diff --git a/Linker/Linker.py b/Linker/Linker.py
index 396639195b6381ad699457dbfe78899af4777e9c..c986c0e1bf3b96cf479a864d4e952a6cf47fa6af 100644
--- a/Linker/Linker.py
+++ b/Linker/Linker.py
@@ -9,6 +9,7 @@ import sys
from torch.optim import AdamW
from torch.utils.data import TensorDataset, random_split
+from torch.utils.tensorboard import SummaryWriter
from transformers import get_cosine_schedule_with_warmup
from Configuration import Configuration
@@ -23,6 +24,19 @@ from Linker.eval import mesure_accuracy, SinkhornLoss
from utils import pad_sequence
+def output_create_dir():
+ """
+ Create le output dir for tensorboard and checkpoint
+ @return: output dir, tensorboard writter
+ """
+ from datetime import datetime
+ outpout_path = 'TensorBoard'
+ training_dir = os.path.join(outpout_path, 'Tranning_' + datetime.today().strftime('%d-%m_%H-%M'))
+ logs_dir = os.path.join(training_dir, 'logs')
+ writer = SummaryWriter(log_dir=logs_dir)
+ return training_dir, writer
+
+
class Linker(Module):
def __init__(self, supertagger):
super(Linker, self).__init__()
@@ -73,18 +87,15 @@ class Linker(Module):
Args:
batch_size : int
df_axiom_links pandas DataFrame
- sentences_tokens
- sentences_mask
validation_rate
Returns:
the training dataloader and the validation dataloader. They contains the list of atoms, their polarities, the axiom links, the sentences tokenized, sentence mask
"""
sentences_batch = df_axiom_links["Sentences"].tolist()
- sentences_tokens, sentences_mask = self.supertagger.sent_tokenizer.fit_transform_tensors(sentences_batch)
+ sentences_tokens, sentences_mask = self.Supertagger.sent_tokenizer.fit_transform_tensors(sentences_batch)
atoms_batch = get_atoms_batch(df_axiom_links["sub_tree"])
- atom_tokenizer = AtomTokenizer(atom_map, self.max_atoms_in_sentence)
- atoms_batch_tokenized = atom_tokenizer.convert_batchs_to_ids(atoms_batch)
+ atoms_batch_tokenized = self.atoms_tokenizer.convert_batchs_to_ids(atoms_batch)
atoms_polarity_batch = find_pos_neg_idexes(self.max_atoms_in_sentence, df_axiom_links["sub_tree"])
@@ -158,27 +169,55 @@ class Linker(Module):
return F.log_softmax(link_weights_per_batch, dim=3)
def train_linker(self, df_axiom_links, validation_rate=0.1, epochs=20,
- batch_size=32, checkpoint=True, validate=True):
+ batch_size=32, checkpoint=True, tensorboard=False):
r"""
Args:
df_axiom_links : pandas dataFrame containing the atoms anoted with _i
- sentences_tokens : sentences tokenized by BERT
- sentences_mask : mask of tokens
validation_rate : float
epochs : int
batch_size : int
checkpoint : boolean
- validate : boolean
+ tensorboard : boolean
Returns:
Final accuracy and final loss
"""
training_dataloader, validation_dataloader = self.__preprocess_data(batch_size, df_axiom_links,
validation_rate)
self.to(self.device)
- for epoch_i in range(0, epochs):
- epoch_acc, epoch_loss = self.train_epoch(training_dataloader, validation_dataloader, checkpoint, validate)
- def train_epoch(self, training_dataloader, validation_dataloader, checkpoint=True, validate=True):
+ if checkpoint or tensorboard:
+ checkpoint_dir, writer = output_create_dir()
+
+ for epoch_i in range(epochs):
+ epoch_acc, epoch_loss = self.train_epoch(training_dataloader)
+
+ print("Average Loss on train dataset : ", epoch_loss)
+ print("Average Accuracy on train dataset : ", epoch_acc)
+
+ if checkpoint:
+ self.__checkpoint_save(
+ path=os.path.join("Output", 'linker' + datetime.today().strftime('%d-%m_%H-%M') + '.pt'))
+
+ if validation_rate > 0.0:
+ with torch.no_grad():
+ accuracy_test, average_test_loss = self.eval_epoch(validation_dataloader, self.cross_entropy_loss)
+ print("Average Loss on test dataset : ", average_test_loss)
+ print("Average Accuracy on test dataset : ", accuracy_test)
+
+ if tensorboard:
+ writer.add_scalars(f'Accuracy', {
+ 'Train': epoch_acc}, epoch_i)
+ writer.add_scalars(f'Loss', {
+ 'Train': epoch_loss}, epoch_i)
+ if validation_rate > 0.0:
+ writer.add_scalars(f'Accuracy', {
+ 'Validation': accuracy_test}, epoch_i)
+ writer.add_scalars(f'Loss', {
+ 'Validation': average_test_loss}, epoch_i)
+
+ print('\n')
+
+ def train_epoch(self, training_dataloader):
r""" Train epoch
Args:
@@ -191,6 +230,7 @@ class Linker(Module):
# Reset the total loss for this epoch.
epoch_loss = 0
+ accuracy_train = 0
self.train()
@@ -223,22 +263,13 @@ class Linker(Module):
self.optimizer.step()
self.scheduler.step()
- avg_train_loss = epoch_loss / len(training_dataloader)
- print("Average Loss on train dataset : ", avg_train_loss)
-
- if checkpoint:
- self.__checkpoint_save(
- path=os.path.join("Output", 'linker' + datetime.today().strftime('%d-%m_%H-%M') + '.pt'))
-
- if validate:
- with torch.no_grad():
- accuracy, average_test_loss = self.eval_epoch(validation_dataloader, self.cross_entropy_loss)
- print("Average Loss on test dataset : ", average_test_loss)
- print("Average Accuracy on test dataset : ", accuracy)
+ pred_axiom_links = torch.argmax(logits_predictions, dim=3)
+ accuracy_train += mesure_accuracy(batch_true_links, pred_axiom_links)
- print('\n')
+ avg_train_loss = epoch_loss / len(training_dataloader)
+ avg_accuracy_train = accuracy_train / len(training_dataloader)
- return accuracy, avg_train_loss
+ return avg_train_loss, avg_accuracy_train
def predict(self, categories, sents_embedding, sents_mask=None):
r"""Prediction from categories output by BERT and hidden_state from BERT
@@ -315,14 +346,12 @@ class Linker(Module):
"""
accuracy_average = 0
loss_average = 0
- compt = 0
for step, batch in enumerate(dataloader):
- compt += 1
accuracy, loss = self.eval_batch(batch, cross_entropy_loss)
accuracy_average += accuracy
loss_average += loss
- return accuracy_average / compt, loss_average / compt
+ return accuracy_average / len(dataloader), loss_average / len(dataloader)
def load_weights(self, model_file):
print("#" * 15)
diff --git a/train.py b/train.py
index 4684d3b7fb816ceb1b04b37aaa40add606db6871..571777ac750eef16c30044cddd4a203abf57e522 100644
--- a/train.py
+++ b/train.py
@@ -21,4 +21,4 @@ print("Linker")
linker = Linker(supertagger)
linker = linker.to(torch.device("cuda" if torch.cuda.is_available() else "cpu"))
print("Linker Training")
-linker.train_linker(df_axiom_links, validation_rate=0.1, epochs=epochs, batch_size=batch_size, checkpoint=True, validate=True)
+linker.train_linker(df_axiom_links, validation_rate=0.1, epochs=epochs, batch_size=batch_size, checkpoint=True)