add processing script to automatically handle initial db process steps

.gitignore

@@ -0,0 +1,111 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+*.sw[op]
+*.pkl
+*pkl.gz
+results/
+*.pth
+*.DS_Store
+*.Rhistory
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# pyenv
+.python-version
+
+# celery beat schedule file
+celerybeat-schedule
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/

README.txt

@@ -15,11 +15,9 @@ Recommended hardware: 4 NVIDIA Tesla P-100 GPUs or 8 NVIDIA Tesla K-80 GPUs
 Instructions for preparing the data:
 1. Download the dataset CUB_200_2011.tgz from http://www.vision.caltech.edu/visipedia/CUB-200-2011.html
 2. Unpack CUB_200_2011.tgz
-3. Crop the images using information from bounding_boxes.txt (included in the dataset)
-4. Split the cropped images into training and test sets, using train_test_split.txt (included in the dataset)
-5. Put the cropped training images in the directory "./datasets/cub200_cropped/train_cropped/"
-6. Put the cropped test images in the directory "./datasets/cub200_cropped/test_cropped/"
-7. Augment the training set using img_aug.py (included in this code package)
+3. Run initial_images_processing.py to process dataset into proper form
+   -- Ex. python initial_images_processing.py /path/to/CUB_200_2011/images/
+4. Augment the training set using img_aug.py (included in this code package)
    -- this will create an augmented training set in the following directory:
       "./datasets/cub200_cropped/train_cropped_augmented/"
 

environment.yml

@@ -0,0 +1,17 @@
+name: protopnet
+dependencies:
+  - python=3.7
+  - ipython
+  - numpy
+  - pandas
+  - pytorch
+  - torchvision
+  - cudatoolkit
+  - scipy
+  - opencv
+  - matplotlib
+  - pip:
+    - Augmentor
+channels:
+  - pytorch
+  - conda-forge

img_aug.py

@@ -1,5 +1,7 @@
 import Augmentor
 import os
+
+
 def makedir(path):
     '''
     if path does not exist in the file system, create it
@@ -17,7 +19,7 @@ def makedir(path):
 
 for i in range(len(folders)):
     fd = folders[i]
-    tfd = target_folders[i]
+    tfd = os.path.abspath(target_folders[i])
     # rotation
     p = Augmentor.Pipeline(source_directory=fd, output_directory=tfd)
     p.rotate(probability=1, max_left_rotation=15, max_right_rotation=15)

initial_images_processing.py

@@ -0,0 +1,62 @@
+import argparse
+from pathlib import Path
+
+import cv2
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('images_dir', help='relative path to the images folder as described in CUB 200-2011')
+    args = parser.parse_args()
+
+    imgs_txt = Path(args.images_dir).joinpath('../images.txt')
+    bb_txt = Path(args.images_dir).joinpath('../bounding_boxes.txt')
+    train_test = Path(args.images_dir).joinpath('../train_test_split.txt')
+    train_dir = Path(__file__).parent.joinpath('datasets/cub200_cropped/train_cropped')
+    test_dir = Path(__file__).parent.joinpath('datasets/cub200_cropped/test_cropped')
+    if not train_dir.exists():
+        train_dir.mkdir(parents=True)
+    if not test_dir.exists():
+        test_dir.mkdir(parents=True)
+
+    imgs_to_data = []
+    with open(str(imgs_txt)) as imgs:
+        img_index = imgs.readlines()
+    with open(str(bb_txt)) as bb:
+        bb_index = bb.readlines()
+    with open(str(train_test)) as tt:
+        tt_index = tt.readlines()
+
+    for i, line in enumerate(img_index):
+        n1, filename = line.strip().split(' ')
+        n2, x, y, width, height = bb_index[i].strip().split(' ')
+        n3, is_train = tt_index[i].strip().split(' ')
+        if n1 != n2 or n2 != n3:
+            raise Exception('something went wrong and indexing on images.txt/bounding_boxes.txt/train_test_split.txt is off')
+        imgs_to_data.append([
+            Path(args.images_dir).joinpath(filename),
+            int(float(x)),
+            int(float(y)),
+            int(float(width)),
+            int(float(height)),
+            bool(int(is_train)),
+        ])
+
+    for path, x, y, w, h, is_train in imgs_to_data:
+        im = cv2.imread(str(path))
+        # crop and save
+        im = im[y:y+h, x:x+w, :]
+
+        if is_train:
+            final_dir = train_dir.joinpath(path.parent.name)
+        else:
+            final_dir = test_dir.joinpath(path.parent.name)
+        final_path = final_dir.joinpath(path.name)
+        if not final_dir.exists():
+            final_dir.mkdir()
+
+        cv2.imwrite(str(final_path), im)
+
+
+if __name__ == "__main__":
+    main()

main.py

@@ -1,5 +1,6 @@
 import os
 import shutil
+import copy
 
 import torch
 import torch.utils.data
@@ -12,12 +13,14 @@
 
 from helpers import makedir
 import model
-import push
+from push import Pusher
 import prune
 import train_and_test as tnt
 import save
 from log import create_logger
 from preprocess import mean, std, preprocess_input_function
+from settings import *
+
 
 parser = argparse.ArgumentParser()
 parser.add_argument('-gpuid', nargs=1, type=str, default='0') # python3 main.py -gpuid=0,1,2,3
@@ -26,8 +29,6 @@
 print(os.environ['CUDA_VISIBLE_DEVICES'])
 
 # book keeping namings and code
-from settings import base_architecture, img_size, prototype_shape, num_classes, \
-                     prototype_activation_function, add_on_layers_type, experiment_run
 
 base_architecture_type = re.match('^[a-z]*', base_architecture).group(0)
 
