Tom Aarsen
commited on
Commit
·
dd9a98e
1
Parent(s):
bc2993c
Update training script to separate dataset loading & training
Browse files
train.py
CHANGED
@@ -20,6 +20,148 @@ logging.basicConfig(
|
|
20 |
random.seed(12)
|
21 |
|
22 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
def main():
|
24 |
# 1. Load a model to finetune with 2. (Optional) model card data
|
25 |
static_embedding = StaticEmbedding(AutoTokenizer.from_pretrained("google-bert/bert-base-uncased"), embedding_dim=1024)
|
@@ -33,129 +175,7 @@ def main():
|
|
33 |
)
|
34 |
|
35 |
# 3. Set up training & evaluation datasets - each dataset is trained with MNRL (with MRL)
|
36 |
-
|
37 |
-
gooaq_dataset = load_dataset("sentence-transformers/gooaq", split="train")
|
38 |
-
gooaq_dataset_dict = gooaq_dataset.train_test_split(test_size=10_000, seed=12)
|
39 |
-
gooaq_train_dataset: Dataset = gooaq_dataset_dict["train"]
|
40 |
-
gooaq_eval_dataset: Dataset = gooaq_dataset_dict["test"]
|
41 |
-
print("Loaded gooaq dataset.")
|
42 |
-
|
43 |
-
print("Loading msmarco dataset...")
|
44 |
-
msmarco_dataset = load_dataset("sentence-transformers/msmarco-co-condenser-margin-mse-sym-mnrl-mean-v1", "triplet", split="train")
|
45 |
-
msmarco_dataset_dict = msmarco_dataset.train_test_split(test_size=10_000, seed=12)
|
46 |
-
msmarco_train_dataset: Dataset = msmarco_dataset_dict["train"]
|
47 |
-
msmarco_eval_dataset: Dataset = msmarco_dataset_dict["test"]
|
48 |
-
print("Loaded msmarco dataset.")
|
49 |
-
|
50 |
-
print("Loading squad dataset...")
|
51 |
-
squad_dataset = load_dataset("sentence-transformers/squad", split="train")
|
52 |
-
squad_dataset_dict = squad_dataset.train_test_split(test_size=10_000, seed=12)
|
53 |
-
squad_train_dataset: Dataset = squad_dataset_dict["train"]
|
54 |
-
squad_eval_dataset: Dataset = squad_dataset_dict["test"]
|
55 |
-
print("Loaded squad dataset.")
|
56 |
-
|
57 |
-
print("Loading s2orc dataset...")
|
58 |
-
s2orc_dataset = load_dataset("sentence-transformers/s2orc", "title-abstract-pair", split="train[:100000]")
|
59 |
-
s2orc_dataset_dict = s2orc_dataset.train_test_split(test_size=10_000, seed=12)
|
60 |
-
s2orc_train_dataset: Dataset = s2orc_dataset_dict["train"]
|
61 |
-
s2orc_eval_dataset: Dataset = s2orc_dataset_dict["test"]
|
62 |
-
print("Loaded s2orc dataset.")
|
63 |
-
|
64 |
-
print("Loading allnli dataset...")
|
65 |
-
allnli_train_dataset = load_dataset("sentence-transformers/all-nli", "triplet", split="train")
|
66 |
-
allnli_eval_dataset = load_dataset("sentence-transformers/all-nli", "triplet", split="dev")
|
67 |
-
print("Loaded allnli dataset.")
|
68 |
-
|
69 |
-
print("Loading paq dataset...")
|
70 |
-
paq_dataset = load_dataset("sentence-transformers/paq", split="train")
|
71 |
-
paq_dataset_dict = paq_dataset.train_test_split(test_size=10_000, seed=12)
|
72 |
-
paq_train_dataset: Dataset = paq_dataset_dict["train"]
|
73 |
-
paq_eval_dataset: Dataset = paq_dataset_dict["test"]
|
74 |
-
print("Loaded paq dataset.")
|
75 |
-
|
76 |
-
print("Loading trivia_qa dataset...")
|
77 |
-
trivia_qa = load_dataset("sentence-transformers/trivia-qa", split="train")
|
78 |
-
trivia_qa_dataset_dict = trivia_qa.train_test_split(test_size=5_000, seed=12)
|
79 |
-
trivia_qa_train_dataset: Dataset = trivia_qa_dataset_dict["train"]
|
80 |
-
trivia_qa_eval_dataset: Dataset = trivia_qa_dataset_dict["test"]
|
81 |
-
print("Loaded trivia_qa dataset.")
