def __init__(self,
experiment_root_directory,
experiment_name,
number_test_of_samples,
batch_size=32,
prefetch_size=32,
dataset=None,
iterator_name="Cifar10BasicIterator",
preprocessed_data_path="preprocessed_data",
train_data_path="train",
validation_data_path="val",
test_data_path="test"):
ImageFeature.__init__(self)
IPreprocessor.__init__(
self,
experiment_name=experiment_name,
preprocessed_data_path=preprocessed_data_path,
experiment_root_directory=experiment_root_directory,
train_data_path=train_data_path,
validation_data_path=validation_data_path,
test_data_path=test_data_path)
IIteratorBase.__init__(
self,
experiment_root_directory=experiment_root_directory,
experiment_name=experiment_name,
batch_size=batch_size,
prefetch_size=prefetch_size,
dataset=dataset)
self._experiment_root_directory = experiment_root_directory
self._experiment_name = experiment_name
self._batch_size = batch_size
self._prefetch_size = prefetch_size
self._dataset = dataset
self._iterator_name = iterator_name
self._preprocessed_data_path = preprocessed_data_path
self._train_data_path = train_data_path
self._validation_data_path = validation_data_path
self._test_data_path = test_data_path
# TODO - make `EXPERIMENT_ROOT_DIR` as local variable
self.EXPERIMENT_ROOT_DIR = os.path.join(
self._experiment_root_directory, self._experiment_name)
self.PREPROCESSED_DATA_OUT_DIR = os.path.join(
self.EXPERIMENT_ROOT_DIR, self._preprocessed_data_path)
self.TRAIN_OUT_PATH = os.path.join(self.PREPROCESSED_DATA_OUT_DIR,
self._train_data_path)
self.TEST_OUT_PATH = os.path.join(self.PREPROCESSED_DATA_OUT_DIR,
self._test_data_path)
self.OUT_DIR = os.path.join(self.EXPERIMENT_ROOT_DIR,
self._iterator_name)
# This rule is assumed to be correct if the previous stage is of IPreprocessor
self.TRAIN_FILES_IN_PATH = os.path.join(self.PREPROCESSED_DATA_OUT_DIR,
"train/")
self.TEST_FILES_IN_PATH = os.path.join(self.PREPROCESSED_DATA_OUT_DIR,
"test/")
check_n_makedirs(self.OUT_DIR)
# Width and height of each image.
self._img_size = 32
# Number of channels in each image, 3 channels: Red, Green, Blue.
self._num_channels = 3
# Length of an image when flattened to a 1-dim array.
self._img_size_flat = self._img_size * self._img_size * self._num_channels
# Number of classes.
self._num_classes = 10
# Number of files for the training-set.
self._num_files_train = 5
# Number of images for each batch-file in the training-set.
self._images_per_file = 10000
# Total number of images in the training-set.
# This is used to pre-allocate arrays for efficiency.
self._num_images_train = self._num_files_train * self._images_per_file
self.images, self.labels = self._load_training_data()
self.images = self.images.astype("float32")
self._number_test_of_samples = number_test_of_samples
def predict_on_test_files(self, predict_fn):
'''
Iterate through the files and use `predict_on_test_file`, for prediction
:param predict_fn:
:return: Creates a folder estimator.model_dir/predictions/ and adds the predicted files
'''
predictions = []
for predict in predict_fn:
predictions.append(predict)
results = []
# Get the files from test folder and zip it with predictions
for each_prediction, file in zip(predictions, self._test_files_path):
df = pd.read_csv(file, sep=self._in_seperator, quoting=csv.QUOTE_NONE)
df.dropna(axis=0, how='any', thresh=None, subset=None, inplace=True)
predicted_id = []
confidence = []
for tag_score in each_prediction["confidence"]:
confidence.append(tag_score)
for tag_id in each_prediction["viterbi_seq"]:
predicted_id.append(self.TAGS_2_ID[tag_id])
top_3_predicted_indices = each_prediction["top_3_indices"]
