.conv4 = paddle.nn.Conv2D(in_channels=16, out_channels=16, kernel_size=[3, 3], padding=1) self.relu4 = paddle.nn.ReLU() self.conv5 = paddle.nn.Conv2D(in_channels=16, out_channels=16, kernel_size=[3, 3], padding=1) self.relu5 = paddle.nn.ReLU() self.conv6 = paddle.nn.Conv2D(in_channels=16, out_channels=16, kernel_size=[3, 3], padding=1) self.relu6 = paddle.nn.ReLU() self.conv7 = paddle.nn.Conv2D(in_channels=16, out_channels=32, kernel_size=[3, 3], stride=2, padding=1) self.pad0 = paddle.nn.Pad2D(mode='constant', padding=[0, 0, 0, 0, 0, 0, 0, 0]) self.relu7 = paddle.nn.ReLU() self.pool0 = paddle.nn.AvgPool2D(kernel_size=[1, 1], stride=2) self.conv8 = paddle.nn.Conv2D(in_channels=32, out_channels=32, kernel_size=[3, 3], padding=1) self.relu8 = paddle.nn.ReLU() self.conv9 = paddle.nn.Conv2D(in_channels=32, out_channels=32, kernel_size=[3, 3], padding=1) self.relu9 = paddle.nn.ReLU() self.conv10 = paddle.nn.Conv2D(in_channels=32, out_channels=32, kernel_size=[3, 3], padding=1) self.relu10 = paddle.nn.ReLU() self.conv11 = paddle.nn.Conv2D(in_channels=32, out_channels=32, kernel_size=[3, 3], padding=1) self.relu11 = paddle.nn.ReLU() self.conv12 = paddle.nn.Conv2D(in_channels=32, out_channels=32, kernel_size=[3, 3], padding=1) self.relu12 = paddle.nn.ReLU() self.conv13 = paddle.nn.Conv2D(in_channels=32, out_channels=64, kernel_size=[3, 3], stride=2, padding=1) self.pad1 = paddle.nn.Pad2D(mode='constant', padding=[0, 0, 0, 0, 0, 0, 0, 0]) self.relu13 = paddle.nn.ReLU() self.pool1 = paddle.nn.AvgPool2D(kernel_size=[1, 1], stride=2) self.conv14 = paddle.nn.Conv2D(in_channels=64, out_channels=64, kernel_size=[3, 3], padding=1) self.relu14 = paddle.nn.ReLU() self.conv15 = paddle.nn.Conv2D(in_channels=64, out_channels=64, kernel_size=[3, 3], padding=1) self.relu15 = paddle.nn.ReLU() self.conv16 = paddle.nn.Conv2D(in_channels=64, out_channels=64, kernel_size=[3, 3], padding=1) self.relu16 = paddle.nn.ReLU() self.conv17 = paddle.nn.Conv2D(in_channels=64, out_channels=64, kernel_size=[3, 3], padding=1) self.relu17 = paddle.nn.ReLU() self.conv18 = paddle.nn.Conv2D(in_channels=64, out_channels=64, kernel_size=[3, 3], padding=1) self.relu18 = paddle.nn.ReLU() self.pad2 = paddle.nn.Pad2D(mode='constant', padding=[0, 0, 0, 0, 0, 0, 0, 0]) self.pool2 = paddle.nn.AvgPool2D(kernel_size=[8, 8], stride=8) def forward(self, x2paddle_input): x2paddle_fc_weight = self.x2paddle_fc_weight x2paddle_fc_bias = self.x2paddle_fc_bias x2paddle_150 = paddle.full(dtype='float32', shape=[1], fill_value=0.0) x2paddle_176 = paddle.full(dtype='float32', shape=[1], fill_value=0.0) x2paddle_204 = self.conv0(x2paddle_input) x2paddle_121 = self.relu0(x2paddle_204) x2paddle_207 = self.conv1(x2paddle_121) x2paddle_124 = self.relu1(x2paddle_207) x2paddle_210 = self.conv2(x2paddle_124) x2paddle_127 = paddle.add(x=x2paddle_121, y=x2paddle_210) x2paddle_128 = self.relu2(x2paddle_127) x2paddle_213 = self.conv3(x2paddle_128) x2paddle_131 = self.relu3(x2paddle_213) x2paddle_216 = self.conv4(x2paddle_131) x2paddle_134 = paddle.add(x=x2paddle_128, y=x2paddle_216) x2paddle_135 = self.relu4(x2paddle_134) x2paddle_219 = self.conv5(x2paddle_135) x2paddle_138 = self.relu5(x2paddle_219) x2paddle_222 = self.conv6(x2paddle_138) x2paddle_141 = paddle.add(x=x2paddle_135, y=x2paddle_222) x2paddle_142 = self.relu6(x2paddle_141) x2paddle_225 = self.conv7(x2paddle_142) x2paddle_148 = self.pad0(x2paddle_142) x2paddle_145 = self.relu7(x2paddle_225) x2paddle_149 = self.pool0(x2paddle_148) x2paddle_228 = self.conv8(x2paddle_145) x2paddle_151 = paddle.multiply(x=x2paddle_149, y=x2paddle_150) x2paddle_152 = paddle.concat(x=[x2paddle_149, x2paddle_151], axis=1) x2paddle_153 = paddle.add(x=x2paddle_152, y=x2paddle_228) x2paddle_154 = self.relu8(x2paddle_153) x2paddle_231 = self.conv9(x2paddle_154) x2paddle_157 = self.relu9(x2paddle_231) x2paddle_234 = self.conv10(x2paddle_157) x2paddle_160 = paddle.add(x=x2paddle_154, y=x2paddle_234) x2paddle_161 = self.relu10(x2paddle_160) x2paddle_237 = self.conv11(x2paddle_161) x2paddle_164 = self.relu11(x2paddle_237) x2paddle_240 = self.conv12(x2paddle_164) x2paddle_167 = paddle.add(x=x2paddle_161, y=x2paddle_240) x2paddle_168 = self.relu12(x2paddle_167) x2paddle_243 = self.conv13(x2paddle_168) x2paddle_174 = self.pad1(x2paddle_168) x2paddle_171 = self.relu13(x2paddle_243) x2paddle_175 = self.pool1(x2paddle_174) x2paddle_246 = self.conv14(x2paddle_171) x2paddle_177 = paddle.multiply(x=x2paddle_175, y=x2paddle_176) x2paddle_178 = paddle.concat(x=[x2paddle_175, x2paddle_177], axis=1) x2paddle_179 = paddle.add(x=x2paddle_178, y=x2paddle_246) x2paddle_180 = self.relu14(x2paddle_179) x2paddle_249 = self.conv15(x2paddle_180) x2paddle_183 = self.relu15(x2paddle_249) x2paddle_252 = self.conv16(x2paddle_183) x2paddle_186 = paddle.add(x=x2paddle_180, y=x2paddle_252) x2paddle_187 = self.relu16(x2paddle_186) x2paddle_255 = self.conv17(x2paddle_187) x2paddle_190 = self.relu17(x2paddle_255) x2paddle_258 = self.conv18(x2paddle_190) x2paddle_193 = paddle.add(x=x2paddle_187, y=x2paddle_258) x2paddle_194 = self.relu18(x2paddle_193) x2paddle_195 = self.pad2(x2paddle_194) x2paddle_196 = self.pool2(x2paddle_195) x2paddle_202 = paddle.reshape(x=x2paddle_196, shape=[1, -1]) x2paddle_output_mm = paddle.matmul(x=x2paddle_202, y=x2paddle_fc_weight, transpose_y=True) x2paddle_output_mm = paddle.scale(x=x2paddle_output_mm) x2paddle_output = paddle.add(x=x2paddle_output_mm, y=x2paddle_fc_bias) return x2paddle_output def main(x2paddle_input): # There are 1 inputs. # x2paddle_input: shape-[1, 3, 32, 32], type-float32. paddle.disable_static() params = paddle.load(r'/home/aistudio/work/doc_scan_pd_model/model.pdparams') model = ONNXModel() model.set_dict(params, use_structured_name=True) model.eval() out = model(x2paddle_input) return out