transformers-rb 0.1.0

data/lib/transformers/image_utils.rb

@@ -0,0 +1,165 @@
+# Copyright 2021 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+module Transformers
+  class ChannelDimension < ExplicitEnum
+    FIRST = "channels_first"
+    LAST = "channels_last"
+  end
+
+  module ImageUtils
+    def self.load_image(image, timeout: nil)
+      Utils.requires_backends(__method__, ["vision"])
+      if image.is_a?(URI)
+        require "open-uri"
+
+        image = Vips::Image.new_from_buffer(image.read(open_timeout: timeout, read_timeout: timeout), "")
+      elsif image.is_a?(String) && File.exist?(image)
+        image = Vips::Image.new_from_file(image)
+      elsif image.is_a?(Vips::Image)
+        image = image
+      else
+        raise ArgumentError, "Incorrect format used for image"
+      end
+      image
+    end
+
+    def self.validate_preprocess_arguments(
+      do_rescale: nil,
+      rescale_factor: nil,
+      do_normalize: nil,
+      image_mean: nil,
+      image_std: nil,
+      do_pad: nil,
+      size_divisibility: nil,
+      do_center_crop: nil,
+      crop_size: nil,
+      do_resize: nil,
+      size: nil,
+      resample:  nil
+    )
+      if do_rescale && rescale_factor.nil?
+        raise ArgumentError, "`rescale_factor` must be specified if `do_rescale` is `true`."
+      end
+
+      if do_pad && size_divisibility.nil?
+        # Here, size_divisor might be passed as the value of size
+        raise ArgumentError, "Depending on the model, `size_divisibility`, `size_divisor`, `pad_size` or `size` must be specified if `do_pad` is `true`."
+      end
+
+      if do_normalize && (image_mean.nil? || image_std.nil?)
+        raise ArgumentError, "`image_mean` and `image_std` must both be specified if `do_normalize` is `true`."
+      end
+
+      if do_center_crop && crop_size.nil?
+        raise ArgumentError, "`crop_size` must be specified if `do_center_crop` is `true`."
+      end
+
+      if do_resize && (size.nil? || resample.nil?)
+        raise ArgumentError, "`size` and `resample` must be specified if `do_resize` is `true`."
+      end
+    end
+
+    def self.make_list_of_images(images, expected_ndims: 3)
+      # TODO improve
+      images.is_a?(Array) ? images : [images]
+    end
+
+    def self.to_numo_array(img)
+      Numo::UInt8.from_binary(img.write_to_memory, [img.height, img.width, img.bands])
+    end
+
+    def self.infer_channel_dimension_format(
+      image, num_channels: nil
+    )
+      num_channels = !num_channels.nil? ? num_channels : [1, 3]
+      num_channels = num_channels.is_a?(Integer) ? [num_channels] : num_channels
+
+      if image.ndim == 3
+        first_dim, last_dim = 0, 2
+      elsif image.ndim == 4
+        first_dim, last_dim = 1, 3
+      else
+        raise ArgumentError, "Unsupported number of image dimensions: #{image.ndim}"
+      end
+
+      if num_channels.include?(image.shape[first_dim]) && num_channels.include?(image.shape[last_dim])
+        Transformers.logger.warn(
+          "The channel dimension is ambiguous. Got image shape #{image.shape}. Assuming channels are the first dimension."
+        )
+        return ChannelDimension::FIRST
+      elsif num_channels.include?(image.shape[first_dim])
+        return ChannelDimension::FIRST
+      elsif num_channels.include?(image.shape[last_dim])
+        return ChannelDimension::LAST
+      end
+      raise ArgumentError, "Unable to infer channel dimension format"
+    end
+
+    def self.get_channel_dimension_axis(
+      image, input_data_format: nil
+    )
+      if input_data_format.nil?
+        input_data_format = infer_channel_dimension_format(image)
+      end
+      if input_data_format == ChannelDimension::FIRST
+        return image.ndim - 3
+      elsif input_data_format == ChannelDimension::LAST
+        return image.ndim - 1
+      end
+      raise ArgumentError, "Unsupported data format: #{input_data_format}"
+    end
+
+    def self.is_vips_image(img)
+      Utils.is_vision_available && img.is_a?(Vips::Image)
+    end
+
+    def self.is_valid_image(img)
+      is_vips_image(img) || is_numo_array(img) || is_torch_tensor(img)
+    end
+
+    def self.valid_images(imgs)
+      # If we have an list of images, make sure every image is valid
+      if imgs.is_a?(Array)
+        imgs.each do |img|
+          if !valid_images(img)
+            return false
+          end
+        end
+      # If not a list of tuple, we have been given a single image or batched tensor of images
+      elsif !is_valid_image(imgs)
+        return false
+      end
+      true
+    end
+
+    def self.is_scaled_image(image)
+      if image.is_a?(Numo::UInt8)
+        return false
+      end
+
+      # It's possible the image has pixel values in [0, 255] but is of floating type
+      image.min >= 0 && image.max <= 1
+    end
+
+    def self.validate_kwargs(valid_processor_keys:, captured_kwargs:)
+      unused_keys = Set.new(captured_kwargs).difference(Set.new(valid_processor_keys))
+      if unused_keys.any?
+        unused_key_str = unused_keys.join(", ")
+        # TODO raise a warning here instead of simply logging?
+        Transformers.logger.warn("Unused or unrecognized kwargs: #{unused_key_str}.")
+      end
+    end
+  end
+end

data/lib/transformers/modeling_outputs.rb

@@ -0,0 +1,81 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+module Transformers
+  class BaseModelOutput < ModelOutput
+    attribute :last_hidden_state
+    attribute :hidden_states
+    attribute :attentions
+  end
+
+  class BaseModelOutputWithPooling < ModelOutput
+    attribute :last_hidden_state
+    attribute :pooler_output
+    attribute :hidden_states
+    attribute :attentions
+  end
+
+  class BaseModelOutputWithPoolingAndCrossAttentions < ModelOutput
+    attribute :last_hidden_state
+    attribute :pooler_output
+    attribute :hidden_states
+    attribute :past_key_values
+    attribute :attentions
+    attribute :cross_attentions
+  end
+
+  class BaseModelOutputWithPastAndCrossAttentions < ModelOutput
+    attribute :last_hidden_state
+    attribute :past_key_values
+    attribute :hidden_states
+    attribute :attentions
+    attribute :cross_attentions
+  end
+
+  class MaskedLMOutput < ModelOutput
+    attribute :loss
+    attribute :logits
+    attribute :hidden_states
+    attribute :attentions
+  end
+
+  class SequenceClassifierOutput < ModelOutput
+    attribute :loss
+    attribute :logits
+    attribute :hidden_states
+    attribute :attentions
+  end
+
+  class TokenClassifierOutput < ModelOutput
+    attribute :loss
+    attribute :logits
+    attribute :hidden_states
+    attribute :attentions
+  end
+
+  class QuestionAnsweringModelOutput < ModelOutput
+    attribute :loss
+    attribute :start_logits
+    attribute :end_logits
+    attribute :hidden_states
+    attribute :attentions
+  end
+
+  class ImageClassifierOutput < ModelOutput
+    attribute :loss
+    attribute :logits
+    attribute :hidden_states
+    attribute :attentions
+  end
+end