top_3_predicted_confidence = each_prediction["top_3_confidence"]
# print(top_3_predicted_indices)
pred_1 = top_3_predicted_indices[:, 0:1].flatten()
pred_1 = list(map(lambda x: self.TAGS_2_ID[x], pred_1))
pred_2 = top_3_predicted_indices[:, 1:2].flatten()
pred_2 = list(map(lambda x: self.TAGS_2_ID[x], pred_2))
pred_3 = top_3_predicted_indices[:, 2:].flatten()
pred_3 = list(map(lambda x: self.TAGS_2_ID[x], pred_3))
pred_1_confidence = top_3_predicted_confidence[:, 0:1]
pred_2_confidence = top_3_predicted_confidence[:, 1:2]
pred_3_confidence = top_3_predicted_confidence[:, 2:]
results.append({
"tokens": df[self._text_col].astype(str).values.tolist(),
self._prediction_col: predicted_id,
"confidence": confidence,
"pred_1": pred_1,
"pred_1_confidence": pred_1_confidence,
"pred_2": pred_2,
"pred_2_confidence": pred_2_confidence,
"pred_3": pred_3,
"pred_3_confidence": pred_3_confidence,
})
df[self._prediction_col] = [j for i, j in
zip(df[self._text_col].astype(str).values.tolist(), predicted_id)]
df["confidence"]= confidence[:len(df)]
df["pred_1"]= pred_1[:len(df)]
df["pred_1_confidence"]= pred_1_confidence[:len(df)]
df["pred_2"]= pred_2[:len(df)]
df["pred_2_confidence"]= pred_2_confidence[:len(df)]
out_dir = os.path.join(self.OUT_DIR, "predictions")
check_n_makedirs(out_dir)
df.to_csv(os.path.join(out_dir, os.path.basename(file)), index=False)
return results
def predict_on_test_files(self, executor: Executor):
"""
Runs the prediction on list of file to be tagged
:return:
"""
model = executor.model
estimator = executor.estimator
data_iterator = executor.data_iterator
predict_fn = estimator.predict(input_fn=lambda: data_iterator.test_input_fn())
# data_iterator.predict_on_test_files(predict_fn)
predictions = []
for predict in predict_fn:
predictions.append(predict)
results = []
# Get the files from test folder and zip it with predictions
for each_prediction, file in zip(predictions, self._test_files_path):
df = pd.read_csv(file, sep=self._in_seperator, quoting=csv.QUOTE_NONE)
df.dropna(axis=0, how='any', thresh=None, subset=None, inplace=True)
predicted_id = []
confidence = []
for tag_score in each_prediction["confidence"]:
confidence.append(tag_score)
for tag_id in each_prediction["viterbi_seq"]:
predicted_id.append(self.TAGS_2_ID[tag_id])
top_3_predicted_indices = each_prediction["top_3_indices"]
top_3_predicted_confidence = each_prediction["top_3_confidence"]
# print(top_3_predicted_indices)
pred_1 = top_3_predicted_indices[:, 0:1].flatten()
pred_1 = list(map(lambda x: self.TAGS_2_ID[x], pred_1))
pred_2 = top_3_predicted_indices[:, 1:2].flatten()
pred_2 = list(map(lambda x: self.TAGS_2_ID[x], pred_2))
pred_3 = top_3_predicted_indices[:, 2:].flatten()
pred_3 = list(map(lambda x: self.TAGS_2_ID[x], pred_3))
pred_1_confidence = top_3_predicted_confidence[:, 0:1].flatten()
pred_2_confidence = top_3_predicted_confidence[:, 1:2].flatten()
pred_3_confidence = top_3_predicted_confidence[:, 2:].flatten()
results.append({
"tokens": df[self._text_col].astype(str).values.tolist(),
self._prediction_col: predicted_id,
"confidence": confidence,
"pred_1": pred_1,
"pred_1_confidence": pred_1_confidence,
"pred_2": pred_2,
"pred_2_confidence": pred_2_confidence,
"pred_3": pred_3,
"pred_3_confidence": pred_3_confidence,
})
df[self._prediction_col] = [j for i, j in
zip(df[self._text_col].astype(str).values.tolist(), predicted_id)]
df["confidence"] = confidence[:len(df)]
df["pred_1"] = pred_1[:len(df)]
df["pred_1_confidence"] = pred_1_confidence[:len(df)]
df["pred_2"] = pred_2[:len(df)]
df["pred_2_confidence"] = pred_2_confidence[:len(df)]
