3636import lfw
3737import os
3838import sys
39- import math
39+ from tensorflow . python . ops import data_flow_ops
4040from sklearn import metrics
4141from scipy .optimize import brentq
4242from scipy import interpolate
@@ -52,44 +52,89 @@ def main(args):
5252
5353 # Get the paths for the corresponding images
5454 paths , actual_issame = lfw .get_paths (os .path .expanduser (args .lfw_dir ), pairs )
55-
56- # Load the model
57- facenet .load_model (args .model )
5855
59- # Get input and output tensors
60- images_placeholder = tf .get_default_graph ().get_tensor_by_name ("input:0" )
56+ image_paths_placeholder = tf .placeholder (tf .string , shape = (None ,1 ), name = 'image_paths' )
57+ labels_placeholder = tf .placeholder (tf .int32 , shape = (None ,1 ), name = 'labels' )
58+ batch_size_placeholder = tf .placeholder (tf .int32 , name = 'batch_size' )
59+ control_placeholder = tf .placeholder (tf .int32 , shape = (None ,1 ), name = 'control' )
60+ phase_train_placeholder = tf .placeholder (tf .bool , name = 'phase_train' )
61+
62+ nrof_preprocess_threads = 4
63+ image_size = (args .image_size , args .image_size )
64+ eval_input_queue = data_flow_ops .FIFOQueue (capacity = 2000000 ,
65+ dtypes = [tf .string , tf .int32 , tf .int32 ],
66+ shapes = [(1 ,), (1 ,), (1 ,)],
67+ shared_name = None , name = None )
68+ eval_enqueue_op = eval_input_queue .enqueue_many ([image_paths_placeholder , labels_placeholder , control_placeholder ], name = 'eval_enqueue_op' )
69+ image_batch , label_batch = facenet .create_input_pipeline (eval_input_queue , image_size , nrof_preprocess_threads , batch_size_placeholder )
70+
71+ # Load the model
72+ input_map = {'image_batch' : image_batch , 'label_batch' : label_batch , 'phase_train' : phase_train_placeholder }
73+ facenet .load_model (args .model , input_map = input_map )
74+
75+ # Get output tensor
6176 embeddings = tf .get_default_graph ().get_tensor_by_name ("embeddings:0" )
62- phase_train_placeholder = tf .get_default_graph ().get_tensor_by_name ("phase_train:0" )
63-
64- #image_size = images_placeholder.get_shape()[1] # For some reason this doesn't work for frozen graphs
65- image_size = args .image_size
66- embedding_size = embeddings .get_shape ()[1 ]
67-
68- # Run forward pass to calculate embeddings
69- print ('Runnning forward pass on LFW images' )
70- batch_size = args .lfw_batch_size
71- nrof_images = len (paths )
72- nrof_batches = int (math .ceil (1.0 * nrof_images / batch_size ))
73- emb_array = np .zeros ((nrof_images , embedding_size ))
74- for i in range (nrof_batches ):
75- start_index = i * batch_size
76- end_index = min ((i + 1 )* batch_size , nrof_images )
77- paths_batch = paths [start_index :end_index ]
78- images = facenet .load_data (paths_batch , False , False , image_size )
79- feed_dict = { images_placeholder :images , phase_train_placeholder :False }
80- emb_array [start_index :end_index ,:] = sess .run (embeddings , feed_dict = feed_dict )
81-
82- tpr , fpr , accuracy , val , val_std , far = lfw .evaluate (emb_array ,
83- actual_issame , nrof_folds = args .lfw_nrof_folds )
77+ #
78+ coord = tf .train .Coordinator ()
79+ tf .train .start_queue_runners (coord = coord , sess = sess )
8480
85- print ('Accuracy: %1.3f+-%1.3f' % (np .mean (accuracy ), np .std (accuracy )))
86- print ('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val , val_std , far ))
81+ evaluate (sess , eval_enqueue_op , image_paths_placeholder , labels_placeholder , phase_train_placeholder , batch_size_placeholder , control_placeholder ,
82+ embeddings , label_batch , paths , actual_issame , args .lfw_batch_size , args .lfw_nrof_folds , args .distance_metric , args .subtract_mean ,
83+ args .use_flipped_images , args .use_fixed_image_standardization )
8784
88- auc = metrics .auc (fpr , tpr )
89- print ('Area Under Curve (AUC): %1.3f' % auc )
90- eer = brentq (lambda x : 1. - x - interpolate .interp1d (fpr , tpr )(x ), 0. , 1. )
91- print ('Equal Error Rate (EER): %1.3f' % eer )
92-
85+
86+ def evaluate (sess , enqueue_op , image_paths_placeholder , labels_placeholder , phase_train_placeholder , batch_size_placeholder , control_placeholder ,
