Add floating point version of hash.

Author: faustomorales

.dockerignore

@@ -1,4 +1,4 @@
 .venv
 ThreatExchange
-!ThreatExchange/hashing/pdq
+!ThreatExchange/hashing/pdq/cpp
 tests

README.md

@@ -1,4 +1,4 @@
-# pdqhash-python
+# pdqhash-python [![CircleCI](https://circleci.com/gh/faustomorales/pdqhash-python.svg?style=svg)](https://circleci.com/gh/faustomorales/pdqhash-python)
 These are Python bindings to the PDQ perceptual hash released by Facebook. Note that the bindings are provided under the MIT license but the PDQ source code is licensed separately under its own license (see the `ThreatExchange/hashing/pdq` folder).
 
 ## Installation
@@ -27,6 +27,9 @@ hash_vector, quality = pdqhash.compute(image)
 # - Rotated 90 degrees and flipped vertically
 # - Rotated 90 degrees and flipped horizontally
 hash_vectors, quality = pdqhash.compute_dihedral(image)
+
+# Get the floating point values of the hash.
+hash_vector_float, quality = pdqhash.compute_float(image)
 ```
 
 ## Contributing

pdqhash/bindings.pyx

@@ -21,6 +21,18 @@ cdef extern from "../ThreatExchange/hashing/pdq/cpp/hashing/pdqhashing.cpp" name
         int& quality
     )
 
+cdef extern from "../ThreatExchange/hashing/pdq/cpp/hashing/pdqhashing.cpp" namespace "facebook::pdq::hashing":
+    void pdqFloat256FromFloatLuma(
+        float* fullBuffer1,
+        float* fullBuffer2,
+        int numRows,
+        int numCols,
+        float buffer64x64[64][64],
+        float buffer16x64[16][64],
+        float output_buffer16x16[16][16],
+        int& quality
+    )
+
 cdef extern from "../ThreatExchange/hashing/pdq/cpp/hashing/pdqhashing.cpp" namespace "facebook::pdq::hashing":
     void pdqDihedralHash256esFromFloatLuma(
         float* fullBuffer1,
@@ -77,6 +89,28 @@ def compute(np.ndarray[char, ndim=3] image):
 
     return hash_to_vector(hash_value.w), quality
 
+def compute_float(np.ndarray[char, ndim=3] image):
+    cdef np.ndarray[float, ndim=2] gray = (image[:, :, 0]*0.299 + image[:, :, 1]*0.587 + image[:, :, 2] * 0.114).astype('float32')
+    cdef np.ndarray[float, ndim=2] placeholder = np.zeros_like(gray)
+    cdef int quality
+    cdef int numRows = gray.shape[0]
+    cdef int numCols = gray.shape[1]
+    cdef float buffer64x64[64][64]
+    cdef float buffer16x64[16][64]
+    cdef float buffer16x16[16][16]
+    cdef float* fullBuffer1 = &gray[0, 0]
+    cdef float* fullBuffer2 = &placeholder[0, 0]
+    pdqFloat256FromFloatLuma(
+        fullBuffer1,
+        fullBuffer2,
+        numRows,
+        numCols,
+        buffer64x64,
+        buffer16x64,
+        buffer16x16,
+        quality)    
+    return np.array(buffer16x16)[:, ::-1].flatten(), quality
+
 def compute_dihedral(np.ndarray[char, ndim=3] image):
     cdef np.ndarray[float, ndim=2] gray = (image[:, :, 0]*0.299 + image[:, :, 1]*0.587 + image[:, :, 2] * 0.114).astype('float32')
     cdef np.ndarray[float, ndim=2] placeholder = np.zeros_like(gray)

tests/test_compute.py

@@ -153,7 +153,10 @@ def test_basic(expected, image_name, threshold):
     image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
     actual, quality = bindings.compute(image)
     assert (actual != expected).sum() <= threshold
-    assert quality == 100
+    
+    actual_float, quality = bindings.compute_float(image)
+    actual_float = actual_float > np.median(actual_float)
+    assert (actual_float != expected).sum() <= threshold
 
     transforms = dict(
         r0=image,