Skip to content

Commit

Permalink
updated reaadme
Browse files Browse the repository at this point in the history
  • Loading branch information
Antoine Miech "WILLOW committed Jun 8, 2017
1 parent e132cc3 commit b3e2afe
Showing 1 changed file with 21 additions and 15 deletions.
36 changes: 21 additions & 15 deletions README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -22,15 +22,20 @@ It might be possible, but rather unlikely, that something is not working properl

Each of the following command lines train a single model. They are scheduled to stop training at the good time.

For the rest of the operations, please replace 'path_to_features' with the folder path which contains all the tensorflow record frame level feature.
Our model were trained on all the training set and almost all the validation set (we only discarded 21k videos). We will however, train the models on all both training and validation set as it was allowed in the kaggle competition. It should not make any huge difference.

Each model takes several days to train, so each command line are separated in order to be run in parallel if possible.
Please replace 'path_to_features' with the folder path which contains all the tensorflow record frame level feature.

```sh
For the rest of the operations, please replace 'path_to_features' with the folder path which contains all the tensorflow record frame level feature.
path_to_features='path_to_features'
```

Training Gated NetVLAD (256 Clusters):

```sh
python train.py --train_data_pattern='path_to_features/*a*??.tfrecord' --model=NetVLADModelLF --train_dir=gatednetvladLF-256k-1024-80-0002-300iter-norelu-basic-gatedmoe --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=80 --base_learning_rate=0.0002 --netvlad_cluster_size=256 --netvlad_hidden_size=1024 --moe_l2=1e-6 --iterations=300 --learning_rate_decay=0.8 --netvlad_relu=False --gating=True --moe_prob_gating=True --max_step=700000
python train.py --train_data_pattern='$path_to_features/*a*??.tfrecord' --model=NetVLADModelLF --train_dir=gatednetvladLF-256k-1024-80-0002-300iter-norelu-basic-gatedmoe --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=80 --base_learning_rate=0.0002 --netvlad_cluster_size=256 --netvlad_hidden_size=1024 --moe_l2=1e-6 --iterations=300 --learning_rate_decay=0.8 --netvlad_relu=False --gating=True --moe_prob_gating=True --max_step=700000
```

Note: The best single model is this model but with flag --max_step=300000. We somehow need it to train longer for better effect on the ensemble.
Expand All @@ -39,37 +44,37 @@ Training Gated NetFV (128 Clusters):


```sh
python train.py --train_data_pattern='path_to_features/*a*??.tfrecord' --model=NetFVModelLF --train_dir=gatednetfvLF-128k-1024-80-0002-300iter-norelu-basic-gatedmoe --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=80 --base_learning_rate=0.0002 --fv_cluster_size=128 --fv_hidden_size=1024 --moe_l2=1e-6 --iterations=300 --learning_rate_decay=0.8 --fv_relu=False --gating=True --moe_prob_gating=True --fv_couple_weights=False --max_step=600000
python train.py --train_data_pattern='$path_to_features/*a*??.tfrecord' --model=NetFVModelLF --train_dir=gatednetfvLF-128k-1024-80-0002-300iter-norelu-basic-gatedmoe --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=80 --base_learning_rate=0.0002 --fv_cluster_size=128 --fv_hidden_size=1024 --moe_l2=1e-6 --iterations=300 --learning_rate_decay=0.8 --fv_relu=False --gating=True --moe_prob_gating=True --fv_couple_weights=False --max_step=600000
```

Training Gated Soft-Dbof (4096 Clusters):

```sh
python train.py --train_data_pattern='path_to_features/*a*??.tfrecord' --model=GatedDbofModelLF --train_dir=gateddboflf-4096-1024-80-0002-300iter --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=80 --base_learning_rate=0.0002 --dbof_cluster_size=4096 --dbof_hidden_size=1024 --moe_l2=1e-6 --iterations=300 --dbof_relu=False --max_step=1000000
python train.py --train_data_pattern='$path_to_features/*a*??.tfrecord' --model=GatedDbofModelLF --train_dir=gateddboflf-4096-1024-80-0002-300iter --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=80 --base_learning_rate=0.0002 --dbof_cluster_size=4096 --dbof_hidden_size=1024 --moe_l2=1e-6 --iterations=300 --dbof_relu=False --max_step=1000000
```

Training Soft-Dbof (8000 Clusters):

```sh
python train.py --train_data_pattern='path_to_features/*a*??.tfrecord' --model=SoftDbofModelLF --train_dir=softdboflf-8000-1024-80-0002-300iter --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=80 --base_learning_rate=0.0002 --dbof_cluster_size=8000 --dbof_hidden_size=1024 --iterations=300 --dbof_relu=False --max_step=800000
python train.py --train_data_pattern='$path_to_features/*a*??.tfrecord' --model=SoftDbofModelLF --train_dir=softdboflf-8000-1024-80-0002-300iter --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=80 --base_learning_rate=0.0002 --dbof_cluster_size=8000 --dbof_hidden_size=1024 --iterations=300 --dbof_relu=False --max_step=800000
```

