diff --git a/onnx2fluid/examples/convert_data_npz.py b/onnx2fluid/examples/convert_data_npz.py
index 0bf613d14e96c232480212fbe6dda158aec07703..d7c8d3f988fd320b165b0cfba458061bb464fea2 100644
--- a/onnx2fluid/examples/convert_data_npz.py
+++ b/onnx2fluid/examples/convert_data_npz.py
@@ -19,12 +19,12 @@ def make_var_name(name):
 
     assert name
 
-    if name[0].isdigit():
-        return 'var_' + name
-    for s in ' \\|/:-':  #
+    for s in ' \\|/:.-':
         name = name.replace(s, '_')
     if name.startswith('_'):
         name = 'var' + name
+    elif name[0].isdigit():
+        name = 'var_' + name
     return name
 
 
diff --git a/onnx2fluid/examples/convert_data_pb.py b/onnx2fluid/examples/convert_data_pb.py
index f48f72e68a2f1fb87bc93ab7c374195235796416..413b4aa3b1cd4401dab3c60890fee9635410a354 100644
--- a/onnx2fluid/examples/convert_data_pb.py
+++ b/onnx2fluid/examples/convert_data_pb.py
@@ -22,12 +22,12 @@ def make_var_name(name):
 
     assert name
 
-    if name[0].isdigit():
-        return 'var_' + name
-    for s in ' \\|/:-':  #
+    for s in ' \\|/:.-':
         name = name.replace(s, '_')
     if name.startswith('_'):
         name = 'var' + name
+    elif name[0].isdigit():
+        name = 'var_' + name
     return name
 
 
diff --git a/onnx2fluid/onnx2fluid/__main__.py b/onnx2fluid/onnx2fluid/__main__.py
index f09f63e331c83a5e6719f2e6396640cb33dba015..97e3e7fa293c967931078dcbc475937112c71d05 100644
--- a/onnx2fluid/onnx2fluid/__main__.py
+++ b/onnx2fluid/onnx2fluid/__main__.py
@@ -64,7 +64,14 @@ parser.add_argument(
     '-x',
     action='store_false',
     dest='pedantic',
-    help='process non-standard ONNX ops, this may lead to fails',
+    help='process non-standard ONNX ops, this may lead to failures',
+)
+parser.add_argument(
+    '--naive',
+    '-n',
+    action='store_true',
+    default=False,
+    help='bypass ONNX op optimizations, especially for training purpose',
 )
 parser.add_argument(
     '--skip-version-conversion',
diff --git a/onnx2fluid/onnx2fluid/cmdline.py b/onnx2fluid/onnx2fluid/cmdline.py
index 4f6a67404327d2e969259d48536097177e9296e9..e435d8cf2817ffcba13b1a47a568e2af4e9868da 100644
--- a/onnx2fluid/onnx2fluid/cmdline.py
+++ b/onnx2fluid/onnx2fluid/cmdline.py
@@ -18,6 +18,8 @@ from __future__ import unicode_literals
 
 import logging, shutil, zipfile
 
+logger = logging.getLogger(__name__)
+
 __all__ = [
     'main',
 ]
@@ -45,6 +47,7 @@ def main(**kwargs):
     model_basename = DEFAULT_MODEL_MODULE + '.py'
     model_func_name = DEFAULT_MODEL_FUNC
     onnx_opset_pedantic = kwargs.pop('pedantic', True)
+    onnx_skip_optimization = kwargs.pop('naive', False)
     skip_version_conversion = kwargs.pop('skip_version_conversion', False)
     onnx_opset_version = None if skip_version_conversion else DEFAULT_ONNX_OPSET_VERSION
 
@@ -55,6 +58,7 @@ def main(**kwargs):
             model_func_name=model_func_name,
             onnx_opset_version=onnx_opset_version,
             onnx_opset_pedantic=onnx_opset_pedantic,
+            onnx_skip_optimization=onnx_skip_optimization,
             **kwargs)
 
     # validate
@@ -65,7 +69,7 @@ def main(**kwargs):
     if golden_data_filename or save_inference_model:
         from .validation import validate
 
