update code and readme

EAS/run_search.py

@@ -34,7 +34,7 @@ def get_config():
 
     parser.add_argument('-problem', default="TSP", type=str, choices=['TSP', 'CVRP'])
     parser.add_argument('-method', default="eas-emb", type=str, choices=['eas-emb', 'eas-lay', 'eas-tab'], help="EAS method")
-    parser.add_argument('-model_path', default="../pretrained/pomo_pretrained/checkpoint-30500.pt", type=str, help="Path of the trained model weights")
+    parser.add_argument('-model_path', default="../pretrained/POMO-TSP/checkpoint-3000-tsp100-instance-norm.pt", type=str, help="Path of the trained model weights")
     parser.add_argument('-instances_path', default="../data/TSP/Size_Distribution/tsp200_rotation.pkl", type=str, help="Path of the instances")
     parser.add_argument('-sol_path', default="../data/TSP/Size_Distribution/concorde/tsp200_rotationoffset0n1000-concorde.pkl", type=str, help="Path of the optimal sol")
     parser.add_argument('-num_instances', default=1000, type=int, help="Maximum number of instances that should be solved")
@@ -46,7 +46,7 @@ def get_config():
     parser.add_argument('-batch_size', default=150, type=int)  # Set to 1 for single instance search
     parser.add_argument('-p_runs', default=1, type=int)  # If batch_size is 1, set this to > 1 to do multiple runs for the instance in parallel
     parser.add_argument('-output_path', default="EAS_results", type=str)
-    parser.add_argument('-norm', default="batch_no_track", choices=['instance', 'batch', 'batch_no_track', 'none'], type=str)
+    parser.add_argument('-norm', default="instance", choices=['instance', 'batch', 'batch_no_track', 'none'], type=str)
     parser.add_argument('-gpu_id', default=0, type=int)
     parser.add_argument('-seed', default=2023, type=int, help="random seed")
 

POMO/CVRP/CVRPTrainer_meta.py

@@ -61,6 +61,7 @@ def __init__(self,
         self.meta_model = Model(**self.model_params)
         self.meta_optimizer = Optimizer(self.meta_model.parameters(), **self.optimizer_params['optimizer'])
         self.alpha = self.meta_params['alpha']  # for reptile
+        self.early_stop = True if self.meta_params['meta_method'] == 'maml_fomaml' else False
         self.task_set = generate_task_set(self.meta_params)
         self.val_data, self.val_opt = {}, {}  # for lkh3_offline
         if self.meta_params["data_type"] == "size_distribution":
@@ -180,6 +181,7 @@ def _train_one_epoch(self, epoch):
         self.meta_optimizer.zero_grad()
         score_AM, loss_AM = AverageMeter(), AverageMeter()
         meta_batch_size = self.meta_params['meta_batch_size']
+        self.meta_params['meta_method'] = 'fomaml' if self.early_stop and epoch > self.meta_params['early_stop_epoch'] else "maml"
 
         # Adaptive task scheduler - Not implemented for "size" and "distribution"
         if self.meta_params['curriculum']:
@@ -212,7 +214,7 @@ def _train_one_epoch(self, epoch):
             if self.meta_params['meta_method'] in ['fomaml', 'reptile']:
                 task_model = copy.deepcopy(self.meta_model)
                 optimizer = Optimizer(task_model.parameters(), **self.optimizer_params['optimizer'])
-                # optimizer.load_state_dict(self.meta_optimizer.state_dict())
+                # optimizer.load_state_dict(self.meta_optimizer.state_dict())  # may cause unstable meta-training for fomaml
             elif self.meta_params['meta_method'] == 'maml':
                 if self.model_params['meta_update_encoder']:
                     fast_weight = OrderedDict(self.meta_model.named_parameters())

