Allow to chain modifications

README.md

@@ -11,6 +11,7 @@ thesis `Testing robustness of DeepFake recognition methods against non-malicious
 * [How it works?](#how-it-works-)
     * [TL;DR:](#tl-dr-)
     * [Tell me more!](#tell-me-more-)
+* [Favourite code snippet](#favourite-code-snippet)
 * [Design](#design)
 * [Installation](#installation)
 * [GPU configuration](#gpu-configuration)
@@ -23,10 +24,9 @@ recognition method used to properly discriminate DeepFake from authentic video w
 ## What is this?
 
 * CLI program allowing to:
-    * Preprocess dataset containing real and fake videos. Chosen part of negatives is modified by selected non-malicious
-      image modifications based on settings assigned by user.
+    * Preprocess dataset containing real and fake videos. Chosen part of negatives is modified by selected modifications based on settings assigned by user.
         * Preprocessing can be done via single command or gradually in steps where single step represents activity such
-          as `extract faces from fake videos.` The letter method allows preprocessing large datasets[^2] on without the
+          as `extract faces from fake videos`. The letter method allows preprocessing large datasets[^2] without the
           need of keeping program on for 24+ hours.
     * Train detection model and evaluate it in assigned settings.
 
@@ -59,28 +59,57 @@ alter negatives are configured via YAML files. Here is a sample:
 ---
 # Example modifications settings.
 # In total half of negatives are altered.
-modifications:
-  - name: RedEyesEffectModification
-  share: 0.125
-  options:
-    brightness_threshold: 50
-                            - name: CLAHEModification
-  share: 0.125
-  options:
-    clip_limit: 2.0
-    grid_width: 8
-    grid_height: 8
-                   - name: HistogramEqualizationModification
-  share: 0.125
-           - name: GaussianBlurModification
-  share: 0.125
-  options:
-    kernel_width: 9
-    kernel_height: 9
+modifications_chains:
+  - share: 0.25
+    modifications:
+      - name: GammaCorrectionModification
+        options:
+          gamma_value: 0.75
+      - name: CLAHEModification
+        options:
+          clip_limit: 2.0
+          grid_width: 8
+          grid_height: 8
+  - share: 0.25
+    modifications:
+      - name: GammaCorrectionModification
+        options:
+          gamma_value: 0.75
+      - name: HistogramEqualizationModification
 ```
 
 The names of supported modifications can be found in [this file](src/dfd/datasets/modifications/register.py).
 
+## Favourite code snippet
+
+```python
+@given(
+    given_specifications=st.lists(
+        st.builds(
+            ModificationStub,
+            name=st.text(min_size=1),
+            no_repeats=st.integers(min_value=1),
+        ),
+        min_size=1,
+        max_size=5,
+    )
+)
+def test_combine_multiple_specifications(given_specifications):
+    # GIVEN
+    image_mock = Mock(spec_set=np.ndarray)
+    # WHEN specifications are combined
+    combined_specification = functools.reduce(operator.and_, given_specifications)
+    # THEN specification names are combined
+    expected_name = "__".join([spec.name for spec in given_specifications])
+    assert combined_specification.name == expected_name
+    # And specifications are performed in order
+    combined_specification.perform(image_mock)
+    expected_calls_in_order = [call.repeat(spec.no_repeats) for spec in given_specifications]
+    image_mock.assert_has_calls(expected_calls_in_order)
+```
+
+It's uses two cool concepts: property-based testing and specification pattern[^4].
+
 ## Design
 
 The application design is loosely inspired
@@ -121,3 +150,4 @@ pip install git+https://github.com/cicheck/dfd.git
 [^1]: Currently the only supported detection method is [Meso-4](https://arxiv.org/abs/1809.00888).
 [^2]: Such as [Celeb-DF](https://github.com/yuezunli/celeb-deepfakeforensics).
 [^3]: Half of negatives modified, 4 modifications used with naive parameters.
