Zalo AI Challenge - Landmark Identification
This code is ranked top-1 Final Leaderboard of the challenge.
Required Packages
$ pip3 install -r requirements.txtData
Data Preparation
Let's assume that the provided dataset is already in data folder:
- data/train_val2018.zip
- data/train_val2018.json
After extracting train_val2018.zip, we will have TrainVal folder containing all images inside data.
We run the code to prepare data for training:
- train/validation with the ratio of 92%/8%.
- splitting it into 10 folds in the cross-validation manner.
$ python3 src/data_prepare.pyAfter finished, all needed files are in data folder and ready for training. The class_dis.txt file contains distribution over all classes, which is also created inside data and later used for weighted loss in training.
Data Augmentation
- Random rotation
- Central cropping
- Color distortion
- Aspect distortion
- Random cropping
- Random horizontal flipping
Training
Model Architectures
- Inception-ResNet
- ResNet-152
Other architectures are also available: ResNet-50, ResNet-101, ResNet-200, Inception-V4, DenseNet-161, NASNet, PNASNet with pre-trained weights can be downloaded from Slim package.
Model Training
First, we need to download pre-trained weights of Inception-ResNet and ResNet-152 from Slim.
$ chmod 700 download.sh
$ ./download.shAll downloaded weights will be stored in pretrained folder.
To trained a new model with Inception-ResNet architecture:
$ python3 src/train.py --net "inception_resnet_v2" --checkpoint_exclude_scopes "InceptionResnetV2/Logits,InceptionResnetV2/AuxLogits" --loss_weighted "linear"With ResNet-152:
$ python3 src/train.py --net "resnet_v2_152" --checkpoint_exclude_scopes "resnet_v2_152/logits" --loss_weighted "linear"optional arguments:
-h, --help show this help message and exit
--data_dir DATA_DIR
Path to data folder (default: data)
--log_dir LOG_DIR
Path to data folder (default: log)
--checkpoint_dir CHECKPOINT_DIR
Path to checkpoint folder (default: checkpoints)
--num_checkpoints NUM_CHECKPOINTS
Number of checkpoints to store (default: 1)
--num_epochs NUM_EPOCHS
Number of training epochs (default: 50)
--num_threads NUM_THREADS
Number of threads for data processing (default: 8)
--batch_size BATCH_SIZE
Batch Size (default: 16)
--image_size IMAGE_SIZE
Size of images after pre-processed (default: 299)
--net NETWORK ARCHITECTURE
Network architecture to use (default: inception_resnet_v2)
--learning_rate LEARNING_RATE
Starting learning rate (default: 0.01)
--train_mode TRAINING_MODE
Training mode (default: 2)
--fold DATA_FOLD
Fold of data to train from (default: -1)
--optimizer OPTIMIZER
Weight update rule (default: momentum)
--lr_decay_rule LR_DECAY_RULE
Decay rule for learning rate (default: step)
--loss_weighted LOSS_WEIGHTED
Cross entropy weighted loss mechanism (default: linear)
--pretrained_dir PRETRAINED_DIR
Path to pre-trained weights (default: pretrained)
--trainable_scopes TRAINABLE_SCOPES
Scopes of variables to be fixed (default: None)
--ignore_mixing_vars IGNORE_MIXING_VARS
When restoring a checkpoint would ignore missing variables (default: False)
--allow_soft_placement ALLOW_SOFT_PLACEMENT
Allow device soft device placement (default: True)
After training, all models will be stored in checkpoints folder.
Final Ensemble Models
- 2 Inception-ResNet and 1 ResNet-152 models trained on 92% and tested on 8% of data.
- 10 Inception-ResNet models trained on 10 folds of cross-validation.
Prediction
Run predict.sh script with an argument which is the path to the data folder containing images. For example, all test images are in data/test:
$ chmod 700 predict.sh
$ ./predict.sh data/testThe final predictions will be in submission.csv file right under current directory.
The code are partially borrowed from Slim package including model architectures and data augmentation with some modifications.