torch-rb 0.2.6 → 0.2.7

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
checksums.yaml CHANGED
@@ -1,7 +1,7 @@
1
1
  ---
2
2
  SHA256:
3
- metadata.gz: da5c88539a2890933e44859af7c1acfe835405c89e82bc6a6fb2df37fdee141a
4
- data.tar.gz: 747ab48ba1b0ba16077ed31cf505f622a4120d18be9c0942ded39810095aa68e
3
+ metadata.gz: 3451d6140ae6a6a9294a73571239df703a9dc753911c5d97a83bcb020b9d878d
4
+ data.tar.gz: 65689090d9fe4d9dee078b2f0f0f56526d76158306390c0988e61b0e2ca98ff1
5
5
  SHA512:
6
- metadata.gz: d2bccf16e7af54d53affbc12030fd89f417fd060afa7057e97765bca00c24b7089f74b9e8aa4bab9180045e466649676f481425dc24ab29533b74138bb03e786
7
- data.tar.gz: 8eb49a743fedb220df4edc39d7d0c492c1827ce63f2c5d9a8d73495da63da5f6055674ca0bd60bb0b81be68db6e5c9300357de2c417e82e7af4fafd1ab6c7ca2
6
+ metadata.gz: 9f2cc800b8c0e7a3a75bbb9c4705e7e306ed68f52a90530d22659a4d23d8ce0126c1cfd9bc7c33612a842bc20199ba8ec7f488bbba591073d8914f108948e084
7
+ data.tar.gz: dbf34592bef6e869a3814f20e891d2d566339080a46d335a5e42f114477a5769f63ee18ca3ee8b8f1d031faf898dfe4f6861064f0cb0773b6d75622b4a663e0f
@@ -1,3 +1,8 @@
1
+ ## 0.2.7 (2020-06-29)
2
+
3
+ - Made tensors enumerable
4
+ - Improved performance of `inspect` method
5
+
1
6
  ## 0.2.6 (2020-06-29)
2
7
 
3
8
  - Added support for indexing with tensors
@@ -1,89 +1,264 @@
1
+ # mirrors _tensor_str.py
1
2
  module Torch
2
3
  module Inspector
3
- # TODO make more performant, especially when summarizing
4
- # how? only read data that will be displayed
5
- def inspect
6
- data =
7
- if numel == 0
8
- "[]"
9
- elsif dim == 0
10
- item
4
+ PRINT_OPTS = {
5
+ precision: 4,
6
+ threshold: 1000,
7
+ edgeitems: 3,
8
+ linewidth: 80,
9
+ sci_mode: nil
10
+ }
11
+
12
+ class Formatter
13
+ def initialize(tensor)
14
+ @floating_dtype = tensor.floating_point?
15
+ @complex_dtype = tensor.complex?
16
+ @int_mode = true
17
+ @sci_mode = false
18
+ @max_width = 1
19
+
20
+ tensor_view = Torch.no_grad { tensor.reshape(-1) }
21
+
22
+ if !@floating_dtype
23
+ tensor_view.each do |value|
24
+ value_str = value.item.to_s
25
+ @max_width = [@max_width, value_str.length].max
26
+ end
11
27
  else
12
- summarize = numel > 1000
28
+ nonzero_finite_vals = Torch.masked_select(tensor_view, Torch.isfinite(tensor_view) & tensor_view.ne(0))
29
+
30
+ # no valid number, do nothing
31
+ return if nonzero_finite_vals.numel == 0
32
+
33
+ # Convert to double for easy calculation. HalfTensor overflows with 1e8, and there's no div() on CPU.
34
+ nonzero_finite_abs = nonzero_finite_vals.abs.double
35
+ nonzero_finite_min = nonzero_finite_abs.min.double
36
+ nonzero_finite_max = nonzero_finite_abs.max.double
37
+
38
+ nonzero_finite_vals.each do |value|
39
+ if value.item != value.item.ceil
40
+ @int_mode = false
41
+ break
42
+ end
43
+ end
13
44
 
