Merge pull request #9 from QVPR/conda_pypi_fixes

Fixes for pypi/conda release

.github/workflows/python-publish.yml

@@ -0,0 +1,29 @@
+# This workflows will upload a Python Package using Twine when a release is created
+# For more information see: https://help.github.com/en/actions/language-and-framework-guides/using-python-with-github-actions#publishing-to-package-registries
+
+name: Upload Python Package
+
+on:
+  release:
+    types: [created]
+
+jobs:
+  deploy:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v2
+      - name: Set up Python
+        uses: actions/setup-python@v2
+        with:
+          python-version: '3.x'
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install setuptools wheel twine
+      - name: Build and publish
+        env:
+          TWINE_USERNAME: __token__
+          TWINE_PASSWORD: ${{ secrets.PYPI_TOKEN }}
+        run: |
+          python setup.py sdist bdist_wheel
+          twine upload dist/*

.gitignore

@@ -4,6 +4,7 @@
 patchnetvlad/pretrained_models/*
 patchnetvlad/results/*
 patchnetvlad/output_features/*
+results/*
 
 ## Python stuff
 # Byte-compiled / optimized / DLL files

feature_extract.py

@@ -164,7 +164,7 @@ def main():
 
     torch.cuda.empty_cache()  # garbage clean GPU memory, a bug can occur when Pytorch doesn't automatically clear the
     # memory after runs
-    print('Done. Finished extracting and saving features')
+    print('\n\nDone. Finished extracting and saving features')
 
 
 if __name__ == "__main__":

feature_match.py

@@ -73,7 +73,9 @@ def compute_recall(gt, predictions, numQ, n_values, recall_str=''):
 def write_kapture_output(opt, eval_set, predictions, outfile_name):
     if not exists(opt.result_save_folder):
         os.mkdir(opt.result_save_folder)
-    with open(join(opt.result_save_folder, outfile_name), 'w') as kap_out:
+    outfile = join(opt.result_save_folder, outfile_name)
+    print('Writing results to', outfile)
+    with open(outfile, 'w') as kap_out:
         kap_out.write('# kapture format: 1.0\n')
         kap_out.write('# query_image, map_image\n')
         image_list_array = np.array(eval_set.images)
@@ -87,7 +89,9 @@ def write_kapture_output(opt, eval_set, predictions, outfile_name):
 def write_recalls_output(opt, recalls_netvlad, recalls_patchnetvlad, n_values):
     if not exists(opt.result_save_folder):
         os.mkdir(opt.result_save_folder)
-    with open(join(opt.result_save_folder, 'recalls.txt'), 'w') as rec_out:
+    outfile = join(opt.result_save_folder, 'recalls.txt')
+    print('Writing recalls to', outfile)
+    with open(outfile, 'w') as rec_out:
         for n in n_values:
             rec_out.write("Recall {}@{}: {:.4f}\n".format('NetVLAD', n, recalls_netvlad[n]))
         for n in n_values:
@@ -120,7 +124,8 @@ def feature_match(eval_set, device, opt, config):
     if config['feature_match']['pred_input_path'] != 'None':
         predictions = np.load(config['feature_match']['pred_input_path'])  # optionally load predictions from a np file
     else:
-        if opt.ground_truth_path.split('/')[1][:-4] == 'tokyo247':
+        if opt.ground_truth_path and 'tokyo247' in opt.ground_truth_path:
+            print('Tokyo24/7: Selecting only one of the 12 cutouts per panorama')
             # followed nnSearchPostprocess in https://github.com/Relja/netvlad/blob/master/datasets/dbTokyo247.m
             # noinspection PyArgumentList
             _, predictions = faiss_index.search(qFeat, max(n_values) * 12)  # 12 cutouts per panorama
@@ -133,7 +138,7 @@ def feature_match(eval_set, device, opt, config):
             predictions = np.array(predictions_new)
         else:
             # noinspection PyArgumentList
-            _, predictions = faiss_index.search(qFeat, max(n_values))
+            _, predictions = faiss_index.search(qFeat, min(len(qFeat), max(n_values)))
 
     reranked_predictions = local_matcher(predictions, eval_set, input_query_local_features_prefix,
                                          input_index_local_features_prefix, config, device)
@@ -142,16 +147,19 @@ def feature_match(eval_set, device, opt, config):
     write_kapture_output(opt, eval_set, predictions, 'NetVLAD_predictions.txt')
     write_kapture_output(opt, eval_set, reranked_predictions, 'PatchNetVLAD_predictions.txt')
 
-    print('Finished matching features. About to eval GT if GT was provided')
+    print('Finished matching features.')
 