-        if save_inference_model:
+        if infer_inputs:
             inference_input_names = infer_inputs.split(',')
         else:
             inference_input_names = None
diff --git a/onnx2fluid/onnx2fluid/conversion.py b/onnx2fluid/onnx2fluid/conversion.py
index 4113b0684349e382a76c04b4af6a94ffc4ee50a7..8c604deeed2f35f2643ccb4680be2aa27d37c01c 100644
--- a/onnx2fluid/onnx2fluid/conversion.py
+++ b/onnx2fluid/onnx2fluid/conversion.py
@@ -24,12 +24,12 @@ def make_var_name(name):
 
     if name == '':
         return ''
-    if name[0].isdigit():
-        return 'var_' + name
     for s in ' \\|/:.-':
         name = name.replace(s, '_')
     if name.startswith('_'):
         name = 'var' + name
+    elif name[0].isdigit():
+        name = 'var_' + name
     return name
 
 
@@ -40,6 +40,7 @@ def convert(onnx_model_filename,
             embed_params=False,
             onnx_opset_version=None,
             onnx_opset_pedantic=True,
+            onnx_skip_optimization=False,
             debug=False,
             **kwargs):
     """
@@ -61,10 +62,10 @@ def convert(onnx_model_filename,
     from .onnx_utils import DEFAULT_OP_DOMAIN
     from .onnx_utils import graph_ops, graph_weights
     from .onnx_utils import inferred_model_value_info
-    from .onnx_utils import polish_model
+    from .onnx_utils import polish_model, optimize_model_strip_initializer
     from .writer import Program, Writer
 
-    logger = logging.getLogger('convert')
+    logger = logging.getLogger('onnx2fluid')
 
     # prepare onnx model
     logger.info('loading model: %s ...', onnx_model_filename)
@@ -90,8 +91,12 @@ def convert(onnx_model_filename,
 
     # onnx model optimization
     logger.info('model has %d ops', len(onnx_model.graph.node))
-    logger.info('optimizing model ...')
-    onnx_model = polish_model(onnx_model, checking=onnx_opset_pedantic)
+    if onnx_skip_optimization:
+        logger.info('stripping model ...')
+        onnx_model = optimize_model_strip_initializer(onnx_model)
+    else:
+        logger.info('optimizing model ...')
+        onnx_model = polish_model(onnx_model, checking=onnx_opset_pedantic)
 
     # prepare filesystem
     shutil.rmtree(save_dir, ignore_errors=True)
@@ -123,7 +128,7 @@ def convert(onnx_model_filename,
     for name, weight in graph_weights(onnx_graph):
         var_name = make_var_name(name)
         value_info = value_infos[var_name]
-        value_info['lod'] = [0]
+        value_info['lod'] = []
         value_info['embedded_as'] = []
         value_info['get_weight'] = (lambda w: lambda: w.tolist())(
             weight)  # lazy getter
@@ -306,7 +311,14 @@ def main():
         '-x',
         action='store_false',
         dest='pedantic',
-        help='process non-standard ONNX ops, this may lead to fails',
+        help='process non-standard ONNX ops, this may lead to failures',
+    )
+    parser.add_argument(
+        '--naive',
+        '-n',
+        action='store_true',
+        default=False,
+        help='bypass ONNX op optimizations, especially for training purpose',
     )
     parser.add_argument(
         '--skip-version-conversion',
@@ -329,13 +341,15 @@ def main():
                 if save_dir else basepath) + shutil.os.sep
     embed_params = args.embed_params
     pedantic = args.pedantic
-    skip_version_conversion = args.skip_version_conversion
+    skip_optimization = args.naive
+    onnx_opset_version = None if args.skip_version_conversion else DEFAULT_ONNX_OPSET_VERSION
 
     convert(model_filename,
             save_dir,
             embed_params=embed_params,
+            onnx_opset_version=onnx_opset_version,
             onnx_opset_pedantic=pedantic,
-            onnx_skip_version_conversion=skip_version_conversion,
+            onnx_skip_optimization=skip_optimization,
             debug=debug)
 
 
diff --git a/onnx2fluid/onnx2fluid/onnx_utils.py b/onnx2fluid/onnx2fluid/onnx_utils.py
index c0c3c66b7f0bd940decf864b1447dbd7bda0a213..759c9bd8323c13ed2481d1680c101690109f9399 100644
--- a/onnx2fluid/onnx2fluid/onnx_utils.py
+++ b/onnx2fluid/onnx2fluid/onnx_utils.py
@@ -356,16 +356,16 @@ def polish_model(model, internals=True, extras=True, checking=True):
 
 
 def polish_and_save(model_filename,
+                    save_filename='',
                     suffix='.polished',
-                    save_filename=None,
                     *args,
                     **kwargs):
     """
     run polish_model and save
     """
 