POMO/CVRP/train.py

@@ -72,9 +72,10 @@
 meta_params = {
     'enable': True,  # whether use meta-learning or not
     'curriculum': True,  # adaptive sample task
-    'meta_method': 'maml',  # choose from ['maml', 'fomaml', 'reptile']
+    'meta_method': 'maml',  # choose from ['maml', 'fomaml', 'maml_fomaml', 'reptile']
     'data_type': 'size_distribution',  # choose from ["size", "distribution", "size_distribution"]
     'epochs': 250000,  # the number of meta-model updates: (250*100000) / (1*5*64)
+    'early_stop_epoch': 50000,  # switch from maml to fomaml
     'B': 1,  # the number of tasks in a mini-batch
     'k': 1,  # gradient decent steps in the inner-loop optimization of meta-learning method
     'L': 0,  # bootstrap steps

POMO/TSP/TSPTrainer_Meta.py

@@ -62,6 +62,7 @@ def __init__(self,
         self.meta_model = Model(**self.model_params)
         self.meta_optimizer = Optimizer(self.meta_model.parameters(), **self.optimizer_params['optimizer'])
         self.alpha = self.meta_params['alpha']  # for reptile
+        self.early_stop = True if self.meta_params['meta_method'] == 'maml_fomaml' else False
         self.task_set = generate_task_set(self.meta_params)
         self.val_data, self.val_opt = {}, {}  # for lkh3_offline
         if self.meta_params["data_type"] == "size_distribution":
@@ -181,6 +182,7 @@ def _train_one_epoch(self, epoch):
         self.meta_optimizer.zero_grad()
         score_AM, loss_AM = AverageMeter(), AverageMeter()
         meta_batch_size = self.meta_params['meta_batch_size']
+        self.meta_params['meta_method'] = 'fomaml' if self.early_stop and epoch > self.meta_params['early_stop_epoch'] else "maml"
 
         # Adaptive task scheduler:
         if self.meta_params['curriculum']:
@@ -213,7 +215,7 @@ def _train_one_epoch(self, epoch):
             if self.meta_params['meta_method'] in ['fomaml', 'reptile']:
                 task_model = copy.deepcopy(self.meta_model)
                 optimizer = Optimizer(task_model.parameters(), **self.optimizer_params['optimizer'])
-                # optimizer.load_state_dict(self.meta_optimizer.state_dict())
+                # optimizer.load_state_dict(self.meta_optimizer.state_dict())  # may cause unstable meta-training for fomaml
             elif self.meta_params['meta_method'] == 'maml':
                 if self.model_params['meta_update_encoder']:
                     fast_weight = OrderedDict(self.meta_model.named_parameters())

POMO/TSP/train.py

@@ -72,9 +72,10 @@
 meta_params = {
     'enable': True,  # whether use meta-learning or not
     'curriculum': True,  # adaptive sample task
-    'meta_method': 'maml',  # choose from ['maml', 'fomaml', 'reptile']
+    'meta_method': 'maml',  # choose from ['maml', 'fomaml', 'maml_fomaml', 'reptile']
     'data_type': 'size_distribution',  # choose from ["size", "distribution", "size_distribution"]
     'epochs': 250000,  # the number of meta-model updates: (250*100000) / (1*5*64)
+    'early_stop_epoch': 50000,  # switch from maml to fomaml
     'B': 1,  # the number of tasks in a mini-batch
     'k': 1,  # gradient decent steps in the inner-loop optimization of meta-learning method
     'L': 0,  # bootstrap steps