|
82 |
-
|
83 |
-
print("Loading msmarco_10m dataset...")
|
84 |
-
msmarco_10m_dataset = load_dataset("bclavie/msmarco-10m-triplets", split="train")
|
85 |
-
msmarco_10m_dataset_dict = msmarco_10m_dataset.train_test_split(test_size=10_000, seed=12)
|
86 |
-
msmarco_10m_train_dataset: Dataset = msmarco_10m_dataset_dict["train"]
|
87 |
-
msmarco_10m_eval_dataset: Dataset = msmarco_10m_dataset_dict["test"]
|
88 |
-
print("Loaded msmarco_10m dataset.")
|
89 |
-
|
90 |
-
print("Loading swim_ir dataset...")
|
91 |
-
swim_ir_dataset = load_dataset("nthakur/swim-ir-monolingual", "en", split="train").select_columns(["query", "text"])
|
92 |
-
swim_ir_dataset_dict = swim_ir_dataset.train_test_split(test_size=10_000, seed=12)
|
93 |
-
swim_ir_train_dataset: Dataset = swim_ir_dataset_dict["train"]
|
94 |
-
swim_ir_eval_dataset: Dataset = swim_ir_dataset_dict["test"]
|
95 |
-
print("Loaded swim_ir dataset.")
|
96 |
-
|
97 |
-
# NOTE: 20 negatives
|
98 |
-
print("Loading pubmedqa dataset...")
|
99 |
-
pubmedqa_dataset = load_dataset("sentence-transformers/pubmedqa", "triplet-20", split="train")
|
100 |
-
pubmedqa_dataset_dict = pubmedqa_dataset.train_test_split(test_size=100, seed=12)
|
101 |
-
pubmedqa_train_dataset: Dataset = pubmedqa_dataset_dict["train"]
|
102 |
-
pubmedqa_eval_dataset: Dataset = pubmedqa_dataset_dict["test"]
|
103 |
-
print("Loaded pubmedqa dataset.")
|
104 |
-
|
105 |
-
# NOTE: A lot of overlap with anchor/positives
|
106 |
-
print("Loading miracl dataset...")
|
107 |
-
miracl_dataset = load_dataset("sentence-transformers/miracl", "en-triplet-all", split="train")
|
108 |
-
miracl_dataset_dict = miracl_dataset.train_test_split(test_size=10_000, seed=12)
|
109 |
-
miracl_train_dataset: Dataset = miracl_dataset_dict["train"]
|
110 |
-
miracl_eval_dataset: Dataset = miracl_dataset_dict["test"]
|
111 |
-
print("Loaded miracl dataset.")
|
112 |
-
|
113 |
-
# NOTE: A lot of overlap with anchor/positives
|
114 |
-
print("Loading mldr dataset...")
|
115 |
-
mldr_dataset = load_dataset("sentence-transformers/mldr", "en-triplet-all", split="train")
|
116 |
-
mldr_dataset_dict = mldr_dataset.train_test_split(test_size=10_000, seed=12)
|
117 |
-
mldr_train_dataset: Dataset = mldr_dataset_dict["train"]
|
118 |
-
mldr_eval_dataset: Dataset = mldr_dataset_dict["test"]
|
119 |
-
print("Loaded mldr dataset.")
|
120 |
-
|
121 |
-
# NOTE: A lot of overlap with anchor/positives
|
122 |
-
print("Loading mr_tydi dataset...")
|
123 |
-
mr_tydi_dataset = load_dataset("sentence-transformers/mr-tydi", "en-triplet-all", split="train")
|
124 |
-
mr_tydi_dataset_dict = mr_tydi_dataset.train_test_split(test_size=10_000, seed=12)
|
125 |
-
mr_tydi_train_dataset: Dataset = mr_tydi_dataset_dict["train"]
|
126 |
-
mr_tydi_eval_dataset: Dataset = mr_tydi_dataset_dict["test"]
|
127 |
-
print("Loaded mr_tydi dataset.")