-    if save_filename is None:
-        save_filename = model_filename.replace('.onnx', suffix + '.onnx')
+    save_filename = save_filename or model_filename.replace(
+        '.onnx', suffix + '.onnx')
 
     model = onnx.load(model_filename)
     model = polish_model(model, *args, **kwargs)
diff --git a/onnx2fluid/onnx2fluid/symbolic.py b/onnx2fluid/onnx2fluid/symbolic.py
index e781d7d415fdec441aeb5580daaa9c0739fb992d..1279e4175d1c6f97b1c824b8e4a1739f8affb3ae 100644
--- a/onnx2fluid/onnx2fluid/symbolic.py
+++ b/onnx2fluid/onnx2fluid/symbolic.py
@@ -18,7 +18,8 @@ import numpy as _np
 from collections import OrderedDict as _dict
 from onnx.mapping import TENSOR_TYPE_TO_NP_TYPE
 
-_logger = _logging.getLogger(__name__)
+# _logger = _logging.getLogger(__name__)
+_logger = _logging.getLogger('onnx2fluid')
 
 ONNX_INT_MAX = 2**63 - 1
 FLUID_INT_MAX = 2**31 - 1  #
@@ -58,8 +59,8 @@ DEFAULT_OP_MAPPING = {
         'Ceil': ['ceil', ['X'], ['Out']],
         'Clip':
             ['clip', ['X'], ['Out'], dict(), dict(
-                    min=(_np.array([255, 255, 127, 255], dtype=_np.uint8).view(_np.float32)),
-                    max=(_np.array([255, 255, 127, 127], dtype=_np.uint8).view(_np.float32)),
+                    min=(_np.asarray([255, 255, 127, 255], dtype=_np.uint8).view(_np.float32)),
+                    max=(_np.asarray([255, 255, 127, 127], dtype=_np.uint8).view(_np.float32)),
                 )],
         'Cos': ['cos', ['X'], ['Out']],
         'Elu': ['elu', ['X'], ['Out'], dict(), dict(alpha=1.)],
@@ -449,7 +450,7 @@ def _pool(prog, pool_type, inputs, outputs, attrs, value_infos, name):
     # I/O
     var_x, = inputs
     var_y, var_indices, = (outputs + [''] * 1)[:2]
-    assert name and var_x and var_y
+    assert name and all(inputs) and var_y
 
     # interpretation
     assert attrs.get(
@@ -512,7 +513,7 @@ def _roi_pool(prog, fluid_op, inputs, outputs, attrs, name):
     # I/O
     var_x, var_rois, = inputs
     var_y, = outputs
-    assert name and var_x and var_rois and var_y
+    assert name and all(inputs) and all(outputs)
 
     # interpretation
     spatial_scale = attrs['spatial_scale']  # required
@@ -565,7 +566,7 @@ def _interpolate(prog, inputs, outputs, attrs, value_infos, name=''):
     # I/O
     var_x, var_scales, = inputs
     var_y, = outputs
-    assert var_x and var_scales and var_y
+    assert all(inputs) and all(outputs)
 
     # interpretation
     # output shape
@@ -701,7 +702,7 @@ def BatchNormalization(prog,
     var_x, var_scale, var_b, var_mean, var_var, = inputs
     var_y, var_mean_, var_var_, var_saved_mean, var_saved_variance, = (
         outputs + [''] * 4)[:5]
-    assert var_x and var_scale and var_b and var_mean and var_var and var_y
+    assert all(inputs) and var_y
     assert var_saved_mean or name
     assert var_saved_variance or name
     var_saved_mean = var_saved_mean or (name + '.saved_mean')  # dummy output
@@ -879,7 +880,8 @@ def Constant(prog, inputs, outputs, attrs, value_infos, *args, **kwargs):
         _logger.warning(
             'in op (Constant -> %s): '
             'attribute "shape" of %s not inferred, '
-            'using value as 1-D tensor may lead to fails', outputs, var_output)
+            'using value as 1-D tensor may lead to failures', outputs,
+            var_output)
 
     # generation
     if not shape or value.size == 1:  # scalar or 1-size
@@ -929,7 +931,7 @@ def ConstantOfShape(prog, inputs, outputs, attrs, value_infos, *args, **kwargs):
         'given shape is neither const value nor deductible from output, '
         'this is not supported')
     attrs = attrs.copy()
-    attrs.setdefault('value', _np.array(0, dtype=_np.float32))
+    attrs.setdefault('value', _np.asarray(0, dtype=_np.float32))
     attrs.update({'shape': shape})  # pass const
 
     prog.Code('# shape: {} = {} # const as literal'.format(var_shape, shape))
@@ -959,7 +961,7 @@ def Conv(prog,
     # I/O
     var_x, var_w, var_b, = (inputs + [''] * 1)[:3]
     var_y, = outputs
-    assert name and var_x and var_w and var_y
+    assert name and var_x and var_w and all(outputs)
 
