diff --git a/bob/paper/nir_patch_pooling/config/patch_pooling_lr.py b/bob/paper/nir_patch_pooling/config/patch_pooling_lr.py
index 17f63564bb7146e5955681c4f6e7f0edee53dcd8..192e9d29c69732d2a8c0add2b13504e662010f70 100644
--- a/bob/paper/nir_patch_pooling/config/patch_pooling_lr.py
+++ b/bob/paper/nir_patch_pooling/config/patch_pooling_lr.py
@@ -1,4 +1,3 @@
-#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
@@ -6,7 +5,7 @@ Configuration file to run PatchPooling + LR classifier for Face PAD
toward detection of mask attacks in NIR.
"""
-#----------------------------------------------------------
+#------------------------------------------------------------------------------
sub_directory = "pooling_lr"
@@ -17,6 +16,7 @@ from bob.pad.face.preprocessor import FaceCropAlign
from bob.bio.video.preprocessor import Wrapper
from bob.bio.video.utils import FrameSelector
+
# parameters and constants
FACE_SIZE = 128
RGB_OUTPUT_FLAG = False
@@ -37,9 +37,10 @@ _image_preprocessor = FaceCropAlign(face_size=FACE_SIZE,
_frame_selector = FrameSelector(selection_style = "all")
-preprocessor = Wrapper(preprocessor = _image_preprocessor, frame_selector = _frame_selector)
+preprocessor = Wrapper(preprocessor = _image_preprocessor,
+ frame_selector = _frame_selector)
-#----------------------------------------------------------
+#------------------------------------------------------------------------------
# define extractor:
@@ -48,16 +49,19 @@ from bob.bio.video.extractor import Wrapper
from bob.extension import rc
import os
-_model_dir = rc.get("LIGHTCNN9_MODEL_DIRECTORY")
+_model_directory = rc["lightcnn9.model.directory"]
_model_name = "LightCNN_9Layers_checkpoint.pth.tar"
-_model_file = os.path.join(_model_dir, _model_name)
+_model_file = os.path.join(_model_directory, _model_name)
+
if not os.path.exists(_model_file):
- print("Error: Could not find the LightCNN-9 model at [{}].\nPlease follow the download instructions from README".format(_model_dir))
+ print("Error: Could not find the LightCNN-9 model [{}].\nPlease follow \
+ the download instructions from README".format(_model_directory))
exit(0)
-extractor = Wrapper(PatchPoolingCNN(model_file=_model_file), frame_selector = _frame_selector)
+extractor = Wrapper(PatchPoolingCNN(model_file=_model_file),
+ frame_selector = _frame_selector)
-#----------------------------------------------------------
+#------------------------------------------------------------------------------
# define algorithm
@@ -67,12 +71,6 @@ C = 1.0
algorithm = LogRegr(C=C, frame_level_scores_flag=True)
-#----------------------------------------------------------
-
-
-
-
-
-
+#------------------------------------------------------------------------------
diff --git a/bob/paper/nir_patch_pooling/extractor/patch_pooling_cnn.py b/bob/paper/nir_patch_pooling/extractor/patch_pooling_cnn.py
index 5941c877c73d2fba8b661d8e241f5bfefb925b72..e98cb5a17177cd79f3c61f769a06c146f1d01e3d 100644
--- a/bob/paper/nir_patch_pooling/extractor/patch_pooling_cnn.py
+++ b/bob/paper/nir_patch_pooling/extractor/patch_pooling_cnn.py
@@ -1,10 +1,8 @@
-#!/usr/bin/env python2
# -*- coding: utf-8 -*-
"""
-Implementation of PCNN feature extractor for LightCNN-9.
+Implementation of Patch Pooling CNN feature extractor with LightCNN-9 backbone
@author: Ketan Kotwal
-
"""
# Imports
@@ -21,14 +19,13 @@ import logging
logger = logging.getLogger(__name__)
logger.setLevel(logging.INFO)
-#----------------------------------------------------------
+#------------------------------------------------------------------------------
class PatchPoolingCNN(Extractor):
"""
- The class implements the feature extraction of LightCNN9 embeddings.
- It has some implementation differences from a similar extractor from
- bob.learn.pytorch.
+ The class implements extraction of patch pooled features from the final
+ convolutional layer of LightCNN9 (MFM5 layer).
"""
def __init__(self, model_file=None, num_classes=79077):
@@ -39,30 +36,36 @@ class PatchPoolingCNN(Extractor):
# load the model into network.
cp = torch.load(model_file, map_location="cpu")
- # checked if pre-trained model was saved using nn.DataParallel ...
