Sign upLogin

tensorflow/models

View on GitHub
Star
official/nlp/modeling/layers/transformer.py

Summary

Maintainability
D
2 days
Test Coverage

File transformer.py has 411 lines of code (exceeds 250 allowed). Consider refactoring.
Open

# Copyright 2024 The TensorFlow Authors. 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
Severity: Minor
Found in official/nlp/modeling/layers/transformer.py - About 5 hrs to fix

    Function __init__ has 21 arguments (exceeds 4 allowed). Consider refactoring.
    Open

      def __init__(self,
    Severity: Major
    Found in official/nlp/modeling/layers/transformer.py - About 2 hrs to fix

      Function __init__ has 19 arguments (exceeds 4 allowed). Consider refactoring.
      Open

        def __init__(self,
      Severity: Major
      Found in official/nlp/modeling/layers/transformer.py - About 2 hrs to fix

        Function call has a Cognitive Complexity of 14 (exceeds 5 allowed). Consider refactoring.
        Open

          def call(self, inputs, cache=None, decode_loop_step=None):
    if self.multi_channel_cross_attention:
      if len(inputs) != 5:
        raise ValueError(
            "TransformerDecoderBlock must have 5 inputs, when it uses "
        Severity: Minor
        Found in official/nlp/modeling/layers/transformer.py - About 1 hr to fix

        Cognitive Complexity

        Cognitive Complexity is a measure of how difficult a unit of code is to intuitively understand. Unlike Cyclomatic Complexity, which determines how difficult your code will be to test, Cognitive Complexity tells you how difficult your code will be to read and comprehend.

        A method's cognitive complexity is based on a few simple rules:

        • Code is not considered more complex when it uses shorthand that the language provides for collapsing multiple statements into one
        • Code is considered more complex for each "break in the linear flow of the code"
        • Code is considered more complex when "flow breaking structures are nested"

        Further reading

        Function __init__ has a Cognitive Complexity of 8 (exceeds 5 allowed). Consider refactoring.
        Open

          def __init__(self,
               num_attention_heads,
               intermediate_size,
               intermediate_activation,
               dropout_rate=0.0,
        Severity: Minor
        Found in official/nlp/modeling/layers/transformer.py - About 45 mins to fix

        Cognitive Complexity

        Cognitive Complexity is a measure of how difficult a unit of code is to intuitively understand. Unlike Cyclomatic Complexity, which determines how difficult your code will be to test, Cognitive Complexity tells you how difficult your code will be to read and comprehend.

        A method's cognitive complexity is based on a few simple rules:

        • Code is not considered more complex when it uses shorthand that the language provides for collapsing multiple statements into one
        • Code is considered more complex for each "break in the linear flow of the code"
        • Code is considered more complex when "flow breaking structures are nested"

        Further reading

        Similar blocks of code found in 2 locations. Consider refactoring.
        Open

            self.output_dense = tf_keras.layers.EinsumDense(
        "abc,cd->abd",
        output_shape=(None, hidden_size),
        bias_axes="d",
        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
        Severity: Major
        Found in official/nlp/modeling/layers/transformer.py and 1 other location - About 2 hrs to fix
        official/nlp/modeling/layers/transformer.py on lines 282..287

        Duplicated Code

        Duplicated code can lead to software that is hard to understand and difficult to change. The Don't Repeat Yourself (DRY) principle states:

        Every piece of knowledge must have a single, unambiguous, authoritative representation within a system.

        When you violate DRY, bugs and maintenance problems are sure to follow. Duplicated code has a tendency to both continue to replicate and also to diverge (leaving bugs as two similar implementations differ in subtle ways).

        Tuning

        This issue has a mass of 52.

        We set useful threshold defaults for the languages we support but you may want to adjust these settings based on your project guidelines.

        The threshold configuration represents the minimum mass a code block must have to be analyzed for duplication. The lower the threshold, the more fine-grained the comparison.

        If the engine is too easily reporting duplication, try raising the threshold. If you suspect that the engine isn't catching enough duplication, try lowering the threshold. The best setting tends to differ from language to language.

