onnxruntime 0.9.2-aarch64-linux → 0.9.4-aarch64-linux

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 52e865156423eb866f07a8e2e13961a7376da543f93ed1b5dbcff08bfc2bf88d
-  data.tar.gz: a5d82e576c5f8a4528dd14efe4ec8bd508933da2a954118f7ff105a8193bd389
+  metadata.gz: 9544eb4f8df1fd611ba42ce3b3687af10d226434202e42caac7e03dd3926162d
+  data.tar.gz: 9176e3659227243408ed5f5eab6508b719ccfbdddb37a2d173f6ebcb26ce490d
 SHA512:
-  metadata.gz: 655c0c752e56372ae2b37710d4e8e2864f67431be48f834588160f61b1a6ecb93874de75f5f892115925d3fbfd78a820df08f5d3cde2405b9807908d80b19e21
-  data.tar.gz: 2b64c39e0ed3c09072b2f70691a4b1ce805616d2b397a563cbf2ed9270e46fb0422cbe3b5b385311c8f8f9ef384ee86545558185049ac1652a9c01dc1f71aa55
+  metadata.gz: 255f60ce409099e1ef8b4bbf820360903785b0690c26fd540511c839185ba8abfde9eee1a67ffb3ff1c38d1a36c1910b9b05626da98eb1b6119eddc1b18036ee
+  data.tar.gz: 177f73d9d97856c34c9f5dccb19a9428776421a54edf2b65349411eae556783ce24d5375f06b1a6ff12449c5c894ce12f0ffd566880e0263ae9ea67aa474ac3b

data/CHANGELOG.md

@@ -1,3 +1,13 @@
+## 0.9.4 (2025-03-08)
+
+- Updated ONNX Runtime to 1.21.0
+
+## 0.9.3 (2024-11-01)
+
+- Updated ONNX Runtime to 1.20.0
+- Added experimental `OrtValue` class
+- Added experimental `run_with_ort_values` method
+
 ## 0.9.2 (2024-09-04)
 
 - Updated ONNX Runtime to 1.19.2

data/README.md

@@ -4,6 +4,8 @@
 
 Check out [an example](https://ankane.org/tensorflow-ruby)
 
+For transformer models, check out [Informers](https://github.com/ankane/informers)
+
 [![Build Status](https://github.com/ankane/onnxruntime-ruby/actions/workflows/build.yml/badge.svg)](https://github.com/ankane/onnxruntime-ruby/actions)
 
 ## Installation

data/lib/onnxruntime/ffi.rb

@@ -11,7 +11,7 @@ module OnnxRuntime
 
     # enums
     TensorElementDataType = enum(:undefined, :float, :uint8, :int8, :uint16, :int16, :int32, :int64, :string, :bool, :float16, :double, :uint32, :uint64, :complex64, :complex128, :bfloat16)
-    OnnxType = enum(:unknown, :tensor, :sequence, :map, :opaque, :sparsetensor)
+    OnnxType = enum(:unknown, :tensor, :sequence, :map, :opaque, :sparsetensor, :optional)
 
     class Api < ::FFI::Struct
       layout \
@@ -247,7 +247,11 @@ module OnnxRuntime
     attach_function :OrtGetApiBase, %i[], ApiBase.by_ref
 
     def self.api
-      @api ||= self.OrtGetApiBase[:GetApi].call(ORT_API_VERSION)
+      @api ||= begin
+        api = self.OrtGetApiBase[:GetApi].call(ORT_API_VERSION)
+        api = Api.by_ref.from_native(api, nil) if RUBY_PLATFORM == "java"
+        api
+      end
     end
 
     if Gem.win_platform?

data/lib/onnxruntime/inference_session.rb

@@ -83,23 +83,36 @@ module OnnxRuntime
       @session = load_session(path_or_bytes, session_options)
       ObjectSpace.define_finalizer(@session, self.class.finalize(read_pointer.to_i))
 
-      @allocator = load_allocator
+      @allocator = Utils.allocator
       @inputs = load_inputs
       @outputs = load_outputs
     ensure
       release :SessionOptions, session_options
     end
 
-    # TODO support logid
     def run(output_names, input_feed, log_severity_level: nil, log_verbosity_level: nil, logid: nil, terminate: nil, output_type: :ruby)
-      # pointer references
-      refs = []
+      if ![:ruby, :numo, :ort_value].include?(output_type)
+        raise ArgumentError, "Invalid output type: #{output_type}"
+      end
 
