Add device selection

Author: matgille

aquilign/align/bertalign/Bertalign.py

@@ -5,18 +5,9 @@
 __author__ = "Jason ([email protected])"
 __version__ = "1.1.0"
 
-from aquilign.align.bertalign.encoder import Encoder
 # from sonar.inference_pipelines.text import TextToEmbeddingModelPipeline
 # See other cross-lingual embedding models at
 # https://www.sbert.net/docs/pretrained_models.html
 print("Chargement 1")
 
-models = {0: "distiluse-base-multilingual-cased-v2", 1: "LaBSE", 2: "Sonar"}
-# as_list = ""
-# for key, value in models.items():
-#     as_list += f"{int(key)}\t{value}\n"
-# print(as_list)
-# nb = input(f'Choose a model:')
-model = Encoder(models[int(1)])
 
-from aquilign.align.bertalign.aligner import Bertalign

aquilign/align/bertalign/aligner.py

@@ -1,13 +1,13 @@
 import numpy as np
 
-from aquilign.align.bertalign.Bertalign import model
 import aquilign.align.bertalign.corelib as core
 import aquilign.align.bertalign.utils as utils
 import torch.nn as nn
 import torch
 
 class Bertalign:
     def __init__(self,
+                 model,
                  src,
                  tgt,
                  max_align=3,
@@ -17,14 +17,17 @@ def __init__(self,
                  margin=True,
                  len_penalty=True,
                  is_split=False,
-               ):
+                 device="cpu"):
 
         self.max_align = max_align
         self.top_k = top_k
         self.win = win
         self.skip = skip
         self.margin = margin
         self.len_penalty = len_penalty
+        self.device = device
+        self.model = model
+        
 
 
 
@@ -38,11 +41,11 @@ def __init__(self,
         assert len(src_sents) != 0, "Problemo"
 
         print("Embedding source and target text using {} ...".format(model.model_name))
-        src_vecs, src_lens = model.transform(src_sents, max_align - 1)
-        tgt_vecs, tgt_lens = model.transform(tgt_sents, max_align - 1)
+        src_vecs, src_lens = self.model.transform(src_sents, max_align - 1)
+        tgt_vecs, tgt_lens = self.model.transform(tgt_sents, max_align - 1)
 
-        self.search_simple_vecs = model.simple_vectorization(src_sents)
-        self.tgt_simple_vecs = model.simple_vectorization(tgt_sents)
+        self.search_simple_vecs = self.model.simple_vectorization(src_sents)
+        self.tgt_simple_vecs = self.model.simple_vectorization(tgt_sents)
 
         char_ratio = np.sum(src_lens[0,]) / np.sum(tgt_lens[0,])
 
@@ -57,15 +60,15 @@ def __init__(self,
         self.tgt_vecs = tgt_vecs
 
     def compute_distance(self):
-        if torch.cuda.is_available():  # GPU version
+        if torch.cuda.is_available() and self.device == 'cuda:0':  # GPU version
             cos = nn.CosineSimilarity(dim=1, eps=1e-6)
             output = cos(torch.from_numpy(self.search_simple_vecs), torch.from_numpy(self.tgt_simple_vecs))
         return output
 
     def align_sents(self, first_alignment_only=False):
 
         print("Performing first-step alignment ...")
-        D, I = core.find_top_k_sents(self.src_vecs[0,:], self.tgt_vecs[0,:], k=self.top_k)
+        D, I = core.find_top_k_sents(self.src_vecs[0,:], self.tgt_vecs[0,:], k=self.top_k, device=self.device)
         first_alignment_types = core.get_alignment_types(2) # 0-1, 1-0, 1-1
         first_w, first_path = core.find_first_search_path(self.src_num, self.tgt_num)
         first_pointers = core.first_pass_align(self.src_num, self.tgt_num, first_w, first_path, first_alignment_types, D, I)

aquilign/align/bertalign/corelib.py

@@ -377,7 +377,7 @@ def get_alignment_types(max_alignment_size):
                 alignment_types.append([x, y])    
     return np.array(alignment_types)
 
-def find_top_k_sents(src_vecs, tgt_vecs, k=3):
+def find_top_k_sents(src_vecs, tgt_vecs, k=3, device='cpu'):
     """
     Find the top_k similar vecs in tgt_vecs for each vec in src_vecs.
     Args:
@@ -389,13 +389,17 @@ def find_top_k_sents(src_vecs, tgt_vecs, k=3):
         I: numpy array. Target index matrix of shape (num_src_sents, k).
     """
     embedding_size = src_vecs.shape[1]
-    if torch.cuda.is_available() and platform == 'linux': # GPU version
+    if torch.cuda.is_available() and platform == 'linux' and device != "cpu": # GPU version
+        print("OK")
+        exit(0)
         res = faiss.StandardGpuResources() 
         index = faiss.IndexFlatIP(embedding_size)
         gpu_index = faiss.index_cpu_to_gpu(res, 0, index)
         gpu_index.add(tgt_vecs) 
         D, I = gpu_index.search(src_vecs, k)
     else: # CPU version
+        print("Not OK")
+        exit(0)
         index = faiss.IndexFlatIP(embedding_size)
         index.add(tgt_vecs)
         D, I = index.search(src_vecs, k)