        See codeclimate-duplication's documentation for more information about tuning the mass threshold in your .codeclimate.yml.

        Refactorings

        Further Reading

        Similar blocks of code found in 2 locations. Consider refactoring.
        Open

            self.self_attention_output_dense = tf_keras.layers.EinsumDense(
        "abc,cd->abd",
        output_shape=(None, hidden_size),
        bias_axes="d",
        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
        Severity: Major
        Found in official/nlp/modeling/layers/transformer.py and 1 other location - About 2 hrs to fix
        official/nlp/modeling/layers/transformer.py on lines 333..338

        Duplicated Code

        Duplicated code can lead to software that is hard to understand and difficult to change. The Don't Repeat Yourself (DRY) principle states:

        Every piece of knowledge must have a single, unambiguous, authoritative representation within a system.

        When you violate DRY, bugs and maintenance problems are sure to follow. Duplicated code has a tendency to both continue to replicate and also to diverge (leaving bugs as two similar implementations differ in subtle ways).

        Tuning

        This issue has a mass of 52.

        We set useful threshold defaults for the languages we support but you may want to adjust these settings based on your project guidelines.

        The threshold configuration represents the minimum mass a code block must have to be analyzed for duplication. The lower the threshold, the more fine-grained the comparison.

        If the engine is too easily reporting duplication, try raising the threshold. If you suspect that the engine isn't catching enough duplication, try lowering the threshold. The best setting tends to differ from language to language.

        See codeclimate-duplication's documentation for more information about tuning the mass threshold in your .codeclimate.yml.

        Refactorings

        Further Reading

        Identical blocks of code found in 2 locations. Consider refactoring.
        Open

          def __init__(self,
               num_attention_heads,
               intermediate_size,
               intermediate_activation,
               dropout_rate=0.0,
        Severity: Minor
        Found in official/nlp/modeling/layers/transformer.py and 1 other location - About 55 mins to fix
        official/nlp/modeling/layers/tn_transformer_expand_condense.py on lines 61..79

        Duplicated Code

        Duplicated code can lead to software that is hard to understand and difficult to change. The Don't Repeat Yourself (DRY) principle states:

        Every piece of knowledge must have a single, unambiguous, authoritative representation within a system.

        When you violate DRY, bugs and maintenance problems are sure to follow. Duplicated code has a tendency to both continue to replicate and also to diverge (leaving bugs as two similar implementations differ in subtle ways).

        Tuning

        This issue has a mass of 37.

        We set useful threshold defaults for the languages we support but you may want to adjust these settings based on your project guidelines.

        The threshold configuration represents the minimum mass a code block must have to be analyzed for duplication. The lower the threshold, the more fine-grained the comparison.

        If the engine is too easily reporting duplication, try raising the threshold. If you suspect that the engine isn't catching enough duplication, try lowering the threshold. The best setting tends to differ from language to language.

        See codeclimate-duplication's documentation for more information about tuning the mass threshold in your .codeclimate.yml.

        Refactorings

        Further Reading

        Similar blocks of code found in 2 locations. Consider refactoring.
        Open

            self.intermediate_dense = tf_keras.layers.EinsumDense(
        "abc,cd->abd",
        output_shape=(None, self.intermediate_size),
        bias_axes="d",
        kernel_initializer=tf_utils.clone_initializer(self._kernel_initializer),
        Severity: Minor
        Found in official/nlp/modeling/layers/transformer.py and 1 other location - About 55 mins to fix
        official/nlp/modeling/layers/reuse_transformer.py on lines 190..195

        Duplicated Code

        Duplicated code can lead to software that is hard to understand and difficult to change. The Don't Repeat Yourself (DRY) principle states:

        Every piece of knowledge must have a single, unambiguous, authoritative representation within a system.

        When you violate DRY, bugs and maintenance problems are sure to follow. Duplicated code has a tendency to both continue to replicate and also to diverge (leaving bugs as two similar implementations differ in subtle ways).

        Tuning

        This issue has a mass of 37.