- saved_with_nnDataParallel = False
+ # checked if pre-trained model was saved using nn.DataParallel
+ saved_with_data_parallel = False
for k, v in cp["state_dict"].items():
if("module" in k):
- saved_with_nnDataParallel = True
+ saved_with_data_parallel = True
break
# if DataParallel format, remove module term
- if(saved_with_nnDataParallel):
+ if(saved_with_data_parallel):
if("state_dict" in cp):
+
from collections import OrderedDict
new_state_dict = OrderedDict()
+
for k, v in cp["state_dict"].items():
name = k[7:]
new_state_dict[name] = v
+
self.network.load_state_dict(new_state_dict)
else:
+
self.network.load_state_dict(cp["state_dict"])
+
self.network.eval()
# image pre-processing
- self.data_transform = transforms.Compose([transforms.Resize(size=128), transforms.ToTensor()])
+ self.data_transform = transforms.Compose([transforms.Resize(size=128),
+ transforms.ToTensor()])
-#----------------------------------------------------------
+#------------------------------------------------------------------------------
def __call__(self, image):
@@ -76,7 +79,7 @@ class PatchPoolingCNN(Extractor):
Returns
-------
feature : :py:class:`numpy.ndarray` (floats)
- The extracted features as a 1d array of size 320
+ The extracted features as a 1d array of size 256
"""
@@ -84,45 +87,49 @@ class PatchPoolingCNN(Extractor):
pil_image = Image.fromarray(image.astype(np.uint8))
input_image = self.data_transform(pil_image)
input_image = input_image.unsqueeze(0)
-
- # to be compliant with the loaded model, where weight and biases are torch.FloatTensor
input_image = input_image.float()
+ # obtain the features (to be pooled) from forward pass of network
_ , features = self.network.forward(Variable(input_image))
+
+ # pool features through patch-level processing
features = self.conv_to_patch(features)
features = features.data.numpy().flatten()
return features.astype(np.float64)
-#----------------------------------------------------------
+#------------------------------------------------------------------------------
def conv_to_patch(self, features):
- logger.debug("Shape of input features: {} {}".format(features.shape, features.squeeze().shape))
-
# parameters for the patch conversion
- stride = 4 # orig:4 #feat.shape[2]/4
- idx = 0 # purely for debugging
- num_patch = features.shape[2]/stride
- feat_patch = torch.zeros(1, stride*stride*features.shape[1])
+ stride = 4 # features.shape[2]/4
+
+ # for debugging
+ # idx = 0
+ # num_patch = features.shape[2]/stride
+
+ pooled_features = torch.zeros(1, stride*stride*features.shape[1])
# obtain patches by tesselation of feature maps
+ # pool linearized version of individual patches
for i in range(0, features.shape[2], stride):
for j in range(0, features.shape[3], stride):
- feat_tmp = features[:, :, i:i+stride, j:j+stride]
- feat_tmp = feat_tmp.contiguous().view(feat_tmp.size(0), -1)
- feat_patch += feat_tmp
- idx += 1
+ feat_temp = features[:, :, i:i+stride, j:j+stride]
+ feat_temp = feat_temp.contiguous().view(feat_temp.size(0), -1)
+ pooled_features += feat_temp
+ # idx += 1
- # normalize the patch vector
- feat_patch = feat_patch/stride/stride
- logger.debug("Feat patch shape: {}".format(feat_patch.shape))
+ # normalize the vector of pooled features
+ pooled_features = pooled_features/stride/stride
+
+ return pooled_features
- return feat_patch
+#------------------------------------------------------------------------------
-#----------------------------------------------------------
+#------------------------------------------------------------------------------
-# class LightCNN9Patch: it inherits the LightCNN-9 class, and returns the last
-# conv layer features; instead of embeddings.
+# class LightCNN9Patch: it inherits the LightCNN-9 class from bob,
+# and returns the last conv layer features and embeddings.
class LightCNN9Patch(LightCNN9):
@@ -131,7 +138,7 @@ class LightCNN9Patch(LightCNN9):
# do not change the init
super(LightCNN9Patch, self).__init__()
-#----------------------------------------------------------
+#------------------------------------------------------------------------------
def forward(self, x):
@@ -147,14 +154,7 @@ class LightCNN9Patch(LightCNN9):
out = self.fc2(x)
return out, conv_out
-#----------------------------------------------------------
-
-
-
-
-
-
-
+#------------------------------------------------------------------------------