torch-rb 0.2.5 → 0.3.2

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 8f6ab78fb5cff27d0d60ddb9c08fb2f526bd60e241dd1011554b21716bdd2f43
-  data.tar.gz: 5568f53d8d5d688e3f29fb55ddbe9457e0b933dc69f59b422035c0cee249e396
+  metadata.gz: 97908e85a67729120f763bb4140323505b77a831d5648e9d2d0961259e3d300c
+  data.tar.gz: f366548f9880dac7dffce6305e192f75a7467526ae55ae13af05d355918375ba
 SHA512:
-  metadata.gz: 9c5dcfbf35382b37678662690677b2b90d0d544e8802703cf83dba6e10483a4df487f9687e4e898c9cc449c568f4e02f9d831daa982b5b3135af6f9ce176ec88
-  data.tar.gz: 34f142d874606e140661ae992a9f8cd4779f95c93c11d9a89a1864dd0bd53c5480c30d9aec5897f1955e2450bd0a3bc56ed0868e3b54d82ff4cdba40af379840
+  metadata.gz: cc32ddbc43131175452a8b62df5d1eac6bc8450eea174018affd5cd073f81e2c9825d613014c62c5f8137cf5dddd1ab6ab6de60a4b3a67a757387446dbc1efad
+  data.tar.gz: c322e0b7ec7f03f12311d737034dad45037d2ad7710974e24250e11f4a0db14e221e870ddc52c0f2b723476be6f41fca8a8719068b0d0b7d8974d2080e9be6dc

data/CHANGELOG.md

@@ -1,3 +1,33 @@
+## 0.3.2 (2020-08-24)
+
+- Added `enable_grad` method
+- Added `random_split` method
+- Added `collate_fn` option to `DataLoader`
+- Added `grad=` method to `Tensor`
+- Fixed error with `grad` method when empty
+- Fixed `EmbeddingBag`
+
+## 0.3.1 (2020-08-17)
+
+- Added `create_graph` and `retain_graph` options to `backward` method
+- Fixed error when `set` not required
+
+## 0.3.0 (2020-07-29)
+
+- Updated LibTorch to 1.6.0
+- Removed `state_dict` method from optimizers until `load_state_dict` is implemented
+
+## 0.2.7 (2020-06-29)
+
+- Made tensors enumerable
+- Improved performance of `inspect` method
+
+## 0.2.6 (2020-06-29)
+
+- Added support for indexing with tensors
+- Added `contiguous` methods
+- Fixed named parameters for nested parameters
+
 ## 0.2.5 (2020-06-07)
 
 - Added `download_url_to_file` and `load_state_dict_from_url` to `Torch::Hub`
@@ -32,7 +62,7 @@
 ## 0.2.0 (2020-04-22)
 
 - No longer experimental
-- Updated libtorch to 1.5.0
+- Updated LibTorch to 1.5.0
 - Added support for GPUs and OpenMP
 - Added adaptive pooling layers
 - Tensor `dtype` is now based on Numo type for `Torch.tensor`
@@ -41,7 +71,7 @@
 
 ## 0.1.8 (2020-01-17)
 
-- Updated libtorch to 1.4.0
+- Updated LibTorch to 1.4.0
 
 ## 0.1.7 (2020-01-10)
 

data/README.md

@@ -42,6 +42,8 @@ This library follows the [PyTorch API](https://pytorch.org/docs/stable/torch.htm
 - Methods that return booleans use `?` instead of `is_`  (`tensor?` instead of `is_tensor`)
 - Numo is used instead of NumPy (`x.numo` instead of `x.numpy()`)
 
+You can follow PyTorch tutorials and convert the code to Ruby in many cases. Feel free to open an issue if you run into problems.
+
 ## Tutorial
 
 Some examples below are from [Deep Learning with PyTorch: A 60 Minutes Blitz](https://pytorch.org/tutorials/beginner/deep_learning_60min_blitz.html)
@@ -409,7 +411,8 @@ Here’s the list of compatible versions.
 
