add the source code

.gitignore

@@ -0,0 +1,15 @@
+
+# for files opened by jupyter notebook, e.g. log files
+.nfs000* 
+
+.DS_Store
+._.DS_Store
+*.pt
+*.pyc
+__pycache__/
+.ipynb_checkpoints
+notebooks/
+log/
+
+# vs code files
+._* 

README.md

@@ -1,4 +1,51 @@
 # Divide and Conquer the Embedding Space for Metric Learning
-***Artsiom Sanakoyeu\*, Vadim Tschernezki\*, Uta Büchler, Björn Ommer*, In CVPR 2019**
 
-The PDF of the paper and the source code are coming soon.
+## About
+
+This repository contains the code for reproducing the results for [Divide and Conquer the Embedding Space for Metric Learning](http://openaccess.thecvf.com/content_CVPR_2019/papers/Sanakoyeu_Divide_and_Conquer_the_Embedding_Space_for_Metric_Learning_CVPR_2019_paper.pdf) (CVPR 2019) with the datasets [In-Shop Clothes](http://mmlab.ie.cuhk.edu.hk/projects/DeepFashion/InShopRetrieval.html), [Stanford Online Products](http://cvgl.stanford.edu/projects/lifted_struct/) and [PKU VehicleID](https://www.pkuml.org/resources/pku-vehicleid.html).
+
+## Requirements
+
+- Python version 3.6.6 or higher
+- SciPy and scikit-learn packages
+- PyTorch ([pytorch.org](http://pytorch.org))
+- Faiss with GPU support ([Faiss](https://github.com/facebookresearch/faiss))
+- download and extract the datasets for [In-Shop Clothes](http://mmlab.ie.cuhk.edu.hk/projects/DeepFashion/InShopRetrieval.html), [Stanford Online Products](http://cvgl.stanford.edu/projects/lifted_struct/) and [PKU VehicleID](https://www.pkuml.org/resources/pku-vehicleid.html)
+
+## Usage
+
+The following command will train the model with Margin loss on the In-Shop Clothes dataset for 200 epochs and a batch size of 80 while splitting the embedding layer with 8 clusters and finetuning the model from epoch 190 on. You can use this command to reproduce the results of the paper for the three datasets by changing simply `--dataset=inshop` to `--dataset=sop` (Stanford Online Products) or `--dataset=vid` (Vehicle-ID).
+
+```
+CUDA_VISIBLE_DEVICES=0 python experiment.py --dataset=inshop \
+--dir=test --exp=0 --random-seed=0 --nb-clusters=8 --nb-epochs=200 \
+--sz-batch=80 --backend=faiss-gpu  --embedding-lr=1e-5 --embedding-wd=1e-4 \
+--backbone-lr=1e-5 --backbone-wd=1e-4 --finetune-epoch=190
+```
+
+The model can be trained without the proposed method by setting the number of clusters to 1 with `--nb-clusters=1`.  
+For faster clustering we run Faiss on GPU. If you installed Faiss without GPU support use flag `--backend=faiss`.
+## Expected Results
+
+The model checkpoints and log files are saved in the selected log-directory. You can print a summary of the results with `python browse_results <log path>`.
+
+You will get slightly higher results than what we have reported in the paper. For SOP, In-Shop and Vehicle-ID the R@1 results should be somewhat around 76.40, 87.36 and 91.54.
+
+## License
+
+You may find out more about the license [here](LICENSE)
+
+## Reference
+
+If you use this code, please cite the following paper:
+
+Artsiom Sanakoyeu, Vadim Tschernezki, Uta Büchler, Björn Ommer. "Divide and Conquer the Embedding Space for Metric Learning", CVPR 2019.
+
+```
+@InProceedings{dcesml,
+  title={Divide and Conquer the Embedding Space for Metric Learning},
+  author={Sanakoyeu, Artsiom and Tschernezki, Vadim and B\"uchler, Uta and Ommer, Bj\"orn},
+  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
+  year={2019},
+}
+```