     # interpretation
     assert attrs.get(
@@ -1066,7 +1068,7 @@ def ConvTranspose(prog,
     # I/O
     var_x, var_w, var_b, = (inputs + [''] * 1)[:3]
     var_y, = outputs
-    assert name and var_x and var_w and var_y
+    assert name and var_x and var_w and all(outputs)
 
     # interpretation
     assert attrs.get(
@@ -1174,7 +1176,7 @@ def Gemm(prog, inputs, outputs, attrs, value_infos, name, *args, **kwargs):
     # due to fluid fc don't support transposed weight, we use matmul + ew_add
     var_a, var_b, var_c, = inputs
     var_y, = outputs
-    assert name and var_a and var_b and var_c and var_y
+    assert name and all(inputs) and all(outputs)
 
     alpha = attrs.get('alpha', 1.)  # optional
     beta = attrs.get('beta', 1.)  # optional
@@ -1794,7 +1796,7 @@ def Pad(prog, inputs, outputs, attrs, value_infos, name='', *args, **kwargs):
             mode)
         fluid_op = 'pad'
         pad2d_attr = ''
-    paddings = _np.array(pads).reshape(
+    paddings = _np.asarray(pads).reshape(
         (-1, 2)).transpose().flatten().tolist()  # SSEE -> SESE
     od_attrs['paddings'] = paddings
     name_attr = ', name={}'.format(repr(name)) if name else ''
@@ -1838,7 +1840,7 @@ def PRelu(prog,
     # I/O
     var_x, var_slope, = inputs
     var_y, = outputs
-    assert name and var_x and var_slope and var_y
+    assert name and all(inputs) and all(outputs)
 
     # interpretation
     mode = 'channel'
@@ -1904,7 +1906,7 @@ def Reshape(prog, inputs, outputs, attrs_, value_infos, name, *args, **kwargs):
     # I/O
     var_data, var_shape, = inputs
     var_reshaped, = outputs
-    assert name and var_data and var_shape and var_reshaped
+    assert name and all(inputs) and all(outputs)
 
     # interpretation
     shape = _const_weight_or_none(value_infos, var_shape)
@@ -2015,7 +2017,7 @@ def Shape(prog, inputs, outputs, attrs_, name, **kwargs):
     # I/O
     var_data, = inputs
     var_shape, = outputs
-    assert name and var_data and var_shape
+    assert name and all(inputs) and all(outputs)
 
     # interpretation
     fluid_op = 'shape'
@@ -2189,7 +2191,7 @@ def Tile(prog, inputs, outputs, attrs_, value_infos, name='', *args, **kwargs):
     # I/O
     var_input, var_repeats, = inputs
     var_output, = outputs
-    assert var_input and var_repeats and var_output
+    assert all(inputs) and all(outputs)
 
     # interpretation
     repeats = _const_weight_or_none(value_infos, var_repeats)
@@ -2227,7 +2229,7 @@ def Transpose(prog, inputs, outputs, attrs, name, *args, **kwargs):
     # I/O
     var_data, = inputs
     var_transposed, = outputs
-    assert name and var_data and var_transposed
+    assert name and all(inputs) and all(outputs)
 
     # interpretation
     fluid_op = 'transpose'
diff --git a/onnx2fluid/onnx2fluid/torch_export_helper.py b/onnx2fluid/onnx2fluid/torch_export_helper.py
index ef13cea9f642459dcf98992108fe9e69cd87cc71..8c49d4cae1ac2db740ea6e8548ecf88453547176 100644
--- a/onnx2fluid/onnx2fluid/torch_export_helper.py
+++ b/onnx2fluid/onnx2fluid/torch_export_helper.py
@@ -138,10 +138,10 @@ def export_onnx_with_validation(
     outputs = torch.onnx.export(model,
                                 torch_inputs,
                                 export_basepath + '.onnx',
-                                input_names=(None if input_names is None else
-                                             flatten_list(input_names)),
-                                output_names=(None if output_names is None else
-                                              flatten_list(output_names)),
+                                input_names=(input_names
+                                             and flatten_list(input_names)),
+                                output_names=(output_names
+                                              and flatten_list(output_names)),
                                 *args,
                                 **kwargs)
     if outputs is None:  # WORKAROUND: for torch.onnx
diff --git a/onnx2fluid/onnx2fluid/validation.py b/onnx2fluid/onnx2fluid/validation.py
index efd5609f1875b3efdf55fd519b62ac2dd939ce9c..6d656ba290f06ab3633577e440d3f6cab52ae185 100644
--- a/onnx2fluid/onnx2fluid/validation.py
+++ b/onnx2fluid/onnx2fluid/validation.py
@@ -90,7 +90,7 @@ def validate(fluid_model_filename,
     import numpy as np
     import paddle.fluid as fluid
 