 Torch.rb | LibTorch
 --- | ---
-0.2.0 | 1.5.0
+0.3.0-0.3.1 | 1.6.0
+0.2.0-0.2.7 | 1.5.0-1.5.1
 0.1.8 | 1.4.0
 0.1.0-0.1.7 | 1.3.1
 

data/ext/torch/ext.cpp

@@ -48,13 +48,14 @@ void Init_ext()
     .define_singleton_method(
       "initial_seed",
       *[]() {
-        return at::detail::getDefaultCPUGenerator()->current_seed();
+        return at::detail::getDefaultCPUGenerator().current_seed();
       })
     .define_singleton_method(
       "seed",
       *[]() {
         // TODO set for CUDA when available
-        return at::detail::getDefaultCPUGenerator()->seed();
+        auto generator = at::detail::getDefaultCPUGenerator();
+        return generator.seed();
       });
 
   // https://pytorch.org/cppdocs/api/structc10_1_1_i_value.html
@@ -329,6 +330,11 @@ void Init_ext()
     .define_method("numel", &torch::Tensor::numel)
     .define_method("element_size", &torch::Tensor::element_size)
     .define_method("requires_grad", &torch::Tensor::requires_grad)
+    .define_method(
+      "contiguous?",
+      *[](Tensor& self) {
+        return self.is_contiguous();
+      })
     .define_method(
       "addcmul!",
       *[](Tensor& self, Scalar value, const Tensor & tensor1, const Tensor & tensor2) {
@@ -346,13 +352,19 @@ void Init_ext()
       })
     .define_method(
       "_backward",
-      *[](Tensor& self, Object gradient) {
-        return gradient.is_nil() ? self.backward() : self.backward(from_ruby<torch::Tensor>(gradient));
+      *[](Tensor& self, OptionalTensor gradient, bool create_graph, bool retain_graph) {
+        return self.backward(gradient, create_graph, retain_graph);
       })
     .define_method(
       "grad",
       *[](Tensor& self) {
-        return self.grad();
+        auto grad = self.grad();
+        return grad.defined() ? to_ruby<torch::Tensor>(grad) : Nil;
+      })
+    .define_method(
+      "grad=",
+      *[](Tensor& self, torch::Tensor& grad) {
+        self.grad() = grad;
       })
     .define_method(
       "_dtype",
@@ -455,7 +467,7 @@ void Init_ext()
     .define_singleton_method(
       "_make_subclass",
       *[](Tensor& rd, bool requires_grad) {
-        auto data = torch::autograd::as_variable_ref(rd).detach();
+        auto data = rd.detach();
         data.unsafeGetTensorImpl()->set_allow_tensor_metadata_change(true);
         auto var = data.set_requires_grad(requires_grad);
         return Parameter(std::move(var));
@@ -574,6 +586,11 @@ void Init_ext()
       *[](Parameter& self) {
         auto grad = self.grad();
         return grad.defined() ? to_ruby<torch::Tensor>(grad) : Nil;
+      })
+    .define_method(
+      "grad=",
+      *[](Parameter& self, torch::Tensor& grad) {
+        self.grad() = grad;
       });
 
   Class rb_cDevice = define_class_under<torch::Device>(rb_mTorch, "Device")

data/ext/torch/extconf.rb

@@ -7,17 +7,16 @@ $CXXFLAGS += " -std=c++14"
 # change to 0 for Linux pre-cxx11 ABI version
 $CXXFLAGS += " -D_GLIBCXX_USE_CXX11_ABI=1"
 
-# TODO check compiler name
-clang = RbConfig::CONFIG["host_os"] =~ /darwin/i
+apple_clang = RbConfig::CONFIG["CC_VERSION_MESSAGE"] =~ /apple clang/i
 
 # check omp first
 if have_library("omp") || have_library("gomp")
   $CXXFLAGS += " -DAT_PARALLEL_OPENMP=1"
-  $CXXFLAGS += " -Xclang" if clang
+  $CXXFLAGS += " -Xclang" if apple_clang
   $CXXFLAGS += " -fopenmp"
 end
 
-if clang
+if apple_clang
   # silence ruby/intern.h warning
   $CXXFLAGS += " -Wno-deprecated-register"
 

data/lib/torch.rb

@@ -4,6 +4,7 @@ require "torch/ext"
 # stdlib
 require "fileutils"
 require "net/http"
+require "set"
 require "tmpdir"
 
 # native functions
@@ -178,8 +179,10 @@ require "torch/nn/functional"
 require "torch/nn/init"
 
 # utils
+require "torch/utils/data"
 require "torch/utils/data/data_loader"
 require "torch/utils/data/dataset"
+require "torch/utils/data/subset"
 require "torch/utils/data/tensor_dataset"
 
