Merge pull request #27 from dblasko/eval-mem-leak

Fix test set evaluation GPU memory leak

Author: ttanida

src/full_model/evaluate_full_model/evaluate_language_model.py

@@ -982,12 +982,19 @@ def update_gen_and_ref_sentences_for_regions(
             index_gen_ref_sentence += 1
 
 
-def get_generated_reports(generated_sentences_for_selected_regions, selected_regions, sentence_tokenizer, bertscore_threshold):
+def get_generated_reports(
+    generated_sentences_for_selected_regions,
+    selected_regions,
+    sentence_tokenizer,
+    bertscore_threshold,
+    bert_score,
+):
     """
     Args:
         generated_sentences_for_selected_regions (List[str]): of length "num_regions_selected_in_batch"
         selected_regions ([batch_size x 29]): boolean array that has exactly "num_regions_selected_in_batch" True values
         sentence_tokenizer: used in remove_duplicate_generated_sentences to separate the generated sentences
+        bert_score: instance of the evaluate bert score evaluation module
 
     Return:
         generated_reports (List[str]): list of length batch_size containing generated reports for every image in batch
@@ -1055,8 +1062,6 @@ def check_gen_sent_in_sents_to_be_removed(gen_sent, similar_generated_sents_to_b
 
         return gen_report_single_image, similar_generated_sents_to_be_removed
 
-    bert_score = evaluate.load("bertscore")
-
     generated_reports = []
     removed_similar_generated_sentences = []
     curr_index = 0

src/full_model/test_set_evaluation.py

@@ -11,6 +11,7 @@
 import pandas as pd
 import spacy
 import torch
+import evaluate
 from torch.utils.data import DataLoader
 import torchmetrics
 from tqdm import tqdm
@@ -268,6 +269,7 @@ def iterate_over_test_loader(test_loader):
 
         # used in function get_generated_reports
         sentence_tokenizer = spacy.load("en_core_web_trf")
+        bert_score = evaluate.load("bertscore")
 
         with torch.no_grad():
             for num_batch, batch in tqdm(enumerate(test_loader)):
@@ -342,6 +344,14 @@ def iterate_over_test_loader(test_loader):
                     selected_regions,
                     sentence_tokenizer,
                     BERTSCORE_SIMILARITY_THRESHOLD
+                generated_reports, removed_similar_generated_sentences = (
+                    get_generated_reports(
+                        generated_sents_for_selected_regions,
+                        selected_regions,
+                        sentence_tokenizer,
+                        BERTSCORE_SIMILARITY_THRESHOLD,
+                        bert_score,
+                    )
                 )
 
                 gen_and_ref_sentences["generated_sentences"].extend(generated_sents_for_selected_regions)