aquilign/align/bertalign/encoder.py

@@ -6,8 +6,8 @@
 
 
 class Encoder:
-    def __init__(self, model_name):
-        device = "cuda:0" if torch.cuda.is_available() else "cpu"
+    def __init__(self, model_name, device):
+        self.device = "cuda:0" if torch.cuda.is_available() and device != "cpu" else "cpu"
         if model_name == "LaBSE":
             self.model = SentenceTransformer(model_name_or_path=model_name, device=device)
             self.model_name = model_name
@@ -21,7 +21,7 @@ def simple_vectorization(self, sents):
         This function produces a simple vectorisation of a sentence, without
         taking into account its lenght as transform does
         """
-        sent_vecs = self.model.encode(sents)
+        sent_vecs = self.model.encode(sents, device=self.device)
         return sent_vecs
 
     def transform(self, sents, num_overlaps):
@@ -30,7 +30,7 @@ def transform(self, sents, num_overlaps):
             overlaps.append(line)
 
         if self.model_name == "LaBSE":
-            sent_vecs = self.model.encode(overlaps)
+            sent_vecs = self.model.encode(overlaps, device=self.device)
         else:
             sents_vecs = self.t2vec_model.predict()
         embedding_dim = sent_vecs.size // (len(sents) * num_overlaps)

main.py

@@ -9,7 +9,9 @@
 import aquilign.align.graph_merge as graph_merge
 import aquilign.align.bertalign.utils as utils
 import aquilign.align.bertalign.syntactic_tokenization as syntactic_tokenization
-from aquilign.align.bertalign.Bertalign import Bertalign
+#from aquilign.align.bertalign.Bertalign import Bertalign
+from aquilign.align.bertalign.encoder import Encoder
+from aquilign.align.bertalign.aligner import Bertalign
 import pandas as pd
 import argparse
 import glob
@@ -60,16 +62,21 @@ class Aligner:
     La classe Aligner initialise le moteur d'alignement, fondé sur Bertalign
     """
 
-    def __init__(self, corpus_size:None, 
+    def __init__(self,
+                 model,
+                 corpus_size:None, 
                  max_align=3, 
                  out_dir="out", 
                  use_punctuation=True, 
                  input_dir="in", 
                  main_wit=None, 
-                 prefix=None):
+                 prefix=None,
+                 device="cpu"):
+        self.model = model
         self.alignment_dict = dict()
         self.text_dict = dict()
         self.files_path = glob.glob(f"{input_dir}/*.txt")
+        self.device = device
         print(input_dir)
         if main_wit is not None:
             self.main_file_index = [index for index, path in enumerate(self.files_path) if main_wit in path][0]
@@ -133,7 +140,14 @@ def parallel_align(self):
             else:
                 margin = False
                 len_penality = True
-            aligner = Bertalign(first_tokenized_text, second_tokenized_text, max_align= self.max_align, win=5, skip=-.2, margin=margin, len_penalty=len_penality)
+            aligner = Bertalign(self.model,
+                                first_tokenized_text, 
+                                second_tokenized_text, 
+                                max_align= self.max_align, 
+                                win=5, skip=-.2, 
+                                margin=margin, 
+                                len_penalty=len_penality, 
+                                device=self.device)
             aligner.align_sents()
 
             # We append the result to the alignment dictionnary
@@ -215,15 +229,25 @@ def run_alignments():
                         help="Pivot witness.")
     parser.add_argument("-p", "--prefix", default=None,
                         help="Prefix for produced files.")
+    parser.add_argument("-d", "--device", default='cpu',
+                        help="Device to be used.")
 
     args = parser.parse_args()
     out_dir = args.out_dir
     input_dir = args.input_dir
     main_wit = args.main_wit
     prefix = args.prefix
+    device = args.device
     use_punctuation = args.use_punctuation
+    
+    # Initialize model 
+    models = {0: "distiluse-base-multilingual-cased-v2", 1: "LaBSE", 2: "Sonar"}
+    model = Encoder(models[int(1)], device=device)
+    
+    
+    
     print(f"Punctuation for tokenization: {use_punctuation}")
-    MyAligner = Aligner(corpus_size=None, max_align=3, out_dir=out_dir, use_punctuation=use_punctuation, input_dir=input_dir, main_wit=main_wit, prefix=prefix)
+    MyAligner = Aligner(model, corpus_size=None, max_align=3, out_dir=out_dir, use_punctuation=use_punctuation, input_dir=input_dir, main_wit=main_wit, prefix=prefix, device=device)
     MyAligner.parallel_align()
     utils.write_json(f"result_dir/{out_dir}/alignment_dict.json", MyAligner.alignment_dict)
     align_dict = utils.read_json(f"result_dir/{out_dir}/alignment_dict.json")