 # hub
@@ -315,6 +318,16 @@ module Torch
       end
     end
 
+    def enable_grad
+      previous_value = grad_enabled?
+      begin
+        _set_grad_enabled(true)
+        yield
+      ensure
+        _set_grad_enabled(previous_value)
+      end
+    end
+
     def device(str)
       Device.new(str)
     end
@@ -470,11 +483,7 @@ module Torch
       when nil
         IValue.new
       when Array
-        if obj.all? { |v| v.is_a?(Tensor) }
-          IValue.from_list(obj.map { |v| IValue.from_tensor(v) })
-        else
-          raise Error, "Unknown list type"
-        end
+        IValue.from_list(obj.map { |v| to_ivalue(v) })
       else
         raise Error, "Unknown type: #{obj.class.name}"
       end

data/lib/torch/hub.rb

@@ -7,25 +7,26 @@ module Torch
 
       def download_url_to_file(url, dst)
         uri = URI(url)
-        tmp = "#{Dir.tmpdir}/#{Time.now.to_f}" # TODO better name
+        tmp = nil
         location = nil
 
+        puts "Downloading #{url}..."
         Net::HTTP.start(uri.host, uri.port, use_ssl: uri.scheme == "https") do |http|
           request = Net::HTTP::Get.new(uri)
 
-          puts "Downloading #{url}..."
-          File.open(tmp, "wb") do |f|
-            http.request(request) do |response|
-              case response
-              when Net::HTTPRedirection
-                location = response["location"]
-              when Net::HTTPSuccess
+          http.request(request) do |response|
+            case response
+            when Net::HTTPRedirection
+              location = response["location"]
+            when Net::HTTPSuccess
+              tmp = "#{Dir.tmpdir}/#{Time.now.to_f}" # TODO better name
+              File.open(tmp, "wb") do |f|
                 response.read_body do |chunk|
                   f.write(chunk)
                 end
-              else
-                raise Error, "Bad response"
               end
+            else
+              raise Error, "Bad response"
             end
           end
         end

data/lib/torch/inspector.rb

@@ -1,89 +1,264 @@
+# mirrors _tensor_str.py
 module Torch
   module Inspector
-    # TODO make more performant, especially when summarizing
-    # how? only read data that will be displayed
-    def inspect
-      data =
-        if numel == 0
-          "[]"
-        elsif dim == 0
-          item
+    PRINT_OPTS = {
+      precision: 4,
+      threshold: 1000,
+      edgeitems: 3,
+      linewidth: 80,
+      sci_mode: nil
+    }
+
+    class Formatter
+      def initialize(tensor)
+        @floating_dtype = tensor.floating_point?
+        @complex_dtype = tensor.complex?
+        @int_mode = true
+        @sci_mode = false
+        @max_width = 1
+
+        tensor_view = Torch.no_grad { tensor.reshape(-1) }
+
+        if !@floating_dtype
+          tensor_view.each do |value|
+            value_str = value.item.to_s
+            @max_width = [@max_width, value_str.length].max
+          end
         else
-          summarize = numel > 1000
+          nonzero_finite_vals = Torch.masked_select(tensor_view, Torch.isfinite(tensor_view) & tensor_view.ne(0))
+
+          # no valid number, do nothing
+          return if nonzero_finite_vals.numel == 0
+
+          # Convert to double for easy calculation. HalfTensor overflows with 1e8, and there's no div() on CPU.
+          nonzero_finite_abs = nonzero_finite_vals.abs.double
+          nonzero_finite_min = nonzero_finite_abs.min.double
+          nonzero_finite_max = nonzero_finite_abs.max.double
+
+          nonzero_finite_vals.each do |value|
+            if value.item != value.item.ceil
+              @int_mode = false
+              break
+            end
+          end
 
-          if dtype == :bool
-            fmt = "%s"
+          if @int_mode
+            # in int_mode for floats, all numbers are integers, and we append a decimal to nonfinites
+            # to indicate that the tensor is of floating type. add 1 to the len to account for this.
+            if nonzero_finite_max / nonzero_finite_min > 1000.0 || nonzero_finite_max > 1.0e8
+              @sci_mode = true
+              nonzero_finite_vals.each do |value|
+                value_str = "%.#{PRINT_OPTS[:precision]}e" % value.item
+                @max_width = [@max_width, value_str.length].max
+              end
+            else
+              nonzero_finite_vals.each do |value|
+                value_str = "%.0f" % value.item
+                @max_width = [@max_width, value_str.length + 1].max
+              end
+            end
           else
-            values = _flat_data
-            abs = values.select { |v| v != 0 }.map(&:abs)
-            max = abs.max || 1
-            min = abs.min || 1
-
-            total = 0
-            if values.any? { |v| v < 0 }
-              total += 1
+            # Check if scientific representation should be used.
+            if nonzero_finite_max / nonzero_finite_min > 1000.0 || nonzero_finite_max > 1.0e8 || nonzero_finite_min < 1.0e-4
+              @sci_mode = true
+              nonzero_finite_vals.each do |value|
+                value_str = "%.#{PRINT_OPTS[:precision]}e" % value.item
+                @max_width = [@max_width, value_str.length].max
+              end
+            else
+              nonzero_finite_vals.each do |value|
+                value_str = "%.#{PRINT_OPTS[:precision]}f" % value.item
+                @max_width = [@max_width, value_str.length].max
+              end
             end
+          end
+        end
 