        We set useful threshold defaults for the languages we support but you may want to adjust these settings based on your project guidelines.

        The threshold configuration represents the minimum mass a code block must have to be analyzed for duplication. The lower the threshold, the more fine-grained the comparison.

        If the engine is too easily reporting duplication, try raising the threshold. If you suspect that the engine isn't catching enough duplication, try lowering the threshold. The best setting tends to differ from language to language.

        See codeclimate-duplication's documentation for more information about tuning the mass threshold in your .codeclimate.yml.

        Refactorings

        Further Reading

        Similar blocks of code found in 7 locations. Consider refactoring.
        Open

            if hidden_size % self.num_attention_heads != 0:
      raise ValueError(
          "The hidden size (%d) is not a multiple of the number of attention "
          "heads (%d)" % (hidden_size, self.num_attention_heads))
        Severity: Major
        Found in official/nlp/modeling/layers/transformer.py and 6 other locations - About 50 mins to fix
        official/nlp/modeling/layers/reuse_transformer.py on lines 154..157
        official/nlp/modeling/layers/tn_transformer_expand_condense.py on lines 127..130
        official/nlp/modeling/layers/transformer_scaffold.py on lines 142..145
        official/nlp/modeling/layers/transformer_xl.py on lines 142..145
        official/projects/detr/modeling/transformer.py on lines 256..259
        official/projects/detr/modeling/transformer.py on lines 670..673

        Duplicated Code

        Duplicated code can lead to software that is hard to understand and difficult to change. The Don't Repeat Yourself (DRY) principle states:

        Every piece of knowledge must have a single, unambiguous, authoritative representation within a system.

        When you violate DRY, bugs and maintenance problems are sure to follow. Duplicated code has a tendency to both continue to replicate and also to diverge (leaving bugs as two similar implementations differ in subtle ways).

        Tuning

        This issue has a mass of 36.

        We set useful threshold defaults for the languages we support but you may want to adjust these settings based on your project guidelines.

        The threshold configuration represents the minimum mass a code block must have to be analyzed for duplication. The lower the threshold, the more fine-grained the comparison.

        If the engine is too easily reporting duplication, try raising the threshold. If you suspect that the engine isn't catching enough duplication, try lowering the threshold. The best setting tends to differ from language to language.

        See codeclimate-duplication's documentation for more information about tuning the mass threshold in your .codeclimate.yml.

        Refactorings

        Further Reading

        Identical blocks of code found in 5 locations. Consider refactoring.
        Open

            if attention_initializer:
      self._attention_initializer = tf_keras.initializers.get(
          attention_initializer)
    else:
      self._attention_initializer = tf_utils.clone_initializer(
        Severity: Major
        Found in official/nlp/modeling/layers/transformer.py and 4 other locations - About 45 mins to fix
        official/nlp/modeling/layers/reuse_transformer.py on lines 132..137
        official/nlp/modeling/layers/tn_transformer_expand_condense.py on lines 100..105
        official/projects/detr/modeling/transformer.py on lines 235..240
        official/projects/detr/modeling/transformer.py on lines 656..661

        Duplicated Code

        Duplicated code can lead to software that is hard to understand and difficult to change. The Don't Repeat Yourself (DRY) principle states:

        Every piece of knowledge must have a single, unambiguous, authoritative representation within a system.

        When you violate DRY, bugs and maintenance problems are sure to follow. Duplicated code has a tendency to both continue to replicate and also to diverge (leaving bugs as two similar implementations differ in subtle ways).

        Tuning

        This issue has a mass of 35.

        We set useful threshold defaults for the languages we support but you may want to adjust these settings based on your project guidelines.

        The threshold configuration represents the minimum mass a code block must have to be analyzed for duplication. The lower the threshold, the more fine-grained the comparison.

        If the engine is too easily reporting duplication, try raising the threshold. If you suspect that the engine isn't catching enough duplication, try lowering the threshold. The best setting tends to differ from language to language.

        See codeclimate-duplication's documentation for more information about tuning the mass threshold in your .codeclimate.yml.

