fix weight decay bugs and apply cross validation

facenet.py

@@ -500,9 +500,9 @@ def calculate_roc(thresholds, embeddings1, embeddings2, actual_issame, nrof_fold
     fprs = np.zeros((nrof_folds,nrof_thresholds))
     accuracy = np.zeros((nrof_folds))
 
-    #diff = np.subtract(embeddings1, embeddings2)
-    #dist = np.sum(np.square(diff),1)
-    dist = (embeddings1 * embeddings2).sum(axis=1)/2.
+    diff = np.subtract(embeddings1, embeddings2)
+    dist = np.sum(np.square(diff),1)
+    #dist = (embeddings1 * embeddings2).sum(axis=1)/2.
     indices = np.arange(nrof_pairs)
 
     best_thres = []
@@ -546,9 +546,9 @@ def calculate_val(thresholds, embeddings1, embeddings2, actual_issame, far_targe
     val = np.zeros(nrof_folds)
     far = np.zeros(nrof_folds)
 
-    #diff = np.subtract(embeddings1, embeddings2)
-    #dist = np.sum(np.square(diff),1)
-    dist = (embeddings1 * embeddings2).sum(axis=1)/2.
+    diff = np.subtract(embeddings1, embeddings2)
+    dist = np.sum(np.square(diff),1)
+    #dist = (embeddings1 * embeddings2).sum(axis=1)/2.
     indices = np.arange(nrof_pairs)
 
     for fold_idx, (train_set, test_set) in enumerate(k_fold.split(indices)):
@@ -572,7 +572,7 @@ def calculate_val(thresholds, embeddings1, embeddings2, actual_issame, far_targe
 
 
 def calculate_val_far(threshold, dist, actual_issame):
-    predict_issame = np.greater(dist, threshold)
+    predict_issame = np.less(dist, threshold)
     true_accept = np.sum(np.logical_and(predict_issame, actual_issame))
     false_accept = np.sum(np.logical_and(predict_issame, np.logical_not(actual_issame)))
     n_same = np.sum(actual_issame)

lfw.py

@@ -33,12 +33,12 @@
 
 def evaluate(embeddings, actual_issame, nrof_folds=10):
     # Calculate evaluation metrics
-    thresholds = np.arange(-1, 1, 0.01)
+    thresholds = np.arange(0, 4, 0.01)
     embeddings1 = embeddings[0::2]
     embeddings2 = embeddings[1::2]
     tpr, fpr, accuracy = facenet.calculate_roc(thresholds, embeddings1, embeddings2,
         np.asarray(actual_issame), nrof_folds=nrof_folds)
-    thresholds = np.arange(-1, 1, 0.001)
+    thresholds = np.arange(0, 4, 0.001)
     val, val_std, far = facenet.calculate_val(thresholds, embeddings1, embeddings2,
         np.asarray(actual_issame), 1e-3, nrof_folds=nrof_folds)
     return tpr, fpr, accuracy, val, val_std, far

models/resface.py

@@ -10,6 +10,7 @@
 def prelu(x):
     with tf.variable_scope('PRelu'):   
         alphas = tf.get_variable(name='prelu_alphas', dtype=tf.float32, initializer=tf.constant_initializer(value=0.25),shape=[x.get_shape()[-1]])
+        #alphas = tf.Variable(tf.constant(0.25,dtype=tf.float32,shape=[x.get_shape()[-1]]),name='prelu_alphas')
         pos = tf.nn.relu(x)
         neg = alphas * (x - abs(x)) * 0.5
         return pos + neg

train.py

@@ -153,14 +153,20 @@ def main(args):
         cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(
             labels=label_batch, logits=AM_logits, name='cross_entropy_per_example')
         cross_entropy_mean = tf.reduce_mean(cross_entropy, name='cross_entropy')
-
+
+        #print('test',tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES))
+
         for weights in slim.get_variables_by_name('kernel'):
             kernel_regularization = tf.contrib.layers.l2_regularizer(args.weight_decay)(weights)
             print(weights)
             tf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES, kernel_regularization)	
 
         regularization_losses = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
-        total_loss = tf.add_n([cross_entropy_mean] + regularization_losses, name='total_loss')
+
+        if args.weight_decay==0:
+            total_loss = tf.add_n([cross_entropy_mean], name='total_loss')
+        else:
+            total_loss = tf.add_n([cross_entropy_mean] + regularization_losses, name='total_loss')
         tf.add_to_collection('losses', total_loss)
 
         #define two saver in case under 'finetuning on different dataset' situation 
@@ -169,7 +175,8 @@ def main(args):
 
         #train_op = facenet.train(total_loss, global_step, args.optimizer, 
         #    learning_rate, args.moving_average_decay, tf.trainable_variables(), args.log_histograms)
-        train_op = tf.train.AdamOptimizer(learning_rate).minimize(total_loss,global_step = global_step,var_list=tf.trainable_variables())
+        #train_op = tf.train.AdamOptimizer(learning_rate).minimize(total_loss,global_step = global_step,var_list=tf.trainable_variables())
+        train_op = tf.train.MomentumOptimizer(learning_rate,momentum=0.9).minimize(total_loss,global_step=global_step,var_list=tf.trainable_variables())
         summary_op = tf.summary.merge_all()
 
         # Start running operations on the Graph.
@@ -199,9 +206,11 @@ def main(args):
 
                 print('validation running...')
                 if args.lfw_dir:
-                    best_accuracy = evaluate_double(sess, enqueue_op, image_paths_placeholder, labels_placeholder, phase_train_placeholder, batch_size_placeholder, embeddings, 
-                    	label_batch, lfw_paths, actual_issame, args.lfw_batch_size, args.lfw_nrof_folds, log_dir, step, summary_writer,best_accuracy,saver_save,model_dir,subdir,image_batch,args)
-
+                    #best_accuracy = evaluate_double(sess, enqueue_op, image_paths_placeholder, labels_placeholder, phase_train_placeholder, batch_size_placeholder, embeddings, 
+                    #	label_batch, lfw_paths, actual_issame, args.lfw_batch_size, args.lfw_nrof_folds, log_dir, step, summary_writer,best_accuracy, saver_save,model_dir,subdir,image_batch,args)
+
+                    best_accuracy = evaluate(sess, enqueue_op, image_paths_placeholder, labels_placeholder, phase_train_placeholder, batch_size_placeholder, embeddings, 
+                        label_batch, lfw_paths, actual_issame, args.lfw_batch_size, args.lfw_nrof_folds, log_dir, step, summary_writer,best_accuracy,saver_save,model_dir,subdir)
     return model_dir
 
 def find_threshold(var, percentile):