POMO/call_tspgen.R

@@ -0,0 +1,89 @@
+# Rscript call_tspgen.R operator point_lower point_upper ins_num seed
+
+library("netgen")
+source("tspgen-master/R/utilities.R")
+source("tspgen-master/R/mutator.explosion.R")
+source("tspgen-master/R/mutator.implosion.R")
+source("tspgen-master/R/mutator.cluster.R")
+source("tspgen-master/R/mutator.compression.R")
+source("tspgen-master/R/mutator.expansion.R")
+source("tspgen-master/R/mutator.grid.R")
+source("tspgen-master/R/mutator.linearprojection.R")
+source("tspgen-master/R/mutator.rotation.R")
+
+library(ggplot2)
+library(gridExtra)
+
+args <- commandArgs(TRUE)
+operator <- toString(args[1])
+points.num <- as.integer(args[2])
+ins.num <- as.integer(args[4])
+seed <- as.integer(args[5])
+choice <- toString(args[6])
+path <- toString(args[7])
+
+#check_internal = ins.num / 5
+
+check_internal = 1
+
+# points.num <- sample(points.lower:points.upper, 1, replace=TRUE)
+
+set.seed(seed)
+
+for (i in 1:ins.num)
+{
+
+    x = generateRandomNetwork(n.points = points.num, lower = 0, upper = 1)
+    rue = x
+    if (operator == "explosion")
+    {
+        x$coordinates = doExplosionMutation(x$coordinates, min.eps=0.3, max.eps=0.3)
+    }
+    if (operator == "implosion")
+    {
+        x$coordinates = doImplosionMutation(x$coordinates, min.eps=0.3, max.eps=0.3)
+    }
+    if (operator == "cluster")
+    {
+        x$coordinates = doClusterMutation(x$coordinates, pm=0.4)
+    }
+    if (operator == "compression")
+    {
+        x$coordinates = doCompressionMutation(x$coordinates, min.eps=0.3, max.eps=0.3)
+    }
+    if (operator == "expansion")
+    {
+        x$coordinates = doExpansionMutation(x$coordinates, min.eps=0.3, max.eps=0.3)
+    }
+    if (operator == "grid")
+    {
+        x$coordinates = doGridMutation(x$coordinates, box.min=0.3, box.max=0.3, p.rot=0, p.jitter=0, jitter.sd=0.05)
+    }
+    if (operator == "linearprojection")
+    {
+        x$coordinates = doLinearProjectionMutation(x$coordinates, pm=0.4, p.jitter=0, jitter.sd=0.05)
+    }
+    if (operator == "rotation")
+    {
+        x$coordinates = doRotationMutation(x$coordinates, pm=0.4)
+    }
+
+    x = rescaleNetwork(x, method = "global2")
+    # x$coordinates = x$coordinates * 1000000
+    # x$coordinates = round(x$coordinates, 0)
+    x$coords = relocateDuplicates(x$coords)
+    x$lower = 0
+    x$upper = 1
+    name = sprintf("data/TSP_test/%s/%d.tsp", operator, i)
+    # if (i %% check_internal == 0)
+    # {
+    #     jpeg(file = paste("data/check/",operator,toString(i), ".jpg", sep = ""), width=1200, height=600)
+    #     grid.arrange(autoplot(x), autoplot(rue), nrow = 1L)
+    #     dev.off()
+    # }
+
+    # print(x$coordinates)
+    cat(x$coordinates, "\n", file=path, append=TRUE)
+
+    # exportToTSPlibFormat(x, name, use.extended.format=FALSE)
+}

README.md

@@ -12,37 +12,84 @@ The PyTorch Implementation of *ICML 2023 Poster -- "Towards Omni-generalizable N
 
 This paper studies a challenging yet realistic setting, which considers generalization across both size and distribution (a.k.a. omni-generalization) of neural methods in VRPs. Technically, a general meta-learning framework is developed to tackle it.
 
