inital file upload

arcface_model/interpret_similarity_statistics.py

@@ -0,0 +1,118 @@
+import numpy as np
+import tensorflow as tf
+import argparse
+import os
+import pandas as pd
+
+from scipy.spatial.distance import cosine
+from tensorflow.keras.models import load_model, Model
+
+from tqdm import tqdm
+from PIL import Image
+
+
+def main(opt):
+    gpus = tf.config.experimental.list_physical_devices('GPU')
+    tf.config.set_visible_devices(gpus[opt.device_id], 'GPU')
+
+    blocks = ['conv5_block3_out',
+              'conv5_block2_out',
+              'conv5_block1_out',
+              'conv4_block6_out',
+              'conv4_block5_out',
+              'conv4_block4_out',
+              'conv4_block3_out',
+              'conv4_block2_out',
+              'conv4_block1_out',
+              'conv3_block4_out',
+              'conv3_block3_out',
+              'conv3_block2_out',
+              'conv3_block1_out',
+              'conv2_block3_out',
+              'conv2_block2_out',
+              'conv2_block1_out',
+              'conv1_relu'
+              ]
+
+    arcface = load_model(opt.arcface_path)
+    extractor = arcface.layers[1]
+    extractor.summary()
+    extractor.trainable = False
+
+    arcface_p = Model(extractor.input, [extractor.get_layer(layer_name).output for layer_name in blocks])
+
+    data_path = opt.data_path
+
+    name = opt.arcface_path.split("/")[0] + "/"
+    print(name)
+
+    if not os.path.isdir(name + "block_distances/" + opt.data_name):
+        os.mkdir(name + "block_distances/")
+        os.mkdir(name + "block_distances/" + opt.data_name)
+
+    # data dicts
+    cos_sim_t_s = {}
+    cos_sim_t_c = {}
+    cos_sim_c_s = {}
+    for block in blocks:
+        cos_sim_t_s[block] = []
+        cos_sim_t_c[block] = []
+        cos_sim_c_s[block] = []
+
+    print("[*] Calculating perceptual similarity between targets, sources and changes...")
+    for file_name in tqdm(os.listdir(data_path + 'target/'), total=len(os.listdir(data_path + 'target/'))):
+        target = np.asarray(Image.open(data_path + 'target/' + file_name).resize((112, 112))) / 255.0
+        source = np.asarray(Image.open(data_path + 'source/' + file_name).resize((112, 112))) / 255.0
+        change = np.asarray(Image.open(data_path + 'change/' + file_name).resize((112, 112))) / 255.0
+
+        target_features = arcface_p.predict(np.expand_dims(target, axis=0))
+        source_features = arcface_p.predict(np.expand_dims(source, axis=0))
+        change_features = arcface_p.predict(np.expand_dims(change, axis=0))
+
+        for t_f, s_f, c_f, block_name in zip(target_features, source_features, change_features, blocks):
+            t_f = tf.reshape(t_f, (1, -1))
+            s_f = tf.reshape(s_f, (1, -1))
+            c_f = tf.reshape(c_f, (1, -1))
+
+            t_s = cosine(t_f, s_f)
+            t_c = cosine(t_f, c_f)
+            c_s = cosine(c_f, s_f)
+
+            cos_sim_t_s[block_name].append(t_s)
+            cos_sim_t_c[block_name].append(t_c)
+            cos_sim_c_s[block_name].append(c_s)
+
+    print("[*] Saving distances as data frame pickles...")
+    for block in blocks:
+        df_t_s = pd.DataFrame(cos_sim_t_s[block], columns=['cos_sim'])
+        df_t_s.to_pickle(name + "block_distances/" + opt.data_name + "/arcface_" + block + "_t_s.pkl")
+
+        df_t_c = pd.DataFrame(cos_sim_t_c[block], columns=['cos_sim'])
+        df_t_c.to_pickle(name + "block_distances/" + opt.data_name + "/arcface_" + block + "_t_c.pkl")
+
+        df_c_s = pd.DataFrame(cos_sim_c_s[block], columns=['cos_sim'])
+        df_c_s.to_pickle(name + "block_distances/" + opt.data_name + "/arcface_" + block + "_c_s.pkl")
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--arcface_path', type=str, default="arcface/arc_res50.h5",
+                        help='path to arcface model')
+    parser.add_argument('--data_path', type=str, default="C:/path/to/facial_recognition_data/",
+                        help='path to data to run comparisons,'
+                             'structure should be: data_path/change, target, source/0.png, 1.png, ... n.png')
+    parser.add_argument('--device_id', type=int, default=0,
+                        help='which device to use')
+    parser.add_argument('--data_name', type=str, default='FF',
+                        help='name of the data set.')
+
+    opt = parser.parse_args()
+
+    main(opt)
+
+
+
+
+
+
+

