bug-fix

Author: js1010

cpp/src/cuw2v/cuw2v.cu

@@ -265,7 +265,7 @@ std::pair<float, float> CuW2V::FeedData(const int* cols, const int* indptr,
   // accumulate loss nume / deno
   std::vector<float> loss_nume(block_cnt_), loss_deno(block_cnt_);
   thrust::copy(dev_loss_nume.begin(), dev_loss_nume.end(), loss_nume.begin());
-  thrust::copy(dev_loss_deno.begin(), dev_loss_deno.end(), loss_nume.begin());
+  thrust::copy(dev_loss_deno.begin(), dev_loss_deno.end(), loss_deno.begin());
   CHECK_CUDA(cudaDeviceSynchronize());
   float loss_nume_sum = std::accumulate(loss_nume.begin(), loss_nume.end(), 0.0f); 
   float loss_deno_sum = std::accumulate(loss_deno.begin(), loss_deno.end(), 0.0f); 

cusim/culda/pyculda.py

@@ -20,12 +20,17 @@
 from cusim.config_pb2 import CuLDAConfigProto
 
 EPS = 1e-10
+WARP_SIZE = 32
 
 class CuLDA:
   def __init__(self, opt=None):
     self.opt = aux.get_opt_as_proto(opt or {}, CuLDAConfigProto)
     self.logger = aux.get_logger("culda", level=self.opt.py_log_level)
 
+    assert self.opt.block_dim <= WARP_SIZE ** 2 and \
+      self.opt.block_dim % WARP_SIZE == 0, \
+      f"invalid block dim ({self.opt.block_dim}, warp size: {WARP_SIZE})"
+
     tmp = tempfile.NamedTemporaryFile(mode='w', delete=False)
     opt_content = json.dumps(aux.proto_to_dict(self.opt), indent=2)
     tmp.write(opt_content)

cusim/cuw2v/pycuw2v.py

@@ -19,12 +19,17 @@
 from cusim.config_pb2 import CuW2VConfigProto
 
 EPS = 1e-10
+WARP_SIZE = 32
 
 class CuW2V:
   def __init__(self, opt=None):
     self.opt = aux.get_opt_as_proto(opt or {}, CuW2VConfigProto)
     self.logger = aux.get_logger("culda", level=self.opt.py_log_level)
 
+    assert self.opt.block_dim <= WARP_SIZE ** 2 and \
+      self.opt.block_dim % WARP_SIZE == 0, \
+      f"invalid block dim ({self.opt.block_dim}, warp size: {WARP_SIZE})"
+
     tmp = tempfile.NamedTemporaryFile(mode='w', delete=False)
     opt_content = json.dumps(aux.proto_to_dict(self.opt), indent=2)
     tmp.write(opt_content)
@@ -61,6 +66,7 @@ def init_model(self):
                                  dtype=np.float32)
     self.word_count = np.power(self.word_count, self.opt.count_power)
     self.num_words = len(self.words)
+    assert len(self.words) == len(self.word_count)
 
     # count number of docs
     h5f = h5py.File(pjoin(data_dir, "token.h5"), "r")
@@ -70,6 +76,12 @@ def init_model(self):
     self.logger.info("number of words: %d, docs: %d",
                      self.num_words, self.num_docs)
 
+    if self.opt.neg:
+      self.obj.build_random_table( \
+        self.word_count, self.opt.random_size, self.opt.num_threads)
+    else:
+      self.obj.build_huffman_tree(self.word_count)
+
     # random initialize alpha and beta
     np.random.seed(self.opt.seed)
     self.emb_in = np.random.normal( \
@@ -86,11 +98,6 @@ def init_model(self):
   def train_model(self):
     self.preprocess_data()
     self.init_model()
-    if self.opt.neg:
-      self.obj.build_random_table( \
-        self.word_count, self.opt.random_size, self.opt.num_threads)
-    else:
-      self.obj.build_huffman_tree(self.word_count)
     h5f = h5py.File(pjoin(self.opt.processed_data_dir, "token.h5"), "r")
     for epoch in range(1, self.opt.epochs + 1):
       self.logger.info("Epoch %d / %d", epoch, self.opt.epochs)

cusim/proto/config.proto

@@ -60,6 +60,5 @@ message CuW2VConfigProto {
   optional bool use_mean = 20 [default = true];
   optional double lr = 21 [default = 0.001];
   optional int32 window_size = 22 [default = 5];
-
-
+  optional int32 num_threads = 23 [default = 4];
 }