def predict_on_test_files(self, estimator, csv_files_path):
failed_csvs = []
for csv_file in tqdm(os.listdir(csv_files_path), desc="predicting"):
csv_file = os.path.join(csv_files_path, csv_file)
if csv_file.endswith(".csv"):
sentence = ""
# try:
print_info("processing ====> {}".format(csv_file))
df = pd.read_csv(csv_file).fillna(UNKNOWN_WORD)
df = self.predict_on_test_file(estimator, df)
out_dir = estimator.model_dir + "/predictions/"
check_n_makedirs(out_dir)
df.to_csv(out_dir + ntpath.basename(csv_file), index=False)
# except Exception as e:
# print_error(traceback.print_exc())
# failed_csvs.append(csv_file)
# print_warn("Failed processing ====> {}".format(csv_file))
# pdb.set_trace()
# print_error(failed_csvs)
return out_dir
def predict_on_csv_files(self, estimator,
csv_files_path):
df = ""
sentence = ""
# for csv_file in tqdm(os.listdir(csv_files_path)):
# csv_file = os.path.join(csv_files_path, csv_file)
csv_file = csv_files_path
if csv_file.endswith(".csv"):
print_info(csv_file)
# print_info("processing ====> {}".format(csv_file))
df = pd.read_csv(csv_file).fillna(UNKNOWN_WORD)
# df = io_2_iob(df, entity_col, entity_iob_col) # removing since we are using preprocessed test folder TODO chain IOB
sentence = (" ".join(df[self.preprocessed_data_info.TEXT_COL].values))
char_ids = [[self.preprocessed_data_info.char_2_id_map.get(c, 0) for c in word] for word in sentence.split(" ")]
char_ids, char_ids_length = self._pad_sequences([char_ids], pad_tok=0, nlevels=2)
# TODO add batch support
predicted_tags, confidence, pred_1, pred_1_confidence, pred_2, pred_2_confidence, \
pred_3, pred_3_confidence = self.get_tags(estimator, sentence, char_ids, self.preprocessed_data_info.ENTITY_VOCAB_FILE)
df["predictions"] = predicted_tags
df["confidence"] = confidence
df["pred_1"] = pred_1
df["pred_1_confidence"] = pred_1_confidence
df["pred_2"] = pred_2
df["pred_2_confidence"] = pred_2_confidence
df["pred_3"] = pred_3
df["pred_3_confidence"] = pred_3_confidence
out_dir = estimator.model_dir +"/predictions/"
check_n_makedirs(out_dir)
df.to_csv(out_dir +ntpath.basename(csv_file), index=False)
return df
def __init__(self, name, experiment_dir, batch_size):
'''
Data Iterators with different features type are expected to
implement this interface, exposing the input functions and their hooks
:param experiment_dir:
:param batch_size:
'''
self.NAME = name
self.EXPERIMENT_ROOT_DIR = experiment_dir
self.OUT_DIR = self.EXPERIMENT_ROOT_DIR + "/" + self.NAME + "/"
self._load_ini()
# self.preprocessed_data_info = PreprocessedDataInfo.load(experiment_dir)
# This rule is assumed to be correct if the previous stage is of IPreprocessorInterface
self.PREPROCESSED_DATA_DIR = self.EXPERIMENT_ROOT_DIR + "/" + self.config.get_item(
"OutputDirectories", "preprocessed_data_dir")
self.TRAIN_FILES_IN_PATH = self.PREPROCESSED_DATA_DIR + "/train/"
self.VAL_FILES_IN_PATH = self.PREPROCESSED_DATA_DIR + "/val/"
self.TEST_FILES_IN_PATH = self.PREPROCESSED_DATA_DIR + "/test/"
self.TEXT_COL = self.config.get_item("Schema", "text_column")
self.ENTITY_COL = self.config.get_item("Schema", "entity_column")
self.WORDS_VOCAB_FILE = self.OUT_DIR + "/" + self.TEXT_COL + "_" + "vocab.tsv"
self.CHARS_VOCAB_FILE = self.OUT_DIR + "/" + self.TEXT_COL + "_" + "chars_vocab.tsv"
self.ENTITY_VOCAB_FILE = self.OUT_DIR + "/" + self.ENTITY_COL + "_vocab.tsv"
