GLiNER
GLiNER copied to clipboard
urchade
Conversion to ONXX
Using this snippet to save pre-trained model locally -
from gliner import GLiNER
model = GLiNER.from_pretrained("urchade/gliner_medium")
model.save_pretrained("gliner_medium")
post this, running script for onxx conversion as per documentation -
python GLiNER/convert_to_onnx.py --model_path gliner_medium/ --save_path gliner_onnx/ --quantize True
getting error on -
OSError: gliner_medium does not appear to have a file named config.json. Checkout 'https://huggingface.co/gliner_medium/tree/None' for available files.
Tried with changing file name from 'gliner_config.json' to 'config.json'
getting new error of -
ValueError: Unrecognized model in gliner_medium. Should have a `model_type` key in its config.json
try the same onnx conversion in notebook version once, https://github.com/urchade/GLiNER/blob/main/examples/convert_to_onnx.ipynb
and try to use earlier gliner version, latest gliner version has some bugs in onnx conversion
thanks @hari-ag00
Also, can you suggest the version for this? Currently, i am using '0.2.2'
Hi @hari-ag00, is the issue with conversion fixed with the latest gliner ?
I am still having issues with attempts to convert to ONNX using the latest v0.2.16.
import os
import onnx
from gliner import GLiNER
chunk_size = 384
gliner_model = GLiNER.from_pretrained("gliner-community/gliner_large-v2.5", max_length=chunk_size)
gliner_model.save_pretrained("gliner_large-v2.5")
onnx_save_path = os.path.join("/dirpath", "gliner-large-v25.onnx")
text = "ONNX is an open-source format designed to enable the interoperability of AI models across various frameworks and tools."
labels = ['format', 'model', 'tool', 'cat']
inputs, _ = gliner_model.prepare_model_inputs([text], labels)
if gliner_model.config.span_mode == 'token_level':
all_inputs = (inputs['input_ids'], inputs['attention_mask'],
inputs['words_mask'], inputs['text_lengths'])
input_names = ['input_ids', 'attention_mask', 'words_mask', 'text_lengths']
dynamic_axes={
"input_ids": {0: "batch_size", 1: "sequence_length"},
"attention_mask": {0: "batch_size", 1: "sequence_length"},
"words_mask": {0: "batch_size", 1: "sequence_length"},
"text_lengths": {0: "batch_size", 1: "value"},
"logits": {0: "position", 1: "batch_size", 2: "sequence_length", 3: "num_classes"},
}
else:
all_inputs = (inputs['input_ids'], inputs['attention_mask'],
inputs['words_mask'], inputs['text_lengths'],
inputs['span_idx'], inputs['span_mask'])
input_names = ['input_ids', 'attention_mask', 'words_mask', 'text_lengths', 'span_idx', 'span_mask']
dynamic_axes={
"input_ids": {0: "batch_size", 1: "sequence_length"},
"attention_mask": {0: "batch_size", 1: "sequence_length"},
"words_mask": {0: "batch_size", 1: "sequence_length"},
"text_lengths": {0: "batch_size", 1: "value"},
"span_idx": {0: "batch_size", 1: "num_spans", 2: "idx"},
"span_mask": {0: "batch_size", 1: "num_spans"},
"logits": {0: "batch_size", 1: "sequence_length", 2: "num_spans", 3: "num_classes"},
}
print('Converting the model...')
torch.onnx.export(
gliner_model.model,
all_inputs,
f=onnx_save_path,
input_names=input_names,
output_names=["logits"],
dynamic_axes=dynamic_axes,
opset_version=14,
)
Returns
Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation.
Converting the model...
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
Cell In[2], line 34
24 dynamic_axes={
25 "input_ids": {0: "batch_size", 1: "sequence_length"},
26 "attention_mask": {0: "batch_size", 1: "sequence_length"},
(...)
31 "logits": {0: "batch_size", 1: "sequence_length", 2: "num_spans", 3: "num_classes"},
32 }
33 print('Converting the model...')
