From 843cc70b6d8ab8e991053d1cfd57505979cfdd52 Mon Sep 17 00:00:00 2001
From: Caroline DE POURTALES <caroline.de-pourtales@irit.fr>
Date: Tue, 28 Mar 2023 18:10:27 +0200
Subject: [PATCH] easier to configure
---
Configuration/config.ini | 9 +++--
SuperTagger/SuperTagger/SentencesTokenizer.py | 4 +-
SuperTagger/SuperTagger/SuperTagger.py | 9 ++++-
SuperTagger/SuperTagger/SymbolTokenizer.py | 1 +
SuperTagger/SuperTagger/eval.py | 1 -
find_config.py | 38 +++++++++++++++++--
train_supertagger.py | 18 ++++++---
utils.py | 12 ++----
8 files changed, 67 insertions(+), 25 deletions(-)
diff --git a/Configuration/config.ini b/Configuration/config.ini
index a695829..4f253f4 100644
--- a/Configuration/config.ini
+++ b/Configuration/config.ini
@@ -4,9 +4,9 @@ transformers = 4.16.2
[DATASET_PARAMS]
symbols_vocab_size = 26
atom_vocab_size = 18
-max_len_sentence = 300
-max_atoms_in_sentence = 900
-max_atoms_in_one_type = 360
+max_len_sentence = 290
+max_atoms_in_sentence = 440
+max_atoms_in_one_type = 180
[MODEL_ENCODER]
dim_encoder = 768
@@ -20,4 +20,5 @@ dim_cat_out = 256
dim_intermediate_ffn = 128
dim_pre_sinkhorn_transfo = 32
dropout = 0.1
-sinkhorn_iters = 5
\ No newline at end of file
+sinkhorn_iters = 5
+
diff --git a/SuperTagger/SuperTagger/SentencesTokenizer.py b/SuperTagger/SuperTagger/SentencesTokenizer.py
index 104577f..dab638a 100644
--- a/SuperTagger/SuperTagger/SentencesTokenizer.py
+++ b/SuperTagger/SuperTagger/SentencesTokenizer.py
@@ -25,7 +25,9 @@ class SentencesTokenizer():
"""
Tokenizes the given sentences
"""
- return self.tokenizer(sents, padding=True)
+ temp = self.tokenizer(sents, padding=True)
+
+ return temp["input_ids"], temp["attention_mask"]
def fit_transform_tensors(self, sents):
"""
diff --git a/SuperTagger/SuperTagger/SuperTagger.py b/SuperTagger/SuperTagger/SuperTagger.py
index 670849d..cabef35 100644
--- a/SuperTagger/SuperTagger/SuperTagger.py
+++ b/SuperTagger/SuperTagger/SuperTagger.py
@@ -5,6 +5,7 @@ import time
import torch
import transformers
from torch.optim import Adam
+from torch.optim.lr_scheduler import StepLR
from torch.utils.data import TensorDataset, random_split
from tqdm import tqdm
from transformers import AutoTokenizer
@@ -70,6 +71,7 @@ class SuperTagger:
self.model = None
self.optimizer = None
+ self.scheduler = None
self.epoch_i = 0
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
@@ -104,6 +106,7 @@ class SuperTagger:
self.max_len_sentence)
self.model.load_state_dict(params['state_dict'])
self.optimizer = params['optimizer']
+ self.scheduler = StepLR(self.optimizer, step_size=10, gamma=0.5)
# self.epoch_i = args['epoch']
print("\n The loading checkpoint was successful ! \n")
print("\tBert model : ", self.bert_name)
@@ -144,7 +147,9 @@ class SuperTagger:
bert_name,
do_lower_case=True),
self.max_len_sentence)
- self.optimizer = Adam(params=self.model.parameters(), lr=1e-3, eps=1e-8)
+
+ self.optimizer = Adam(params=self.model.parameters(), lr=1e-5, eps=1e-8)
+ self.scheduler = StepLR(self.optimizer, step_size=10, gamma=0.9)
self.tags_tokenizer = SymbolTokenizer(index_to_tags)
self.trainable = True
self.model_load = True
@@ -333,6 +338,8 @@ class SuperTagger:
self.optimizer.step()
i += 1
+
+ self.scheduler.step()