-    logger = logging.getLogger('validate')
+    logger = logging.getLogger('onnx2fluid')
 
     place = fluid.CPUPlace()
     exe = fluid.Executor(place)
@@ -126,6 +126,7 @@ def validate(fluid_model_filename,
         logger.info('import passed')
 
         prog = fluid.default_main_program()
+        prog = prog.clone(for_test=True)  # force inference mode
         fluid.io.load_persistables(executor=exe,
                                    dirname=fluid_model_dir,
                                    main_program=prog)
@@ -160,8 +161,7 @@ def validate(fluid_model_filename,
         logger.info('with %d inputs and %d outputs', len(input_data),
                     len(output_data))
     elif save_inference_model:
-        assert inference_input_names is not None, (
-            'input names required for type-shape inference')
+        assert inference_input_names, 'input names required for type-shape inference'
 
         input_names = inference_input_names
         logger.info('using input names: %s', ', '.join(input_names))
@@ -185,6 +185,7 @@ def validate(fluid_model_filename,
     # execute
     outputs = exe.run(prog, feed=input_data,
                       fetch_list=out_names)  # out_names can be vars
+    exe.close()
     logger.info('execution passed')
 
     # validate
@@ -264,7 +265,7 @@ def main():
     atol, rtol = args.atol, args.rtol
     save_inference_model = args.infer_inputs is not None
     inference_input_names = args.infer_inputs.split(
-        ',') if args.infer_inputs else None
+        ',') if save_inference_model else None
 
     validate(fluid_model_filename,
              golden_data_filename=golden_data_filename,
diff --git a/onnx2fluid/onnx2fluid/writer.py b/onnx2fluid/onnx2fluid/writer.py
index 32b91b515a571dc50a3702ab76adff2bb52ea16e..7f958d83d0c829b0c38da72d05bb236903d302a6 100644
--- a/onnx2fluid/onnx2fluid/writer.py
+++ b/onnx2fluid/onnx2fluid/writer.py
@@ -372,7 +372,7 @@ class Writer(object):
             prog.Code('# input {}'.format(name))
             prog.Code((
                 '{} = layers.data(name={}, shape={}, dtype={}, '
-                'append_batch_size={})'  # , stop_gradient=True
+                'append_batch_size={}, lod_level=1)'  # , stop_gradient=True
             ).format(
                 name,
                 repr(name),
@@ -427,20 +427,28 @@ class Writer(object):
         assert lod is None or isinstance(lod,
                                          list), 'lod should be None or list'
 
-        if lod is None:
-            lod = [0]
+        lod = lod or []
 
         tensor_desc = framework_pb2.VarType.TensorDesc()
         tensor_desc.data_type = Program.Dtype(weight.dtype)
         tensor_desc.dims.extend(weight.shape)
 
         fp = open(filename, 'wb')
-        np.array([0], dtype=np.int32).tofile(fp)  # version
-        np.array(lod, dtype=np.int64).tofile(fp)  # LOD level
-        np.array([0], dtype=np.int32).tofile(fp)  # tensor version
-        np.array([tensor_desc.ByteSize()], dtype=np.int32).tofile(fp)
+
+        # lod_tensor.cc: SerializeToStream
+        np.asarray([0], dtype=np.uint32).tofile(fp)  # version
+        np.asarray([len(lod)], dtype=np.int64).tofile(fp)  # LOD levels
+        for level in lod:
+            np.asarray([len(level)], dtype=np.int64).tofile(fp)  # level size
+            np.asarray(level, dtype=np.uint64).tofile(fp)  # LOD: size_t
+
+        # tensor_util.cc: TensorToStream
+        np.asarray([0], dtype=np.uint32).tofile(fp)  # tensor version
+        np.asarray([tensor_desc.ByteSize()], dtype=np.int32).tofile(fp)
         fp.write(tensor_desc.SerializeToString())
+
         weight.tofile(fp)
+
         fp.close()
 
     @staticmethod