---> 34 torch.onnx.export(
35 gliner_model.model,
36 all_inputs,
37 f=onnx_save_path,
38 input_names=input_names,
39 output_names=["logits"],
40 dynamic_axes=dynamic_axes,
41 opset_version=14,
42 )
File .../python3.11/site-packages/torch/onnx/__init__.py:375, in export(model, args, f, kwargs, export_params, verbose, input_names, output_names, opset_version, dynamic_axes, keep_initializers_as_inputs, dynamo, external_data, dynamic_shapes, report, verify, profile, dump_exported_program, artifacts_dir, fallback, training, operator_export_type, do_constant_folding, custom_opsets, export_modules_as_functions, autograd_inlining, **_)
369 if dynamic_shapes:
370 raise ValueError(
371 "The exporter only supports dynamic shapes "
372 "through parameter dynamic_axes when dynamo=False."
373 )
--> 375 export(
376 model,
377 args,
378 f, # type: ignore[arg-type]
379 kwargs=kwargs,
380 export_params=export_params,
381 verbose=verbose is True,
382 input_names=input_names,
383 output_names=output_names,
384 opset_version=opset_version,
385 dynamic_axes=dynamic_axes,
386 keep_initializers_as_inputs=keep_initializers_as_inputs,
387 training=training,
388 operator_export_type=operator_export_type,
389 do_constant_folding=do_constant_folding,
390 custom_opsets=custom_opsets,
391 export_modules_as_functions=export_modules_as_functions,
392 autograd_inlining=autograd_inlining,
393 )
394 return None
File .../python3.11/site-packages/torch/onnx/utils.py:502, in export(model, args, f, kwargs, 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, autograd_inlining)
499 if kwargs is not None:
500 args = args + (kwargs,)
--> 502 _export(
503 model,
504 args,
505 f,
506 export_params,
507 verbose,
508 training,
509 input_names,
510 output_names,
511 operator_export_type=operator_export_type,
512 opset_version=opset_version,
513 do_constant_folding=do_constant_folding,
514 dynamic_axes=dynamic_axes,
515 keep_initializers_as_inputs=keep_initializers_as_inputs,
516 custom_opsets=custom_opsets,
517 export_modules_as_functions=export_modules_as_functions,
518 autograd_inlining=autograd_inlining,
519 )
521 return None
File .../lib/python3.11/site-packages/torch/onnx/utils.py:1564, 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, autograd_inlining)
1561 dynamic_axes = {}
1562 _validate_dynamic_axes(dynamic_axes, model, input_names, output_names)
-> 1564 graph, params_dict, torch_out = _model_to_graph(
1565 model,
1566 args,
1567 verbose,
1568 input_names,
1569 output_names,
1570 operator_export_type,
1571 val_do_constant_folding,
1572 fixed_batch_size=fixed_batch_size,
1573 training=training,
1574 dynamic_axes=dynamic_axes,
1575 )
1577 # TODO: Don't allocate a in-memory string for the protobuf
1578 defer_weight_export = (
1579 export_type is not _exporter_states.ExportTypes.PROTOBUF_FILE
1580 )
File .../lib/python3.11/site-packages/torch/onnx/utils.py:1113, 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)
1110 args = (args,)
1112 model = _pre_trace_quant_model(model, args)
-> 1113 graph, params, torch_out, module = _create_jit_graph(model, args)
1114 params_dict = _get_named_param_dict(graph, params)
1116 try:
File .../lib/python3.11/site-packages/torch/onnx/utils.py:997, in _create_jit_graph(model, args)
992 graph = _C._propagate_and_assign_input_shapes(
993 graph, flattened_args, param_count_list, False, False
994 )
995 return graph, params, torch_out, None
--> 997 graph, torch_out = _trace_and_get_graph_from_model(model, args)
998 _C._jit_pass_onnx_lint(graph)
999 state_dict = torch.jit._unique_state_dict(model)
File .../lib/python3.11/site-packages/torch/onnx/utils.py:904, in _trace_and_get_graph_from_model(model, args)
902 prev_autocast_cache_enabled = torch.is_autocast_cache_enabled()
903 torch.set_autocast_cache_enabled(False)
--> 904 trace_graph, torch_out, inputs_states = torch.jit._get_trace_graph(
905 model,
906 args,
907 strict=False,
908 _force_outplace=False,
909 _return_inputs_states=True,
910 )
911 torch.set_autocast_cache_enabled(prev_autocast_cache_enabled)
913 warn_on_static_input_change(inputs_states)
File .../lib/python3.11/site-packages/torch/jit/_trace.py:1500, in _get_trace_graph(f, args, kwargs, strict, _force_outplace, return_inputs, _return_inputs_states)
1498 if not isinstance(args, tuple):
1499 args = (args,)
-> 1500 outs = ONNXTracedModule(
1501 f, strict, _force_outplace, return_inputs, _return_inputs_states
1502 )(*args, **kwargs)
1503 return outs
File .../lib/python3.11/site-packages/torch/nn/modules/module.py:1736, in Module._wrapped_call_impl(self, *args, **kwargs)
1734 return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc]
1735 else:
-> 1736 return self._call_impl(*args, **kwargs)
File .../lib/python3.11/site-packages/torch/nn/modules/module.py:1747, in Module._call_impl(self, *args, **kwargs)
1742 # If we don't have any hooks, we want to skip the rest of the logic in
1743 # this function, and just call forward.
