fix window stride and size
Browse files- iwslt2011.py +10 -26
iwslt2011.py
CHANGED
@@ -65,44 +65,25 @@ class TaggingTask:
|
|
65 |
"""Treat punctuation prediction as a sequence tagging problem."""
|
66 |
|
67 |
def __init__(
|
68 |
-
self, window_size=120,
|
69 |
):
|
70 |
self.window_size = window_size
|
71 |
-
self.
|
72 |
-
self.include_reference = include_reference
|
73 |
|
74 |
def __eq__(self, other):
|
75 |
return Task.TAGGING == other
|
76 |
|
77 |
-
class DecodingStrategy:
|
78 |
-
"""Strategy used to decode results."""
|
79 |
-
|
80 |
-
def __init__(
|
81 |
-
self, task: Union[TaggingTask]
|
82 |
-
):
|
83 |
-
self.task = task
|
84 |
-
|
85 |
-
@abstractmethod
|
86 |
-
def decode():
|
87 |
-
pass
|
88 |
-
|
89 |
-
class AverageDecodingStrategy:
|
90 |
-
"""Averages predictions together."""
|
91 |
-
|
92 |
-
def decode():
|
93 |
-
pass
|
94 |
-
|
95 |
class IWSLT11Config(datasets.BuilderConfig):
|
96 |
"""The IWSLT11 Dataset."""
|
97 |
|
98 |
def __init__(
|
99 |
self,
|
100 |
-
task: Union[TaggingTask] = TaggingTask(),
|
101 |
segmented: bool = False,
|
102 |
asr_or_ref: str = "ref",
|
103 |
-
decoder: DecodingStrategy = AverageDecodingStrategy(),
|
104 |
tokenizer = None,
|
105 |
label_subword = LabelSubword.IGNORE,
|
|
|
|
|
106 |
**kwargs
|
107 |
):
|
108 |
"""BuilderConfig for IWSLT2011.
|
@@ -111,10 +92,13 @@ class IWSLT11Config(datasets.BuilderConfig):
|
|
111 |
segmented: if segmentation present in IWSLT2011 should be respected. removes segmenation by default.
|
112 |
**kwargs: keyword arguments forwarded to super.
|
113 |
"""
|
114 |
-
self.task =
|
|
|
|
|
|
|
|
|
115 |
self.segmented = segmented
|
116 |
self.asr_or_ref = asr_or_ref
|
117 |
-
self.decoder = decoder
|
118 |
self.punctuation = [
|
119 |
Punctuation.NONE,
|
120 |
Punctuation.PERIOD,
|
@@ -258,7 +242,7 @@ class IWSLT11(datasets.GeneratorBasedBuilder):
|
|
258 |
return window(
|
259 |
l,
|
260 |
self.config.task.window_size,
|
261 |
-
|
262 |
)
|
263 |
ids = apply_window(np.arange(len(tokens)))
|
264 |
tokens = apply_window(tokens)
|
|
|
65 |
"""Treat punctuation prediction as a sequence tagging problem."""
|
66 |
|
67 |
def __init__(
|
68 |
+
self, window_size=120, window_stride=60
|
69 |
):
|
70 |
self.window_size = window_size
|
71 |
+
self.window_stride = window_stride
|
|
|
72 |
|
73 |
def __eq__(self, other):
|
74 |
return Task.TAGGING == other
|
75 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
76 |
class IWSLT11Config(datasets.BuilderConfig):
|
77 |
"""The IWSLT11 Dataset."""
|
78 |
|
79 |
def __init__(
|
80 |
self,
|
|
|
81 |
segmented: bool = False,
|
82 |
asr_or_ref: str = "ref",
|
|
|
83 |
tokenizer = None,
|
84 |
label_subword = LabelSubword.IGNORE,
|
85 |
+
window_size = None,
|
86 |
+
window_stride = None,
|
87 |
**kwargs
|
88 |
):
|
89 |
"""BuilderConfig for IWSLT2011.
|
|
|
92 |
segmented: if segmentation present in IWSLT2011 should be respected. removes segmenation by default.
|
93 |
**kwargs: keyword arguments forwarded to super.
|
94 |
"""
|
95 |
+
self.task = TaggingTask()
|
96 |
+
if window_size is not None:
|
97 |
+
self.task.window_size = window_size
|
98 |
+
if window_stride is not None:
|
99 |
+
self.task.window_stride = window_stride
|
100 |
self.segmented = segmented
|
101 |
self.asr_or_ref = asr_or_ref
|
|
|
102 |
self.punctuation = [
|
103 |
Punctuation.NONE,
|
104 |
Punctuation.PERIOD,
|
|
|
242 |
return window(
|
243 |
l,
|
244 |
self.config.task.window_size,
|
245 |
+
self.config.task.window_stride
|
246 |
)
|
247 |
ids = apply_window(np.arange(len(tokens)))
|
248 |
tokens = apply_window(tokens)
|