+[^4]: Or rather design loosely inspired by specification pattern :stuck_out_tongue:

docs/diagrams/app_architecture.puml

@@ -24,7 +24,7 @@ package "Business Logic (Core)" #skyblue {
         }
 
         package Modifications {
-            interface ModificationInterface {
+            abstract ModificationSpecification {
                 + name
                 --
                 + perform()
@@ -100,8 +100,8 @@ MesoNet -up-- ModelRegistry : register >
 
 MesoNet -up--|> ModelInterface
 
-GammaCorrectionModification -up--|> ModificationInterface
-CLAHEModification -up--|> ModificationInterface
+GammaCorrectionModification -up--|> ModificationSpecification
+CLAHEModification -up--|> ModificationSpecification
 
 GammaCorrectionModification -up-- ModificationRegistry : register >
 CLAHEModification -up-- ModificationRegistry : register >

example_settings.yaml

@@ -1,25 +1,19 @@
 ---
 # Example ModificationGenerator settings file
-modifications:
-  - name: RedEyesEffectModification
-    share: 0.125
-    options:
-      brightness_threshold: 50
-      face_landmarks_detector_path: "/media/cicheck/Extreme Pro/\
-        models/shape_predictor_68_face_landmarks.dat"
-  - name: CLAHEModification
-    share: 0.125
-    options:
-      clip_limit: 2.0
-      grid_width: 8
-      grid_height: 8
-  - name: HistogramEqualizationModification
-    share: 0.125
-  - name: GammaCorrectionModification
-    share: 0.0625
-    options:
-      gamma_value: 0.75
-  - name: GammaCorrectionModification
-    share: 0.0625
-    options:
-      gamma_value: 1.25
+modifications_chains:
+  - share: 0.25
+    modifications:
+      - name: GammaCorrectionModification
+        options:
+          gamma_value: 0.75
+      - name: CLAHEModification
+        options:
+          clip_limit: 2.0
+          grid_width: 8
+          grid_height: 8
+  - share: 0.25
+    modifications:
+      - name: GammaCorrectionModification
+        options:
+          gamma_value: 0.75
+      - name: HistogramEqualizationModification

setup.cfg

@@ -47,6 +47,7 @@ tests =
     pytest-cov==2.12.1
     pytest>=6.2.4
     coverage[toml]
+    hypothesis==6.34.1
 style =
     darglint>=1.8.0
     flake8>=3.9.2

src/dfd/datasets/frames_generators/modification.py

@@ -1,23 +1,24 @@
 """Generate new frames after performing set non malicious modifications on original frames."""
 import functools
 import itertools
+import operator
 import pathlib
-from typing import Generator, List, NamedTuple, Optional
+from typing import Generator, List, NamedTuple, Optional, Sequence
 
 import cv2 as cv
 import numpy as np
 
-from dfd.exceptions import DfdError
 from dfd.datasets.modifications.definitions import IdentityModification
-from dfd.datasets.modifications.interfaces import ModificationInterface
 from dfd.datasets.modifications.register import ModificationRegister
-from dfd.datasets.settings import GeneratorSettings
+from dfd.datasets.modifications.specification import ModificationSpecification
+from dfd.datasets.settings import GeneratorSettings, ModificationSettings
+from dfd.exceptions import DfdError
 
 
 class ModificationShare(NamedTuple):
     """Share of frames on which modification will be performed."""
 
-    modification: ModificationInterface
+    modification: ModificationSpecification
     share: float
 
 
@@ -31,7 +32,7 @@ class ModificationRange(NamedTuple):
 
     """
 
-    modification: ModificationInterface
+    modification: ModificationSpecification
     lower_bound: int
     upper_bound: int
 
@@ -93,22 +94,17 @@ def from_directory(
             input_frame = cv.imread(str(input_frame_path))
             modified_frame = modification.perform(input_frame)
             yield ModifiedFrame(
-                modification_used=str(modification),
+                modification_used=modification.name,
                 frame=modified_frame,
                 original_path=input_frame_path,
             )
 
     @functools.lru_cache(maxsize=1)
     def _get_modifications_share(self) -> List[ModificationShare]:
         modifications_share: List[ModificationShare] = []
-        for modification_settings in self._setting.modifications:
-            mame = modification_settings.name
-            options = modification_settings.options
-            share = modification_settings.share
-
-            modification_class = self._register.get_modification_class(mame)
-            # TODO: fix typing
-            modification = modification_class(**options)  # type: ignore
+        for modification_chain_settings in self._setting.modifications_chains:
+            share = modification_chain_settings.share
+            modification = self._chain_modifications(modification_chain_settings.modifications)
             modifications_share.append(ModificationShare(modification, share))
 
         self._check_modifications_are_unique(
@@ -117,7 +113,7 @@ def _get_modifications_share(self) -> List[ModificationShare]:
         return modifications_share
 
     @staticmethod
-    def _check_modifications_are_unique(modifications: List[ModificationInterface]):
+    def _check_modifications_are_unique(modifications: List[ModificationSpecification]):
         """Check if modifications are unique.