1744 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
1745 or _global_backward_pre_hooks or _global_backward_hooks
1746 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1747 return forward_call(*args, **kwargs)
1749 result = None
1750 called_always_called_hooks = set()
File .../lib/python3.11/site-packages/torch/jit/_trace.py:139, in ONNXTracedModule.forward(self, *args)
136 else:
137 return tuple(out_vars)
--> 139 graph, out = torch._C._create_graph_by_tracing(
140 wrapper,
141 in_vars + module_state,
142 _create_interpreter_name_lookup_fn(),
143 self.strict,
144 self._force_outplace,
145 )
147 if self._return_inputs:
148 return graph, outs[0], ret_inputs[0]
File .../lib/python3.11/site-packages/torch/jit/_trace.py:130, in ONNXTracedModule.forward.<locals>.wrapper(*args)
128 if self._return_inputs_states:
129 inputs_states.append(_unflatten(in_args, in_desc))
--> 130 outs.append(self.inner(*trace_inputs))
131 if self._return_inputs_states:
132 inputs_states[0] = (inputs_states[0], trace_inputs)
File .../lib/python3.11/site-packages/torch/nn/modules/module.py:1736, in Module._wrapped_call_impl(self, *args, **kwargs)
1734 return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc]
1735 else:
-> 1736 return self._call_impl(*args, **kwargs)
File .../nlp/lib/python3.11/site-packages/torch/nn/modules/module.py:1747, in Module._call_impl(self, *args, **kwargs)
1742 # If we don't have any hooks, we want to skip the rest of the logic in
1743 # this function, and just call forward.
1744 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
1745 or _global_backward_pre_hooks or _global_backward_hooks
1746 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1747 return forward_call(*args, **kwargs)
1749 result = None
1750 called_always_called_hooks = set()
File .../python3.11/site-packages/torch/nn/modules/module.py:1726, in Module._slow_forward(self, *input, **kwargs)
1724 recording_scopes = False
1725 try:
-> 1726 result = self.forward(*input, **kwargs)
1727 finally:
1728 if recording_scopes:
File .../lib/python3.11/site-packages/gliner/modeling/base.py:233, in SpanModel.forward(self, input_ids, attention_mask, labels_embeddings, labels_input_ids, labels_attention_mask, words_embedding, mask, prompts_embedding, prompts_embedding_mask, words_mask, text_lengths, span_idx, span_mask, labels, **kwargs)
215 def forward(self,
216 input_ids: Optional[torch.FloatTensor] = None,
217 attention_mask: Optional[torch.LongTensor] = None,
(...)