87+ embeddings , labels , image_paths , actual_issame , batch_size , nrof_folds , distance_metric , subtract_mean , use_flipped_images , use_fixed_image_standardization ):
88+ # Run forward pass to calculate embeddings
89+ print ('Runnning forward pass on LFW images' )
90+
91+ # Enqueue one epoch of image paths and labels
92+ nrof_embeddings = len (actual_issame )* 2 # nrof_pairs * nrof_images_per_pair
93+ nrof_flips = 2 if use_flipped_images else 1
94+ nrof_images = nrof_embeddings * nrof_flips
95+ labels_array = np .expand_dims (np .arange (0 ,nrof_images ),1 )
96+ image_paths_array = np .expand_dims (np .repeat (np .array (image_paths ),nrof_flips ),1 )
97+ control_array = np .zeros_like (labels_array , np .int32 )
98+ if use_fixed_image_standardization :
99+ control_array += np .ones_like (labels_array )* facenet .FIXED_STANDARDIZATION
100+ if use_flipped_images :
101+ # Flip every second image
102+ control_array += (labels_array % 2 )* facenet .FLIP
103+ sess .run (enqueue_op , {image_paths_placeholder : image_paths_array , labels_placeholder : labels_array , control_placeholder : control_array })
104+
105+ embedding_size = int (embeddings .get_shape ()[1 ])
106+ assert nrof_images % batch_size == 0 , 'The number of LFW images must be an integer multiple of the LFW batch size'
107+ nrof_batches = nrof_images // batch_size
108+ emb_array = np .zeros ((nrof_images , embedding_size ))
109+ lab_array = np .zeros ((nrof_images ,))
110+ for i in range (nrof_batches ):
111+ feed_dict = {phase_train_placeholder :False , batch_size_placeholder :batch_size }
112+ emb , lab = sess .run ([embeddings , labels ], feed_dict = feed_dict )
113+ lab_array [lab ] = lab
114+ emb_array [lab , :] = emb
115+ if i % 10 == 9 :
116+ print ('.' , end = '' )
117+ sys .stdout .flush ()
118+ print ('' )
119+ embeddings = np .zeros ((nrof_embeddings , embedding_size * nrof_flips ))
120+ if use_flipped_images :
121+ # Concatenate embeddings for flipped and non flipped iversion of the images
122+ embeddings [:,:embedding_size ] = emb_array [0 ::2 ,:]
123+ embeddings [:,embedding_size :] = emb_array [1 ::2 ,:]
124+ else :
125+ embeddings = emb_array
126+
127+ assert np .array_equal (lab_array , np .arange (nrof_images ))== True , 'Wrong labels used for evaluation, possibly caused by training examples left in the input pipeline'
128+ tpr , fpr , accuracy , val , val_std , far = lfw .evaluate (embeddings , actual_issame , nrof_folds = nrof_folds , distance_metric = distance_metric , subtract_mean = subtract_mean )
129+
130+ print ('Accuracy: %2.5f+-%2.5f' % (np .mean (accuracy ), np .std (accuracy )))
131+ print ('Validation rate: %2.5f+-%2.5f @ FAR=%2.5f' % (val , val_std , far ))
132+
133+ auc = metrics .auc (fpr , tpr )
134+ print ('Area Under Curve (AUC): %1.3f' % auc )
135+ eer = brentq (lambda x : 1. - x - interpolate .interp1d (fpr , tpr )(x ), 0. , 1. )
136+ print ('Equal Error Rate (EER): %1.3f' % eer )
137+
93138def parse_arguments (argv ):
94139 parser = argparse .ArgumentParser ()
95140
@@ -105,6 +150,14 @@ def parse_arguments(argv):
105150 help = 'The file containing the pairs to use for validation.' , default = 'data/pairs.txt' )
106151 parser .add_argument ('--lfw_nrof_folds' , type = int ,
107152 help = 'Number of folds to use for cross validation. Mainly used for testing.' , default = 10 )
153+ parser .add_argument ('--distance_metric' , type = int ,
154+ help = 'Distance metric 0:euclidian, 1:cosine similarity.' , default = 0 )
155+ parser .add_argument ('--use_flipped_images' ,
156+ help = 'Concatenates embeddings for the image and its horizontally flipped counterpart.' , action = 'store_true' )
157+ parser .add_argument ('--subtract_mean' ,
158+ help = 'Subtract feature mean before calculating distance.' , action = 'store_true' )
159+ parser .add_argument ('--use_fixed_image_standardization' ,
160+ help = 'Performs fixed standardization of images.' , action = 'store_true' )
108161 return parser .parse_args (argv )
109162
110163if __name__ == '__main__' :
0 commit comments