Training Gated RVLAD (256 Clusters):

```sh
python train.py --train_data_pattern='path_to_features/*a*??.tfrecord' --model=NetVLADModelLF --train_dir=gatedlightvladLF-256k-1024-80-0002-300iter-norelu-basic-gatedmoe --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=80 --base_learning_rate=0.0002 --netvlad_cluster_size=256 --netvlad_hidden_size=1024 --moe_l2=1e-6 --iterations=300 --learning_rate_decay=0.8 --netvlad_relu=False --gating=True --moe_prob_gating=True --lightvlad=True --max_step=600000
python train.py --train_data_pattern='$path_to_features/*a*??.tfrecord' --model=NetVLADModelLF --train_dir=gatedlightvladLF-256k-1024-80-0002-300iter-norelu-basic-gatedmoe --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=80 --base_learning_rate=0.0002 --netvlad_cluster_size=256 --netvlad_hidden_size=1024 --moe_l2=1e-6 --iterations=300 --learning_rate_decay=0.8 --netvlad_relu=False --gating=True --moe_prob_gating=True --lightvlad=True --max_step=600000
```

Training GRU (2 layers):

```sh
python train.py --train_data_pattern='path_to_features/*a*??.tfrecord' --model=GruModel --train_dir=GRU-0002-1200 --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=128 --base_learning_rate=0.0002 --gru_cells=1200 --learning_rate_decay=0.9 --moe_l2=1e-6 --max_step=300000
python train.py --train_data_pattern='$path_to_features/*a*??.tfrecord' --model=GruModel --train_dir=GRU-0002-1200 --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=128 --base_learning_rate=0.0002 --gru_cells=1200 --learning_rate_decay=0.9 --moe_l2=1e-6 --max_step=300000
```

Training LSTM (2 layers):

```sh
python train.py --train_data_pattern='path_to_features/*a*??.tfrecord' --model=LstmModel --train_dir=lstm-0002-val-150-random-2 --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=128 --base_learning_rate=0.0002 --iterations=150 --lstm_random_sequence=True --max_step=400000
python train.py --train_data_pattern='$path_to_features/*a*??.tfrecord' --model=LstmModel --train_dir=lstm-0002-val-150-random-2 --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=128 --base_learning_rate=0.0002 --iterations=150 --lstm_random_sequence=True --max_step=400000
```