-      input_tensor = create_input_tensor(input_feed, refs)
+      ort_values = input_feed.keys.zip(create_input_tensor(input_feed)).to_h
+
+      outputs = run_with_ort_values(output_names, ort_values, log_severity_level: log_severity_level, log_verbosity_level: log_verbosity_level, logid: logid, terminate: terminate)
+
+      outputs.map { |v| output_type == :numo ? v.numo : (output_type == :ort_value ? v : v.to_ruby) }
+    end
+
+    # TODO support logid
+    def run_with_ort_values(output_names, input_feed, log_severity_level: nil, log_verbosity_level: nil, logid: nil, terminate: nil)
+      input_tensor = ::FFI::MemoryPointer.new(:pointer, input_feed.size)
+      input_feed.each_with_index do |(_, input), i|
+        input_tensor[i].write_pointer(input.to_ptr)
+      end
 
       output_names ||= @outputs.map { |v| v[:name] }
 
       output_tensor = ::FFI::MemoryPointer.new(:pointer, outputs.size)
+      refs = []
       input_node_names = create_node_names(input_feed.keys.map(&:to_s), refs)
       output_node_names = create_node_names(output_names.map(&:to_s), refs)
 
@@ -113,17 +126,9 @@ module OnnxRuntime
 
       check_status api[:Run].call(read_pointer, run_options.read_pointer, input_node_names, input_tensor, input_feed.size, output_node_names, output_names.size, output_tensor)
 
-      output_names.size.times.map do |i|
-        create_from_onnx_value(output_tensor[i].read_pointer, output_type)
-      end
+      output_names.size.times.map { |i| OrtValue.new(output_tensor[i]) }
     ensure
       release :RunOptions, run_options
-      if input_tensor
-        input_feed.size.times do |i|
-          release :Value, input_tensor[i]
-        end
-      end
-      # output values released in create_from_onnx_value
     end
 
     def modelmeta
@@ -221,12 +226,6 @@ module OnnxRuntime
       session
     end
 
-    def load_allocator
-      allocator = ::FFI::MemoryPointer.new(:pointer)
-      check_status api[:GetAllocatorWithDefaultOptions].call(allocator)
-      allocator
-    end
-
     def load_inputs
       inputs = []
       num_input_nodes = ::FFI::MemoryPointer.new(:size_t)
@@ -237,7 +236,7 @@ module OnnxRuntime
         # freed in node_info
         typeinfo = ::FFI::MemoryPointer.new(:pointer)
         check_status api[:SessionGetInputTypeInfo].call(read_pointer, i, typeinfo)
-        inputs << {name: name_ptr.read_pointer.read_string}.merge(node_info(typeinfo))
+        inputs << {name: name_ptr.read_pointer.read_string}.merge(Utils.node_info(typeinfo))
         allocator_free name_ptr
       end
       inputs
@@ -253,91 +252,28 @@ module OnnxRuntime
         # freed in node_info
         typeinfo = ::FFI::MemoryPointer.new(:pointer)
         check_status api[:SessionGetOutputTypeInfo].call(read_pointer, i, typeinfo)
-        outputs << {name: name_ptr.read_pointer.read_string}.merge(node_info(typeinfo))
+        outputs << {name: name_ptr.read_pointer.read_string}.merge(Utils.node_info(typeinfo))
         allocator_free name_ptr
       end
       outputs
     end
 
-    def create_input_tensor(input_feed, refs)
-      allocator_info = ::FFI::MemoryPointer.new(:pointer)
-      check_status api[:CreateCpuMemoryInfo].call(1, 0, allocator_info)
-      input_tensor = ::FFI::MemoryPointer.new(:pointer, input_feed.size)
-
-      input_feed.each_with_index do |(input_name, input), idx|
+    def create_input_tensor(input_feed)
+      input_feed.map do |input_name, input|
         # TODO support more types
         inp = @inputs.find { |i| i[:name] == input_name.to_s }
         raise Error, "Unknown input: #{input_name}" unless inp
 
