How to use for the fintuned roberta model for text classification

[pradeepdev-1995]

I have finetuned ROBERTA model(TFAutomodelforsequenceclassification busased ) in my local path How can I perform text classification inferencing from that model using deepsparse? I am using GPU and CUDA 12 Please share the sample code

[mgoin]

Hey @pradeepdev-1995 this document would be a good starting point. Just make sure to export to ONNX using SparseML https://github.com/neuralmagic/deepsparse/blob/main/docs/use-cases/nlp/text-classification.md

[pradeepdev-1995]

The /data/onnx_model directory contain the onnx model file and tokenizer files then tried this code

from deepsparse import Pipeline

# download onnx from sparsezoo and compile with batch size 1
sparsezoo_stub = "/data/onnx_model"
pipeline = Pipeline.create(
  task="text-classification",
  model_path=sparsezoo_stub,   # sparsezoo stub or path to local ONNX
  batch_size=1                 # default batch size is 1
)

# run inference
sequences = [[
  "Which is the best gaming laptop under 40k?"
]]
prediction = pipeline(sequences)
print(prediction)

but following error comes

---------------------------------------------------------------------------
DecodeError                               Traceback (most recent call last)
Cell In[2], line 5
      3 # download onnx from sparsezoo and compile with batch size 1
      4 sparsezoo_stub = "/data/onnx_model"
----> 5 pipeline = Pipeline.create(
      6   task="text-classification",
      7   model_path=sparsezoo_stub,   # sparsezoo stub or path to local ONNX
      8   batch_size=1                 # default batch size is 1
      9 )
     11 # run inference
     12 sequences = [[
     13   "Which is the best gaming laptop under 40k?"
     14 ]]

File /tmp/pip_packages/deepsparse/base_pipeline.py:210, in BasePipeline.create(task, **kwargs)
    204     buckets = pipeline_constructor.create_pipeline_buckets(
    205         task=task,
    206         **kwargs,
    207     )
    208     return BucketingPipeline(pipelines=buckets)
--> 210 return pipeline_constructor(**kwargs)

File /tmp/pip_packages/deepsparse/transformers/pipelines/text_classification.py:152, in TextClassificationPipeline.__init__(self, top_k, return_all_scores, **kwargs)
    145 def __init__(
    146     self,
    147     *,
   (...)
    150     **kwargs,
    151 ):
--> 152     super().__init__(**kwargs)
    154     self._top_k = _get_top_k(top_k, return_all_scores, self.config.num_labels)
    155     self._return_all_scores = return_all_scores

File /tmp/pip_packages/deepsparse/transformers/pipelines/pipeline.py:110, in TransformersPipeline.__init__(self, sequence_length, trust_remote_code, config, tokenizer, **kwargs)
    105 self._delay_overwriting_inputs = (
    106     kwargs.pop("_delay_overwriting_inputs", None) or False
    107 )
    108 self._temp_model_directory = None
--> 110 super().__init__(**kwargs)

File /tmp/pip_packages/deepsparse/pipeline.py:199, in Pipeline.__init__(self, model_path, engine_type, batch_size, num_cores, num_streams, scheduler, input_shapes, context, executor, benchmark, _delay_engine_initialize, **kwargs)
    196     self._engine_args["scheduler"] = scheduler
    197     self._engine_args["num_streams"] = num_streams
--> 199 self.onnx_file_path = self.setup_onnx_file_path()
    201 if _delay_engine_initialize:
    202     self.engine = None

File /tmp/pip_packages/deepsparse/transformers/pipelines/pipeline.py:153, in TransformersPipeline.setup_onnx_file_path(self)
    141 self.tokenizer = AutoTokenizer.from_pretrained(
    142     deployment_path,
    143     trust_remote_code=self._trust_remote_code,
    144     model_max_length=self.sequence_length,
    145 )
    147 if not self._delay_overwriting_inputs:
    148     # overwrite onnx graph to given required input shape
    149     (
    150         onnx_path,
    151         self.onnx_input_names,
    152         self._temp_model_directory,
--> 153     ) = overwrite_transformer_onnx_model_inputs(
    154         onnx_path, max_length=self.sequence_length
    155     )
    157 if not self.config or not self.tokenizer:
    158     raise RuntimeError(
    159         "Invalid config or tokenizer provided. Please provide "
    160         "paths to the files or ensure they exist in the `model_path` provided. "
    161         "See `tokenizer` and `config` arguments for details."
    162     )

File /tmp/pip_packages/deepsparse/transformers/helpers.py:110, in overwrite_transformer_onnx_model_inputs(path, batch_size, max_length, inplace)
     92 """
     93 Overrides an ONNX model's inputs to have the given batch size and sequence lengths.
     94 Assumes that these are the first and second shape indices of the given model inputs
   (...)
    105     `inplace=False`, else None)
    106 """
    107 # overwrite input shapes
    108 # if > 2Gb model is to be modified in-place, operate
    109 # exclusively on the model graph
--> 110 model = onnx.load(path, load_external_data=not inplace)
    111 initializer_input_names = set([node.name for node in model.graph.initializer])
    112 external_inputs = [
    113     inp for inp in model.graph.input if inp.name not in initializer_input_names
    114 ]

File /tmp/pip_packages/onnx/__init__.py:170, in load_model(f, format, load_external_data)
    156 """Loads a serialized ModelProto into memory.
    157 
    158 Args:
   (...)
    167     Loaded in-memory ModelProto.
    168 """
    169 s = _load_bytes(f)
--> 170 model = load_model_from_string(s, format=format)
    172 if load_external_data:
    173     model_filepath = _get_file_path(f)

File /tmp/pip_packages/onnx/__init__.py:212, in load_model_from_string(***failed resolving arguments***)
    202 """Loads a binary string (bytes) that contains serialized ModelProto.
    203 
    204 Args:
   (...)
    209     Loaded in-memory ModelProto.
    210 """
    211 del format  # Unused
--> 212 return _deserialize(s, ModelProto())

File /tmp/pip_packages/onnx/__init__.py:143, in _deserialize(s, proto)
    140 if not (hasattr(proto, "ParseFromString") and callable(proto.ParseFromString)):
    141     raise TypeError(f"No ParseFromString method is detected. Type is {type(proto)}")
--> 143 decoded = typing.cast(Optional[int], proto.ParseFromString(s))
    144 if decoded is not None and decoded != len(s):
    145     raise google.protobuf.message.DecodeError(
    146         f"Protobuf decoding consumed too few bytes: {decoded} out of {len(s)}"
    147     )

DecodeError: Error parsing message

@mgoin

[mgoin]

@pradeepdev-1995 it seems like the model was downloaded partially or incorrectly, could you try redownloading it?