     # for each query get those within threshold distance
     if opt.ground_truth_path is not None:
+        print('Calculating recalls using ground truth.')
         gt = eval_set.get_positives()
 
         global_recalls = compute_recall(gt, predictions, eval_set.numQ, n_values, 'NetVLAD')
         local_recalls = compute_recall(gt, reranked_predictions, eval_set.numQ, n_values, 'PatchNetVLAD')
 
         write_recalls_output(opt, global_recalls, local_recalls, n_values)
+    else:
+        print('No ground truth was provided; not calculating recalls.')
 
 
 def main():

match_two.py

@@ -158,7 +158,7 @@ def match_two(model, device, opt, config):
     scores, inlier_keypoints_one, inlier_keypoints_two = matcher.match(local_feats_one, local_feats_two)
     score = -normalise_func(scores, len(patch_sizes), patch_weights)
 
-    print("Similarity score between the two images is: '{:.5f}'. In this example, a larger score indicates a better match.".format(score))
+    print(f"Similarity score between the two images is: {score:.5f}. Larger scores indicate better matches.")
 
     if config['feature_match']['matcher'] == 'RANSAC':
         tqdm.write('====> Plotting Local Features')

patchnetvlad/models/local_matcher.py

@@ -149,17 +149,17 @@ def local_matcher(predictions, eval_set, input_query_local_features_prefix,
 
     for q_idx, pred in enumerate(tqdm(predictions, leave=False, desc='Patch compare pred')):
         diffs = np.zeros((predictions.shape[1], len(patch_sizes)))
-        image_name = os.path.splitext(os.path.basename(eval_set.images[eval_set.numDb + q_idx]))[0]
+        image_name_query = os.path.splitext(os.path.basename(eval_set.images[eval_set.numDb + q_idx]))[0]
         qfeat = []
         for patch_size in patch_sizes:
-            qfilename = input_query_local_features_prefix + '_' + 'psize{}_'.format(patch_size) + image_name + '.npy'
+            qfilename = input_query_local_features_prefix + '_' + 'psize{}_'.format(patch_size) + image_name_query + '.npy'
             qfeat.append(torch.transpose(torch.tensor(np.load(qfilename), device=device), 0, 1))
             # we pre-transpose here to save compute speed
         for k, candidate in enumerate(pred):
-            image_name = os.path.splitext(os.path.basename(eval_set.images[candidate]))[0]
+            image_name_index = os.path.splitext(os.path.basename(eval_set.images[candidate]))[0]
             dbfeat = []
             for patch_size in patch_sizes:
-                dbfilename = input_index_local_features_prefix + '_' + 'psize{}_'.format(patch_size) + image_name + '.npy'
+                dbfilename = input_index_local_features_prefix + '_' + 'psize{}_'.format(patch_size) + image_name_index + '.npy'
                 dbfeat.append(torch.tensor(np.load(dbfilename), device=device))
 
             diffs[k, :], _, _ = matcher.match(qfeat, dbfeat)

patchnetvlad/tools/datasets.py

@@ -33,6 +33,8 @@
 from PIL import Image
 from sklearn.neighbors import NearestNeighbors
 
+from patchnetvlad.tools import PATCHNETVLAD_ROOT_DIR
+
 
 class PlaceDataset(data.Dataset):
     def __init__(self, query_file_path, index_file_path, dataset_root_dir, ground_truth_path, config):
@@ -52,6 +54,10 @@ def __init__(self, query_file_path, index_file_path, dataset_root_dir, ground_tr
             self.images = self.database
 
         self.images = [os.path.join(dataset_root_dir, image) for image in self.images]
+        # check if images are relative to root dir
+        if not os.path.isfile(self.images[0]):
+            if os.path.isfile(os.path.join(PATCHNETVLAD_ROOT_DIR, self.images[0])):
+                self.images = [os.path.join(PATCHNETVLAD_ROOT_DIR, image) for image in self.images]
 
         self.positives = None
         self.distances = None

setup.py

@@ -1,4 +1,4 @@
-import os
+import os, sys
 from setuptools import setup, find_packages
 
 
@@ -9,8 +9,19 @@
     long_description = f.read()
 
 
+install_require_list = [
+    'numpy', 'torch', 'torchvision',
+    'tqdm', 'scipy', 'Pillow', 'scikit-learn',
+    'faiss', 'natsort']
+
+# workaround as opencv-python does not show up in "pip list" within a conda environment
+# we do not care as conda recipe has py-opencv requirement anyhow
+is_conda = os.path.exists(os.path.join(sys.prefix, 'conda-meta'))
+if not is_conda:
+    install_require_list.append('opencv-python')
+
 setup(name='patchnetvlad',
-      version='0.1.0',
+      version='0.1.1',
       description='Patch-NetVLAD: An open-source Python implementation of the CVPR2021 paper',
       long_description = long_description,
       long_description_content_type='text/markdown',
@@ -37,10 +48,7 @@
         'Programming Language :: Python :: 3.9',
       ],
       python_requires='>=3.6',
-      install_requires=[
-          'numpy', 'torch', 'torchvision',
-          'tqdm', 'scipy', 'Pillow', 'scikit-learn',
-          'faiss', 'natsort', 'opencv-python'],
+      install_requires=install_require_list,
       packages=find_packages(),
       keywords=[
           'python', 'place recognition', 'image retrieval', 'computer vision', 'robotics'