-        input = input.to_a unless input.is_a?(Array) || numo_array?(input)
-        shape = input_shape(input)
-
-        input_node_dims = ::FFI::MemoryPointer.new(:int64, shape.size)
-        input_node_dims.write_array_of_int64(shape)
-
-        if inp[:type] == "tensor(string)"
-          type_enum = FFI::TensorElementDataType[:string]
-          check_status api[:CreateTensorAsOrtValue].call(@allocator.read_pointer, input_node_dims, shape.size, type_enum, input_tensor[idx])
-
-          # keep reference to _str_ptrs until FillStringTensor call
-          input_tensor_values, _str_ptrs = create_input_strings(input)
-          check_status api[:FillStringTensor].call(input_tensor[idx].read_pointer, input_tensor_values, input_tensor_values.size / input_tensor_values.type_size)
+        if input.is_a?(OrtValue)
+          input
+        elsif inp[:type] == "tensor(string)"
+          OrtValue.from_array(input, element_type: :string)
         elsif (tensor_type = tensor_types[inp[:type]])
-          input_tensor_values = create_input_data(input, tensor_type)
-          type_enum = FFI::TensorElementDataType[tensor_type]
-          check_status api[:CreateTensorWithDataAsOrtValue].call(allocator_info.read_pointer, input_tensor_values, input_tensor_values.size, input_node_dims, shape.size, type_enum, input_tensor[idx])
-
-          refs << input_tensor_values
+          OrtValue.from_array(input, element_type: tensor_type)
         else
-          unsupported_type("input", inp[:type])
+          Utils.unsupported_type("input", inp[:type])
         end
       end
-
-      input_tensor
-    ensure
-      release :MemoryInfo, allocator_info
-    end
-
-    def input_shape(input)
-      if numo_array?(input)
-        input.shape
-      else
-        shape = []
-        s = input
-        while s.is_a?(Array)
-          shape << s.size
-          s = s.first
-        end
-        shape
-      end
-    end
-
-    def create_input_strings(input)
-      str_ptrs =
-        if numo_array?(input)
-          input.size.times.map { |i| ::FFI::MemoryPointer.from_string(input[i]) }
-        else
-          input.flatten.map { |v| ::FFI::MemoryPointer.from_string(v) }
-        end
-
-      input_tensor_values = ::FFI::MemoryPointer.new(:pointer, str_ptrs.size)
-      input_tensor_values.write_array_of_pointer(str_ptrs)
-      [input_tensor_values, str_ptrs]
-    end
-
-    def create_input_data(input, tensor_type)
-      if numo_array?(input)
-        input.cast_to(numo_types[tensor_type]).to_binary
-      else
-        flat_input = input.flatten.to_a
-        input_tensor_values = ::FFI::MemoryPointer.new(tensor_type, flat_input.size)
-        if tensor_type == :bool
-          input_tensor_values.write_array_of_uint8(flat_input.map { |v| v ? 1 : 0 })
-        else
-          input_tensor_values.send("write_array_of_#{tensor_type}", flat_input)
-        end
-        input_tensor_values
-      end
     end
 
     def create_node_names(names, refs)
@@ -349,230 +285,18 @@ module OnnxRuntime
       ptr
     end
 