browse_results.py

@@ -0,0 +1,83 @@
+import shelve
+from collections import defaultdict
+import sys
+import os
+import numpy as np
+import pandas as pd
+import time
+import glob
+import argparse
+
+parser = argparse.ArgumentParser()
+parser.add_argument('exp_dir', type = str)
+parser.add_argument('-cw', '--col-width', type=int, default=100)
+args = parser.parse_args()
+print(args)
+
+print('exp_dir=', args.exp_dir)
+
+files = sorted(list(map(lambda x: x[:-4], glob.glob(os.path.join(args.exp_dir, '*.dat')))))
+
+
+results = defaultdict(list)
+ks = [1, 2, 4, 8, 10, 20, 30, 50]
+columns=[
+        'epoch',
+         *['R@{}'.format(i) for i in ks],
+        ]
+
+last_modified = None
+
+for p in files:
+    try:
+        db = shelve.open(p)
+        log_path = p + '.log'
+        assert os.path.exists(log_path), log_path
+        last_modified = (time.time() - os.path.getmtime(p + '.log')) / 60
+    except:
+        print('Failed to open', p)
+    try:
+        p = os.path.basename(p)
+        cur_results_t = np.array([(epoch, *d['score']['recall'])
+                             for (epoch, d) in db['metrics'].items()])
+        cur_results = np.zeros((cur_results_t.shape[0], 1 + len(ks)), dtype=float)
+        cur_results[:, :] = np.nan
+        cur_results[:, :2] = cur_results_t[:, :2]
+        # TODO: maybe rename args to config
+        if db['config']['dataset_selected'] == 'inshop':
+           cur_results[:, 5:] = cur_results_t[:, 2:]
+        else:
+           cur_results[:, 2:5] = cur_results_t[:, 2:]
+
+    except Exception as e:
+        print(p, e)
+        print(db['config'])
+
+    idx_max_recall = cur_results[:, 1].argmax()
+    best_epoch_results = cur_results[idx_max_recall]
+    max_epoch = cur_results[:, 0].max()
+    best_epoch_results = best_epoch_results.tolist()
+    best_epoch_results[0] = '{:02}/{:02}'.format(int(best_epoch_results[0]), int(max_epoch))
+    assert len(best_epoch_results) == len(columns)
+
+    for i, col_name in enumerate(columns):
+        results[col_name].append(best_epoch_results[i])
+
+    # if the file was last modified < 10 minute ago; than print Running status
+    if last_modified is None:
+        results['S'].append('?')
+    elif last_modified > 10:
+        results['S'].append('-')
+    else:
+        results['S'].append('[R]')
+
+
+df = pd.DataFrame(index=list(map(os.path.basename, files)),
+                  data=results)
+
+pd.set_option('display.max_rows', 10000)
+pd.set_option('display.max_columns', 10000)
+pd.set_option('display.max_colwidth', args.col_width)
+pd.set_option('display.width', 1000000)
+df.sort_values(by=['R@1'], inplace=True)
+print(df)

config.json

@@ -0,0 +1,71 @@
+{
+    "random_seed": 0,
+    "verbose": false,
+    "save_model": true,
+    "cuda_device": 0,
+    "sz_embedding": 128,
+    "backend": "faiss-gpu",
+    "nb_epochs": 200,
+    "nb_clusters": 8,
+    "finetune_epoch": 100,
+    "dataset_selected": "sop",
+    "dataset": {
+        "vid": {
+            "root": "/export/home/vtschern/data/vehicle-id/VehicleID_V1.0",
+            "classes": {
+                "train": "range(0, 13164)",
+                "init": "range(0, 13164)",
+                "_note": "small: 800, mid: +1600, large: +2400",
+                "eval": "range(13164, 13164 + 2400)"
+            }
+        },
+        "inshop": {
+            "root": "/export/home/vtschern/data/in-shop",
+            "classes": {
+                "train": "range(0, 3997)",
+                "init": "range(0, 3997)",
+                "eval": "range(0, 3985)"
+            }
+        },
+        "sop": {
+            "root": "/export/home/vtschern/data/sop",
+            "classes": {
+                "train": "range(0, 11318)",
+                "init": "range(0, 11318)",
+                "eval": "range(11318, 22634)"
+            }
+        }
+    },
+    "log": {
+        "path": "log/default",
+        "name": "sop-K-8-M-2-exp-0"
+    },
+    "dataloader": {
+        "num_workers": 4,
+        "drop_last": false,
+        "shuffle": false,
+        "pin_memory": true,
+        "batch_size": 80
+    },
+    "opt": {
+        "backbone": {
+            "lr": 1e-5,
+            "weight_decay": 1e-4
+        },
+        "embedding": {
+            "lr": 1e-5,
+            "weight_decay": 1e-4
+        }
+    },
+    "recluster": {
+        "enabled": true,
+        "mod_epoch": 2
+    },
+    "transform_parameters": {
+        "rgb_to_bgr": false,
+        "intensity_scale": [[0, 1], [0, 1]],
+        "mean": [0.485, 0.456, 0.406],
+        "std": [0.229, 0.224, 0.225],
+        "sz_crop": 224
+    }
+}