14
- if dtype == :bool
15
- fmt = "%s"
45
+ if @int_mode
46
+ # in int_mode for floats, all numbers are integers, and we append a decimal to nonfinites
47
+ # to indicate that the tensor is of floating type. add 1 to the len to account for this.
48
+ if nonzero_finite_max / nonzero_finite_min > 1000.0 || nonzero_finite_max > 1.0e8
49
+ @sci_mode = true
50
+ nonzero_finite_vals.each do |value|
51
+ value_str = "%.#{PRINT_OPTS[:precision]}e" % value.item
52
+ @max_width = [@max_width, value_str.length].max
53
+ end
54
+ else
55
+ nonzero_finite_vals.each do |value|
56
+ value_str = "%.0f" % value.item
57
+ @max_width = [@max_width, value_str.length + 1].max
58
+ end
59
+ end
16
60
  else
17
- values = _flat_data
18
- abs = values.select { |v| v != 0 }.map(&:abs)
19
- max = abs.max || 1
20
- min = abs.min || 1
21
-
22
- total = 0
23
- if values.any? { |v| v < 0 }
24
- total += 1
61
+ # Check if scientific representation should be used.
62
+ if nonzero_finite_max / nonzero_finite_min > 1000.0 || nonzero_finite_max > 1.0e8 || nonzero_finite_min < 1.0e-4
63
+ @sci_mode = true
64
+ nonzero_finite_vals.each do |value|
65
+ value_str = "%.#{PRINT_OPTS[:precision]}e" % value.item
66
+ @max_width = [@max_width, value_str.length].max
67
+ end
68
+ else
69
+ nonzero_finite_vals.each do |value|
70
+ value_str = "%.#{PRINT_OPTS[:precision]}f" % value.item
71
+ @max_width = [@max_width, value_str.length].max
72
+ end
25
73
  end
74
+ end
75
+ end
26
76
 
27
- if floating_point?
28
- sci = max > 1e8 || max < 1e-4
77
+ @sci_mode = PRINT_OPTS[:sci_mode] unless PRINT_OPTS[:sci_mode].nil?
78
+ end
29
79
 
30
- all_int = values.all? { |v| v.finite? && v == v.to_i }
31
- decimal = all_int ? 1 : 4
80
+ def width
81
+ @max_width
82
+ end
32
83
 
33
- total += sci ? 10 : decimal + 1 + max.to_i.to_s.size
84
+ def format(value)
85
+ value = value.item
34
86
 
35
- if sci
36
- fmt = "%#{total}.4e"
37
- else
38
- fmt = "%#{total}.#{decimal}f"
39
- end
40
- else
41
- total += max.to_s.size
42
- fmt = "%#{total}d"
87
+ if @floating_dtype
88
+ if @sci_mode
89
+ ret = "%#{@max_width}.#{PRINT_OPTS[:precision]}e" % value
90
+ elsif @int_mode
91
+ ret = String.new("%.0f" % value)
92
+ unless value.infinite? || value.nan?
93
+ ret += "."
43
94
  end
95
+ else
96
+ ret = "%.#{PRINT_OPTS[:precision]}f" % value
44
97
  end
98
+ elsif @complex_dtype
99
+ p = PRINT_OPTS[:precision]
100
+ raise NotImplementedYet
101
+ else
102
+ ret = value.to_s
103
+ end
104
+ # Ruby throws error when negative, Python doesn't
105
+ " " * [@max_width - ret.size, 0].max + ret
106
+ end
107
+ end
108
+
109
+ def inspect
110
+ Torch.no_grad do
111
+ str_intern(self)
112
+ end
113
+ rescue => e
114
+ # prevent stack error
115
+ puts e.backtrace.join("\n")
116
+ "Error inspecting tensor: #{e.inspect}"
117
+ end
118
+
119
+ private
120
+
121
+ # TODO update
122
+ def str_intern(slf)
123
+ prefix = "tensor("
124
+ indent = prefix.length
125
+ suffixes = []
126
+
127
+ has_default_dtype = [:float32, :int64, :bool].include?(slf.dtype)
128
+
129
+ if slf.numel == 0 && !slf.sparse?
130
+ # Explicitly print the shape if it is not (0,), to match NumPy behavior
131
+ if slf.dim != 1
132
+ suffixes << "size: #{shape.inspect}"
133
+ end
45
134
 
46
- inspect_level(to_a, fmt, dim - 1, 0, summarize)
135
+ # In an empty tensor, there are no elements to infer if the dtype
136
+ # should be int64, so it must be shown explicitly.
137
+ if slf.dtype != :int64
138
+ suffixes << "dtype: #{slf.dtype.inspect}"
47
139
  end
140
+ tensor_str = "[]"
141
+ else
142
+ if !has_default_dtype
143
+ suffixes << "dtype: #{slf.dtype.inspect}"
144
+ end
145
+
146
+ if slf.layout != :strided
147
+ tensor_str = tensor_str(slf.to_dense, indent)
148
+ else
149
+ tensor_str = tensor_str(slf, indent)
150
+ end
151
+ end
48
152
 