-    def create_from_onnx_value(out_ptr, output_type)
-      out_type = ::FFI::MemoryPointer.new(:int)
-      check_status api[:GetValueType].call(out_ptr, out_type)
-      type = FFI::OnnxType[out_type.read_int]
-
-      case type
-      when :tensor
-        typeinfo = ::FFI::MemoryPointer.new(:pointer)
-        check_status api[:GetTensorTypeAndShape].call(out_ptr, typeinfo)
-
-        type, shape = tensor_type_and_shape(typeinfo)
-
-        tensor_data = ::FFI::MemoryPointer.new(:pointer)
-        check_status api[:GetTensorMutableData].call(out_ptr, tensor_data)
-
-        out_size = ::FFI::MemoryPointer.new(:size_t)
-        check_status api[:GetTensorShapeElementCount].call(typeinfo.read_pointer, out_size)
-        output_tensor_size = out_size.read(:size_t)
-
-        release :TensorTypeAndShapeInfo, typeinfo
-
-        # TODO support more types
-        type = FFI::TensorElementDataType[type]
-
-        case output_type
-        when :numo
-          case type
-          when :string
-            result = Numo::RObject.new(shape)
-            result.allocate
-            create_strings_from_onnx_value(out_ptr, output_tensor_size, result)
-          else
-            numo_type = numo_types[type]
-            unsupported_type("element", type) unless numo_type
-            numo_type.from_binary(tensor_data.read_pointer.read_bytes(output_tensor_size * numo_type::ELEMENT_BYTE_SIZE), shape)
-          end
-        when :ruby
-          arr =
-            case type
-            when :float, :uint8, :int8, :uint16, :int16, :int32, :int64, :double, :uint32, :uint64
-              tensor_data.read_pointer.send("read_array_of_#{type}", output_tensor_size)
-            when :bool
-              tensor_data.read_pointer.read_array_of_uint8(output_tensor_size).map { |v| v == 1 }
-            when :string
-              create_strings_from_onnx_value(out_ptr, output_tensor_size, [])
-            else
-              unsupported_type("element", type)
-            end
-
-          Utils.reshape(arr, shape)
-        else
-          raise ArgumentError, "Invalid output type: #{output_type}"
-        end
-      when :sequence
-        out = ::FFI::MemoryPointer.new(:size_t)
-        check_status api[:GetValueCount].call(out_ptr, out)
-
-        out.read(:size_t).times.map do |i|
-          seq = ::FFI::MemoryPointer.new(:pointer)
-          check_status api[:GetValue].call(out_ptr, i, @allocator.read_pointer, seq)
-          create_from_onnx_value(seq.read_pointer, output_type)
-        end
-      when :map
-        type_shape = ::FFI::MemoryPointer.new(:pointer)
-        map_keys = ::FFI::MemoryPointer.new(:pointer)
-        map_values = ::FFI::MemoryPointer.new(:pointer)
-        elem_type = ::FFI::MemoryPointer.new(:int)
-
-        check_status api[:GetValue].call(out_ptr, 0, @allocator.read_pointer, map_keys)
-        check_status api[:GetValue].call(out_ptr, 1, @allocator.read_pointer, map_values)
-        check_status api[:GetTensorTypeAndShape].call(map_keys.read_pointer, type_shape)
-        check_status api[:GetTensorElementType].call(type_shape.read_pointer, elem_type)
-        release :TensorTypeAndShapeInfo, type_shape
-
-        # TODO support more types
-        elem_type = FFI::TensorElementDataType[elem_type.read_int]
-        case elem_type
-        when :int64
-          ret = {}
-          keys = create_from_onnx_value(map_keys.read_pointer, output_type)
-          values = create_from_onnx_value(map_values.read_pointer, output_type)
-          keys.zip(values).each do |k, v|
-            ret[k] = v
-          end
-          ret
-        else
-          unsupported_type("element", elem_type)
-        end
-      else
-        unsupported_type("ONNX", type)
-      end
-    ensure
-      api[:ReleaseValue].call(out_ptr) unless out_ptr.null?
-    end
-
-    def create_strings_from_onnx_value(out_ptr, output_tensor_size, result)
-      len = ::FFI::MemoryPointer.new(:size_t)
-      check_status api[:GetStringTensorDataLength].call(out_ptr, len)
-
-      s_len = len.read(:size_t)
-      s = ::FFI::MemoryPointer.new(:uchar, s_len)
-      offsets = ::FFI::MemoryPointer.new(:size_t, output_tensor_size)
-      check_status api[:GetStringTensorContent].call(out_ptr, s, s_len, offsets, output_tensor_size)
-
-      offsets = output_tensor_size.times.map { |i| offsets[i].read(:size_t) }
-      offsets << s_len
-      output_tensor_size.times do |i|
-        result[i] = s.get_bytes(offsets[i], offsets[i + 1] - offsets[i])
-      end
-      result
-    end
-
     def read_pointer
       @session.read_pointer
     end
 