 
         Raises:
@@ -160,7 +156,7 @@ def _get_frames_permutation(self, no_frames: int) -> np.ndarray:
 
     def _choose_modification(
         self, frame_index: int, input_frame_path: pathlib.Path, no_frames: int
-    ) -> ModificationInterface:
+    ) -> ModificationSpecification:
         frames_permutation = self._get_frames_permutation(no_frames)
         modifications_range = self._get_modifications_range(no_frames)
         permuted_index = frames_permutation[frame_index]
@@ -170,3 +166,14 @@ def _choose_modification(
         # TODO: log error
         # This should never happen
         raise DfdError("Could not select modification.")
+
+    def _chain_modifications(
+        self, modifications_settings: Sequence[ModificationSettings]
+    ) -> ModificationSpecification:
+        modifications: List[ModificationSpecification] = []
+        for modification_settings in modifications_settings:
+            modification_class = self._register.get_modification_class(modification_settings.name)
+            # TODO: fix typing
+            modification = modification_class(**modification_settings.options)  # type: ignore
+            modifications.append(modification)
+        return functools.reduce(operator.and_, modifications)

src/dfd/datasets/modifications/definitions/clahe.py

@@ -3,10 +3,10 @@
 import cv2 as cv
 import numpy as np
 
-from dfd.datasets.modifications.interfaces import ModificationInterface
+from dfd.datasets.modifications.specification import ModificationSpecification
 
 
-class CLAHEModification(ModificationInterface):
+class CLAHEModification(ModificationSpecification):
     """Modification CLAHE (Contrast Limited Adaptive Histogram Equalization)"""
 
     def __init__(self, clip_limit: float, grid_width: int, grid_height: int) -> None:
@@ -21,6 +21,17 @@ def __init__(self, clip_limit: float, grid_width: int, grid_height: int) -> None
         self._clip_limit = clip_limit
         self._title_grid_size = (grid_width, grid_height)
 
+    @property
+    def name(self) -> str:
+        """Get specification name.
+
+        Returns:
+            The name of specification.
+
+        """
+        width, height = self._title_grid_size
+        return f"clahe_{width}_{height}_{self._clip_limit}"
+
     def perform(self, image: np.ndarray) -> np.ndarray:
         """Perform CLAHE on image.
 
@@ -44,5 +55,4 @@ def perform(self, image: np.ndarray) -> np.ndarray:
         return cv.cvtColor(ycrcb_image, cv.COLOR_YCrCb2BGR)
 
     def __str__(self) -> str:
-        width, height = self._title_grid_size
-        return f"clahe_{width}_{height}_{self._clip_limit}"
+        return self.name

src/dfd/datasets/modifications/definitions/gamma_correction.py

@@ -2,10 +2,10 @@
 import cv2 as cv
 import numpy as np
 
-from dfd.datasets.modifications.interfaces import ModificationInterface
+from dfd.datasets.modifications.specification import ModificationSpecification
 
 
-class GammaCorrectionModification(ModificationInterface):
+class GammaCorrectionModification(ModificationSpecification):
     """Modification Gamma Correction."""
 
     def __init__(self, gamma_value: float) -> None:
@@ -17,6 +17,16 @@ def __init__(self, gamma_value: float) -> None:
         """
         self._gamma_value = gamma_value
 
+    @property
+    def name(self) -> str:
+        """Get specification name.
+
+        Returns:
+            The name of specification.
+
+        """
+        return f"gamma_correction_{self._gamma_value}"
+
     def perform(self, image: np.ndarray) -> np.ndarray:
         """Perform gamma correction on provided image.
 
@@ -38,4 +48,4 @@ def perform(self, image: np.ndarray) -> np.ndarray:
         return cv.LUT(image, look_up_table)
 
     def __str__(self) -> str:
-        return f"gamma_correction_{self._gamma_value}"
+        return self.name

src/dfd/datasets/modifications/definitions/gaussian_blur.py

@@ -3,10 +3,10 @@
 import cv2 as cv
 import numpy as np
 
-from dfd.datasets.modifications.interfaces import ModificationInterface
+from dfd.datasets.modifications.specification import ModificationSpecification
 
 
-class GaussianBlurModification(ModificationInterface):
+class GaussianBlurModification(ModificationSpecification):
     """Modification Gaussian blur (AKA Gaussian smoothing)."""
 
     def __init__(
@@ -30,6 +30,16 @@ def __init__(
         self._sigma_y = sigma_y
 
     def __str__(self) -> str:
+        return self.name
+
+    @property
+    def name(self) -> str:
+        """Get specification name.
+
+        Returns:
+            The name of specification.
+
+        """
         width, height = self._kernel_size
         return f"gaussian_blur{width}_{height}_{self._sigma_x}_{self._sigma_y}"