|
128 |
-
|
129 |
-
train_dataset = DatasetDict({
|
130 |
-
"gooaq": gooaq_train_dataset,
|
131 |
-
"msmarco": msmarco_train_dataset,
|
132 |
-
"squad": squad_train_dataset,
|
133 |
-
"s2orc": s2orc_train_dataset,
|
134 |
-
"allnli": allnli_train_dataset,
|
135 |
-
"paq": paq_train_dataset,
|
136 |
-
"trivia_qa": trivia_qa_train_dataset,
|
137 |
-
"msmarco_10m": msmarco_10m_train_dataset,
|
138 |
-
"swim_ir": swim_ir_train_dataset,
|
139 |
-
"pubmedqa": pubmedqa_train_dataset,
|
140 |
-
"miracl": miracl_train_dataset,
|
141 |
-
"mldr": mldr_train_dataset,
|
142 |
-
"mr_tydi": mr_tydi_train_dataset,
|
143 |
-
})
|
144 |
-
eval_dataset = {
|
145 |
-
"gooaq": gooaq_eval_dataset,
|
146 |
-
"msmarco": msmarco_eval_dataset,
|
147 |
-
"squad": squad_eval_dataset,
|
148 |
-
"s2orc": s2orc_eval_dataset,
|
149 |
-
"allnli": allnli_eval_dataset,
|
150 |
-
"paq": paq_eval_dataset,
|
151 |
-
"trivia_qa": trivia_qa_eval_dataset,
|
152 |
-
"msmarco_10m": msmarco_10m_eval_dataset,
|
153 |
-
"swim_ir": swim_ir_eval_dataset,
|
154 |
-
"pubmedqa": pubmedqa_eval_dataset,
|
155 |
-
"miracl": miracl_eval_dataset,
|
156 |
-
"mldr": mldr_eval_dataset,
|
157 |
-
"mr_tydi": mr_tydi_eval_dataset,
|
158 |
-
}
|
159 |
print(train_dataset)
|
160 |
|
161 |
# 4. Define a loss function
|
|
|
20 |
random.seed(12)
|
21 |
|
22 |
|
23 |
+
def load_train_eval_datasets():
|
24 |
+
"""
|
25 |
+
Either load the train and eval datasets from disk or load them from the datasets library & save them to disk.
|
26 |
+
|
27 |
+
Upon saving to disk, we quit() to ensure that the datasets are not loaded into memory before training.
|
28 |
+
"""
|
29 |
+
try:
|
30 |
+
train_dataset = DatasetDict.load_from_disk("datasets/train_dataset")
|
31 |
+
eval_dataset = DatasetDict.load_from_disk("datasets/eval_dataset")
|
32 |
+
return train_dataset, eval_dataset
|
33 |
+
except FileNotFoundError:
|
34 |
+
print("Loading gooaq dataset...")
|
35 |
+
gooaq_dataset = load_dataset("sentence-transformers/gooaq", split="train")
|
36 |
+
gooaq_dataset_dict = gooaq_dataset.train_test_split(test_size=10_000, seed=12)
|
37 |
+
gooaq_train_dataset: Dataset = gooaq_dataset_dict["train"]
|
38 |
+
gooaq_eval_dataset: Dataset = gooaq_dataset_dict["test"]
|
39 |
+
print("Loaded gooaq dataset.")
|
40 |
+
|
41 |
+
print("Loading msmarco dataset...")
|
42 |
+
msmarco_dataset = load_dataset("sentence-transformers/msmarco-co-condenser-margin-mse-sym-mnrl-mean-v1", "triplet", split="train")
|
43 |
+
msmarco_dataset_dict = msmarco_dataset.train_test_split(test_size=10_000, seed=12)
|
44 |
+
msmarco_train_dataset: Dataset = msmarco_dataset_dict["train"]
|
45 |
+
msmarco_eval_dataset: Dataset = msmarco_dataset_dict["test"]
|
46 |
+
print("Loaded msmarco dataset.")
|
47 |
+
|
48 |
+
print("Loading squad dataset...")
|
49 |
+
squad_dataset = load_dataset("sentence-transformers/squad", split="train")
|
50 |
+
squad_dataset_dict = squad_dataset.train_test_split(test_size=10_000, seed=12)
|
51 |
+
squad_train_dataset: Dataset = squad_dataset_dict["train"]
|
52 |
+
squad_eval_dataset: Dataset = squad_dataset_dict["test"]
|
53 |
+
print("Loaded squad dataset.")
|
54 |
+
|
55 |
+
print("Loading s2orc dataset...")