230 **kwargs
231 ):
--> 233 prompts_embedding, prompts_embedding_mask, words_embedding, mask = self.get_representations(input_ids, attention_mask,
234 labels_embeddings, labels_input_ids, labels_attention_mask,
235 text_lengths, words_mask)
236 span_idx = span_idx*span_mask.unsqueeze(-1)
238 span_rep = self.span_rep_layer(words_embedding, span_idx)
File .../lib/python3.11/site-packages/gliner/modeling/base.py:182, in BaseModel.get_representations(self, input_ids, attention_mask, labels_embeddings, labels_input_ids, labels_attention_mask, text_lengths, words_mask, **kwargs)
177 prompts_embedding, prompts_embedding_mask, words_embedding, mask = self.get_bi_representations(
178 input_ids, attention_mask, labels_embeddings, labels_input_ids, labels_attention_mask,
179 text_lengths, words_mask, **kwargs
180 )
181 else:
--> 182 prompts_embedding, prompts_embedding_mask, words_embedding, mask = self.get_uni_representations(
183 input_ids, attention_mask, text_lengths, words_mask, **kwargs
184 )
185 return prompts_embedding, prompts_embedding_mask, words_embedding, mask
File .../lib/python3.11/site-packages/gliner/modeling/base.py:128, in BaseModel.get_uni_representations(self, input_ids, attention_mask, text_lengths, words_mask, **kwargs)
119 def get_uni_representations(self,
120 input_ids: Optional[torch.FloatTensor] = None,
121 attention_mask: Optional[torch.LongTensor] = None,
122 text_lengths: Optional[torch.Tensor] = None,
123 words_mask: Optional[torch.LongTensor] = None,
124 **kwargs):
126 token_embeds = self.token_rep_layer(input_ids, attention_mask, **kwargs)
--> 128 prompts_embedding, prompts_embedding_mask, words_embedding, mask = self._extract_prompt_features_and_word_embeddings(token_embeds, input_ids, attention_mask,
129 text_lengths, words_mask)
131 if self.config.has_rnn:
132 words_embedding = self.rnn(words_embedding, mask)
File .../lib/python3.11/site-packages/gliner/modeling/base.py:110, in BaseModel._extract_prompt_features_and_word_embeddings(self, token_embeds, input_ids, attention_mask, text_lengths, words_mask)
108 def _extract_prompt_features_and_word_embeddings(self, token_embeds, input_ids, attention_mask,
109 text_lengths, words_mask):
--> 110 prompts_embedding, prompts_embedding_mask, words_embedding, mask = extract_prompt_features_and_word_embeddings(self.config,
111 token_embeds,
112 input_ids,
113 attention_mask,
114 text_lengths,
115 words_mask,
116 self.config.embed_ent_token)
117 return prompts_embedding, prompts_embedding_mask, words_embedding, mask
File .../python3.11/site-packages/gliner/modeling/base.py:55, in extract_prompt_features_and_word_embeddings(config, token_embeds, input_ids, attention_mask, text_lengths, words_mask, embed_ent_token, **kwargs)
52 num_class_tokens = torch.sum(class_token_mask, dim=-1, keepdim=True)
54 max_embed_dim = num_class_tokens.max()
---> 55 max_text_length = text_lengths.max()
56 aranged_class_idx = torch.arange(max_embed_dim,
57 dtype=attention_mask.dtype,
58 device=token_embeds.device).expand(batch_size, -1)
60 batch_indices, target_class_idx = torch.where(aranged_class_idx<num_class_tokens)
AttributeError: 'NoneType' object has no attribute 'max'
I am still having issues with attempts to convert to ONNX using the latest v0.2.16.
import os import onnx from gliner import GLiNER chunk_size = 384 gliner_model = GLiNER.from_pretrained("gliner-community/gliner_large-v2.5", max_length=chunk_size) gliner_model.save_pretrained("gliner_large-v2.5") onnx_save_path = os.path.join("/dirpath", "gliner-large-v25.onnx") text = "ONNX is an open-source format designed to enable the interoperability of AI models across various frameworks and tools." labels = ['format', 'model', 'tool', 'cat'] inputs, _ = gliner_model.prepare_model_inputs([text], labels) if gliner_model.config.span_mode == 'token_level': all_inputs = (inputs['input_ids'], inputs['attention_mask'], inputs['words_mask'], inputs['text_lengths']) input_names = ['input_ids', 'attention_mask', 'words_mask', 'text_lengths'] dynamic_axes={ "input_ids": {0: "batch_size", 1: "sequence_length"}, "attention_mask": {0: "batch_size", 1: "sequence_length"}, "words_mask": {0: "batch_size", 1: "sequence_length"}, "text_lengths": {0: "batch_size", 1: "value"}, "logits": {0: "position", 1: "batch_size", 2: "sequence_length", 3: "num_classes"}, } else: all_inputs = (inputs['input_ids'], inputs['attention_mask'], inputs['words_mask'], inputs['text_lengths'], inputs['span_idx'], inputs['span_mask']) input_names = ['input_ids', 'attention_mask', 'words_mask', 'text_lengths', 'span_idx', 'span_mask'] dynamic_axes={ "input_ids": {0: "batch_size", 1: "sequence_length"}, "attention_mask": {0: "batch_size", 1: "sequence_length"}, "words_mask": {0: "batch_size", 1: "sequence_length"}, "text_lengths": {0: "batch_size", 1: "value"}, "span_idx": {0: "batch_size", 1: "num_spans", 2: "idx"}, "span_mask": {0: "batch_size", 1: "num_spans"}, "logits": {0: "batch_size", 1: "sequence_length", 2: "num_spans", 3: "num_classes"}, } print('Converting the model...') torch.onnx.export( gliner_model.model, all_inputs, f=onnx_save_path, input_names=input_names, output_names=["logits"], dynamic_axes=dynamic_axes, opset_version=14, )Returns
Asking to truncate to max_length but no maximum length is provided and the model has no predefined maximum length. Default to no truncation. Converting the model... --------------------------------------------------------------------------- AttributeError Traceback (most recent call last) Cell In[2], line 34 24 dynamic_axes={ 25 "input_ids": {0: "batch_size", 1: "sequence_length"}, 26 "attention_mask": {0: "batch_size", 1: "sequence_length"}, (...) 31 "logits": {0: "batch_size", 1: "sequence_length", 2: "num_spans", 3: "num_classes"}, 32 } 33 print('Converting the model...') ---> 34 torch.onnx.export( 35 gliner_model.model, 36 all_inputs, 37 f=onnx_save_path, 38 input_names=input_names, 39 output_names=["logits"], 40 dynamic_axes=dynamic_axes, 41 opset_version=14, 42 ) File .../python3.11/site-packages/torch/onnx/__init__.py:375, in export(model, args, f, kwargs, export_params, verbose, input_names, output_names, opset_version, dynamic_axes, keep_initializers_as_inputs, dynamo, external_data, dynamic_shapes, report, verify, profile, dump_exported_program, artifacts_dir, fallback, training, operator_export_type, do_constant_folding, custom_opsets, export_modules_as_functions, autograd_inlining, **_) 369 if dynamic_shapes: 370 raise ValueError( 371 "The exporter only supports dynamic shapes " 372 "through parameter dynamic_axes when dynamo=False." 373 ) --> 375 export( 376 model, 377 args, 378 f, # type: ignore[arg-type] 379 kwargs=kwargs, 380 export_params=export_params, 381 verbose=verbose is True, 382 input_names=input_names, 383 output_names=output_names, 384 opset_version=opset_version, 385 dynamic_axes=dynamic_axes, 386 keep_initializers_as_inputs=keep_initializers_as_inputs, 387 training=training, 388 operator_export_type=operator_export_type, 389 do_constant_folding=do_constant_folding, 390 custom_opsets=custom_opsets, 391 export_modules_as_functions=export_modules_as_functions, 392 autograd_inlining=autograd_inlining, 393 ) 394 return None File .../python3.11/site-packages/torch/onnx/utils.py:502, in export(model, args, f, kwargs, 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, autograd_inlining) 499 if kwargs is not None: 500 args = args + (kwargs,) --> 502 _export( 503 model, 504 args, 505 f, 506 export_params, 507 verbose, 508 training, 509 input_names, 510 output_names, 511 operator_export_type=operator_export_type, 512 opset_version=opset_version, 513 do_constant_folding=do_constant_folding, 514 dynamic_axes=dynamic_axes, 515 keep_initializers_as_inputs=keep_initializers_as_inputs, 516 custom_opsets=custom_opsets, 517 export_modules_as_functions=export_modules_as_functions, 518 autograd_inlining=autograd_inlining, 519 ) 521 return None File .../lib/python3.11/site-packages/torch/onnx/utils.py:1564, 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, autograd_inlining) 1561 dynamic_axes = {} 1562 _validate_dynamic_axes(dynamic_axes, model, input_names, output_names) -> 1564 graph, params_dict, torch_out = _model_to_graph( 1565 model, 1566 args, 1567 verbose, 1568 input_names, 1569 output_names, 1570 