arcface_model/plot_block_distances.py

@@ -0,0 +1,96 @@
+import pandas as pd
+import numpy as np
+import argparse
+import os
+from matplotlib import pyplot as plt
+
+
+def main(opt):
+    pd.set_option('display.max_rows', 1000)
+    pd.set_option('display.max_columns', 151)
+
+    dataset = opt.data_name
+    name = opt.arcface_path.split("/")[0] + "/"
+
+    meta_dict = {'af': 'ArcFace', 'fn': 'FaceNet', 'l2': 'L2',
+                 'sl2': 'squared L2', 'cos': 'cosine similarity', 'l1': 'L1'}
+    columns = ['cos_sim']
+
+    blocks = ['conv5_block3_out',
+              'conv5_block2_out',
+              'conv5_block1_out',
+              'conv4_block6_out',
+              'conv4_block5_out',
+              'conv4_block4_out',
+              'conv4_block3_out',
+              'conv4_block2_out',
+              'conv4_block1_out',
+              'conv3_block4_out',
+              'conv3_block3_out',
+              'conv3_block2_out',
+              'conv3_block1_out',
+              'conv2_block3_out',
+              'conv2_block2_out',
+              'conv2_block1_out',
+              'conv1_relu'
+              ]
+
+    t_c_margins = []
+
+    if not os.path.isdir(name + "block_distances_plots/"):
+        os.mkdir(name + "block_distances_plots")
+    if not os.path.isdir(name + "block_distances_plots/" + opt.data_name):
+        os.mkdir(name + "block_distances_plots/" + opt.data_name)
+
+    for block in blocks:
+
+        #positive_samples = pd.read_pickle('block_distances/imagenet_pos_neg_c2t_c2s/arcface_' + block + '_distance_df_pos.pkl')
+        t_s = pd.read_pickle(name + 'block_distances/' + dataset + '/arcface_' + block + '_t_s.pkl')
+        t_c = pd.read_pickle(name + 'block_distances/' + dataset + '/arcface_' + block + '_t_c.pkl')
+        c_s = pd.read_pickle(name + 'block_distances/' + dataset + '/arcface_' + block + '_c_s.pkl')
+
+        print(block)
+        print(t_s.mean())
+        print(t_c.mean())
+        print(c_s.mean())
+        print('==========================================================')
+        print()
+
+        for col in columns:
+            bins = np.linspace(0, 1.5, 100)
+
+            fig = plt.figure(dpi=500)
+            t_s_s, _, _ = plt.hist(t_s[col].values, bins, alpha=0.5, label='y')
+            t_c_s, _, _ = plt.hist(t_c[col].values, bins, alpha=0.5, label='y')
+            c_s_s, _, _ = plt.hist(c_s[col].values, bins, alpha=0.5, label='y')
+            plt.legend(['t2s', 't2c', 'c2s'], loc='upper right')
+
+            plt.axvline(t_c[col].values.mean(), color='k', linestyle='dashed', linewidth=1)
+
+            min_ylim, max_ylim = plt.ylim()
+            plt.text(t_c[col].values.mean() * 1.1, max_ylim * 0.9, 'mean {:.2f}'.format(t_c[col].values.mean()))
+
+            plt.title('ArcFace ResNet50 - ' + block + ' feature cosine similarity distance.')
+            plt.xlabel('distance')
+            plt.ylabel('facial comparisons')
+            plt.savefig(name + 'block_distances_plots/' + dataset + '/' + block + '.png')
+            plt.close()
+
+            t_c_margins.append(t_c[col].values.mean())
+
+    np.save(name + 'block_distances_plots/' + dataset + '/perceptual_similarity_margins.npy', np.asarray(t_c_margins))
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--arcface_path', type=str, default="arcface/arc_res50.h5",
+                        help='path to arcface model')
+    parser.add_argument('--data_path', type=str, default="D:/forensic_face_swap_data/SimSwap/",
+                        help='path to data to run comparisons,'
+                             'structure should be: data_path/change, target, source/0.png, 1.png, ... n.png')
+    parser.add_argument('--data_name', type=str, default='FF',
+                        help='name of the data set.')
+
+    opt = parser.parse_args()
+
+    main(opt)