experiment.py

@@ -0,0 +1,67 @@
+from __future__ import print_function
+
+import argparse
+import math
+import matplotlib
+import sys
+
+import train
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--nb-clusters', required = True, type = int)
+    parser.add_argument('--dataset', dest = 'dataset_selected',
+        choices=['sop', 'inshop', 'vid'], required = True
+    )
+    parser.add_argument('--nb-epochs', type = int, default=200)
+    parser.add_argument('--finetune-epoch', type = int, default=190)
+    parser.add_argument('--mod-epoch', type = int, default=2)
+    parser.add_argument('--num-workers', default=4, type=int)
+    parser.add_argument('--sz-batch', type=int, default=80)
+    parser.add_argument('--sz-embedding', default=128, type=int)
+    parser.add_argument('--cuda-device', default = 0, type = int)
+    parser.add_argument('--exp', default='0', type=str, help='experiment identifier')
+    parser.add_argument('--dir', default='default', type=str)
+    parser.add_argument('--backend', default='faiss',
+        choices=('torch+sklearn', 'faiss', 'faiss-gpu')
+    )
+    parser.add_argument('--random-seed', default = 0, type = int)
+    parser.add_argument('--backbone-wd', default=1e-4, type=float)
+    parser.add_argument('--backbone-lr', default=1e-5, type=float)
+    parser.add_argument('--embedding-lr', default=1e-5, type=float)
+    parser.add_argument('--embedding-wd', default=1e-4, type=float)
+    parser.add_argument('--verbose', action = 'store_true')
+    args = vars(parser.parse_args())
+
+    config = train.load_config(config_name = 'config.json')
+
+    config['dataloader']['batch_size'] = args.pop('sz_batch')
+    config['dataloader']['num_workers'] = args.pop('num_workers')
+    config['recluster']['mod_epoch'] = args.pop('mod_epoch')
+    config['opt']['backbone']['lr'] = args.pop('backbone_lr')
+    config['opt']['backbone']['weight_decay'] = args.pop('backbone_wd')
+    config['opt']['embedding']['lr'] = args.pop('embedding_lr')
+    config['opt']['embedding']['weight_decay'] = args.pop('embedding_wd')
+
+    for k in args:
+        if k in config:
+            config[k] = args[k]
+
+    if config['nb_clusters'] == 1:
+        config['recluster']['enabled'] = False
+
+    config['log'] = {
+        'name': '{}-K-{}-M-{}-exp-{}'.format(
+            config['dataset_selected'],
+            config['nb_clusters'],
+            config['recluster']['mod_epoch'],
+            args['exp']
+        ),
+        'path': 'log/{}'.format(args['dir'])
+    }
+
+    # tkinter not installed on this system, use non-GUI backend
+    matplotlib.use('agg')
+    train.start(config)
+

lib/.gitignore

@@ -0,0 +1,8 @@
+
+.DS_Store
+._.DS_Store
+__pycache__/
+.ipynb_checkpoints/
+._*
+.nfs*
+

lib/__init__.py

@@ -0,0 +1,8 @@
+from . import loss
+from . import utils
+from . import evaluation
+from . import similarity
+from . import model
+from . import data
+from . import clustering
+