operator_export_type, 1571 val_do_constant_folding, 1572 fixed_batch_size=fixed_batch_size, 1573 training=training, 1574 dynamic_axes=dynamic_axes, 1575 ) 1577 # TODO: Don't allocate a in-memory string for the protobuf 1578 defer_weight_export = ( 1579 export_type is not _exporter_states.ExportTypes.PROTOBUF_FILE 1580 ) File .../lib/python3.11/site-packages/torch/onnx/utils.py:1113, 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) 1110 args = (args,) 1112 model = _pre_trace_quant_model(model, args) -> 1113 graph, params, torch_out, module = _create_jit_graph(model, args) 1114 params_dict = _get_named_param_dict(graph, params) 1116 try: File .../lib/python3.11/site-packages/torch/onnx/utils.py:997, in _create_jit_graph(model, args) 992 graph = _C._propagate_and_assign_input_shapes( 993 graph, flattened_args, param_count_list, False, False 994 ) 995 return graph, params, torch_out, None --> 997 graph, torch_out = _trace_and_get_graph_from_model(model, args) 998 _C._jit_pass_onnx_lint(graph) 999 state_dict = torch.jit._unique_state_dict(model) File .../lib/python3.11/site-packages/torch/onnx/utils.py:904, in _trace_and_get_graph_from_model(model, args) 902 prev_autocast_cache_enabled = torch.is_autocast_cache_enabled() 903 torch.set_autocast_cache_enabled(False) --> 904 trace_graph, torch_out, inputs_states = torch.jit._get_trace_graph( 905 model, 906 args, 907 strict=False, 908 _force_outplace=False, 909 _return_inputs_states=True, 910 ) 911 torch.set_autocast_cache_enabled(prev_autocast_cache_enabled) 913 warn_on_static_input_change(inputs_states) File .../lib/python3.11/site-packages/torch/jit/_trace.py:1500, in _get_trace_graph(f, args, kwargs, strict, _force_outplace, return_inputs, _return_inputs_states) 1498 if not isinstance(args, tuple): 1499 args = (args,) -> 1500 outs = ONNXTracedModule( 1501 f, strict, _force_outplace, return_inputs, _return_inputs_states 1502 )(*args, **kwargs) 1503 return outs File .../lib/python3.11/site-packages/torch/nn/modules/module.py:1736, in Module._wrapped_call_impl(self, *args, **kwargs) 1734 return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc] 1735 else: -> 1736 return self._call_impl(*args, **kwargs) File .../lib/python3.11/site-packages/torch/nn/modules/module.py:1747, in Module._call_impl(self, *args, **kwargs) 1742 # If we don't have any hooks, we want to skip the rest of the logic in 1743 # this function, and just call forward. 1744 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks 1745 or _global_backward_pre_hooks or _global_backward_hooks 1746 or _global_forward_hooks or _global_forward_pre_hooks): -> 1747 return forward_call(*args, **kwargs) 1749 result = None 1750 called_always_called_hooks = set() File .../lib/python3.11/site-packages/torch/jit/_trace.py:139, in ONNXTracedModule.forward(self, *args) 136 else: 137 return tuple(out_vars) --> 139 graph, out = torch._C._create_graph_by_tracing( 140 wrapper, 141 in_vars + module_state, 142 _create_interpreter_name_lookup_fn(), 143 self.strict, 144 self._force_outplace, 145 ) 147 if self._return_inputs: 148 return graph, outs[0], ret_inputs[0] File .../lib/python3.11/site-packages/torch/jit/_trace.py:130, in ONNXTracedModule.forward.<locals>.wrapper(*args) 128 if self._return_inputs_states: 129 inputs_states.append(_unflatten(in_args, in_desc)) --> 130 outs.append(self.inner(*trace_inputs)) 131 if self._return_inputs_states: 132 inputs_states[0] = (inputs_states[0], trace_inputs) File .../lib/python3.11/site-packages/torch/nn/modules/module.py:1736, in Module._wrapped_call_impl(self, *args, **kwargs) 1734 return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc] 1735 else: -> 1736 return self._call_impl(*args, **kwargs) File .../nlp/lib/python3.11/site-packages/torch/nn/modules/module.py:1747, in Module._call_impl(self, *args, **kwargs) 1742 # If we don't have any hooks, we want to skip the rest of the logic in 1743 # this function, and just call forward. 