Class: Transformers::Pipeline

Inherits:
Object
  • Object
show all
Defined in:
lib/transformers/pipelines/base.rb

Instance Method Summary collapse

Constructor Details

#initialize(model, tokenizer: nil, feature_extractor: nil, image_processor: nil, modelcard: nil, framework: nil, task: "", device: nil, **kwargs) ⇒ Pipeline

Returns a new instance of Pipeline.



94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
 
# File 'lib/transformers/pipelines/base.rb', line 94

def initialize(
  model,
  tokenizer: nil,
  feature_extractor: nil,
  image_processor: nil,
  modelcard: nil,
  framework: nil,
  task: "",
  device: nil,
  **kwargs
)
  if framework.nil?
    raise Todo
  end

  @task = task
  @model = model
  @tokenizer = tokenizer
  @feature_extractor = feature_extractor
  @image_processor = image_processor
  @modelcard = modelcard
  @framework = framework

  if device.nil?
    if Torch::CUDA.available? || Torch::Backends::MPS.available?
      Transformers.logger.warn(
        "Hardware accelerator e.g. GPU is available in the environment, but no `device` argument" +
        " is passed to the `Pipeline` object. Model will be on CPU."
      )
    end
  end

  if @framework == "pt"
    if device.is_a?(String)
      @device = Torch.device(device)
    else
      # TODO update default in 0.2.0
      @device = @model.device
    end
  else
    raise Todo
  end

  # TODO Fix eql? for Torch::Device in Torch.rb
  if @device.type != @model.device.type || @device.index != @model.device.index
    @model.to(@device)
  end

  @call_count = 0
  @batch_size = kwargs.delete(:batch_size)
  @num_workers = kwargs.delete(:num_workers)
  @preprocess_params, @forward_params, @postprocess_params = _sanitize_parameters(**kwargs)
end

Instance Method Details

#_ensure_tensor_on_device(inputs, device) ⇒ Object 



152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
 
# File 'lib/transformers/pipelines/base.rb', line 152

def _ensure_tensor_on_device(inputs, device)
  # TODO move
  inputs = inputs.to_h if inputs.is_a?(BatchEncoding)

  if inputs.is_a?(ModelOutput)
    inputs.instance_variable_get(:@data).transform_values! { |tensor| _ensure_tensor_on_device(tensor, device) }
    inputs
  elsif inputs.is_a?(Hash)
    inputs.transform_values { |tensor| _ensure_tensor_on_device(tensor, device) }
  elsif inputs.is_a?(Array)
    inputs.map { |item| _ensure_tensor_on_device(item, device) }
  elsif inputs.is_a?(Torch::Tensor)
    inputs.to(device)
  else
    inputs
  end
end

#call(inputs, *args, num_workers: nil, batch_size: nil, **kwargs) ⇒ Object 



217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
 
# File 'lib/transformers/pipelines/base.rb', line 217

def call(inputs, *args, num_workers: nil, batch_size: nil, **kwargs)
  if args.any?
    Transformers.logger.warn("Ignoring args : #{args}")
  end

  if num_workers.nil?
    if @num_workers.nil?
      num_workers = 0
    else
      num_workers = @num_workers
    end
  end
  if batch_size.nil?
    if @batch_size.nil?
      batch_size = 1
    else
      batch_size = @batch_size
    end
  end

  preprocess_params, forward_params, postprocess_params = _sanitize_parameters(**kwargs)

  preprocess_params = @preprocess_params.merge(preprocess_params)
  forward_params = @forward_params.merge(forward_params)
  postprocess_params = @postprocess_params.merge(postprocess_params)

  @call_count += 1
  if @call_count > 10 && @framework == "pt" && @device.type == "cuda"
    Transformers.logger.warn(
      "You seem to be using the pipelines sequentially on GPU. In order to maximize efficiency please use a" +
      " dataset"
    )
  end

  is_dataset = inputs.is_a?(Torch::Utils::Data::Dataset)
  is_generator = inputs.is_a?(Enumerable)
  is_list = inputs.is_a?(Array)

  _is_iterable = is_dataset || is_generator || is_list

  # TODO make the get_iterator work also for `tf` (and `flax`).
  can_use_iterator = @framework == "pt" && (is_dataset || is_generator || is_list)

  if is_list
    if can_use_iterator
      final_iterator = get_iterator(
        inputs, num_workers, batch_size, preprocess_params, forward_params, postprocess_params
      )
      outputs = final_iterator.to_a
      outputs
    else
      run_multi(inputs, preprocess_params, forward_params, postprocess_params)
    end
  else
    run_single(inputs, preprocess_params, forward_params, postprocess_params)
  end
end

#check_model_type(supported_models) ⇒ Object 



170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
 
# File 'lib/transformers/pipelines/base.rb', line 170

def check_model_type(supported_models)
  if !supported_models.is_a?(Array)
    supported_models_names = []
    supported_models.each do |_, model_name|
      # Mapping can now contain tuples of models for the same configuration.
      if model_name.is_a?(Array)
        supported_models_names.concat(model_name)
      else
        supported_models_names << model_name
      end
    end
    supported_models = supported_models_names
  end
  if !supported_models.include?(@model.class.name.split("::").last)
    Transformers.logger.error(
      "The model '#{@model.class.name}' is not supported for #{@task}. Supported models are" +
      " #{supported_models}."
    )
  end
end

#get_iterator(inputs, num_workers, batch_size, preprocess_params, forward_params, postprocess_params) ⇒ Object 



191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
 
# File 'lib/transformers/pipelines/base.rb', line 191

def get_iterator(
  inputs, num_workers, batch_size, preprocess_params, forward_params, postprocess_params
)
  if inputs.respond_to?(:size)
    dataset = PipelineDataset.new(inputs, method(:preprocess), preprocess_params)
  else
    if num_workers > 1
      Transformers.logger.warn(
        "For iterable dataset using num_workers>1 is likely to result" +
        " in errors since everything is iterable, setting `num_workers: 1`" +
        " to guarantee correctness."
      )
      num_workers = 1
    end
    dataset = PipelineIterator.new(inputs, method(:preprocess), preprocess_params)
  end

  # TODO hack by collating feature_extractor and image_processor
  feature_extractor = !@feature_extractor.nil? ? @feature_extractor : @image_processor
  collate_fn = batch_size == 1 ? method(:no_collate_fn) : pad_collate_fn(@tokenizer, feature_extractor)
  dataloader = Torch::Utils::Data::DataLoader.new(dataset, batch_size: batch_size, collate_fn: collate_fn) # num_workers: num_workers,
  model_iterator = PipelineIterator.new(dataloader, method(:forward), forward_params, loader_batch_size: batch_size)
  final_iterator = PipelineIterator.new(model_iterator, method(:postprocess), postprocess_params)
  final_iterator
end

#torch_dtype ⇒ Object 



148
149
150
 
# File 'lib/transformers/pipelines/base.rb', line 148

def torch_dtype
  @model.dtype
end