|
56 |
+
s2orc_dataset = load_dataset("sentence-transformers/s2orc", "title-abstract-pair", split="train[:100000]")
|
57 |
+
s2orc_dataset_dict = s2orc_dataset.train_test_split(test_size=10_000, seed=12)
|
58 |
+
s2orc_train_dataset: Dataset = s2orc_dataset_dict["train"]
|
59 |
+
s2orc_eval_dataset: Dataset = s2orc_dataset_dict["test"]
|
60 |
+
print("Loaded s2orc dataset.")
|
61 |
+
|
62 |
+
print("Loading allnli dataset...")
|
63 |
+
allnli_train_dataset = load_dataset("sentence-transformers/all-nli", "triplet", split="train")
|
64 |
+
allnli_eval_dataset = load_dataset("sentence-transformers/all-nli", "triplet", split="dev")
|
65 |
+
print("Loaded allnli dataset.")
|
66 |
+
|
67 |
+
print("Loading paq dataset...")
|
68 |
+
paq_dataset = load_dataset("sentence-transformers/paq", split="train")
|
69 |
+
paq_dataset_dict = paq_dataset.train_test_split(test_size=10_000, seed=12)
|
70 |
+
paq_train_dataset: Dataset = paq_dataset_dict["train"]
|
71 |
+
paq_eval_dataset: Dataset = paq_dataset_dict["test"]
|
72 |
+
print("Loaded paq dataset.")
|
73 |
+
|
74 |
+
print("Loading trivia_qa dataset...")
|
75 |
+
trivia_qa = load_dataset("sentence-transformers/trivia-qa", split="train")
|
76 |
+
trivia_qa_dataset_dict = trivia_qa.train_test_split(test_size=5_000, seed=12)
|
77 |
+
trivia_qa_train_dataset: Dataset = trivia_qa_dataset_dict["train"]
|
78 |
+
trivia_qa_eval_dataset: Dataset = trivia_qa_dataset_dict["test"]
|
79 |
+
print("Loaded trivia_qa dataset.")
|
80 |
+
|
81 |
+
print("Loading msmarco_10m dataset...")
|
82 |
+
msmarco_10m_dataset = load_dataset("bclavie/msmarco-10m-triplets", split="train")
|
83 |
+
msmarco_10m_dataset_dict = msmarco_10m_dataset.train_test_split(test_size=10_000, seed=12)
|
84 |
+
msmarco_10m_train_dataset: Dataset = msmarco_10m_dataset_dict["train"]
|
85 |
+
msmarco_10m_eval_dataset: Dataset = msmarco_10m_dataset_dict["test"]
|
86 |
+
print("Loaded msmarco_10m dataset.")
|
87 |
+
|
88 |
+
print("Loading swim_ir dataset...")
|
89 |
+
swim_ir_dataset = load_dataset("nthakur/swim-ir-monolingual", "en", split="train").select_columns(["query", "text"])
|
90 |
+
swim_ir_dataset_dict = swim_ir_dataset.train_test_split(test_size=10_000, seed=12)
|
91 |
+
swim_ir_train_dataset: Dataset = swim_ir_dataset_dict["train"]
|
92 |
+
swim_ir_eval_dataset: Dataset = swim_ir_dataset_dict["test"]
|
93 |
+
print("Loaded swim_ir dataset.")
|
94 |
+
|
95 |
+
# NOTE: 20 negatives
|
96 |
+
print("Loading pubmedqa dataset...")
|
97 |
+
pubmedqa_dataset = load_dataset("sentence-transformers/pubmedqa", "triplet-20", split="train")
|
98 |
+
pubmedqa_dataset_dict = pubmedqa_dataset.train_test_split(test_size=100, seed=12)
|
99 |
+
pubmedqa_train_dataset: Dataset = pubmedqa_dataset_dict["train"]
|
100 |
+
pubmedqa_eval_dataset: Dataset = pubmedqa_dataset_dict["test"]
|
101 |
+
print("Loaded pubmedqa dataset.")
|
102 |
+
|
103 |
+
# NOTE: A lot of overlap with anchor/positives
|
104 |
+
print("Loading miracl dataset...")
|
105 |
+
miracl_dataset = load_dataset("sentence-transformers/miracl", "en-triplet-all", split="train")
|
106 |
+
miracl_dataset_dict = miracl_dataset.train_test_split(test_size=10_000, seed=12)
|
107 |
+
miracl_train_dataset: Dataset = miracl_dataset_dict["train"]
|
108 |
+
miracl_eval_dataset: Dataset = miracl_dataset_dict["test"]
|
109 |
+
print("Loaded miracl dataset.")
