I wanted to graphically visualise a very complex network using Netron. I can’t export the architecture to onnx format. This is my architecture.
>>>fc_arch
DGMArch(
(fc_start): FCLayer(
(linear): WeightNormLinear(in_features=2, out_features=1, bias=True)
)
(dgm_layers): ModuleList(
(0): ModuleDict(
(z): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(g): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(r): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(h): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
)
(1): ModuleDict(
(z): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(g): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(r): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(h): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
)
(2): ModuleDict(
(z): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(g): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(r): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(h): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
)
(3): ModuleDict(
(z): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(g): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(r): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(h): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
)
(4): ModuleDict(
(z): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(g): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(r): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(h): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
)
)
(fc_end): FCLayer(
(linear): Linear(in_features=1, out_features=1, bias=True)
)
)
The architecture contains 3 components.
>>> fc_arch._modules.keys()
odict_keys(['fc_start', 'dgm_layers', 'fc_end'])
I can easily export fc_start.
>>> fc_arch.fc_start
FCLayer(
(linear): WeightNormLinear(in_features=2, out_features=1, bias=True)
)
using this command.
torch.onnx.export(fc_arch._modules['fc_start'], dummy_input, 'fc_start.onnx', input_names=['X','Y'], output_names=['u'])
but the problem starts when I export dgm_layer. This is the DGM layer.
>>> fc_arch.dgm_layers
ModuleList(
(0): ModuleDict(
(z): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(g): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(r): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(h): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
)
(1): ModuleDict(
(z): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(g): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(r): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(h): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
)
(2): ModuleDict(
(z): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(g): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(r): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(h): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
)
(3): ModuleDict(
(z): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(g): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(r): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(h): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
)
(4): ModuleDict(
(z): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(g): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(r): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
(h): DGMLayer(
(linear_1): WeightNormLinear(in_features=2, out_features=1, bias=False)
(linear_2): WeightNormLinear(in_features=1, out_features=1, bias=False)
)
)
)
Here the dgm_layer is
>>> type(fc_arch.dgm_layers)
torch.nn.modules.container.ModuleList
When I try to export, it gives me the following error:
>>> torch.onnx.export(fc_arch.dgm_layers, dummy_input, 'dgm_layers.onnx', input_names=['X','Y'], output_names=['u'])
---------------------------------------------------------------------------
NotImplementedError Traceback (most recent call last)
/tmp/ipykernel_47794/1172123868.py in <cell line: 1>()
----> 1 torch.onnx.export(fc_arch.dgm_layers, dummy_input, 'dgm_layers.onnx', input_names=['X','Y'], output_names=['u'])
/scratch/s.1915438/env/modulus/lib/python3.9/site-packages/torch/onnx/__init__.py in export(model, args, f, export_params, verbose, training, input_names, output_names, operator_export_type, opset_version, do_constant_folding, dynamic_axes, keep_initializers_as_inputs, custom_opsets, export_modules_as_functions)
303
304 from torch.onnx import utils
--> 305 return utils.export(model, args, f, export_params, verbose, training,
306 input_names, output_names, operator_export_type, opset_version,
307 do_constant_folding, dynamic_axes,
/scratch/s.1915438/env/modulus/lib/python3.9/site-packages/torch/onnx/utils.py in export(model, args, f, export_params, verbose, training, input_names, output_names, operator_export_type, opset_version, do_constant_folding, dynamic_axes, keep_initializers_as_inputs, custom_opsets, export_modules_as_functions)
116 operator_export_type = OperatorExportTypes.ONNX
117
--> 118 _export(model, args, f, export_params, verbose, training, input_names, output_names,
119 operator_export_type=operator_export_type, opset_version=opset_version,
120 do_constant_folding=do_constant_folding, dynamic_axes=dynamic_axes,
/scratch/s.1915438/env/modulus/lib/python3.9/site-packages/torch/onnx/utils.py in _export(model, args, f, export_params, verbose, training, input_names, output_names, operator_export_type, export_type, opset_version, do_constant_folding, dynamic_axes, keep_initializers_as_inputs, fixed_batch_size, custom_opsets, add_node_names, onnx_shape_inference, export_modules_as_functions)
717
718 graph, params_dict, torch_out = \
--> 719 _model_to_graph(model, args, verbose, input_names,
720 output_names, operator_export_type,
721 val_do_constant_folding,
/scratch/s.1915438/env/modulus/lib/python3.9/site-packages/torch/onnx/utils.py in _model_to_graph(model, args, verbose, input_names, output_names, operator_export_type, do_constant_folding, _disable_torch_constant_prop, fixed_batch_size, training, dynamic_axes)
497 args = (args, )
498
--> 499 graph, params, torch_out, module = _create_jit_graph(model, args)
500
501 params_dict = _get_named_param_dict(graph, params)
/scratch/s.1915438/env/modulus/lib/python3.9/site-packages/torch/onnx/utils.py in _create_jit_graph(model, args)