        Refactorings

        Further Reading

        Similar blocks of code found in 10 locations. Consider refactoring.
        Open

            if self._norm_first:
      layer_output = source_attention_output + layer_output
    else:
      layer_output = self.output_layer_norm(layer_output + attention_output)
        Severity: Major
        Found in official/nlp/modeling/layers/transformer.py and 9 other locations - About 45 mins to fix
        official/nlp/modeling/layers/reuse_transformer.py on lines 338..341
        official/nlp/modeling/layers/tn_transformer_expand_condense.py on lines 235..238
        official/nlp/modeling/layers/tn_transformer_expand_condense.py on lines 250..253
        official/nlp/modeling/layers/transformer.py on lines 419..423
        official/nlp/modeling/layers/transformer.py on lines 438..442
        official/projects/detr/modeling/transformer.py on lines 402..405
        official/projects/detr/modeling/transformer.py on lines 814..818
        official/projects/detr/modeling/transformer.py on lines 830..834
        official/projects/detr/modeling/transformer.py on lines 845..848

        Duplicated Code

        Duplicated code can lead to software that is hard to understand and difficult to change. The Don't Repeat Yourself (DRY) principle states:

        Every piece of knowledge must have a single, unambiguous, authoritative representation within a system.

        When you violate DRY, bugs and maintenance problems are sure to follow. Duplicated code has a tendency to both continue to replicate and also to diverge (leaving bugs as two similar implementations differ in subtle ways).

        Tuning

        This issue has a mass of 35.

        We set useful threshold defaults for the languages we support but you may want to adjust these settings based on your project guidelines.

        The threshold configuration represents the minimum mass a code block must have to be analyzed for duplication. The lower the threshold, the more fine-grained the comparison.

        If the engine is too easily reporting duplication, try raising the threshold. If you suspect that the engine isn't catching enough duplication, try lowering the threshold. The best setting tends to differ from language to language.

        See codeclimate-duplication's documentation for more information about tuning the mass threshold in your .codeclimate.yml.

        Refactorings

        Further Reading

        Similar blocks of code found in 10 locations. Consider refactoring.
        Open

            if self._norm_first:
      self_attention_output = source_tensor + self_attention_output
    else:
      self_attention_output = self.self_attention_layer_norm(
          input_tensor + self_attention_output)
        Severity: Major
        Found in official/nlp/modeling/layers/transformer.py and 9 other locations - About 45 mins to fix
        official/nlp/modeling/layers/reuse_transformer.py on lines 338..341
        official/nlp/modeling/layers/tn_transformer_expand_condense.py on lines 235..238
        official/nlp/modeling/layers/tn_transformer_expand_condense.py on lines 250..253
        official/nlp/modeling/layers/transformer.py on lines 438..442
        official/nlp/modeling/layers/transformer.py on lines 453..456
        official/projects/detr/modeling/transformer.py on lines 402..405
        official/projects/detr/modeling/transformer.py on lines 814..818
        official/projects/detr/modeling/transformer.py on lines 830..834
        official/projects/detr/modeling/transformer.py on lines 845..848

        Duplicated Code

        Duplicated code can lead to software that is hard to understand and difficult to change. The Don't Repeat Yourself (DRY) principle states:

        Every piece of knowledge must have a single, unambiguous, authoritative representation within a system.

        When you violate DRY, bugs and maintenance problems are sure to follow. Duplicated code has a tendency to both continue to replicate and also to diverge (leaving bugs as two similar implementations differ in subtle ways).

        Tuning

        This issue has a mass of 35.

        We set useful threshold defaults for the languages we support but you may want to adjust these settings based on your project guidelines.

        The threshold configuration represents the minimum mass a code block must have to be analyzed for duplication. The lower the threshold, the more fine-grained the comparison.

        If the engine is too easily reporting duplication, try raising the threshold. If you suspect that the engine isn't catching enough duplication, try lowering the threshold. The best setting tends to differ from language to language.

        See codeclimate-duplication's documentation for more information about tuning the mass threshold in your .codeclimate.yml.