49
- attributes = []
50
- if requires_grad
51
- attributes << "requires_grad: true"
153
+ if slf.layout != :strided
154
+ suffixes << "layout: #{slf.layout.inspect}"
52
155
  end
53
- if ![:float32, :int64, :bool].include?(dtype)
54
- attributes << "dtype: #{dtype.inspect}"
156
+
157
+ # TODO show grad_fn
158
+ if slf.requires_grad?
159
+ suffixes << "requires_grad: true"
55
160
  end
56
161
 
57
- "tensor(#{data}#{attributes.map { |a| ", #{a}" }.join("")})"
162
+ add_suffixes(prefix + tensor_str, suffixes, indent, slf.sparse?)
58
163
  end
59
164
 
60
- private
165
+ def add_suffixes(tensor_str, suffixes, indent, force_newline)
166
+ tensor_strs = [tensor_str]
167
+ # rfind in Python returns -1 when not found
168
+ last_line_len = tensor_str.length - (tensor_str.rindex("\n") || -1) + 1
169
+ suffixes.each do |suffix|
170
+ suffix_len = suffix.length
171
+ if force_newline || last_line_len + suffix_len + 2 > PRINT_OPTS[:linewidth]
172
+ tensor_strs << ",\n" + " " * indent + suffix
173
+ last_line_len = indent + suffix_len
174
+ force_newline = false
175
+ else
176
+ tensor_strs.append(", " + suffix)
177
+ last_line_len += suffix_len + 2
178
+ end
179
+ end
180
+ tensor_strs.append(")")
181
+ tensor_strs.join("")
182
+ end
61
183
 
62
- # TODO DRY code
63
- def inspect_level(arr, fmt, total, level, summarize)
64
- if level == total
65
- cols =
66
- if summarize && arr.size > 7
67
- arr[0..2].map { |v| fmt % v } +
68
- ["..."] +
69
- arr[-3..-1].map { |v| fmt % v }
70
- else
71
- arr.map { |v| fmt % v }
72
- end
184
+ def tensor_str(slf, indent)
185
+ return "[]" if slf.numel == 0
186
+
187
+ summarize = slf.numel > PRINT_OPTS[:threshold]
188
+
189
+ if slf.dtype == :float16 || slf.dtype == :bfloat16
190
+ slf = slf.float
191
+ end
192
+ formatter = Formatter.new(summarize ? summarized_data(slf) : slf)
193
+ tensor_str_with_formatter(slf, indent, formatter, summarize)
194
+ end
195
+
196
+ def summarized_data(slf)
197
+ edgeitems = PRINT_OPTS[:edgeitems]
73
198
 
74
- "[#{cols.join(", ")}]"
199
+ dim = slf.dim
200
+ if dim == 0
201
+ slf
202
+ elsif dim == 1
203
+ if size(0) > 2 * edgeitems
204
+ Torch.cat([slf[0...edgeitems], slf[-edgeitems..-1]])
205
+ else
206
+ slf
207
+ end
208
+ elsif slf.size(0) > 2 * edgeitems
209
+ start = edgeitems.times.map { |i| slf[i] }
210
+ finish = (slf.length - edgeitems).upto(slf.length - 1).map { |i| slf[i] }
211
+ Torch.stack((start + finish).map { |x| summarized_data(x) })
75
212
  else
76
- rows =
77
- if summarize && arr.size > 7
78
- arr[0..2].map { |row| inspect_level(row, fmt, total, level + 1, summarize) } +
79
- ["..."] +
80
- arr[-3..-1].map { |row| inspect_level(row, fmt, total, level + 1, summarize) }
81
- else
82
- arr.map { |row| inspect_level(row, fmt, total, level + 1, summarize) }
83
- end
213
+ Torch.stack(slf.map { |x| summarized_data(x) })
214
+ end
215
+ end
216
+
217
+ def tensor_str_with_formatter(slf, indent, formatter, summarize)
218
+ edgeitems = PRINT_OPTS[:edgeitems]
219
+
220
+ dim = slf.dim
84
221
 