1744 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks 1745 or _global_backward_pre_hooks or _global_backward_hooks 1746 or _global_forward_hooks or _global_forward_pre_hooks): -> 1747 return forward_call(*args, **kwargs) 1749 result = None 1750 called_always_called_hooks = set() File .../python3.11/site-packages/torch/nn/modules/module.py:1726, in Module._slow_forward(self, *input, **kwargs) 1724 recording_scopes = False 1725 try: -> 1726 result = self.forward(*input, **kwargs) 1727 finally: 1728 if recording_scopes: File .../lib/python3.11/site-packages/gliner/modeling/base.py:233, in SpanModel.forward(self, input_ids, attention_mask, labels_embeddings, labels_input_ids, labels_attention_mask, words_embedding, mask, prompts_embedding, prompts_embedding_mask, words_mask, text_lengths, span_idx, span_mask, labels, **kwargs) 215 def forward(self, 216 input_ids: Optional[torch.FloatTensor] = None, 217 attention_mask: Optional[torch.LongTensor] = None, (...) 230 **kwargs 231 ): --> 233 prompts_embedding, prompts_embedding_mask, words_embedding, mask = self.get_representations(input_ids, attention_mask, 234 labels_embeddings, labels_input_ids, labels_attention_mask, 235 text_lengths, words_mask) 236 span_idx = span_idx*span_mask.unsqueeze(-1) 238 span_rep = self.span_rep_layer(words_embedding, span_idx) File .../lib/python3.11/site-packages/gliner/modeling/base.py:182, in BaseModel.get_representations(self, input_ids, attention_mask, labels_embeddings, labels_input_ids, labels_attention_mask, text_lengths, words_mask, **kwargs) 177 prompts_embedding, prompts_embedding_mask, words_embedding, mask = self.get_bi_representations( 178 input_ids, attention_mask, labels_embeddings, labels_input_ids, labels_attention_mask, 179 text_lengths, words_mask, **kwargs 180 ) 181 else: --> 182 prompts_embedding, prompts_embedding_mask, words_embedding, mask = self.get_uni_representations( 183 input_ids, attention_mask, text_lengths, words_mask, **kwargs 184 ) 185 return prompts_embedding, prompts_embedding_mask, words_embedding, mask File .../lib/python3.11/site-packages/gliner/modeling/base.py:128, in BaseModel.get_uni_representations(self, input_ids, attention_mask, text_lengths, words_mask, **kwargs) 119 def get_uni_representations(self, 120 input_ids: Optional[torch.FloatTensor] = None, 121 attention_mask: Optional[torch.LongTensor] = None, 122 text_lengths: Optional[torch.Tensor] = None, 123 words_mask: Optional[torch.LongTensor] = None, 124 **kwargs): 126 token_embeds = self.token_rep_layer(input_ids, attention_mask, **kwargs) --> 128 prompts_embedding, prompts_embedding_mask, words_embedding, mask = self._extract_prompt_features_and_word_embeddings(token_embeds, input_ids, attention_mask, 129 text_lengths, words_mask) 131 if self.config.has_rnn: 132 words_embedding = self.rnn(words_embedding, mask) File .../lib/python3.11/site-packages/gliner/modeling/base.py:110, in BaseModel._extract_prompt_features_and_word_embeddings(self, token_embeds, input_ids, attention_mask, text_lengths, words_mask) 108 def _extract_prompt_features_and_word_embeddings(self, token_embeds, input_ids, attention_mask, 109 text_lengths, words_mask): --> 110 prompts_embedding, prompts_embedding_mask, words_embedding, mask = extract_prompt_features_and_word_embeddings(self.config, 111 token_embeds, 112 input_ids, 113 attention_mask, 114 text_lengths, 115 words_mask, 116 self.config.embed_ent_token) 117 return prompts_embedding, prompts_embedding_mask, words_embedding, mask File .../python3.11/site-packages/gliner/modeling/base.py:55, in extract_prompt_features_and_word_embeddings(config, token_embeds, input_ids, attention_mask, text_lengths, words_mask, embed_ent_token, **kwargs) 52 num_class_tokens = torch.sum(class_token_mask, dim=-1, keepdim=True) 54 max_embed_dim = num_class_tokens.max() ---> 55 max_text_length = text_lengths.max() 56 aranged_class_idx = torch.arange(max_embed_dim, 57 dtype=attention_mask.dtype, 58 device=token_embeds.device).expand(batch_size, -1) 60 batch_indices, target_class_idx = torch.where(aranged_class_idx<num_class_tokens) AttributeError: 'NoneType' object has no attribute 'max'
Check my sample notebook, I have used my fine-tuned Gliner model and tried converting ONNX, it's working https://colab.research.google.com/drive/1lwGSOAOo_hVs2K8cMkRks2MEJho7H18f?usp=sharing