        Refactorings

        Further Reading

        Similar blocks of code found in 10 locations. Consider refactoring.
        Open

            if self._norm_first:
      attention_output = source_self_attention_output + attention_output
    else:
      attention_output = self.encdec_attention_layer_norm(
          self_attention_output + attention_output)
        Severity: Major
        Found in official/nlp/modeling/layers/transformer.py and 9 other locations - About 45 mins to fix
        official/nlp/modeling/layers/reuse_transformer.py on lines 338..341
        official/nlp/modeling/layers/tn_transformer_expand_condense.py on lines 235..238
        official/nlp/modeling/layers/tn_transformer_expand_condense.py on lines 250..253
        official/nlp/modeling/layers/transformer.py on lines 419..423
        official/nlp/modeling/layers/transformer.py on lines 453..456
        official/projects/detr/modeling/transformer.py on lines 402..405
        official/projects/detr/modeling/transformer.py on lines 814..818
        official/projects/detr/modeling/transformer.py on lines 830..834
        official/projects/detr/modeling/transformer.py on lines 845..848

        Duplicated Code

        Duplicated code can lead to software that is hard to understand and difficult to change. The Don't Repeat Yourself (DRY) principle states:

        Every piece of knowledge must have a single, unambiguous, authoritative representation within a system.

        When you violate DRY, bugs and maintenance problems are sure to follow. Duplicated code has a tendency to both continue to replicate and also to diverge (leaving bugs as two similar implementations differ in subtle ways).

        Tuning

        This issue has a mass of 35.

        We set useful threshold defaults for the languages we support but you may want to adjust these settings based on your project guidelines.

        The threshold configuration represents the minimum mass a code block must have to be analyzed for duplication. The lower the threshold, the more fine-grained the comparison.

        If the engine is too easily reporting duplication, try raising the threshold. If you suspect that the engine isn't catching enough duplication, try lowering the threshold. The best setting tends to differ from language to language.

        See codeclimate-duplication's documentation for more information about tuning the mass threshold in your .codeclimate.yml.

        Refactorings

        Further Reading

        Identical blocks of code found in 7 locations. Consider refactoring.
        Open

            common_kwargs = dict(
        kernel_regularizer=self._kernel_regularizer,
        bias_regularizer=self._bias_regularizer,
        activity_regularizer=self._activity_regularizer,
        kernel_constraint=self._kernel_constraint,
        Severity: Major
        Found in official/nlp/modeling/layers/transformer.py and 6 other locations - About 35 mins to fix
        official/nlp/modeling/layers/gated_feedforward.py on lines 103..108
        official/nlp/modeling/layers/multi_channel_attention.py on lines 63..68
        official/nlp/modeling/layers/reuse_attention.py on lines 351..356
        official/nlp/modeling/layers/reuse_transformer.py on lines 161..166
        official/nlp/modeling/layers/tn_transformer_expand_condense.py on lines 132..137
        official/nlp/modeling/layers/transformer_scaffold.py on lines 148..153

        Duplicated Code

        Duplicated code can lead to software that is hard to understand and difficult to change. The Don't Repeat Yourself (DRY) principle states:

        Every piece of knowledge must have a single, unambiguous, authoritative representation within a system.

        When you violate DRY, bugs and maintenance problems are sure to follow. Duplicated code has a tendency to both continue to replicate and also to diverge (leaving bugs as two similar implementations differ in subtle ways).

        Tuning

        This issue has a mass of 33.

        We set useful threshold defaults for the languages we support but you may want to adjust these settings based on your project guidelines.

        The threshold configuration represents the minimum mass a code block must have to be analyzed for duplication. The lower the threshold, the more fine-grained the comparison.

        If the engine is too easily reporting duplication, try raising the threshold. If you suspect that the engine isn't catching enough duplication, try lowering the threshold. The best setting tends to differ from language to language.

        See codeclimate-duplication's documentation for more information about tuning the mass threshold in your .codeclimate.yml.

        Refactorings

        Further Reading

        There are no issues that match your filters.

        Category
        Status