|
110 |
+
|
111 |
+
# NOTE: A lot of overlap with anchor/positives
|
112 |
+
print("Loading mldr dataset...")
|
113 |
+
mldr_dataset = load_dataset("sentence-transformers/mldr", "en-triplet-all", split="train")
|
114 |
+
mldr_dataset_dict = mldr_dataset.train_test_split(test_size=10_000, seed=12)
|
115 |
+
mldr_train_dataset: Dataset = mldr_dataset_dict["train"]
|
116 |
+
mldr_eval_dataset: Dataset = mldr_dataset_dict["test"]
|
117 |
+
print("Loaded mldr dataset.")
|
118 |
+
|
119 |
+
# NOTE: A lot of overlap with anchor/positives
|
120 |
+
print("Loading mr_tydi dataset...")
|
121 |
+
mr_tydi_dataset = load_dataset("sentence-transformers/mr-tydi", "en-triplet-all", split="train")
|
122 |
+
mr_tydi_dataset_dict = mr_tydi_dataset.train_test_split(test_size=10_000, seed=12)
|
123 |
+
mr_tydi_train_dataset: Dataset = mr_tydi_dataset_dict["train"]
|
124 |
+
mr_tydi_eval_dataset: Dataset = mr_tydi_dataset_dict["test"]
|
125 |
+
print("Loaded mr_tydi dataset.")
|
126 |
+
|
127 |
+
train_dataset = DatasetDict({
|
128 |
+
"gooaq": gooaq_train_dataset,
|
129 |
+
"msmarco": msmarco_train_dataset,
|
130 |
+
"squad": squad_train_dataset,
|
131 |
+
"s2orc": s2orc_train_dataset,
|
132 |
+
"allnli": allnli_train_dataset,
|
133 |
+
"paq": paq_train_dataset,
|
134 |
+
"trivia_qa": trivia_qa_train_dataset,
|
135 |
+
"msmarco_10m": msmarco_10m_train_dataset,
|
136 |
+
"swim_ir": swim_ir_train_dataset,
|
137 |
+
"pubmedqa": pubmedqa_train_dataset,
|
138 |
+
"miracl": miracl_train_dataset,
|
139 |
+
"mldr": mldr_train_dataset,
|
140 |
+
"mr_tydi": mr_tydi_train_dataset,
|
141 |
+
})
|
142 |
+
eval_dataset = DatasetDict({
|
143 |
+
"gooaq": gooaq_eval_dataset,
|
144 |
+
"msmarco": msmarco_eval_dataset,
|
145 |
+
"squad": squad_eval_dataset,
|
146 |
+
"s2orc": s2orc_eval_dataset,
|
147 |
+
"allnli": allnli_eval_dataset,
|
148 |
+
"paq": paq_eval_dataset,
|
149 |
+
"trivia_qa": trivia_qa_eval_dataset,
|
150 |
+
"msmarco_10m": msmarco_10m_eval_dataset,
|
151 |
+
"swim_ir": swim_ir_eval_dataset,
|
152 |
+
"pubmedqa": pubmedqa_eval_dataset,
|
153 |
+
"miracl": miracl_eval_dataset,
|
154 |
+
"mldr": mldr_eval_dataset,
|
155 |
+
"mr_tydi": mr_tydi_eval_dataset,
|
156 |
+
})
|
157 |
+
|
158 |
+
train_dataset.save_to_disk("datasets/train_dataset")
|
159 |
+
eval_dataset.save_to_disk("datasets/eval_dataset")
|
160 |
+
|
161 |
+
# The `train_test_split` calls have put a lot of the datasets in memory, while we want it to just be on disk
|
162 |
+
quit()
|
163 |
+
|
164 |
+
|
165 |
def main():
|
166 |
# 1. Load a model to finetune with 2. (Optional) model card data
|
167 |
static_embedding = StaticEmbedding(AutoTokenizer.from_pretrained("google-bert/bert-base-uncased"), embedding_dim=1024)
|
|
|
175 |
)
|
176 |
|
177 |
# 3. Set up training & evaluation datasets - each dataset is trained with MNRL (with MRL)
|
178 |
+
train_dataset, eval_dataset = load_train_eval_datasets()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
179 |
print(train_dataset)
|
180 |
|
181 |
# 4. Define a loss function
|