Expand All @@ -79,17 +84,17 @@ We will write the predictions into 7 different files and then ensemble them.
Run each one of this command to run inference for a each model.

```sh
python inference.py --output_file=test-lstm-0002-val-150-random.csv --input_data_pattern='path_to_features/test*.tfrecord' --model=LstmModel --train_dir=lstm-0002-val-150-random --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=1024 --base_learning_rate=0.0002 --iterations=150 --lstm_random_sequence=True --run_once=True --top_k=50
python inference.py --output_file=test-lstm-0002-val-150-random.csv --input_data_pattern='$path_to_features/test*.tfrecord' --model=LstmModel --train_dir=lstm-0002-val-150-random --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=1024 --base_learning_rate=0.0002 --iterations=150 --lstm_random_sequence=True --run_once=True --top_k=50

python inference.py --output_file=test-gatedlightvladLF-256k-1024-80-0002-300iter-norelu-basic-gatedmoe.csv --input_data_pattern='path_to_features/test*.tfrecord' --model=NetVLADModelLF --train_dir=gatedlightvladLF-256k-1024-80-0002-300iter-norelu-basic-gatedmoe --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=1024 --base_learning_rate=0.0002 --netvlad_cluster_size=256 --netvlad_hidden_size=1024 --moe_l2=1e-6 --iterations=300 --learning_rate_decay=0.8 --netvlad_relu=False --gating=True --moe_prob_gating=True --lightvlad=True --run_once=True --top_k=50
python inference.py --output_file=test-gatedlightvladLF-256k-1024-80-0002-300iter-norelu-basic-gatedmoe.csv --input_data_pattern='$path_to_features/test*.tfrecord' --model=NetVLADModelLF --train_dir=gatedlightvladLF-256k-1024-80-0002-300iter-norelu-basic-gatedmoe --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=1024 --base_learning_rate=0.0002 --netvlad_cluster_size=256 --netvlad_hidden_size=1024 --moe_l2=1e-6 --iterations=300 --learning_rate_decay=0.8 --netvlad_relu=False --gating=True --moe_prob_gating=True --lightvlad=True --run_once=True --top_k=50

python inference.py --output_file=test-gateddboflf-4096-1024-80-0002-300iter-gatedmoe.csv --input_data_pattern='path_to_features/test*.tfrecord' --model=GatedDbofModelLF --train_dir=gateddboflf-4096-1024-80-0002-300iter-gatedmoe --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=512 --base_learning_rate=0.0002 --dbof_cluster_size=4096 --dbof_hidden_size=1024 --moe_l2=1e-6 --iterations=300 --dbof_relu=False --moe_prob_gating=True --run_once=True --top_k=50
python inference.py --output_file=test-gateddboflf-4096-1024-80-0002-300iter-gatedmoe.csv --input_data_pattern='$path_to_features/test*.tfrecord' --model=GatedDbofModelLF --train_dir=gateddboflf-4096-1024-80-0002-300iter-gatedmoe --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=512 --base_learning_rate=0.0002 --dbof_cluster_size=4096 --dbof_hidden_size=1024 --moe_l2=1e-6 --iterations=300 --dbof_relu=False --moe_prob_gating=True --run_once=True --top_k=50

python inference.py --output_file=test-gatednetfvLF-128k-1024-80-0002-300iter-norelu-basic-gatedmoe.csv --input_data_pattern='path_to_features/test*.tfrecord' --model=NetFVModelLF --train_dir=gatednetfvLF-128k-1024-80-0002-300iter-norelu-basic-gatedmoe --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=1024 --base_learning_rate=0.0002 --fv_cluster_size=128 --fv_hidden_size=1024 --moe_l2=1e-6 --iterations=300 --learning_rate_decay=0.8 --fv_relu=False --gating=True --moe_prob_gating=True --fv_couple_weights=False --top_k=50
python inference.py --output_file=test-gatednetfvLF-128k-1024-80-0002-300iter-norelu-basic-gatedmoe.csv --input_data_pattern='$path_to_features/test*.tfrecord' --model=NetFVModelLF --train_dir=gatednetfvLF-128k-1024-80-0002-300iter-norelu-basic-gatedmoe --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=1024 --base_learning_rate=0.0002 --fv_cluster_size=128 --fv_hidden_size=1024 --moe_l2=1e-6 --iterations=300 --learning_rate_decay=0.8 --fv_relu=False --gating=True --moe_prob_gating=True --fv_couple_weights=False --top_k=50

python inference.py --output_file=test-GRU-0002-1200.csv --input_data_pattern='path_to_features/test*.tfrecord' --model=GruModel --train_dir=GRU-0002-1200 --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=1024 --base_learning_rate=0.0002 --gru_cells=1200 --learning_rate_decay=0.9 --moe_l2=1e-6 --run_once=True --top_k=50
python inference.py --output_file=test-GRU-0002-1200.csv --input_data_pattern='$path_to_features/test*.tfrecord' --model=GruModel --train_dir=GRU-0002-1200 --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=1024 --base_learning_rate=0.0002 --gru_cells=1200 --learning_rate_decay=0.9 --moe_l2=1e-6 --run_once=True --top_k=50

python inference.py --output_file=test-gatednetvladLF-256k-1024-80-0002-300iter-norelu-basic-gatedmoe.csv --input_data_pattern='path_to_features/test*.tfrecord' --model=NetVLADModelLF --train_dir=gatednetvladLF-256k-1024-80-0002-300iter-norelu-basic-gatedmoe --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=1024 --base_learning_rate=0.0002 --netvlad_cluster_size=256 --netvlad_hidden_size=1024 --moe_l2=1e-6 --iterations=300 --learning_rate_decay=0.8 --netvlad_relu=False --gating=True --moe_prob_gating=True --run_once=True --top_k=50
python inference.py --output_file=test-gatednetvladLF-256k-1024-80-0002-300iter-norelu-basic-gatedmoe.csv --input_data_pattern='$path_to_features/test*.tfrecord' --model=NetVLADModelLF --train_dir=gatednetvladLF-256k-1024-80-0002-300iter-norelu-basic-gatedmoe --frame_features=True --feature_names="rgb,audio" --feature_sizes="1024,128" --batch_size=1024 --base_learning_rate=0.0002 --netvlad_cluster_size=256 --netvlad_hidden_size=1024 --moe_l2=1e-6 --iterations=300 --learning_rate_decay=0.8 --netvlad_relu=False --gating=True --moe_prob_gating=True --run_once=True --top_k=50
```

## Averaging the models
Expand All @@ -101,6 +106,7 @@ After inference done for all the models just run:
python final_averaging.py
```

After waiting several minutes the submission file will be written in WILLOW_submission.csv
It will just take you the time to make a coffee and the submission file will be written in WILLOW_submission.csv
as soon as you finished to drink it :D.

Antoine Miech

0 comments on commit b3e2afe

Please sign in to comment.