changing directory name for windows compatibility

Author: spmallick

PyTorch-Multi-Label-Image-Classification-Image-Tagging/Pipeline.ipynb

@@ -0,0 +1,57 @@
+# Setup
+Before installation create and activate virtual environment
+```bash
+python3 -m venv venv
+source venv/bin/activate
+```
+Install the dependencies
+```bash
+pip install -r requirements.txt
+```
+
+# Training
+For training run [jupyter notebook](Pipeline.ipynb)
+
+# Additional instructions
+## Data preparation
+We use [the NUS-WIDE dataset](https://lms.comp.nus.edu.sg/wp-content/uploads/2019/research/nuswide/NUS-WIDE.html) for this tutorial. 
+Instead of parsing Flickr for image downloading we use [a dump](https://drive.google.com/open?id=0B7IzDz-4yH_HSmpjSTlFeUlSS00) from [this github repository](https://github.com/thuml/HashNet/tree/master/pytorch#datasets)
+Download and extract it.
+
+Also, we added pre-processed annotations:  
+```nus_wide/train.json```  
+```nus_wide/test.json``` 
+
+If you want to create them yourself, run the command:
+```bash
+python split_data_nus.py -i images
+```
+where ``` -i images``` is the path to the folder with extracted images
+
+## Subset creation
+You can train the model for the entire data set, but it takes a lot of time. For this tutorial we use part of this data.
+
+For subset creation run the command: 
+```bash
+python create_subset.py -i images
+```
+where ``` -i images``` is the path to the folder with extracted images
+
+Additional options:
+```bash
+python create_subset.py -h
+usage: Subset creation [-h] -i IMG_PATH [-v VAL_SIZE] [-t TRAIN_SIZE]
+                       [--shuffle] [-l LABELS [LABELS ...]]
+
+optional arguments:
+  -h, --help            show this help message and exit
+  -i IMG_PATH, --img-path IMG_PATH
+                        Path to the "images" folder
+  -v VAL_SIZE, --val-size VAL_SIZE
+                        Size of the validation data
+  -t TRAIN_SIZE, --train-size TRAIN_SIZE
+                        Size of the train data
+  --shuffle             Shuffle samples before splitting
+  -l LABELS [LABELS ...], --labels LABELS [LABELS ...]
+                        Subset labels
+```

PyTorch-Multi-Label-Image-Classification-Image-Tagging/README.md

@@ -0,0 +1,74 @@
+import argparse
+import json
+import os
+from random import shuffle
+
+
+def main():
+    parser = argparse.ArgumentParser('Subset creation')
+    parser.add_argument("-i", "--img-path", required=True, type=str, help='Path to the "images" folder')
+    parser.add_argument("-v", "--val-size", default=1000, type=int, help='Size of the validation data')
+    parser.add_argument("-t", "--train-size", default=5000, type=int, help='Size of the train data')
+    parser.add_argument('--shuffle', action='store_true', help='Shuffle samples before splitting')
+    parser.add_argument("-l", "--labels", nargs='+', default=['house', 'birds', 'sun', 'valley',
+                                                              'nighttime', 'boats', 'mountain', 'tree', 'snow',
+                                                              'beach', 'vehicle', 'rocks',
+                                                              'reflection', 'sunset', 'road', 'flowers', 'ocean',
+                                                              'lake', 'window', 'plants',
+                                                              'buildings', 'grass', 'water', 'animal', 'person',
+                                                              'clouds', 'sky'], help='Subset labels')
+    args = parser.parse_args()
+    img_path = args.img_path
+    labels = args.labels
+
+    with open('nus_wide/cats') as l_f:
+        possible_labels = l_f.readlines()
+        possible_labels = [i.strip() for i in possible_labels]
+
+    for label in labels:
+        if label not in possible_labels:
+            print('Label:', label, "is unknown. Possible labels:", ', '.join(possible_labels))
+            exit(-1)
+
+    with open(os.path.join(img_path, 'test.json')) as fp:
+        test_data = json.load(fp)
+    test_samples = test_data['samples']
+
+    with open(os.path.join(img_path, 'train.json')) as fp:
+        train_data = json.load(fp)
+    train_samples = train_data['samples']
+
+    if args.shuffle:
+        shuffle(test_samples)
+        shuffle(train_samples)
+
+    train_size = args.train_size
+    test_size = args.val_size
+
+    small_train = []
+    i = 0
+    while len(small_train) < train_size:
+        sample_img_path, sample_labels = train_samples[i]['image_name'], train_samples[i]['image_labels']
+        sample_labels = [label for label in sample_labels if label in labels]
+        if len(sample_labels):
+            small_train.append({'image_name': sample_img_path, 'image_labels': sample_labels})
+        i += 1
+
+    small_test = []
+    i = 0
+    while len(small_test) < test_size:
+        sample_img_path, sample_labels = test_samples[i]['image_name'], test_samples[i]['image_labels']
+        sample_labels = [label for label in sample_labels if label in labels]
+        if len(sample_labels):
+            small_test.append({'image_name': sample_img_path, 'image_labels': sample_labels})
+        i += 1
+
+    with open(os.path.join(img_path, 'small_train.json'), 'w') as fp:
+        json.dump({'samples': small_train, 'labels': labels}, fp, indent=3)
+
+    with open(os.path.join(img_path, 'small_test.json'), 'w') as fp:
+        json.dump({'samples': small_test, 'labels': labels}, fp, indent=3)
+
+
+if __name__ == '__main__':
+    main()