438 return graph, params, torch_out, None
439 else:
--> 440 graph, torch_out = _trace_and_get_graph_from_model(model, args)
441 torch._C._jit_pass_onnx_lint(graph)
442 state_dict = _unique_state_dict(model)
/scratch/s.1915438/env/modulus/lib/python3.9/site-packages/torch/onnx/utils.py in _trace_and_get_graph_from_model(model, args)
389
390 trace_graph, torch_out, inputs_states = \
--> 391 torch.jit._get_trace_graph(model, args, strict=False, _force_outplace=False, _return_inputs_states=True)
392 warn_on_static_input_change(inputs_states)
393
/scratch/s.1915438/env/modulus/lib/python3.9/site-packages/torch/jit/_trace.py in _get_trace_graph(f, args, kwargs, strict, _force_outplace, return_inputs, _return_inputs_states)
1164 if not isinstance(args, tuple):
1165 args = (args,)
-> 1166 outs = ONNXTracedModule(f, strict, _force_outplace, return_inputs, _return_inputs_states)(*args, **kwargs)
1167 return outs
/scratch/s.1915438/env/modulus/lib/python3.9/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1108 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1109 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1110 return forward_call(*input, **kwargs)
1111 # Do not call functions when jit is used
1112 full_backward_hooks, non_full_backward_hooks = [], []
/scratch/s.1915438/env/modulus/lib/python3.9/site-packages/torch/jit/_trace.py in forward(self, *args)
125 return tuple(out_vars)
126
--> 127 graph, out = torch._C._create_graph_by_tracing(
128 wrapper,
129 in_vars + module_state,
/scratch/s.1915438/env/modulus/lib/python3.9/site-packages/torch/jit/_trace.py in wrapper(*args)
116 if self._return_inputs_states:
117 inputs_states.append(_unflatten(in_args, in_desc))
--> 118 outs.append(self.inner(*trace_inputs))
119 if self._return_inputs_states:
120 inputs_states[0] = (inputs_states[0], trace_inputs)
/scratch/s.1915438/env/modulus/lib/python3.9/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1108 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1109 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1110 return forward_call(*input, **kwargs)
1111 # Do not call functions when jit is used
1112 full_backward_hooks, non_full_backward_hooks = [], []
/scratch/s.1915438/env/modulus/lib/python3.9/site-packages/torch/nn/modules/module.py in _slow_forward(self, *input, **kwargs)
1096 recording_scopes = False
1097 try:
-> 1098 result = self.forward(*input, **kwargs)
1099 finally:
1100 if recording_scopes:
/scratch/s.1915438/env/modulus/lib/python3.9/site-packages/torch/nn/modules/module.py in _forward_unimplemented(self, *input)
199 registered hooks while the latter silently ignores them.
200 """
--> 201 raise NotImplementedError
202
203
NotImplementedError:
torch.onnx.export(fc_arch._modules['fc_0'], dummy_input, '0.onnx', input_names=['X','Y'], output_names=['u'])
torch.onnx.export(fc_arch.fc_layers._modules, dummy_input, 'fc_layers.onnx', input_names=['X','Y'], output_names=['u']) # just fully connected layer
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
/tmp/ipykernel_42623/2301621278.py in <cell line: 1>()
----> 1 torch.onnx.export(fc_arch.fc_layers._modules, dummy_input, 'fc_layers.onnx', input_names=['X','Y'], output_names=['u']) # just fully connected layer
/scratch/s.1915438/env/modulus/lib/python3.9/site-packages/torch/onnx/__init__.py in export(model, args, f, export_params, verbose, training, input_names, output_names, operator_export_type, opset_version, do_constant_folding, dynamic_axes, keep_initializers_as_inputs, custom_opsets, export_modules_as_functions)
303
304 from torch.onnx import utils
--> 305 return utils.export(model, args, f, export_params, verbose, training,
306 input_names, output_names, operator_export_type, opset_version,
307 do_constant_folding, dynamic_axes,
/scratch/s.1915438/env/modulus/lib/python3.9/site-packages/torch/onnx/utils.py in export(model, args, f, export_params, verbose, training, input_names, output_names, operator_export_type, opset_version, do_constant_folding, dynamic_axes, keep_initializers_as_inputs, custom_opsets, export_modules_as_functions)
116 operator_export_type = OperatorExportTypes.ONNX
117
--> 118 _export(model, args, f, export_params, verbose, training, input_names, output_names,
119 operator_export_type=operator_export_type, opset_version=opset_version,
120 do_constant_folding=do_constant_folding, dynamic_axes=dynamic_axes,
/scratch/s.1915438/env/modulus/lib/python3.9/site-packages/torch/onnx/utils.py in _export(model, args, f, export_params, verbose, training, input_names, output_names, operator_export_type, export_type, opset_version, do_constant_folding, dynamic_axes, keep_initializers_as_inputs, fixed_batch_size, custom_opsets, add_node_names, onnx_shape_inference, export_modules_as_functions)
699 _set_opset_version(opset_version)
700 _set_operator_export_type(operator_export_type)
--> 701 with exporter_context(model, training):
702 val_keep_init_as_ip = _decide_keep_init_as_input(keep_initializers_as_inputs,
703 operator_export_type,
/scratch/s.1915438/modulus/lib/python3.9/contextlib.py in __enter__(self)
117 del self.args, self.kwds, self.func
118 try:
--> 119 return next(self.gen)
120 except StopIteration:
121 raise RuntimeError("generator didn't yield") from None
/scratch/s.1915438/env/modulus/lib/python3.9/site-packages/torch/onnx/utils.py in exporter_context(model, mode)
100 @contextlib.contextmanager
101 def exporter_context(model, mode):
--> 102 with select_model_mode_for_export(model, mode) as mode_ctx, \
103 disable_apex_o2_state_dict_hook(model) as apex_ctx:
104 yield (mode_ctx, apex_ctx)
/scratch/s.1915438/modulus/lib/python3.9/contextlib.py in __enter__(self)
117 del self.args, self.kwds, self.func
118 try:
--> 119 return next(self.gen)
120 except StopIteration:
121 raise RuntimeError("generator didn't yield") from None
/scratch/s.1915438/env/modulus/lib/python3.9/site-packages/torch/onnx/utils.py in select_model_mode_for_export(model, mode)
39 def select_model_mode_for_export(model, mode):
40 if not isinstance(model, torch.jit.ScriptFunction):
---> 41 is_originally_training = model.training
42
43 if mode is None:
AttributeError: 'collections.OrderedDict' object has no attribute 'training'