check_n_makedirs(self.OUT_DIR)
self.BATCH_SIZE = batch_size
self.NUM_TAGS = None
self.VOCAB_SIZE = None
self.CHAR_VOCAB_SIZE = None
self._train_data_input_fn = None
self._train_data_init_hook = None
self._val_data_input_fn = None
self._val_data_init_hook = None
self._test_data_input_fn = None
self._test_data_init_hook = None
def predict_on_test_files(self, estimator, csv_files_path):
failed_csvs = []
out_dir = estimator.model_dir + "/predictions/"
check_n_makedirs(out_dir)
files = [
file for file in os.listdir(csv_files_path)
if file.endswith('.csv')
]
batchsize = 12
index = 0
remaining = len(files)
progress_bar = tqdm(total=len(files))
while remaining > 0:
batch = min(remaining, batchsize)
print('NEW BATCH\n')
dfs = []
for csv_file in files[index:index + batch]:
df = pd.read_csv(os.path.join(csv_files_path,
csv_file)).fillna(UNKNOWN_WORD)
df.file_name = csv_file
dfs.append(df)
dfs = self.predict_on_dataframes(estimator, dfs)
for predicted_df in dfs:
print_info(predicted_df.file_name)
predicted_df.to_csv(out_dir +
ntpath.basename(predicted_df.file_name),
index=False)
index += batch
remaining -= batch
progress_bar.update(index)
progress_bar.close()
return out_dir
def create_target_directories(self):
if os.path.exists(self.OUT_DIR):
if self.OVER_WRITE == "yes":
print_info("Deletingls data folder: {}".format(self.OUT_DIR))
shutil.rmtree(self.OUT_DIR)
print_info("Recreating data folder: {}".format(self.OUT_DIR))
os.makedirs(self.OUT_DIR)
check_n_makedirs(self.TRAIN_CSV_INTERMEDIATE_PATH)
check_n_makedirs(self.VAL_CSV_INTERMEDIATE_PATH)
check_n_makedirs(self.TEST_CSV_INTERMEDIATE_PATH)
else:
print_info(
"Skipping preprocessing step, since the data is already available"
)
return "skip"
else:
print_info("Creating data folder: {}".format(self.OUT_DIR))
os.makedirs(self.OUT_DIR)
check_n_makedirs(self.TRAIN_CSV_INTERMEDIATE_PATH)
check_n_makedirs(self.VAL_CSV_INTERMEDIATE_PATH)
check_n_makedirs(self.TEST_CSV_INTERMEDIATE_PATH)
def with_user_hyperparamaters(experiment_root_dir, data_dir):
# model_root_dir,
# vocab_size,
# char_vocab_size,
# number_tags,
# unknown_word,
# pad_word,
# tags_vocab_file,
# words_vocab_file,
# chars_vocab_file):
preprocessed_data_info = PreprocessedDataInfo.load(data_dir)
use_crf = "y" # TODO
use_char_embedding = False
char_level_lstm_hidden_size = 128 # default
char_emd_size = 128 # default
if use_crf == 'y':
use_crf = True
else:
use_crf = False
use_char_embedding_option = input("use_char_embedding (y/n)") or "y"
learning_rate = float(input("learning_rate (0.001): ")) or 0.001
num_lstm_layers = int(input("num_word_lstm_layers (2): ")) or 2
if use_char_embedding_option == 'y':
use_char_embedding = True
char_level_lstm_hidden_size = int(
input("char_level_lstm_hidden_size (32): ")) or 32
char_emd_size = int(input("char_emd_size (32): ")) or 32
else:
use_char_embedding = False
word_level_lstm_hidden_size = int(
input("word_level_lstm_hidden_size (32): ")) or 32
word_emd_size = int(input("word_emd_size (128): ")) or 32
out_keep_propability = float(
input("out_keep_propability(0.5) : ")) or 0.5
# Does this sound logical? review please
model_dir = experiment_root_dir + "/bilstm_crf_v1/" + \
"charembd_{}_lr_{}_lstmsize_{}-{}-{}_wemb_{}_cemb_{}_outprob_{}".format(
str(use_char_embedding),
learning_rate,
num_lstm_layers,
word_level_lstm_hidden_size,
char_level_lstm_hidden_size,
word_emd_size,
char_emd_size,
