[KG][Optimization] Soft relation partition (dmlc#1252)

* Several optimizations on DGL-KG:
1. Sorted positive edges for sampling which can reduce random
   memory access during positive sampling
2. Asynchronous node embedding update
3. Balanced Relation Partition that gives balanced number of
   edges in each partition. When there is no cross partition
   relation, relation embedding can be pin into GPU memory
4. tunable neg_sample_size instead of fixed neg_sample_size

* Fix test

* Fix test and eval.py

* Now TransR is OK

* Fix single GPU with mix_cpu_gpu

* Add app tests

* Fix test script

* fix mxnet

* Fix sample

* Add docstrings

* Fix

* Default value for num_workers

* Add soft relation part

* Upd

* Some fix

* upd

* Now work

* Fix TransR

* Fix eval and add some doc string

* triger

* upd

* Add some training scripts for freebase multi-gpu

* upd

* upd

* upd

apps/kg/config/best_config.sh

@@ -104,3 +104,20 @@ DGLBACKEND=pytorch python3 train.py --model ComplEx --dataset Freebase --batch_s
     --neg_sample_size 256 --hidden_dim 400 --gamma 500.0     --lr 0.1 --max_step 50000 \
     --batch_size_eval 128 --test -adv --eval_interval 300000 \
     --neg_sample_size_test 100000 --eval_percent 0.02 --num_proc 64
+
+# Freebase multi-gpu
+# TransE_l2 8 GPU
+DGLBACKEND=pytorch python3 train.py --model TransE_l2 --dataset Freebase --batch_size 1024 \
+    --neg_sample_size 256 --hidden_dim 400 --gamma 10 --lr 0.1 --batch_size_eval 1000 \
+    --valid --test -adv --mix_cpu_gpu --neg_deg_sample_eval --neg_sample_size_test 1000 \
+    --num_proc 8 --gpu 0 1 2 3 4 5 6 7 --num_worker 4 --regularization_coef 1e-9 \
+    --no_eval_filter --max_step 400000 --rel_part --eval_interval 100000 --log_interval 10000 \
+    --no_eval_filter --async_update --neg_deg_sample --force_sync_interval 1000
+
+# TransE_l2 16 GPU
+DGLBACKEND=pytorch python3 train.py --model TransE_l2 --dataset Freebase --batch_size 1024 \
+    --neg_sample_size 256 --hidden_dim 400 --gamma 10 --lr 0.1 --batch_size_eval 1000 \
+    --valid --test -adv --mix_cpu_gpu --neg_deg_sample_eval --neg_sample_size_test 1000 \
+    --num_proc 16 --gpu 0 1 2 3 4 5 6 7 --num_worker 4 --regularization_coef 1e-9 \
+    --no_eval_filter --max_step 200000 --soft_rel_part --eval_interval 100000 --log_interval 10000 \
+    --no_eval_filter --async_update --neg_deg_sample --force_sync_interval 1000