-### TODO
-
-- [ ] Finish Dependencies & How to Run & Take-home Messages.
-- [ ] Camera-ready.
-- [ ] Slide and Poster.
-- [ ] Release Review.
-
-### Dependencies
+### Env Setup
+
+```shell
+conda create -n omni_vrp python=3.8
+conda activate omni_vrp
+# install pytorch (we use V.1.12.1), see here: https://pytorch.org/get-started/previous-versions
+conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch
+conda install pytz tqdm scikit-learn matplotlib
+# for L2D env setup, see here: https://github.com/mit-wu-lab/learning-to-delegate
+```
 
+### How to Run
 
+We take the meta-training on POMO as an example. See [here](https://github.com/RoyalSkye/Omni-VRP/tree/main/L2D) for L2D.
+
+```shell
+# 1. training
+# a. second-order
+nohup python -u train.py 2>&1 &
+# b. early-stopping
+meta_params['meta_method'] = "maml_fomaml"
+# c. first-order
+meta_params['meta_method'] = "fomaml"
+# 2. testing
+# a. zero-shot on datasets (.pkl), including 1K test instances and 1K fine-tuning instances.
+nohup python -u test.py 2>&1 &
+# b. few-shot on datasets (.pkl)
+fine_tune_params['enable'] = True
+# c. zero-shot on benchmark instances (.tsp or .vrp)
+tester_params['test_set_path'] = "../../data/TSP/tsplib"
+# 3. Traditional VRP solvers ["concorde","lkh"] for TSP; ["hgs", "lkh"] for CVRP
+nohup python -u TSP_baseline.py --method "lkh" --cpus 32 --no_cuda --disable_cache 2>&1 &
+# 4. EAS
+# Note: change -instances_path to the folder path if conducting EAS on benchmark instances
+# Note: Pls carefully check parameters, e.g., -norm, -round_distances, etc.
+nohup python -u run_search.py 2>&1 &
+```
 
-### How to Run
+#### Options
+
+```shell
+# Modify the default value in train.py
+# 1. Bootstrapped Meta-Learning - ICLR 2022
+meta_params['L'] = X (X > 0)
+# 2. ANIL (Almost No Inner Loop) - ICLR 2020
+model_params["meta_update_encoder"] = False
+# 3. Meta-training on the pretrained model
+# Note: The type of normalization layers should match (search for model_params["norm"])
+# Supported normalization layers: ["no", "batch", "batch_no_track", "instance", "rezero"]
+trainer_params['pretrain_load'] = True
+# 4. Resume meta-training 
+trainer_params['model_load'] = True
+# 5. No task scheduler
+meta_params['curriculum'] = False
+# 6. Baselines
+# Note: For AMDKD-POMO, refer to: https://github.com/jieyibi/AMDKD
+meta_params['enable'] = False  # POMO
+meta_params['reptile'] = 'reptile'  # Meta-POMO
+```
 
+### Discussions
 
+In summary: 1) Normalization layers matter in AM-based models (see [here](https://github.com/RoyalSkye/Omni-VRP/blob/main/POMO/TSP/TSPTrainer_Meta.py#L58)); 2) The training efficiency and scalability heavily depend on the base model and meta-learning algorithm; 3) It may be better to conduct meta-training on the pretrained model. For further discussions, refer to Appendix E.
 
-### Take-home Messages
+### Acknowledgments
 
+* We would like to thank the anonymous reviewers and (S)ACs of ICML 2023 for their constructive comments and dedicated service to the community. The reviews and meta-review are available [here](https://github.com/RoyalSkye/Omni-VRP/blob/main/Reviews_ICML23.md).
 
+* We also would like to thank the following repositories, which are baselines of our code:
 
-### Reviews
+  * https://github.com/wouterkool/attention-learn-to-route
 
-We would like to thank the anonymous reviewers and (S)ACs of ICML 2023 for their constructive comments and recommendation. We will share the reviews later.
+  * https://github.com/yd-kwon/POMO
 
-### Acknowledgments
+  * https://github.com/mit-wu-lab/learning-to-delegate
 
-Thank the following repositories, which are baselines of our code:
+  * "On the generalization of neural combinatorial optimization heuristics"
 
-* https://github.com/wouterkool/attention-learn-to-route
-* https://github.com/yd-kwon/POMO
-* https://github.com/mit-wu-lab/learning-to-delegate
-* "On the generalization of neural combinatorial optimization heuristics"
 
 ### Citation