def get_run_components(run_dir):
# Load args
config = utils.load_json(os.path.join(run_dir, 'config.json'))
args = Namespace(**config)
# Load tokenizers
X_tokenizer = data.Tokenizer.load(
fp=os.path.join(run_dir, 'X_tokenizer.json'))
y_tokenizer = data.LabelEncoder.load(
fp=os.path.join(run_dir, 'y_tokenizer.json'))
# Load model
model = models.TextCNN(embedding_dim=args.embedding_dim,
vocab_size=len(X_tokenizer) + 1,
num_filters=args.num_filters,
filter_sizes=args.filter_sizes,
hidden_dim=args.hidden_dim,
dropout_p=args.dropout_p,
num_classes=len(y_tokenizer.classes))
model.load_state_dict(torch.load(os.path.join(run_dir, 'model.pt')))
device = torch.device('cuda' if (
torch.cuda.is_available() and args.cuda) else 'cpu')
model = model.to(device)
return args, model, X_tokenizer, y_tokenizer
def predict(experiment_id, text):
"""Predict the class for a text using
a trained model from an experiment."""
# Get experiment config
experiment_dir = os.path.join(config.EXPERIMENTS_DIR, experiment_id)
experiment_config = utilities.load_json(
os.path.join(experiment_dir, 'config.json'))
args = Namespace(**experiment_config)
# Preprocess
texts = [text]
X_tokenizer = data.Tokenizer.load(
fp=os.path.join(experiment_dir, 'X_tokenizer.json'))
y_tokenizer = data.LabelEncoder.load(
fp=os.path.join(experiment_dir, 'y_tokenizer.json'))
preprocessed_texts = data.preprocess_texts(
texts, lower=args.lower, filters=args.filters)
# Create dataset
X_infer = np.array(X_tokenizer.texts_to_sequences(preprocessed_texts))
y_filler = np.array([0]*len(X_infer))
infer_set = data.TextDataset(
X=X_infer, y=y_filler, batch_size=args.batch_size,
max_filter_size=max(args.filter_sizes))
# Load model
model = models.TextCNN(
embedding_dim=args.embedding_dim, vocab_size=len(X_tokenizer),
num_filters=args.num_filters, filter_sizes=args.filter_sizes,
hidden_dim=args.hidden_dim, dropout_p=args.dropout_p,
num_classes=len(y_tokenizer.classes))
model.load_state_dict(torch.load(os.path.join(experiment_dir, 'model.h5')))
device = torch.device('cuda' if (
torch.cuda.is_available() and args.cuda) else 'cpu')
model = model.to(device)
# Predict
results = []
y_prob, conv_outputs = predict_step(
model=model, dataset=infer_set, filter_sizes=args.filter_sizes, device=device)
for index in range(len(X_infer)):
results.append({
'raw_input': texts[index],
'preprocessed_input': X_tokenizer.sequences_to_texts([X_infer[index]])[0],
'probabilities': get_probability_distribution(y_prob[index], y_tokenizer.classes),
'top_n_grams': get_top_n_grams(tokens=preprocessed_texts[index].split(' '),
conv_outputs={
k: v[index] for k, v in conv_outputs.items()},
filter_sizes=args.filter_sizes)})
return results
def get_run_components(run_dir):
# Load args
config = utils.load_json(os.path.join(run_dir, 'config.json'))
args = Namespace(**config)
# Load tokenizers
with open(os.path.join(run_dir, 'X_tokenizer.json'), 'r') as fp:
X_tokenizer = tokenizer_from_json(json.load(fp))
y_tokenizer = LabelEncoder()
y_tokenizer.classes_ = np.load(os.path.join(run_dir,
'y_tokenizer.npy'),
allow_pickle=True)
# Load model
model = models.TextCNN(embedding_dim=args.embedding_dim,