out_keep_propability)
model_config = BiLSTMCRFConfigV1(
model_dir=model_dir,
vocab_size=preprocessed_data_info.VOCAB_SIZE,
char_vocab_size=preprocessed_data_info.CHAR_VOCAB_SIZE,
number_tags=preprocessed_data_info.NUM_TAGS,
unknown_word=UNKNOWN_WORD,
pad_word=PAD_WORD,
tags_vocab_file=preprocessed_data_info.ENTITY_VOCAB_FILE,
words_vocab_file=preprocessed_data_info.WORDS_VOCAB_FILE,
chars_vocab_file=preprocessed_data_info.WORDS_VOCAB_FILE,
# hyper parameters
use_char_embedding=use_char_embedding,
learning_rate=learning_rate,
word_level_lstm_hidden_size=word_level_lstm_hidden_size,
char_level_lstm_hidden_size=char_level_lstm_hidden_size,
word_emd_size=word_emd_size,
char_emd_size=char_emd_size,
num_lstm_layers=num_lstm_layers,
out_keep_propability=out_keep_propability,
use_crf=True)
check_n_makedirs(model_dir)
IModelConfig.save(model_dir=model_dir, config=model_config)
return model_config
def with_user_hyperparamaters(experiment_root_dir, data_iterator):
use_crf = "y" # TODO
use_char_embedding = False
char_level_lstm_hidden_size = 32 # default
char_emd_size = 32 # default
if use_crf == 'y':
use_crf = True
else:
use_crf = False
use_char_embedding_option = input("use_char_embedding (y/n): ") or "y"
learning_rate = input("learning_rate (0.001): ") or 0.001
learning_rate = float(learning_rate)
num_lstm_layers = input("num_word_lstm_layers (2): ") or 2
num_lstm_layers = int(num_lstm_layers)
if use_char_embedding_option == 'y':
use_char_embedding = True
char_level_lstm_hidden_size = input(
"char_level_lstm_hidden_size (48): ") or 32
char_level_lstm_hidden_size = int(char_level_lstm_hidden_size)
char_emd_size = input("char_emd_size (32): ") or 32
char_emd_size = int(char_emd_size)
else:
use_char_embedding = False
word_level_lstm_hidden_size = input(
"word_level_lstm_hidden_size (48): ") or 48
word_level_lstm_hidden_size = int(word_level_lstm_hidden_size)
word_emd_size = input("word_emd_size (48): ") or 48
word_emd_size = int(word_emd_size)
out_keep_propability = input("out_keep_propability(0.5) : ") or 0.5
out_keep_propability = float(out_keep_propability)
# Does this sound logical? review please
'''
experiment_root_dir/
- data_iterator/
- model_name/
- user_hyper_params/
'''
model_dir = experiment_root_dir + "/" + data_iterator.NAME + "/bilstm_crf_v0/" + \
"charembd_{}_lr_{}_lstmsize_{}-{}-{}_wemb_{}_cemb_{}_outprob_{}".format(
str(use_char_embedding),
learning_rate,
num_lstm_layers,
word_level_lstm_hidden_size,
char_level_lstm_hidden_size,
word_emd_size,
char_emd_size,
out_keep_propability)
model_config = BiLSTMCRFConfigV0(
model_dir=model_dir,
vocab_size=data_iterator.VOCAB_SIZE,
char_vocab_size=data_iterator.CHAR_VOCAB_SIZE,
number_tags=data_iterator.NUM_TAGS,
unknown_word=UNKNOWN_WORD,
pad_word=PAD_WORD,
tags_vocab_file=data_iterator.ENTITY_VOCAB_FILE,
words_vocab_file=data_iterator.WORDS_VOCAB_FILE,
chars_vocab_file=data_iterator.CHAR_VOCAB_SIZE,
# hyper parameters
use_char_embedding=use_char_embedding,
learning_rate=learning_rate,
word_level_lstm_hidden_size=word_level_lstm_hidden_size,
char_level_lstm_hidden_size=char_level_lstm_hidden_size,
word_emd_size=word_emd_size,
char_emd_size=char_emd_size,
num_lstm_layers=num_lstm_layers,
out_keep_propability=out_keep_propability,
use_crf=True)
check_n_makedirs(model_dir)
IModelConfig.save(model_dir=model_dir, config=model_config)
return model_config
