Merge branch 'v3'

Author: wouterkool

README.md

@@ -21,7 +21,7 @@ For more details, please see our paper [Attention, Learn to Solve Routing Proble
 * Python>=3.6
 * NumPy
 * SciPy
-* [PyTorch](http://pytorch.org/)=0.4
+* [PyTorch](http://pytorch.org/)>=1.1
 * tqdm
 * [tensorboard_logger](https://github.com/TeamHG-Memex/tensorboard_logger)
 * Matplotlib (optional, only for plotting)

nets/attention_model.py

@@ -1,6 +1,5 @@
 import torch
 from torch import nn
-import torch.nn.functional as F
 from torch.utils.checkpoint import checkpoint
 import math
 from typing import NamedTuple
@@ -131,7 +130,7 @@ def forward(self, input, return_pi=False):
         :return:
         """
 
-        if self.checkpoint_encoder:
+        if self.checkpoint_encoder and self.training:  # Only checkpoint if we need gradients
             embeddings, _ = checkpoint(self.embedder, self._init_embed(input))
         else:
             embeddings, _ = self.embedder(self._init_embed(input))
@@ -360,7 +359,7 @@ def _get_log_p(self, fixed, state, normalize=True):
         log_p, glimpse = self._one_to_many_logits(query, glimpse_K, glimpse_V, logit_K, mask)
 
         if normalize:
-            log_p = F.log_softmax(log_p / self.temp, dim=-1)
+            log_p = torch.log_softmax(log_p / self.temp, dim=-1)
 
         assert not torch.isnan(log_p).any()
 
@@ -465,7 +464,7 @@ def _one_to_many_logits(self, query, glimpse_K, glimpse_V, logit_K, mask):
             compatibility[mask[None, :, :, None, :].expand_as(compatibility)] = -math.inf
 
         # Batch matrix multiplication to compute heads (n_heads, batch_size, num_steps, val_size)
-        heads = torch.matmul(F.softmax(compatibility, dim=-1), glimpse_V)
+        heads = torch.matmul(torch.softmax(compatibility, dim=-1), glimpse_V)
 
         # Project to get glimpse/updated context node embedding (batch_size, num_steps, embedding_dim)
         glimpse = self.project_out(
@@ -480,7 +479,7 @@ def _one_to_many_logits(self, query, glimpse_K, glimpse_V, logit_K, mask):
 
         # From the logits compute the probabilities by clipping, masking and softmax
         if self.tanh_clipping > 0:
-            logits = F.tanh(logits) * self.tanh_clipping
+            logits = torch.tanh(logits) * self.tanh_clipping
         if self.mask_logits:
             logits[mask] = -math.inf
 

nets/graph_encoder.py

@@ -1,5 +1,4 @@
 import torch
-import torch.nn.functional as F
 import numpy as np
 from torch import nn
 import math
@@ -95,7 +94,7 @@ def forward(self, q, h=None, mask=None):
             mask = mask.view(1, batch_size, n_query, graph_size).expand_as(compatibility)
             compatibility[mask] = -np.inf
 
-        attn = F.softmax(compatibility, dim=-1)
+        attn = torch.softmax(compatibility, dim=-1)
 
         # If there are nodes with no neighbours then softmax returns nan so we fix them to 0
         if mask is not None:

nets/pointer_network.py

@@ -1,7 +1,6 @@
 import torch
 import torch.nn as nn
 from torch.autograd import Variable
-import torch.nn.functional as F
 import math
 import numpy as np
 
@@ -105,7 +104,7 @@ def recurrence(self, x, h_in, prev_mask, prev_idxs, step, context):
         logits, h_out = self.calc_logits(x, h_in, logit_mask, context, self.mask_glimpses, self.mask_logits)
 
         # Calculate log_softmax for better numerical stability
-        log_p = F.log_softmax(logits, dim=1)
+        log_p = torch.log_softmax(logits, dim=1)
         probs = log_p.exp()
 
         if not self.mask_logits:

problems/op/state_op.py

@@ -142,17 +142,15 @@ def get_mask(self):
         :return:
         """
 