     def check_status(status)
-      unless status.null?
-        message = api[:GetErrorMessage].call(status).read_string
-        api[:ReleaseStatus].call(status)
-        raise Error, message
-      end
-    end
-
-    def node_info(typeinfo)
-      onnx_type = ::FFI::MemoryPointer.new(:int)
-      check_status api[:GetOnnxTypeFromTypeInfo].call(typeinfo.read_pointer, onnx_type)
-
-      type = FFI::OnnxType[onnx_type.read_int]
-      case type
-      when :tensor
-        tensor_info = ::FFI::MemoryPointer.new(:pointer)
-        # don't free tensor_info
-        check_status api[:CastTypeInfoToTensorInfo].call(typeinfo.read_pointer, tensor_info)
-
-        type, shape = tensor_type_and_shape(tensor_info)
-        {
-          type: "tensor(#{FFI::TensorElementDataType[type]})",
-          shape: shape
-        }
-      when :sequence
-        sequence_type_info = ::FFI::MemoryPointer.new(:pointer)
-        check_status api[:CastTypeInfoToSequenceTypeInfo].call(typeinfo.read_pointer, sequence_type_info)
-        nested_type_info = ::FFI::MemoryPointer.new(:pointer)
-        check_status api[:GetSequenceElementType].call(sequence_type_info.read_pointer, nested_type_info)
-        v = node_info(nested_type_info)[:type]
-
-        {
-          type: "seq(#{v})",
-          shape: []
-        }
-      when :map
-        map_type_info = ::FFI::MemoryPointer.new(:pointer)
-        check_status api[:CastTypeInfoToMapTypeInfo].call(typeinfo.read_pointer, map_type_info)
-
-        # key
-        key_type = ::FFI::MemoryPointer.new(:int)
-        check_status api[:GetMapKeyType].call(map_type_info.read_pointer, key_type)
-        k = FFI::TensorElementDataType[key_type.read_int]
-
-        # value
-        value_type_info = ::FFI::MemoryPointer.new(:pointer)
-        check_status api[:GetMapValueType].call(map_type_info.read_pointer, value_type_info)
-        v = node_info(value_type_info)[:type]
-
-        {
-          type: "map(#{k},#{v})",
-          shape: []
-        }
-      else
-        unsupported_type("ONNX", type)
-      end
-    ensure
-      release :TypeInfo, typeinfo
-    end
-
-    def tensor_type_and_shape(tensor_info)
-      type = ::FFI::MemoryPointer.new(:int)
-      check_status api[:GetTensorElementType].call(tensor_info.read_pointer, type)
-
-      num_dims_ptr = ::FFI::MemoryPointer.new(:size_t)
-      check_status api[:GetDimensionsCount].call(tensor_info.read_pointer, num_dims_ptr)
-      num_dims = num_dims_ptr.read(:size_t)
-
-      node_dims = ::FFI::MemoryPointer.new(:int64, num_dims)
-      check_status api[:GetDimensions].call(tensor_info.read_pointer, node_dims, num_dims)
-      dims = node_dims.read_array_of_int64(num_dims)
-
-      symbolic_dims = ::FFI::MemoryPointer.new(:pointer, num_dims)
-      check_status api[:GetSymbolicDimensions].call(tensor_info.read_pointer, symbolic_dims, num_dims)
-      named_dims = num_dims.times.map { |i| symbolic_dims[i].read_pointer.read_string }
-      dims = named_dims.zip(dims).map { |n, d| n.empty? ? d : n }
-
-      [type.read_int, dims]
-    end
-
-    def unsupported_type(name, type)
-      raise Error, "Unsupported #{name} type: #{type}"
+      Utils.check_status(status)
     end
 
     def tensor_types
       @tensor_types ||= [:float, :uint8, :int8, :uint16, :int16, :int32, :int64, :bool, :double, :uint32, :uint64].map { |v| ["tensor(#{v})", v] }.to_h
     end
 
-    def numo_array?(obj)
-      defined?(Numo::NArray) && obj.is_a?(Numo::NArray)
-    end
-
-    def numo_types
-      @numo_types ||= {
-        float: Numo::SFloat,
-        uint8: Numo::UInt8,
-        int8: Numo::Int8,
-        uint16: Numo::UInt16,
-        int16: Numo::Int16,
-        int32: Numo::Int32,
-        int64: Numo::Int64,
-        bool: Numo::UInt8,
-        double: Numo::DFloat,
-        uint32: Numo::UInt32,
-        uint64: Numo::UInt64
-      }
-    end
-
     def api
       self.class.api
     end
@@ -590,7 +314,7 @@ module OnnxRuntime
     end
 
     def self.release(type, pointer)
-      api[:"Release#{type}"].call(pointer.read_pointer) if pointer && !pointer.null?
+      Utils.release(type, pointer)
     end
 
     def self.finalize(addr)