@@ -48,8 +49,14 @@
 proto_bound_boxes_filename_prefix = 'bb'
 
 # load the data
-from settings import train_dir, test_dir, train_push_dir, \
-                     train_batch_size, test_batch_size, train_push_batch_size
+
+
+def perform_push(pusher, epoch_number):
+    if use_protobank:
+        pusher.push_protobank(epoch_number)
+    else:
+        pusher.push_orig(epoch_number)
+
 
 normalize = transforms.Normalize(mean=mean,
                                  std=std)
@@ -100,43 +107,63 @@
                               prototype_shape=prototype_shape,
                               num_classes=num_classes,
                               prototype_activation_function=prototype_activation_function,
-                              add_on_layers_type=add_on_layers_type)
+                              add_on_layers_type=add_on_layers_type,
+                              bank_size=bank_size)
 #if prototype_activation_function == 'linear':
 #    ppnet.set_last_layer_incorrect_connection(incorrect_strength=0)
 ppnet = ppnet.cuda()
 ppnet_multi = torch.nn.DataParallel(ppnet)
 class_specific = True
 
 # define optimizer
-from settings import joint_optimizer_lrs, joint_lr_step_size
+if use_protobank:
+    prototype_params = {
+        'params': ppnet.protobank_tensor,
+        'lr': joint_optimizer_lrs['prototype_vectors']
+    }
+else:
+    prototype_params = {
+        'params': ppnet.prototype_vectors,
+        'lr': joint_optimizer_lrs['prototype_vectors']
+    }
+
 joint_optimizer_specs = \
 [{'params': ppnet.features.parameters(), 'lr': joint_optimizer_lrs['features'], 'weight_decay': 1e-3}, # bias are now also being regularized
  {'params': ppnet.add_on_layers.parameters(), 'lr': joint_optimizer_lrs['add_on_layers'], 'weight_decay': 1e-3},
- {'params': ppnet.prototype_vectors, 'lr': joint_optimizer_lrs['prototype_vectors']},
+ prototype_params,
 ]
 joint_optimizer = torch.optim.Adam(joint_optimizer_specs)
 joint_lr_scheduler = torch.optim.lr_scheduler.StepLR(joint_optimizer, step_size=joint_lr_step_size, gamma=0.1)
 
-from settings import warm_optimizer_lrs
 warm_optimizer_specs = \
 [{'params': ppnet.add_on_layers.parameters(), 'lr': warm_optimizer_lrs['add_on_layers'], 'weight_decay': 1e-3},
- {'params': ppnet.prototype_vectors, 'lr': warm_optimizer_lrs['prototype_vectors']},
+ prototype_params,
 ]
 warm_optimizer = torch.optim.Adam(warm_optimizer_specs)
 
-from settings import last_layer_optimizer_lr
 last_layer_optimizer_specs = [{'params': ppnet.last_layer.parameters(), 'lr': last_layer_optimizer_lr}]
 last_layer_optimizer = torch.optim.Adam(last_layer_optimizer_specs)
 
-# weighting of different training losses
-from settings import coefs
-
-# number of training epochs, number of warm epochs, push start epoch, push epochs
-from settings import num_train_epochs, num_warm_epochs, push_start, push_epochs
-
 # train the model
 log('start training')
-import copy
+pusher = Pusher(
+    train_push_loader,
+    prototype_network_parallel=ppnet_multi,
+    bank_size=bank_size,
+    class_specific=class_specific,
+    preprocess_input_function=preprocess_input_function, # normalize if needed
+    prototype_layer_stride=1,
+    dir_for_saving_prototypes=img_dir, # if not None, prototypes will be saved here
+    prototype_img_filename_prefix=prototype_img_filename_prefix,
+    prototype_self_act_filename_prefix=prototype_self_act_filename_prefix,
+    proto_bound_boxes_filename_prefix=proto_bound_boxes_filename_prefix,
+    save_prototype_class_identity=True,
+    log=log
+)
+
+if do_initial_push:
+    perform_push(pusher, 0)
+
 for epoch in range(num_train_epochs):
     log('epoch: \t{0}'.format(epoch))
 
@@ -156,19 +183,7 @@
                                 target_accu=0.70, log=log)
 
     if epoch >= push_start and epoch in push_epochs:
-        push.push_prototypes(
-            train_push_loader, # pytorch dataloader (must be unnormalized in [0,1])
-            prototype_network_parallel=ppnet_multi, # pytorch network with prototype_vectors
-            class_specific=class_specific,
-            preprocess_input_function=preprocess_input_function, # normalize if needed
-            prototype_layer_stride=1,
-            root_dir_for_saving_prototypes=img_dir, # if not None, prototypes will be saved here
-            epoch_number=epoch, # if not provided, prototypes saved previously will be overwritten
-            prototype_img_filename_prefix=prototype_img_filename_prefix,
-            prototype_self_act_filename_prefix=prototype_self_act_filename_prefix,
-            proto_bound_boxes_filename_prefix=proto_bound_boxes_filename_prefix,
-            save_prototype_class_identity=True,
-            log=log)
+        perform_push(pusher, epoch)
         accu = tnt.test(model=ppnet_multi, dataloader=test_loader,
                         class_specific=class_specific, log=log)
         save.save_model_w_condition(model=ppnet, model_dir=model_dir, model_name=str(epoch) + 'push', accu=accu,
@@ -184,6 +199,5 @@
                                 class_specific=class_specific, log=log)
                 save.save_model_w_condition(model=ppnet, model_dir=model_dir, model_name=str(epoch) + '_' + str(i) + 'push', accu=accu,
                                             target_accu=0.70, log=log)
-
-logclose()
 
+logclose()