# Measure how long this epoch took.
training_time = format_time(time.time() - t0)
diff --git a/SuperTagger/SuperTagger/SymbolTokenizer.py b/SuperTagger/SuperTagger/SymbolTokenizer.py
index dafcc12..032da97 100644
--- a/SuperTagger/SuperTagger/SymbolTokenizer.py
+++ b/SuperTagger/SuperTagger/SymbolTokenizer.py
@@ -56,3 +56,4 @@ class SymbolTokenizer():
return labels
+
diff --git a/SuperTagger/SuperTagger/eval.py b/SuperTagger/SuperTagger/eval.py
index 1fd1b97..84e2c6c 100644
--- a/SuperTagger/SuperTagger/eval.py
+++ b/SuperTagger/SuperTagger/eval.py
@@ -16,4 +16,3 @@ def categorical_accuracy(preds, truth):
good_label += 1
nb_label += 1
return good_label / nb_label
-
\ No newline at end of file
diff --git a/find_config.py b/find_config.py
index 42d5e19..050ad80 100644
--- a/find_config.py
+++ b/find_config.py
@@ -6,10 +6,38 @@ import torch
from Linker.atom_map import atom_map_redux
from Linker.utils_linker import get_GOAL, get_atoms_links_batch, get_atoms_batch
from SuperTagger.SuperTagger.SuperTagger import SuperTagger
-from utils import read_csv_pgbar, pad_sequence
+from utils import read_links_csv, read_supertags_csv, pad_sequence, load_obj
+def configurate_supertagger(dataset, index_to_super_path, model_tagger, nb_sentences=1000000000):
+ print("#" * 20)
+ print("#" * 20)
+ print("Configuration with dataset\n")
+ config = configparser.ConfigParser()
+ config.read('Configuration/config.ini')
+
+ df = read_supertags_csv(dataset, nb_sentences)
+ index_to_super = load_obj(index_to_super_path)
-def configurate(dataset, model_tagger, nb_sentences=1000000000):
+ supertagger = SuperTagger()
+ supertagger.create_new_model(len(index_to_super),model_tagger,index_to_super)
+
+ sentences_batch = df["X"].str.strip().tolist()
+ sentences_tokens, sentences_mask = supertagger.sent_tokenizer.fit_transform(sentences_batch)
+ max_len_sentence = 0
+ for sentence in sentences_tokens:
+ if len(sentence) > max_len_sentence:
+ max_len_sentence = len(sentence)
+ print("Configure parameter max len sentence to ", max_len_sentence)
+ config.set('DATASET_PARAMS', 'max_len_sentence', str(max_len_sentence))
+
+ with open('Configuration/config.ini', 'w') as configfile: # save
+ config.write(configfile)
+
+ print("#" * 20)
+ print("#" * 20)
+
+
+def configurate_linker(dataset, model_tagger, nb_sentences=1000000000):
print("#" * 20)
print("#" * 20)
print("Configuration with dataset\n")
@@ -17,12 +45,13 @@ def configurate(dataset, model_tagger, nb_sentences=1000000000):
config.read('Configuration/config.ini')
file_path_axiom_links = dataset
- df_axiom_links = read_csv_pgbar(file_path_axiom_links, nb_sentences)
+ df_axiom_links = read_links_csv(file_path_axiom_links, nb_sentences)
supertagger = SuperTagger()
supertagger.load_weights(model_tagger)
+
sentences_batch = df_axiom_links["X"].str.strip().tolist()
- sentences_tokens, sentences_mask = supertagger.sent_tokenizer.fit_transform_tensors(sentences_batch)
+ sentences_tokens, sentences_mask = supertagger.sent_tokenizer.fit_transform(sentences_batch)
max_len_sentence = 0
for sentence in sentences_tokens:
if len(sentence) > max_len_sentence:
@@ -59,3 +88,4 @@ def configurate(dataset, model_tagger, nb_sentences=1000000000):
print("#" * 20)
print("#" * 20)
+
diff --git a/train_supertagger.py b/train_supertagger.py
index 571edc6..67d8d7c 100644
--- a/train_supertagger.py
+++ b/train_supertagger.py
@@ -1,24 +1,30 @@
from SuperTagger.SuperTagger.SuperTagger import SuperTagger
from utils import read_supertags_csv, load_obj
+from find_config import configurate_supertagger
+
import torch
torch.cuda.empty_cache()
+dataset = 'SuperTagger/Datasets/m2_dataset_V2.csv'
+index_to_super_path = 'SuperTagger/Datasets/index_to_super'
+bert_model = "flaubert/flaubert_base_cased"
+
+configurate_supertagger(dataset, index_to_super_path, bert_model, nb_sentences=1000000000)
+
# region data
-file_path = 'SuperTagger/Datasets/m2_dataset_V2.csv'
-df = read_supertags_csv(file_path)
+df = read_supertags_csv(dataset)
texts = df['X'].tolist()
tags = df['Z'].tolist()
-index_to_super = load_obj('SuperTagger/Datasets/index_to_super')
+index_to_super = load_obj(index_to_super_path)
# endregion
-
# region model
tagger = SuperTagger()
-tagger.create_new_model(len(index_to_super),"flaubert/flaubert_base_cased",index_to_super)
+tagger.create_new_model(len(index_to_super),bert_model,index_to_super)
## If you want to upload a pretrained model
# tagger.load_weights("models/model_check.pt")
-tagger.train(texts, tags, epochs=60, batch_size=16, validation_rate=0.1,
+tagger.train(texts, tags, epochs=70, batch_size=16, validation_rate=0.1,
tensorboard=True, checkpoint=True)
# endregion
diff --git a/utils.py b/utils.py
index b1c9f85..95217aa 100644
--- a/utils.py
+++ b/utils.py
@@ -20,13 +20,9 @@ def read_links_csv(csv_path, nrows=float('inf'), chunksize=100):
print("\n" + "#" * 20)
print("Loading csv...")
- rows = sum(1 for _ in open(csv_path, 'r', encoding="utf8")) - 1 # minus the header
chunk_list = []
- if rows > nrows:
- rows = nrows
-
- with tqdm(total=nrows, desc='Rows read: ') as bar:
+ with tqdm(total=rows, desc='Rows read: ') as bar:
for chunk in pd.read_csv(csv_path, header=0, converters={'Y': pd.eval, 'Z': pd.eval},
chunksize=chunksize, nrows=nrows):
chunk_list.append(chunk)
@@ -55,9 +51,9 @@ def read_supertags_csv(csv_path, nrows=float('inf'), chunksize=100):
if rows > nrows:
rows = nrows
- with tqdm(total=nrows, desc='Rows read: ') as bar:
- for chunk in pd.read_csv(csv_path, header=0, converters={'Y1': pd.eval, 'Y2': pd.eval, 'Z': pd.eval},
- chunksize=chunksize, nrows=nrows):
+ with tqdm(total=rows, desc='Rows read: ') as bar:
+ for chunk in pd.read_csv(csv_path, converters={'Y1': pd.eval, 'Y2': pd.eval, 'Z': pd.eval}, chunksize=chunksize,
+ nrows=rows):
chunk_list.append(chunk)
bar.update(len(chunk))
--
GitLab