vocab_size=len(X_tokenizer.word_index) + 1,
num_filters=args.num_filters,
filter_sizes=args.filter_sizes,
hidden_dim=args.hidden_dim,
dropout_p=args.dropout_p,
num_classes=len(y_tokenizer.classes_))
model.summary(input_shape=(10, )) # build it
model_path = os.path.join(run_dir, 'model/cp.ckpt')
model.load_weights(model_path)
# Conv output model
conv_outputs_model = models.ConvOutputsModel(
vocab_size=len(X_tokenizer.word_index) + 1,
embedding_dim=args.embedding_dim,
filter_sizes=args.filter_sizes,
num_filters=args.num_filters)
conv_outputs_model.summary(input_shape=(10, )) # build it
# Set weights
conv_outputs_model.layers[0].set_weights(model.layers[0].get_weights())
conv_layer_start_num = 1
for layer_num in range(conv_layer_start_num,
conv_layer_start_num + len(args.filter_sizes)):
conv_outputs_model.layers[layer_num].set_weights(
model.layers[layer_num].get_weights())
return args, model, conv_outputs_model, X_tokenizer, y_tokenizer
y_tokenizer = data.LabelEncoder.load(
fp=os.path.join(experiment_dir, 'y_tokenizer.json'))
preprocessed_texts = data.preprocess_texts(
texts, lower=args.lower, filters=args.filters)
# Create dataset
X_infer = np.array(X_tokenizer.texts_to_sequences(preprocessed_texts))
y_filler = np.array([0]*len(X_infer))
infer_set = data.TextDataset(
X=X_infer, y=y_filler, batch_size=args.batch_size,
max_filter_size=max(args.filter_sizes))
# Load model
model = models.TextCNN(
embedding_dim=args.embedding_dim, vocab_size=len(X_tokenizer),
num_filters=args.num_filters, filter_sizes=args.filter_sizes,
hidden_dim=args.hidden_dim, dropout_p=args.dropout_p,
num_classes=len(y_tokenizer.classes))
model.load_state_dict(torch.load(os.path.join(experiment_dir, 'model.h5')))
device = torch.device('cuda' if (
torch.cuda.is_available() and args.cuda) else 'cpu')
model = model.to(device)
# Predict
results = []
y_prob, conv_outputs = predict_step(
model=model, dataset=infer_set, filter_sizes=args.filter_sizes, device=device)
for index in range(len(X_infer)):
results.append({
'raw_input': texts[index],
'preprocessed_input': X_tokenizer.sequences_to_texts([X_infer[index]])[0],
embeddings_file = os.path.join(config.EMBEDDINGS_DIR,
f'glove.6B.{args.embedding_dim}d.txt')
glove_embeddings = utils.load_glove_embeddings(
embeddings_file=embeddings_file)
embedding_matrix = utils.make_embeddings_matrix(
embeddings=glove_embeddings,
token_to_index=X_tokenizer.token_to_index,
embedding_dim=args.embedding_dim)
config.logger.info("→ GloVe Embeddings:\n" f"{embedding_matrix.shape}")
# Initialize model
model = models.TextCNN(embedding_dim=args.embedding_dim,
vocab_size=len(X_tokenizer),
num_filters=args.num_filters,
filter_sizes=args.filter_sizes,
hidden_dim=args.hidden_dim,
dropout_p=args.dropout_p,
num_classes=len(y_tokenizer.classes),
pretrained_embeddings=embedding_matrix,
freeze_embeddings=args.freeze_embeddings)
model = model.to(device)
config.logger.info("→ Model:\n" f" {model.named_parameters}")
# Define optimizer & scheduler
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
mode='min',
factor=0.1,
patience=3)
# Model dir
"Embedding dim must be in (50, 100, 200, 300) is using GloVe.")