85
- "[#{rows.join(",#{"\n" * (total - level)}#{" " * (level + 8)}")}]"
222
+ return scalar_str(slf, formatter) if dim == 0
223
+ return vector_str(slf, indent, formatter, summarize) if dim == 1
224
+
225
+ if summarize && slf.size(0) > 2 * edgeitems
226
+ slices = (
227
+ [edgeitems.times.map { |i| tensor_str_with_formatter(slf[i], indent + 1, formatter, summarize) }] +
228
+ ["..."] +
229
+ [((slf.length - edgeitems)...slf.length).map { |i| tensor_str_with_formatter(slf[i], indent + 1, formatter, summarize) }]
230
+ )
231
+ else
232
+ slices = slf.size(0).times.map { |i| tensor_str_with_formatter(slf[i], indent + 1, formatter, summarize) }
86
233
  end
234
+
235
+ tensor_str = slices.join("," + "\n" * (dim - 1) + " " * (indent + 1))
236
+ "[" + tensor_str + "]"
237
+ end
238
+
239
+ def scalar_str(slf, formatter)
240
+ formatter.format(slf)
241
+ end
242
+
243
+ def vector_str(slf, indent, formatter, summarize)
244
+ # length includes spaces and comma between elements
245
+ element_length = formatter.width + 2
246
+ elements_per_line = [1, ((PRINT_OPTS[:linewidth] - indent) / element_length.to_f).floor.to_i].max
247
+ char_per_line = element_length * elements_per_line
248
+
249
+ if summarize && slf.size(0) > 2 * PRINT_OPTS[:edgeitems]
250
+ data = (
251
+ [slf[0...PRINT_OPTS[:edgeitems]].map { |val| formatter.format(val) }] +
252
+ [" ..."] +
253
+ [slf[-PRINT_OPTS[:edgeitems]..-1].map { |val| formatter.format(val) }]
254
+ )
255
+ else
256
+ data = slf.map { |val| formatter.format(val) }
257
+ end
258
+
259
+ data_lines = (0...data.length).step(elements_per_line).map { |i| data[i...(i + elements_per_line)] }
260
+ lines = data_lines.map { |line| line.join(", ") }
261
+ "[" + lines.join("," + "\n" + " " * (indent + 1)) + "]"
87
262
  end
88
263
  end
89
264
  end
@@ -286,8 +286,11 @@ module Torch
286
286
  str % vars
287
287
  end
288
288
 
289
+ # used for format
290
+ # remove tensors for performance
291
+ # so we can skip call to inspect
289
292
  def dict
290
- instance_variables.map { |k| [k[1..-1].to_sym, instance_variable_get(k)] }.to_h
293
+ instance_variables.reject { |k| instance_variable_get(k).is_a?(Tensor) }.map { |k| [k[1..-1].to_sym, instance_variable_get(k)] }.to_h
291
294
  end
292
295
  end
293
296
  end
@@ -1,6 +1,7 @@
1
1
  module Torch
2
2
  class Tensor
3
3
  include Comparable
4
+ include Enumerable
4
5
  include Inspector
5
6
 
6
7
  alias_method :requires_grad?, :requires_grad
@@ -25,6 +26,14 @@ module Torch
25
26
  inspect
26
27
  end
27
28
 
29
+ def each
30
+ return enum_for(:each) unless block_given?
31
+
32
+ size(0).times do |i|
33
+ yield self[i]
34
+ end
35
+ end
36
+
28
37
  # TODO make more performant
29
38
  def to_a
30
39
  arr = _flat_data
@@ -153,6 +162,18 @@ module Torch
153
162
  neg
154
163
  end
155
164
 
165
+ def &(other)
166
+ logical_and(other)
167
+ end
168
+
169
+ def |(other)
170
+ logical_or(other)
171
+ end
172
+
173
+ def ^(other)
174
+ logical_xor(other)
175
+ end
176
+
156
177
  # TODO better compare?
157
178
  def <=>(other)
158
179
  item <=> other
@@ -1,3 +1,3 @@
1
1
  module Torch
2
- VERSION = "0.2.6"
2
+ VERSION = "0.2.7"
3
3
  end
metadata CHANGED
@@ -1,14 +1,14 @@
1
1
  --- !ruby/object:Gem::Specification
2
2
  name: torch-rb
3
3
  version: !ruby/object:Gem::Version
4
- version: 0.2.6
4
+ version: 0.2.7
5
5
  platform: ruby
6
6
  authors:
7
7
  - Andrew Kane
8
8
  autorequire:
9
9
  bindir: bin
10
10
  cert_chain: []
11
- date: 2020-06-29 00:00:00.000000000 Z
11
+ date: 2020-06-30 00:00:00.000000000 Z
12
12
  dependencies:
13
13
  - !ruby/object:Gem::Dependency
14
14
  name: rice