I’ve followed the tutorials as best I can for Feature Extraction of text, and that pipeline seems to work really well. The code is in the multimodal section of the pipeline documentation, but I can’t see what kind of input the model is expecting for an image. The filepath and a PIL image both have given errors, though the PIL version seems to work better than the filepath.
My current code is below:
from transformers import pipeline
import torch
from PIL import Image
extractor = pipeline(model=“google/vit-base-patch16-224”, task=“feature-extraction”,device=0)
img = Image.open(“0531d216-4a12-4836-8389-efbbfac21a2b.png”)
result = extractor(inputs=img, return_tensors=True)
print(result.shape)
This gives an error:
Cell In[18], line 7
5 extractor = pipeline(model=“google/vit-base-patch16-224”, task=“feature-extraction”,device=0)
6 img = Image.open(“0531d216-4a12-4836-8389-efbbfac21a2b.png”)
----> 7 result = extractor(inputs=img, return_tensors=True)
8 result.shape # This is a tensor of shape [1, sequence_lenth, hidden_dimension] representing the input string.
File ~/.local/lib/python3.10/site-packages/transformers/pipelines/feature_extraction.py:107, in FeatureExtractionPipeline.call(self, *args, **kwargs)
97 def call(self, *args, **kwargs):
98 “”"
99 Extract the features of the input(s).
100
(…)
105 A nested list of float: The features computed by the model.
106 “”"
→ 107 return super().call(*args, **kwargs)
File ~/.local/lib/python3.10/site-packages/transformers/pipelines/base.py:1122, in Pipeline.call(self, inputs, num_workers, batch_size, *args, **kwargs)
1114 return next(
1115 iter(
1116 self.get_iterator(
(…)
1119 )
1120 )
1121 else:
→ 1122 return self.run_single(inputs, preprocess_params, forward_params, postprocess_params)
File ~/.local/lib/python3.10/site-packages/transformers/pipelines/base.py:1128, in Pipeline.run_single(self, inputs, preprocess_params, forward_params, postprocess_params)
1127 def run_single(self, inputs, preprocess_params, forward_params, postprocess_params):
→ 1128 model_inputs = self.preprocess(inputs, **preprocess_params)
1129 model_outputs = self.forward(model_inputs, **forward_params)
1130 outputs = self.postprocess(model_outputs, **postprocess_params)
File ~/.local/lib/python3.10/site-packages/transformers/pipelines/feature_extraction.py:81, in FeatureExtractionPipeline.preprocess(self, inputs, **tokenize_kwargs)
79 def preprocess(self, inputs, **tokenize_kwargs) → Dict[str, GenericTensor]:
80 return_tensors = self.framework
—> 81 model_inputs = self.tokenizer(inputs, return_tensors=return_tensors, **tokenize_kwargs)
82 return model_inputs
TypeError: ‘NoneType’ object is not callable
I think that’s referring to the tokenizer - which I’ve not currently set, but they all seem to be text tokenizers? what am I doing wrong here?