+        exceeds_length = (
+            self.lengths[:, :, None] + (self.coords[self.ids, :, :] - self.cur_coord[:, :, None, :]).norm(p=2, dim=-1)
+            > self.max_length[self.ids, :]
+        )
         # Note: this always allows going to the depot, but that should always be suboptimal so be ok
         # Cannot visit if already visited or if length that would be upon arrival is too large to return to depot
         # If the depot has already been visited then we cannot visit anymore
-        visited_ = self.visited
-        mask = (
-            visited_ | visited_[:, :, 0:1] |
-            (
-                self.lengths[:, :, None] + (self.coords[self.ids, :, :] - self.cur_coord[:, :, None, :]).norm(p=2, dim=-1)
-                > self.max_length[self.ids, :]
-            )
-        )
+        visited_ = self.visited.to(exceeds_length.dtype)
+        mask = visited_ | visited_[:, :, 0:1] | exceeds_length
         # Depot can always be visited
         # (so we do not hardcode knowledge that this is strictly suboptimal if other options are available)
         mask[:, :, 0] = 0

problems/pctsp/state_pctsp.py

@@ -162,7 +162,7 @@ def get_mask(self):
         # Cannot visit depot if not yet collected 1 total prize and there are unvisited nodes
         mask[:, :, 0] = (self.cur_total_prize < 1.) & (visited_[:, :, 1:].int().sum(-1) < visited_[:, :, 1:].size(-1))
 
-        return mask
+        return mask > 0  # Hacky way to return bool or uint8 depending on pytorch version
 
     def construct_solutions(self, actions):
         return actions

problems/tsp/state_tsp.py

@@ -106,7 +106,7 @@ def get_current_node(self):
         return self.prev_a
 
     def get_mask(self):
-        return self.visited
+        return self.visited > 0  # Hacky way to return bool or uint8 depending on pytorch version
 
     def get_nn(self, k=None):
         # Insert step dimension

problems/vrp/state_cvrp.py

@@ -143,12 +143,10 @@ def get_mask(self):
         else:
             visited_loc = mask_long2bool(self.visited_, n=self.demand.size(-1))
 
+        # For demand steps_dim is inserted by indexing with id, for used_capacity insert node dim for broadcasting
+        exceeds_cap = (self.demand[self.ids, :] + self.used_capacity[:, :, None] > self.VEHICLE_CAPACITY)
         # Nodes that cannot be visited are already visited or too much demand to be served now
-        mask_loc = (
-            visited_loc |
-            # For demand steps_dim is inserted by indexing with id, for used_capacity insert node dim for broadcasting
-            (self.demand[self.ids, :] + self.used_capacity[:, :, None] > self.VEHICLE_CAPACITY)
-        )
+        mask_loc = visited_loc.to(exceeds_cap.dtype) | exceeds_cap
 
         # Cannot visit the depot if just visited and still unserved nodes
         mask_depot = (self.prev_a == 0) & ((mask_loc == 0).int().sum(-1) > 0)

train.py

@@ -68,7 +68,6 @@ def train_epoch(model, optimizer, baseline, lr_scheduler, epoch, val_dataset, pr
     print("Start train epoch {}, lr={} for run {}".format(epoch, optimizer.param_groups[0]['lr'], opts.run_name))
     step = epoch * (opts.epoch_size // opts.batch_size)
     start_time = time.time()
-    lr_scheduler.step(epoch)
 
     if not opts.no_tensorboard:
         tb_logger.log_value('learnrate_pg0', optimizer.param_groups[0]['lr'], step)
@@ -121,6 +120,9 @@ def train_epoch(model, optimizer, baseline, lr_scheduler, epoch, val_dataset, pr
 
     baseline.epoch_callback(model, epoch)
 
+    # lr_scheduler should be called at end of epoch
+    lr_scheduler.step()
+
 
 def train_batch(
         model,