Skip to content

Latest commit

 

History

History
286 lines (256 loc) · 7.26 KB

JSSP.md

File metadata and controls

286 lines (256 loc) · 7.26 KB
Raw

Solving JSSP Problems

Quickstart

Visdom

Launch the visdom logging server:

python -m visdom.server

Trainings are displayed on localhost. More information about visdom here.

Training

Launch a training run:

python3 -m jssp.train\
    --batch_size 245\
    --clip_range 0.20\
    --custom_heuristic_names SPT MWKR MOPNR FDD/MWKR\
    --device cuda:0\
    --duration_type deterministic\
    --ent_coef 0.05\
    --exp_name_appendix QUICKSTART_RUN\
    --fe_type dgl\
    --fixed_validation\
    --gae_lambda 0.99\
    --gamma 1.00\
    --graph_has_relu\
    --graph_pooling max\
    --hidden_dim_actor 32\
    --hidden_dim_critic 32\
    --hidden_dim_features_extractor 64\
    --layer_pooling last\
    --lr 1e-4\
    --max_n_j 100\
    --max_n_m 30\
    --mlp_act gelu\
    --n_epochs 3\
    --n_j 10\
    --n_layers_features_extractor 10\
    --n_m 10\
    --n_mlp_layers_actor 1\
    --n_mlp_layers_critic 1\
    --n_mlp_layers_features_extractor 1\
    --n_steps_episode 9800\
    --n_validation_env 100\
    --n_workers 1\
    --optimizer adamw\
    --ortools_strategy realistic\
    --residual_gnn\
    --seed 0\
    --target_kl 0.04\
    --total_timesteps 10_000_000\
    --validation_freq 3\
    --vf_coef 2.0

The important parameters here are --n_j and --n_m, to specify the number of jobs and machines of the JSSP instances the model will be trained on. --max_n_j and --max_n_m is used to define the starting feature space of the model. Once trained, the model can be used to solve instances of maximal sizes $maxn_j \times maxn_m$.

Here the problems are deterministic (--duration_type), but you can choose to train the model on stochastic problems (see section Stochastic 10x10).

The rest of the arguments either define the overall GNN model or the training dynamics.

Inference

Once a model is trained, you can use it to solve new problems using:

python3 -m jssp.solve\
    --path "./PATH/TO/EXPERIMENT/"\
    --load_problem "./PATH/TO/INSTANCE.TXT"\
    [--first_machine_id_is_one]  # Optional, if you load taillard problems, you should tell that the index starts at one.

Check jssp/solve.py to see how to load and use the model in your own python scripts.

Examples

Unless specified, all training and validation problems are randomly generated.

Stochastic 10x10

python3 -m jssp.train\
    --batch_size 245\
    --clip_range 0.20\
    --custom_heuristic_names SPT MWKR MOPNR FDD/MWKR\
    --device cuda:0\
    --duration_type stochastic\
    --ent_coef 0.05\
    --exp_name_appendix stochastic-10x10\
    --fe_type dgl\
    --fixed_validation\
    --gae_lambda 0.99\
    --gamma 1.00\
    --generate_duration_bounds 0.05 0.1\
    --graph_has_relu\
    --graph_pooling max\
    --hidden_dim_actor 32\
    --hidden_dim_critic 32\
    --hidden_dim_features_extractor 64\
    --layer_pooling last\
    --lr 1e-4\
    --max_n_j 100\
    --max_n_m 30\
    --mlp_act gelu\
    --n_epochs 3\
    --n_j 10\
    --n_layers_features_extractor 10\
    --n_m 10\
    --n_mlp_layers_actor 1\
    --n_mlp_layers_critic 1\
    --n_mlp_layers_features_extractor 1\
    --n_steps_episode 9800\
    --n_validation_env 100\
    --n_workers 1\
    --optimizer adamw\
    --ortools_strategy realistic averagistic\
    --residual_gnn\
    --seed 0\
    --target_kl 0.04\
    --total_timesteps 10_000_000\
    --validation_freq 3\
    --vf_coef 2.0

