def _perform_training(self,
data_csv_path,
model,
num_epochs,
trained_epochs=0,
plot_metrics=False):
train_df, validation_df, _ = TrainingUtils.split_train_val_test(
data_csv_path)
dataloader = retrieve_train_dataloader(train_df,
self.vocab,
batch_size=self.batch_size)
loss_func = nn.CrossEntropyLoss(ignore_index=self.vocab.stoi["<PAD>"])
optimizer = optim.Adam(model.parameters(), lr=3e-4)
vocab_size = len(self.vocab)
train_losses, val_losses = [], []
train_levenshteins, val_levenshteins = [], []
for epoch in tqdm(range(trained_epochs + 1,
trained_epochs + num_epochs + 1),
position=0,
leave=True):
print("\n Epoch: ", epoch)
for image, captions in tqdm(dataloader):
image, captions = image.to(device), captions.to(device)
optimizer.zero_grad()
outputs = model(image, 'train', captions)
targets = captions[:, 1:] # Remove <SOS> token
loss = loss_func(outputs.view(-1, vocab_size),
targets.reshape(-1))
loss.backward()
optimizer.step()
model.eval()
print(f'Training set evaluation:')
train_loss, train_levenshtein = TrainingUtils.evaluate_model_on_dataset(
model, train_df, self.sequence_length, self.batch_size,
self.vocab, 'eval')
train_losses.append(train_loss)
train_levenshteins.append(train_levenshtein)
print(f'Validation set evaluation:')
val_loss, val_levenshtein = TrainingUtils.evaluate_model_on_dataset(
model, validation_df, self.sequence_length, self.batch_size,
self.vocab, 'eval')
val_losses.append(val_loss)
val_levenshteins.append(val_levenshtein)
model.train()
if plot_metrics and train_losses and train_levenshteins and val_losses and val_levenshteins:
TrainingUtils.plot_metrics(train_losses, train_levenshteins,
val_losses, val_levenshteins)
self._save_model(model, epoch)
def _perform_training(self,
data_csv_path,
model,
num_epochs,
trained_epochs=0,
plot_metrics=False):
train_df, validation_df, _ = TrainingUtils.split_train_val_test(
data_csv_path)
dataloader = retrieve_train_dataloader(train_df,
self.vocab,
batch_size=self.batch_size)
loss_func = nn.CrossEntropyLoss(ignore_index=self.vocab.stoi["<PAD>"])
optimizer = optim.Adam(model.parameters(), self.LEARNING_RATE)
train_losses, val_losses = [], []
train_levenshteins, val_levenshteins = [], []
for epoch in tqdm(range(trained_epochs + 1,
trained_epochs + num_epochs + 1),
position=0,
leave=True):
print("\nEpoch: ", epoch)
for images, captions in tqdm(dataloader):
images, captions = images.to(device), captions.to(device)
optimizer.zero_grad()
targets = captions[:, 1:]
preds = model(images, 'train', captions)
loss = loss_func(preds.reshape(-1, self.vocab_size),
targets.reshape(-1))
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1)
optimizer.step()
model.eval()
print(f'Training set evaluation:')
train_loss, train_levenshtein = TrainingUtils.evaluate_model_on_dataset(
model, train_df, None, self.batch_size, self.vocab, 'train')
train_losses.append(train_loss)
train_levenshteins.append(train_levenshtein)
print(f'Validation set evaluation:')
val_loss, val_levenshtein = TrainingUtils.evaluate_model_on_dataset(
model, validation_df, None, self.batch_size, self.vocab,
'train')
val_losses.append(val_loss)
val_levenshteins.append(val_levenshtein)
model.train()
if plot_metrics and train_losses and train_levenshteins and val_losses and val_levenshteins:
TrainingUtils.plot_metrics(train_losses, train_levenshteins,
val_losses, val_levenshteins)
self._save_model(model, epoch)
def evaluate(self, data_csv_path: str, model_state_file):
torch.cuda.empty_cache()
saved_params = torch.load(model_state_file)
model = EncoderDecoderVisualTransformer(
sequence_length=self.sequence_length, vocab=self.vocab).to(device)
model.load_state_dict(saved_params['state_dict'])
model.eval()
dataframe = pd.read_csv(data_csv_path)
TrainingUtils.evaluate_model_levenshtein(model, dataframe,
self.sequence_length,
self.batch_size, self.vocab)
def predict(self, data_csv_path: str, model_state_file):
torch.cuda.empty_cache()
saved_params = torch.load(model_state_file)
model = EncoderDecoderVisualTransformer(
sequence_length=self.sequence_length, vocab=self.vocab).to(device)
model.load_state_dict(saved_params['state_dict'])
model.eval()
dataframe = pd.read_csv(data_csv_path)
results_file_name = data_csv_path.replace(".csv", "_results.csv")
TrainingUtils.predict_on_dataset(model, dataframe, self.batch_size,
self.vocab, results_file_name)
def evaluate(self, data_csv_path: str, model_state_file):
torch.cuda.empty_cache()
saved_params = torch.load(model_state_file)
self.embed_size = saved_params['embed_size']
self.attention_dim = saved_params['attention_dim']
self.encoder_dim = saved_params['encoder_dim']
self.decoder_dim = saved_params['decoder_dim']
model = EncoderDecoder(embed_size=self.embed_size,
vocab_size=len(self.vocab),
attention_dim=self.attention_dim,
encoder_dim=self.encoder_dim,
decoder_dim=self.decoder_dim,
sequence_length=self.sequence_length,
vocab=self.vocab).to(device)
model.load_state_dict(saved_params['state_dict'])
model.eval()
dataframe = pd.read_csv(data_csv_path)
TrainingUtils.evaluate_model_levenshtein(model, dataframe,
self.sequence_length,
self.batch_size, self.vocab)
def predict(self, data_csv_path: str, model_state_file):
torch.cuda.empty_cache()
saved_params = torch.load(model_state_file)
self.embed_size = saved_params['embed_size']
self.attention_dim = saved_params['attention_dim']
self.encoder_dim = saved_params['encoder_dim']
self.decoder_dim = saved_params['decoder_dim']
model = EncoderDecoder(embed_size=self.embed_size,
vocab_size=len(self.vocab),
attention_dim=self.attention_dim,
encoder_dim=self.encoder_dim,
decoder_dim=self.decoder_dim,
sequence_length=self.sequence_length,
vocab=self.vocab).to(device)
model.load_state_dict(saved_params['state_dict'])
model.eval()
dataframe = pd.read_csv(data_csv_path)
results_file_name = data_csv_path.replace(".csv", "_results.csv")
TrainingUtils.predict_on_dataset(model, dataframe, self.batch_size,
self.vocab, results_file_name)