apps/kg/dataloader/sampler.py

@@ -8,6 +8,119 @@
 import pickle
 import time
 
+def SoftRelationPartition(edges, n, threshold=0.05):
+    """This partitions a list of edges to n partitions according to their
+    relation types. For any relation with number of edges larger than the
+    threshold, its edges will be evenly distributed into all partitions.
+    For any relation with number of edges smaller than the threshold, its
+    edges will be put into one single partition.
+
+    Algo:
+    For r in relations:
+        if r.size() > threadold
+            Evenly divide edges of r into n parts and put into each relation.
+        else
+            Find partition with fewest edges, and put edges of r into 
+            this partition.
+
+    Parameters
+    ----------
+    edges : (heads, rels, tails) triple
+        Edge list to partition
+    n : int
+        Number of partitions
+    threshold : float
+        The threshold of whether a relation is LARGE or SMALL
+        Default: 5%
+
+    Returns
+    -------
+    List of np.array
+        Edges of each partition
+    List of np.array
+        Edge types of each partition
+    bool
+        Whether there exists some relations belongs to multiple partitions
+    """
+    heads, rels, tails = edges
+    print('relation partition {} edges into {} parts'.format(len(heads), n))
+    uniq, cnts = np.unique(rels, return_counts=True)
+    idx = np.flip(np.argsort(cnts))
+    cnts = cnts[idx]
+    uniq = uniq[idx]
+    assert cnts[0] > cnts[-1]
+    edge_cnts = np.zeros(shape=(n,), dtype=np.int64)
+    rel_cnts = np.zeros(shape=(n,), dtype=np.int64)
+    rel_dict = {}
+    rel_parts = []
+    cross_rel_part = []
+    for _ in range(n):
+        rel_parts.append([])
+
+    large_threshold = int(len(rels) * threshold)
+    capacity_per_partition = int(len(rels) / n)
+    # ensure any relation larger than the partition capacity will be split
+    large_threshold = capacity_per_partition if capacity_per_partition < large_threshold \
+                      else large_threshold
+    num_cross_part = 0
+    for i in range(len(cnts)):
+        cnt = cnts[i]
+        r = uniq[i]
+        r_parts = []
+        if cnt > large_threshold:
+            avg_part_cnt = (cnt // n) + 1
+            num_cross_part += 1
+            for j in range(n):
+                part_cnt = avg_part_cnt if cnt > avg_part_cnt else cnt
+                r_parts.append([j, part_cnt])
+                rel_parts[j].append(r)
+                edge_cnts[j] += part_cnt
+                rel_cnts[j] += 1
+                cnt -= part_cnt
+            cross_rel_part.append(r)
+        else:
+            idx = np.argmin(edge_cnts)
+            r_parts.append([idx, cnt])
+            rel_parts[idx].append(r)
+            edge_cnts[idx] += cnt
+            rel_cnts[idx] += 1
+        rel_dict[r] = r_parts
+
+    for i, edge_cnt in enumerate(edge_cnts):
+        print('part {} has {} edges and {} relations'.format(i, edge_cnt, rel_cnts[i]))
+    print('{}/{} duplicated relation across partitions'.format(num_cross_part, len(cnts)))
+
+    parts = []
+    for i in range(n):
+        parts.append([])
+        rel_parts[i] = np.array(rel_parts[i])
+
+    for i, r in enumerate(rels):
+        r_part = rel_dict[r][0]
+        part_idx = r_part[0]
+        cnt = r_part[1]
+        parts[part_idx].append(i)
+        cnt -= 1
+        if cnt == 0:
+            rel_dict[r].pop(0)
+        else:
+            rel_dict[r][0][1] = cnt
+
+    for i, part in enumerate(parts):
+        parts[i] = np.array(part, dtype=np.int64)
+    shuffle_idx = np.concatenate(parts)
+    heads[:] = heads[shuffle_idx]
+    rels[:] = rels[shuffle_idx]
+    tails[:] = tails[shuffle_idx]
+
+    off = 0
+    for i, part in enumerate(parts):
+        parts[i] = np.arange(off, off + len(part))
+        off += len(part)
+    cross_rel_part = np.array(cross_rel_part)
+
+    return parts, rel_parts, num_cross_part > 0, cross_rel_part
+
 def BalancedRelationPartition(edges, n):
     """This partitions a list of edges based on relations to make sure
     each partition has roughly the same number of edges and relations.
@@ -184,7 +297,10 @@ def __init__(self, dataset, args, ranks=64):
         num_train = len(triples[0])
         print('|Train|:', num_train)
 
-        if ranks > 1 and args.rel_part:
+        if ranks > 1 and args.soft_rel_part:
+            self.edge_parts, self.rel_parts, self.cross_part, self.cross_rels = \
+            SoftRelationPartition(triples, ranks)
+        elif ranks > 1 and args.rel_part:
             self.edge_parts, self.rel_parts, self.cross_part = \
                 BalancedRelationPartition(triples, ranks)
         elif ranks > 1:

apps/kg/eval.py

@@ -101,6 +101,7 @@ def main(args):
     args.valid = False
     args.test = True
     args.strict_rel_part = False
+    args.soft_rel_part = False
     args.async_update = False
     args.batch_size_eval = args.batch_size
 