Deterministic 20x20

python3 -m jssp.train\
    --batch_size 64\
    --clip_range 0.20\
    --custom_heuristic_names SPT MWKR MOPNR FDD/MWKR\
    --device cuda:0\
    --duration_type deterministic\
    --ent_coef 0.05\
    --exp_name_appendix deterministic-20x20\
    --fe_type dgl\
    --fixed_validation\
    --gae_lambda 0.99\
    --gamma 1.00\
    --graph_has_relu\
    --graph_pooling max\
    --hidden_dim_actor 32\
    --hidden_dim_critic 32\
    --hidden_dim_features_extractor 64\
    --layer_pooling last\
    --lr 1e-5\
    --max_n_j 100\
    --max_n_m 30\
    --mlp_act gelu\
    --n_epochs 1\
    --n_j 20\
    --n_layers_features_extractor 10\
    --n_m 20\
    --n_mlp_layers_actor 1\
    --n_mlp_layers_critic 1\
    --n_mlp_layers_features_extractor 1\
    --n_steps_episode 9800\
    --n_validation_env 100\
    --n_workers 1\
    --optimizer adamw\
    --ortools_strategy realistic\
    --residual_gnn\
    --seed 0\
    --target_kl 0.04\
    --total_timesteps 10_000_000\
    --validation_freq 10\
    --vf_coef 2.0

The same as previous examples but with a lower batch size so that it fits in memory. Note that big --max_n_j and --max_n_m increase the GPU memory requirements. Also do the validation evaluation every 10 epochs to save some time.

Stochastic 100x20 with sub sampling

Note: the subsampling feature exists but has not been tested thoroughly.

python3 -m jssp.train\
    --batch_size 64\
    --clip_range 0.20\
    --custom_heuristic_names SPT MWKR MOPNR FDD/MWKR\
    --device cuda:0\
    --duration_type deterministic\
    --ent_coef 0.05\
    --exp_name_appendix deterministic-subsampling-100x20\
    --fe_type dgl\
    --fixed_validation\
    --gae_lambda 0.99\
    --gamma 1.00\
    --graph_has_relu\
    --graph_pooling max\
    --hidden_dim_actor 32\
    --hidden_dim_critic 32\
    --hidden_dim_features_extractor 64\
    --layer_pooling last\
    --lr 1e-4\
    --max_n_j 100\
    --max_n_m 30\
    --mlp_act gelu\
    --n_epochs 1\
    --n_j 100\
    --n_layers_features_extractor 10\
    --n_m 20\
    --n_mlp_layers_actor 1\
    --n_mlp_layers_critic 1\
    --n_mlp_layers_features_extractor 1\
    --n_steps_episode 9800\
    --n_validation_env 100\
    --n_workers 1\
    --optimizer adamw\
    --ortools_strategy realistic\
    --residual_gnn\
    --sample_n_jobs 10\
    --seed 0\
    --target_kl 0.04\
    --total_timesteps 10_000_000\
    --validate_on_total_data\
    --validation_freq 10\
    --vf_coef 2.0

Make use of --sample_n_jobs and --validate_on_total_data.

Single fixed problem

python3 -m jssp.train\
    --batch_size 245\
    --clip_range 0.20\
    --custom_heuristic_names SPT MWKR MOPNR FDD/MWKR\
    --device cuda:0\
    --duration_type deterministic\
    --ent_coef 0.05\
    --exp_name_appendix single-problem\
    --fe_type dgl\
    --first_machine_id_is_one\
    --fixed_problem\
    --fixed_validation\
    --gae_lambda 0.99\
    --gamma 1.00\
    --graph_has_relu\
    --graph_pooling max\
    --hidden_dim_actor 32\
    --hidden_dim_critic 32\
    --hidden_dim_features_extractor 64\
    --layer_pooling last\
    --load_problem instances/taillard/ta57.txt\
    --lr 1e-4\
    --mlp_act gelu\
    --n_epochs 3\
    --n_j 10\
    --n_layers_features_extractor 10\
    --n_m 10\
    --n_mlp_layers_actor 1\
    --n_mlp_layers_critic 1\
    --n_mlp_layers_features_extractor 1\
    --n_steps_episode 9800\
    --n_validation_env 1\
    --n_workers 1\
    --optimizer adamw\
    --ortools_strategy realistic\
    --residual_gnn\
    --seed 0\
    --target_kl 0.04\
    --total_timesteps 10_000_000\
    --vf_coef 2.0

Load the taillard problem and use it as the single training and validation instance. This is useful if you have a specific problem you want to solve. You can set the durations to be stochastic so that the training instances will be randomly generated around the durations bound.