def compare_with_pytorch(self,
pytorch_model,
input_shape_with_batch,
compare_result=True,
rtol=1e-6,
atol=1e-6):
input_shape_with_batch = to_list(input_shape_with_batch)
onnx_path = self.dump_pytorch_to_onnx(pytorch_model,
input_shape_with_batch)
onnx_model = onnx.load(onnx_path)
# TODO: we only consider single output for now.
input_data_with_batch = [
np.random.uniform(0, 1, shape).astype(np.float32)
for shape in input_shape_with_batch
]
# coutput = caffe2.python.onnx.backend.run_model(onnx_model, input_data_with_batch)[0]
pytorch_out = pytorch_model.forward(
self._convert_ndarray_to_tensor(input_data_with_batch))
zmodel = OnnxLoader(onnx_model.graph).to_keras()
zoutput = zmodel.forward(input_data_with_batch[0] if len(
input_data_with_batch) == 1 else input_data_with_batch)
if compare_result:
self.assert_allclose(pytorch_out.detach().numpy(), zoutput, rtol,
atol)
def compare_with_pytorch(self, pytorch_model, input_shape_with_batch):
onnx_path = self.dump_pytorch_to_onnx(pytorch_model, input_shape_with_batch)
onnx_model = onnx.load(onnx_path)
# TODO: we only consider single input and output for now.
input_data_with_batch = np.random.uniform(0, 1, input_shape_with_batch).astype(np.float32)
# coutput = caffe2.python.onnx.backend.run_model(onnx_model, input_data_with_batch)[0]
pytorch_out = pytorch_model.forward(torch.from_numpy(input_data_with_batch))
zmodel = OnnxLoader(onnx_model.graph).to_keras()
zoutput = zmodel.forward(input_data_with_batch)
self.assert_allclose(pytorch_out.detach().numpy(), zoutput)