apps/kg/models/general_models.py

@@ -82,7 +82,8 @@ def __init__(self, args, model_name, n_entities, n_relations, hidden_dim, gamma,
         self.rel_dim = rel_dim
         self.entity_dim = entity_dim
         self.strict_rel_part = args.strict_rel_part
-        if not self.strict_rel_part:
+        self.soft_rel_part = args.soft_rel_part
+        if not self.strict_rel_part and not self.soft_rel_part:
             self.relation_emb = ExternalEmbedding(args, n_relations, rel_dim,
                                                   F.cpu() if args.mix_cpu_gpu else device)
         else:
@@ -120,7 +121,7 @@ def share_memory(self):
         """Use torch.tensor.share_memory_() to allow cross process embeddings access.
         """
         self.entity_emb.share_memory()
-        if self.strict_rel_part:
+        if self.strict_rel_part or self.soft_rel_part:
             self.global_relation_emb.share_memory()
         else:
             self.relation_emb.share_memory()
@@ -139,7 +140,7 @@ def save_emb(self, path, dataset):
             Dataset name as prefix to the saved embeddings.
         """
         self.entity_emb.save(path, dataset+'_'+self.model_name+'_entity')
-        if self.strict_rel_part:
+        if self.strict_rel_part or self.soft_rel_part:
             self.global_relation_emb.save(path, dataset+'_'+self.model_name+'_relation')
         else:
             self.relation_emb.save(path, dataset+'_'+self.model_name+'_relation')   
@@ -165,8 +166,10 @@ def reset_parameters(self):
         """
         self.entity_emb.init(self.emb_init)
         self.score_func.reset_parameters()
-        if not self.strict_rel_part:
+        if (not self.strict_rel_part) and (not self.soft_rel_part):
             self.relation_emb.init(self.emb_init)
+        else:
+            self.global_relation_emb.init(self.emb_init)
 
     def predict_score(self, g):
         """Predict the positive score.
@@ -415,6 +418,11 @@ def prepare_relation(self, device=None):
             self.score_func.prepare_local_emb(local_projection_emb)
             self.score_func.reset_parameters()
 
+    def prepare_cross_rels(self, cross_rels):
+        self.relation_emb.setup_cross_rels(cross_rels, self.global_relation_emb)
+        if self.model_name == 'TransR':
+            self.score_func.prepare_cross_rels(cross_rels)
+
     def writeback_relation(self, rank=0, rel_parts=None):
         """ Writeback relation embeddings in a specific process to global relation embedding.
         Used in multi-process multi-gpu training model.
@@ -425,6 +433,8 @@ def writeback_relation(self, rank=0, rel_parts=None):
             List of tensor stroing edge types of each partition.
         """
         idx = rel_parts[rank]
+        if self.soft_rel_part:
+            idx = self.relation_emb.get_noncross_idx(idx)
         self.global_relation_emb.emb[idx] = F.copy_to(self.relation_emb.emb, F.cpu())[idx]
         if self.model_name == 'TransR':
             self.score_func.writeback_local_emb(idx)

apps/kg/models/pytorch/score_fun.py

@@ -157,6 +157,9 @@ def prepare_local_emb(self, projection_emb):
         self.global_projection_emb = self.projection_emb
         self.projection_emb = projection_emb
 
+    def prepare_cross_rels(self, cross_rels):
+        self.projection_emb.setup_cross_rels(cross_rels, self.global_projection_emb)
+
     def writeback_local_emb(self, idx):
         self.global_projection_emb.emb[idx] = self.projection_emb.emb.cpu()[idx]
 

apps/kg/models/pytorch/tensor_models.py

@@ -117,11 +117,13 @@ class ExternalEmbedding:
     def __init__(self, args, num, dim, device):
         self.gpu = args.gpu
         self.args = args
+        self.num = num
         self.trace = []
 
         self.emb = th.empty(num, dim, dtype=th.float32, device=device)
         self.state_sum = self.emb.new().resize_(self.emb.size(0)).zero_()
         self.state_step = 0
+        self.has_cross_rel = False
         # queue used by asynchronous update
         self.async_q = None
         # asynchronous update process
@@ -138,6 +140,19 @@ def init(self, emb_init):
         INIT.uniform_(self.emb, -emb_init, emb_init)
         INIT.zeros_(self.state_sum)
 