-            if floating_point?
-              sci = max > 1e8 || max < 1e-4
+        @sci_mode = PRINT_OPTS[:sci_mode] unless PRINT_OPTS[:sci_mode].nil?
+      end
 
-              all_int = values.all? { |v| v.finite? && v == v.to_i }
-              decimal = all_int ? 1 : 4
+      def width
+        @max_width
+      end
 
-              total += sci ? 10 : decimal + 1 + max.to_i.to_s.size
+      def format(value)
+        value = value.item
 
-              if sci
-                fmt = "%#{total}.4e"
-              else
-                fmt = "%#{total}.#{decimal}f"
-              end
-            else
-              total += max.to_s.size
-              fmt = "%#{total}d"
+        if @floating_dtype
+          if @sci_mode
+            ret = "%#{@max_width}.#{PRINT_OPTS[:precision]}e" % value
+          elsif @int_mode
+            ret = String.new("%.0f" % value)
+            unless value.infinite? || value.nan?
+              ret += "."
             end
+          else
+            ret = "%.#{PRINT_OPTS[:precision]}f" % value
           end
+        elsif @complex_dtype
+          p = PRINT_OPTS[:precision]
+          raise NotImplementedYet
+        else
+          ret = value.to_s
+        end
+        # Ruby throws error when negative, Python doesn't
+        " " * [@max_width - ret.size, 0].max + ret
+      end
+    end
+
+    def inspect
+      Torch.no_grad do
+        str_intern(self)
+      end
+    rescue => e
+      # prevent stack error
+      puts e.backtrace.join("\n")
+      "Error inspecting tensor: #{e.inspect}"
+    end
+
+    private
+
+    # TODO update
+    def str_intern(slf)
+      prefix = "tensor("
+      indent = prefix.length
+      suffixes = []
+
+      has_default_dtype = [:float32, :int64, :bool].include?(slf.dtype)
+
+      if slf.numel == 0 && !slf.sparse?
+        # Explicitly print the shape if it is not (0,), to match NumPy behavior
+        if slf.dim != 1
+          suffixes << "size: #{shape.inspect}"
+        end
 
-          inspect_level(to_a, fmt, dim - 1, 0, summarize)
+        # In an empty tensor, there are no elements to infer if the dtype
+        # should be int64, so it must be shown explicitly.
+        if slf.dtype != :int64
+          suffixes << "dtype: #{slf.dtype.inspect}"
         end
+        tensor_str = "[]"
+      else
+        if !has_default_dtype
+          suffixes << "dtype: #{slf.dtype.inspect}"
+        end
+
+        if slf.layout != :strided
+          tensor_str = tensor_str(slf.to_dense, indent)
+        else
+          tensor_str = tensor_str(slf, indent)
+        end
+      end
 
-      attributes = []
-      if requires_grad
-        attributes << "requires_grad: true"
+      if slf.layout != :strided
+        suffixes << "layout: #{slf.layout.inspect}"
       end
-      if ![:float32, :int64, :bool].include?(dtype)
-        attributes << "dtype: #{dtype.inspect}"
+
+      # TODO show grad_fn
+      if slf.requires_grad?
+        suffixes << "requires_grad: true"
       end
 
-      "tensor(#{data}#{attributes.map { |a| ", #{a}" }.join("")})"
+      add_suffixes(prefix + tensor_str, suffixes, indent, slf.sparse?)
     end
 
-    private
+    def add_suffixes(tensor_str, suffixes, indent, force_newline)
+      tensor_strs = [tensor_str]
+      # rfind in Python returns -1 when not found
+      last_line_len = tensor_str.length - (tensor_str.rindex("\n") || -1) + 1
+      suffixes.each do |suffix|
+        suffix_len = suffix.length
+        if force_newline || last_line_len + suffix_len + 2 > PRINT_OPTS[:linewidth]
+          tensor_strs << ",\n" + " " * indent + suffix
+          last_line_len = indent + suffix_len
+          force_newline = false
+        else
+          tensor_strs.append(", " + suffix)
+          last_line_len += suffix_len + 2
+        end
+      end
+      tensor_strs.append(")")
+      tensor_strs.join("")
+    end
 