embeddings_file = os.path.join(config.EMBEDDINGS_DIR,
f'glove.6B.{args.embedding_dim}d.txt')
glove_embeddings = utils.load_glove_embeddings(
embeddings_file=embeddings_file)
embedding_matrix = utils.make_embeddings_matrix(
embeddings=glove_embeddings,
token_to_index=X_tokenizer.word_index,
embedding_dim=args.embedding_dim)
config.logger.info("→ Embeddings:\n" f"{embedding_matrix.shape}")
# Initialize model
model = models.TextCNN(vocab_size=vocab_size,
embedding_dim=args.embedding_dim,
filter_sizes=args.filter_sizes,
num_filters=args.num_filters,
hidden_dim=args.hidden_dim,
dropout_p=args.dropout_p,
num_classes=len(y_tokenizer.classes_),
freeze_embeddings=args.freeze_embeddings)
model.summary(input_shape=(10, )) # build it
# Set GloVe embeddings
if args.use_glove:
model.layers[0].set_weights([embedding_matrix])
# Model dir
experiment_id = f'TextCNN_{datetime.now().strftime("%Y-%m-%d-%H:%M:%S")}'
experiment_dir = os.path.join(config.EXPERIMENTS_DIR, experiment_id)
utilities.create_dirs(dirpath=experiment_dir)
model_path = os.path.join(experiment_dir, 'model/cp.ckpt')
def predict(experiment_id, text):
"""Predict the class for a text using
a trained model from an experiment."""
# Get experiment config
experiment_dir = os.path.join(config.EXPERIMENTS_DIR, experiment_id)
experiment_config = utilities.load_json(
os.path.join(experiment_dir, 'config.json'))
args = Namespace(**experiment_config)
# Tokenizers
texts = [text]
with open(os.path.join(experiment_dir, 'X_tokenizer.json'), 'r') as fp:
X_tokenizer = tokenizer_from_json(json.load(fp))
y_tokenizer = LabelEncoder()
y_tokenizer.classes_ = np.load(os.path.join(experiment_dir,
'y_tokenizer.npy'),
allow_pickle=True)
# Create dataset generator
X_infer = np.array(X_tokenizer.texts_to_sequences(texts))
preprocessed_texts = X_tokenizer.sequences_to_texts(X_infer),
y_filler = np.array([0] * len(X_infer))
inference_generator = data.DataGenerator(X=X_infer,
y=y_filler,
batch_size=args.batch_size,
max_filter_size=max(
args.filter_sizes))
# Load model
model = models.TextCNN(embedding_dim=args.embedding_dim,
vocab_size=len(X_tokenizer.word_index) + 1,
num_filters=args.num_filters,
filter_sizes=args.filter_sizes,
hidden_dim=args.hidden_dim,
dropout_p=args.dropout_p,
num_classes=len(y_tokenizer.classes_))
model.summary(input_shape=(10, )) # build it
model_path = os.path.join(experiment_dir, 'model/cp.ckpt')
model.load_weights(model_path)
# Conv output model
conv_outputs_model = models.ConvOutputsModel(
vocab_size=len(X_tokenizer.word_index) + 1,
embedding_dim=args.embedding_dim,
filter_sizes=args.filter_sizes,
num_filters=args.num_filters)
conv_outputs_model.summary(input_shape=(10, )) # build it
# Set weights
conv_outputs_model.layers[0].set_weights(model.layers[0].get_weights())
conv_layer_start_num = 1
for layer_num in range(conv_layer_start_num,
conv_layer_start_num + len(args.filter_sizes)):
conv_outputs_model.layers[layer_num].set_weights(
model.layers[layer_num].get_weights())
# Predict
results = []
y_prob = model.predict(x=inference_generator, verbose=1)
conv_outputs = conv_outputs_model.predict(x=inference_generator, verbose=1)
for index in range(len(X_infer)):
results.append({
'raw_input':
texts[index],
'preprocessed_input':
preprocessed_texts[index][0],
'probabilities':
get_probability_distribution(y_prob[index], y_tokenizer.classes_),
'top_n_grams':
get_top_n_grams(tokens=preprocessed_texts[index][0].split(' '),
conv_outputs=conv_outputs,
filter_sizes=args.filter_sizes)
})
return results