sampling GLAMI1-M dataset

dataManager/gFunDataset.py

@@ -108,30 +108,13 @@ class gFunDataset:
         return dataset, labels, data_langs
 
     def _load_glami(self, dataset_dir, nrows):
-        # TODO: a better way to get a stratified sampling of the dataset (see: groupby + sample)
-        def _balanced_sample(data, n, remainder=0):
-            import pandas as pd
-
-            langs = sorted(data.geo.unique().tolist())
-            dict_n = {lang: n for lang in langs}
-            dict_n[langs[0]] += remainder
-
-            sampled = []
-            for lang in langs:
-                sampled.append(data[data.geo == lang].sample(n=dict_n[lang]))
-
-            return pd.concat(sampled, axis=0)
-
-        # TODO: set this sampling as determinsitic/dependeing on the seed
-        lang_nrows = (
-            nrows // 13 if self.data_langs is None else nrows // len(self.data_langs)
-        )  # GLAMI 1-M has 13 languages
-        remainder = (
-            nrows % 13 if self.data_langs is None else nrows % len(self.data_langs)
+        train_split = get_dataframe("train", dataset_dir=dataset_dir).sample(n=nrows)
+        test_split = get_dataframe("test", dataset_dir=dataset_dir).sample(
+            n=int(nrows / 10)
         )
 
-        train_split = get_dataframe("train", dataset_dir=dataset_dir)
-        train_split = _balanced_sample(train_split, lang_nrows, remainder=remainder)
+        gb_train = train_split.groupby("geo")
+        gb_test = test_split.groupby("geo")
 
         if self.data_langs is None:
             data_langs = sorted(train_split.geo.unique().tolist())
@@ -139,14 +122,6 @@ class gFunDataset:
         if self.labels is None:
             labels = train_split.category_name.unique().tolist()
 
-        # TODO: atm test data should contain same languages as train data
-        test_split = get_dataframe("test", dataset_dir=dataset_dir)
-        # TODO: atm we're using 1:1 train-test
-        test_split = _balanced_sample(test_split, lang_nrows, remainder=remainder)
-
-        gb_train = train_split.groupby("geo")
-        gb_test = test_split.groupby("geo")
-
         def _format_glami(data_df):
             text = (data_df.name + " " + data_df.description).tolist()
             image = data_df.image_file.tolist()

gfun/vgfs/commons.py

@@ -340,7 +340,7 @@ class EarlyStopping:
         self.experiment_name = experiment_name
 
     def __call__(self, validation, model, epoch):
-        if validation > self.best_score:
+        if validation >= self.best_score:
             if self.verbose:
                 print(
                     f"- earlystopping: Validation score improved from {self.best_score:.3f} to {validation:.3f}"

gfun/vgfs/textualTransformerGen.py

@@ -100,7 +100,7 @@ class TextualTransformerGen(ViewGen, TransformerGen):
             return "bert-base-uncased"
         elif "mbert" == model_name:
             return "bert-base-multilingual-uncased"
-        elif "xlm" == model_name:
+        elif "xlm-roberta" == model_name:
             return "xlm-roberta-base"
         elif "mt5" == model_name:
             return "google/mt5-small"
@@ -270,8 +270,8 @@ class TextualTransformerGen(ViewGen, TransformerGen):
         elif "bert" in model_name:
             if "multilingual" in model_name:
                 return "mbert"
-            elif "xlm" in model_name:
-                return "xlm"
+            elif "xlm-roberta" in model_name:
+                return "xlm-roberta"
         else:
             return model_name
 

main.py

@@ -13,6 +13,7 @@ from gfun.generalizedFunnelling import GeneralizedFunnelling
 """
 TODO:
     - Transformers VGFs:
+        - scheduler with warmup and cosine
         - freeze params method
     - General:
         [!] zero-shot setup
@@ -177,17 +178,17 @@ if __name__ == "__main__":
     parser.add_argument("--features", action="store_false")
     parser.add_argument("--aggfunc", type=str, default="mean")
     # transformer parameters ---------------
+    parser.add_argument("--epochs", type=int, default=100)
     parser.add_argument("--textual_trf_name", type=str, default="mbert")
     parser.add_argument("--batch_size", type=int, default=32)
     parser.add_argument("--eval_batch_size", type=int, default=128)
-    parser.add_argument("--epochs", type=int, default=100)
-    parser.add_argument("--textual_lr", type=float, default=1e-5)
-    parser.add_argument("--visual_lr", type=float, default=1e-5)
+    parser.add_argument("--textual_lr", type=float, default=1e-4)
     parser.add_argument("--max_length", type=int, default=128)
     parser.add_argument("--patience", type=int, default=5)
     parser.add_argument("--evaluate_step", type=int, default=10)
     # Visual Transformer parameters --------------
     parser.add_argument("--visual_trf_name", type=str, default="vit")
+    parser.add_argument("--visual_lr", type=float, default=1e-4)
 
     args = parser.parse_args()