-    # TODO DRY code
-    def inspect_level(arr, fmt, total, level, summarize)
-      if level == total
-        cols =
-          if summarize && arr.size > 7
-            arr[0..2].map { |v| fmt % v } +
-            ["..."] +
-            arr[-3..-1].map { |v| fmt % v }
-          else
-            arr.map { |v| fmt % v }
-          end
+    def tensor_str(slf, indent)
+      return "[]" if slf.numel == 0
+
+      summarize = slf.numel > PRINT_OPTS[:threshold]
+
+      if slf.dtype == :float16 || slf.dtype == :bfloat16
+        slf = slf.float
+      end
+      formatter = Formatter.new(summarize ? summarized_data(slf) : slf)
+      tensor_str_with_formatter(slf, indent, formatter, summarize)
+    end
+
+    def summarized_data(slf)
+      edgeitems = PRINT_OPTS[:edgeitems]
 
-        "[#{cols.join(", ")}]"
+      dim = slf.dim
+      if dim == 0
+        slf
+      elsif dim == 1
+        if size(0) > 2 * edgeitems
+          Torch.cat([slf[0...edgeitems], slf[-edgeitems..-1]])
+        else
+          slf
+        end
+      elsif slf.size(0) > 2 * edgeitems
+        start = edgeitems.times.map { |i| slf[i] }
+        finish = (slf.length - edgeitems).upto(slf.length - 1).map { |i| slf[i] }
+        Torch.stack((start + finish).map { |x| summarized_data(x) })
       else
-        rows =
-          if summarize && arr.size > 7
-            arr[0..2].map { |row| inspect_level(row, fmt, total, level + 1, summarize) } +
-            ["..."] +
-            arr[-3..-1].map { |row| inspect_level(row, fmt, total, level + 1, summarize) }
-          else
-            arr.map { |row| inspect_level(row, fmt, total, level + 1, summarize) }
-          end
+        Torch.stack(slf.map { |x| summarized_data(x) })
+      end
+    end
+
+    def tensor_str_with_formatter(slf, indent, formatter, summarize)
+      edgeitems = PRINT_OPTS[:edgeitems]
+
+      dim = slf.dim
 
-        "[#{rows.join(",#{"\n" * (total - level)}#{" " * (level + 8)}")}]"
+      return scalar_str(slf, formatter) if dim == 0
+      return vector_str(slf, indent, formatter, summarize) if dim == 1
+
+      if summarize && slf.size(0) > 2 * edgeitems
+        slices = (
+          [edgeitems.times.map { |i| tensor_str_with_formatter(slf[i], indent + 1, formatter, summarize) }] +
+          ["..."] +
+          [((slf.length - edgeitems)...slf.length).map { |i| tensor_str_with_formatter(slf[i], indent + 1, formatter, summarize) }]
+        )
+      else
+        slices = slf.size(0).times.map { |i| tensor_str_with_formatter(slf[i], indent + 1, formatter, summarize) }
       end
+
+      tensor_str = slices.join("," + "\n" * (dim - 1) + " " * (indent + 1))
+      "[" + tensor_str + "]"
+    end
+
+    def scalar_str(slf, formatter)
+      formatter.format(slf)
+    end
+
+    def vector_str(slf, indent, formatter, summarize)
+      # length includes spaces and comma between elements
+      element_length = formatter.width + 2
+      elements_per_line = [1, ((PRINT_OPTS[:linewidth] - indent) / element_length.to_f).floor.to_i].max
+      char_per_line = element_length * elements_per_line
+
+      if summarize && slf.size(0) > 2 * PRINT_OPTS[:edgeitems]
+        data = (
+          [slf[0...PRINT_OPTS[:edgeitems]].map { |val| formatter.format(val) }] +
+          [" ..."] +
+          [slf[-PRINT_OPTS[:edgeitems]..-1].map { |val| formatter.format(val) }]
+        )
+      else
+        data = slf.map { |val| formatter.format(val) }
+      end
+
+      data_lines = (0...data.length).step(elements_per_line).map { |i| data[i...(i + elements_per_line)] }
+      lines = data_lines.map { |line| line.join(", ") }
+      "[" + lines.join("," + "\n" + " " * (indent + 1)) + "]"
     end
   end
 end