+    def setup_cross_rels(self, cross_rels, global_emb):
+        cpu_bitmap = th.zeros((self.num,), dtype=th.bool)
+        for i, rel in enumerate(cross_rels):
+            cpu_bitmap[rel] = 1
+        self.cpu_bitmap = cpu_bitmap
+        self.has_cross_rel = True
+        self.global_emb = global_emb
+
+    def get_noncross_idx(self, idx):
+        cpu_mask = self.cpu_bitmap[idx]
+        gpu_mask = ~cpu_mask
+        return idx[gpu_mask]
+
     def share_memory(self):
         """Use torch.tensor.share_memory_() to allow cross process tensor access
         """
@@ -158,6 +173,14 @@ def __call__(self, idx, gpu_id=-1, trace=True):
             If False, do not trace the computation.
             Default: True
         """
+        if self.has_cross_rel:
+            cpu_idx = idx.cpu()
+            cpu_mask = self.cpu_bitmap[cpu_idx]
+            cpu_idx = cpu_idx[cpu_mask]
+            cpu_idx = th.unique(cpu_idx)
+            if cpu_idx.shape[0] != 0:
+                cpu_emb = self.global_emb.emb[cpu_idx]
+                self.emb[cpu_idx] = cpu_emb.cuda(gpu_id)
         s = self.emb[idx]
         if gpu_id >= 0:
             s = s.cuda(gpu_id)
@@ -202,6 +225,22 @@ def update(self, gpu_id=-1):
                         grad_indices = grad_indices.to(device)
                     if device != grad_sum.device:
                         grad_sum = grad_sum.to(device)
+
+                    if self.has_cross_rel:
+                        cpu_mask = self.cpu_bitmap[grad_indices]
+                        cpu_idx = grad_indices[cpu_mask]
+                        if cpu_idx.shape[0] > 0:
+                            cpu_grad = grad_values[cpu_mask]
+                            cpu_sum = grad_sum[cpu_mask].cpu()
+                            cpu_idx = cpu_idx.cpu()
+                            self.global_emb.state_sum.index_add_(0, cpu_idx, cpu_sum)
+                            std = self.global_emb.state_sum[cpu_idx]
+                            if gpu_id >= 0:
+                                std = std.cuda(gpu_id)
+                            std_values = std.sqrt_().add_(1e-10).unsqueeze(1)
+                            tmp = (-clr * cpu_grad / std_values)
+                            tmp = tmp.cpu()
+                            self.global_emb.emb.index_add_(0, cpu_idx, tmp)
                     self.state_sum.index_add_(0, grad_indices, grad_sum)
                     std = self.state_sum[grad_indices]  # _sparse_mask
                     if gpu_id >= 0:

apps/kg/train.py

@@ -112,6 +112,8 @@ def __init__(self):
                           help='number of process used')
         self.add_argument('--rel_part', action='store_true',
                           help='enable relation partitioning')
+        self.add_argument('--soft_rel_part', action='store_true',
+                          help='enable soft relation partition')
         self.add_argument('--nomp_thread_per_process', type=int, default=-1,
                           help='num of omp threads used per process in multi-process training')
         self.add_argument('--async_update', action='store_true',
@@ -170,7 +172,9 @@ def run(args, logger):
 
     num_workers = args.num_worker
     train_data = TrainDataset(dataset, args, ranks=args.num_proc)
+    # if there is no cross partition relaiton, we fall back to strict_rel_part
     args.strict_rel_part = args.mix_cpu_gpu and (train_data.cross_part == False)
+    args.soft_rel_part = args.mix_cpu_gpu and args.soft_rel_part and train_data.cross_part
 
     # Automatically set number of OMP threads for each process if it is not provided
     # The value for GPU is evaluated in AWS p3.16xlarge
@@ -322,7 +326,8 @@ def run(args, logger):
 
     # train
     start = time.time()
-    rel_parts = train_data.rel_parts if args.strict_rel_part else None
+    rel_parts = train_data.rel_parts if args.strict_rel_part or args.soft_rel_part else None
+    cross_rels = train_data.cross_rels if args.soft_rel_part else None
     if args.num_proc > 1:
         procs = []
         barrier = mp.Barrier(args.num_proc)
@@ -334,6 +339,7 @@ def run(args, logger):
                                                      valid_sampler,
                                                      i,
                                                      rel_parts,
+                                                     cross_rels,
                                                      barrier))
             procs.append(proc)
             proc.start()