SCRATCH - image KNN with Xception network

.gitignore

@@ -0,0 +1 @@
+.ipynb_checkpoints

scratch/es-lsh-images/get_imagenet_vectors_labels.py

@@ -1,5 +1,5 @@
 from glob import glob
-from keras.applications import MobileNet
+from keras.applications import Xception
 from keras.applications.imagenet_utils import preprocess_input, decode_predictions
 from keras.models import Model
 from scipy.misc import imread, imresize, imshow
@@ -8,25 +8,9 @@
 import pdb
 
 
-imgs_dir = "/home/alex/Downloads/ILSVRC/Data/DET/test/"
-vecs_path = 'imagenet_vectors.npy'
-labels_path = 'imagenet_labels.txt'
-
-img_paths = sorted(glob('%s/*.JPEG' % imgs_dir))
-fp_labels = open(labels_path, 'w')
+def get_img(img_path, dim=224):
 
-batch_size = 500
-imgs_batch = np.zeros((batch_size, 224, 224, 3))
-
-# Instantiate model and chop off some layers.
-vector_layer = "conv_preds"
-model = MobileNet()
-model2 = Model(inputs=model.input, outputs=model.get_layer(vector_layer).output)
-model.summary()
-
-vecs = np.zeros((len(img_paths), model.output_shape[-1]))
-
-for i, img_path in enumerate(img_paths):
+    assert dim % 2 == 0
 
     img = imread(img_path)
 
@@ -38,40 +22,65 @@
         tmp[:, :, 2] = img
         img = tmp
 
-    # Resize and crop to fit network.
+    # Resize image.
     h0, w0, d = img.shape
     h1, w1 = h0, w0
     if h0 < w0:
-        h1 = 224
-        w1 = int(224 * w0 / h0)
+        h1 = dim
+        w1 = int(dim * w0 / h0)
     else:
-        w1 = 224
-        h1 = int(224 * h0 / w0)
+        w1 = dim
+        h1 = int(dim * h0 / w0)
     assert abs((h0 / w0) - (h1 / w1)) <= 0.1, "%d %d %d %d" % (h0, w0, h1, w1)
 
     img = imresize(img, (h1, w1, d))
 
-    # imgs_batch = img[np.newaxis, :224, :224, :].astype(np.float32)
-    # imgs_batch = preprocess_input(imgs_batch, mode='tf')
-    # prds_batch = model.predict(imgs_batch)
-    # print(decode_predictions(prds_batch))
-    # imshow(img)
+    # Crop image at the center.
+    # Width > height.
+    if w1 > h1:
+        c = int(w1 / 2)
+        o = int(dim / 2)
+        img = img[:, c - o: c + o, :]
+    # Height > width.
+    elif h1 > w1:
+        c = int(h1 / 2)
+        o = int(dim / 2)
+        img = img[c - o: c + o, :, :]
+
+    assert img.shape == (dim, dim, 3), '%s, %s' % (img_path, img.shape)
+
+    return img
+
+
+imgs_dir = "/home/alex/Downloads/ILSVRC/Data/DET/test/"
+vecs_path = 'imagenet_vectors.npy'
+
+img_paths = sorted(glob('%s/*.JPEG' % imgs_dir))
+fp_paths = open('imagenet_paths.txt', 'w')
+
+dim = 224
+batch_size = 500
+imgs_batch = np.zeros((batch_size, dim, dim, 3))
+
+# Instantiate model and chop off some layers.
+vector_layer = "avg_pool"
+m1 = Xception()
+m2 = Model(inputs=m1.input, outputs=m1.get_layer(vector_layer).output)
 
-    imgs_batch[i % batch_size] = img[:224, :224, :]
+vecs = np.zeros((len(img_paths), m2.output_shape[-1]))
 
-    if (i % batch_size == 0 and i > 0):
-        imgs_batch = preprocess_input(imgs_batch, mode='tf')
+for i in range(0, len(img_paths), batch_size):
 
-        labels = decode_predictions(model.predict(imgs_batch))
-        s = ""
-        for img_path, line in zip(img_paths[i - batch_size:i], labels):
-            s += img_path.split('/')[-1] + ' '
-            s += ','.join(map(lambda x: x[1], line)) + '\n'
-        fp_labels.write(s)
+    for j in range(batch_size):
+        imgs_batch[j] = get_img(img_paths[i + j], dim)
 
-        prd = model2.predict(imgs_batch)
-        vecs[i - batch_size:i] = prd.reshape(prd.shape[0], prd.shape[-1])
-        print('%d %.3lf %.3lf %.3lf' % (i, prd.min(), prd.mean(), prd.max()))
+    imgs_batch = preprocess_input(imgs_batch, mode='tf')
+    prds_batch = m2.predict(imgs_batch)
+    vecs[i:i + batch_size] = prds_batch
+    fp_paths.write('\n'.join(img_paths[i:i + batch_size]) + '\n')
 
+    print('%d-%d %.3lf %.3lf %.3lf' % (
+        i, i + batch_size, prds_batch.min(),
+        np.median(prds_batch), prds_batch.max()))
 
 np.save(vecs_path, vecs)