out_dir = os.path.join(self.OUT_DIR, "predictions")
check_n_makedirs(out_dir)
df.to_csv(os.path.join(out_dir, os.path.basename(file)), index=False)
prev_tag = None
doc_text = ""
text = ""
enter = False
print(file)
for index, row in df.iterrows():
print(row[self._text_col],row[self._prediction_col])
# check if the first row
if row[self._prediction_col] != "O":
if index == 0 or not enter:
text = row[self._text_col]
prev_tag = row[self._prediction_col]
enter = True
else:
# second index onwards
if is_new_tag(prev_tag, row[self._prediction_col]):
doc_text = doc_text + text + "~" + strip_iob(prev_tag) + "\n"
text = row[self._text_col]
else:
text = text + " " + row[self._text_col]
prev_tag = row[self._prediction_col]
else:
doc_text = doc_text + text + "~" + strip_iob(row[self._prediction_col]) + "\n"
prev_tag = row[self._prediction_col]
doc_text = doc_text + text + "~" + strip_iob(prev_tag) + "\n"
print(doc_text)
post_out_dir = os.path.join(self.OUT_DIR, "postprocessed")
check_n_makedirs(post_out_dir)
# with open(os.path.join(post_out_dir, os.path.basename(file)), "w") as post_file:
# post_file.write("Item~Tag\n")
# post_file.write(doc_text)
# for loop ends
# Compute accuracy metric for the all the predictions
get_naive_metrics(predicted_csvs_path=out_dir,
ner_tag_vocab_file=self.ENTITY_VOCAB_FILE,
entity_col_name=self._entity_col,
prediction_col_name=self._prediction_col,
out_dir=out_dir)
def __init__(self,
experiment_root_directory,
experiment_name,
number_test_of_samples=4,
batch_size=32,
prefetch_size=32,
iterator_name="Cifar10BasicIterator",
preprocessed_data_path="preprocessed_data",
train_data_path="train",
validation_data_path="val",
test_data_path="test",
dataset=None,
text_col=0,
entity_col=3,
prediction_col="prediction",
in_seperator="~",
out_seperator="~",
quotechar="^",
max_word_length=20,
use_char_embd=False):
'''
:param hparams:
:param dataset:
'''
IIteratorBase.__init__(self,
experiment_root_directory=experiment_root_directory,
experiment_name=experiment_name,
batch_size=batch_size,
prefetch_size=prefetch_size,
dataset=dataset)
ITextFeature.__init__(self)
self._experiment_root_directory = experiment_root_directory
self._experiment_name = experiment_name
self._prefetch_size = prefetch_size
self._dataset = dataset
self._iterator_name = iterator_name
self._preprocessed_data_path = preprocessed_data_path
self._train_data_path = train_data_path
self._validation_data_path = validation_data_path
self._test_data_path = test_data_path
self._number_test_of_samples = number_test_of_samples
self._text_col = text_col
self._entity_col = entity_col
self._prediction_col = prediction_col
self._batch_size = batch_size
self._in_seperator = in_seperator
self._out_seperator = out_seperator
self._quotechar = quotechar
self._max_word_length = max_word_length
self._use_char_embd = use_char_embd
self.EXPERIMENT_ROOT_DIR = os.path.join(self._experiment_root_directory,
self._experiment_name)
self.PREPROCESSED_DATA_OUT_DIR = os.path.join(self.EXPERIMENT_ROOT_DIR,
self._preprocessed_data_path)
self.TRAIN_OUT_PATH = os.path.join(self.PREPROCESSED_DATA_OUT_DIR,
self._train_data_path)
self.VAL_OUT_PATH = os.path.join(self.PREPROCESSED_DATA_OUT_DIR,
self._validation_data_path)
self.TEST_OUT_PATH = os.path.join(self.PREPROCESSED_DATA_OUT_DIR,
self._test_data_path)
self.OUT_DIR = os.path.join(self.EXPERIMENT_ROOT_DIR,
self._iterator_name)
# This rule is assumed to be correct if the previous stage is of IPreprocessor
self.TRAIN_FILES_IN_PATH = os.path.join(self.PREPROCESSED_DATA_OUT_DIR, "train/")
self.VAL_FILES_IN_PATH = os.path.join(self.PREPROCESSED_DATA_OUT_DIR, "val/")
self.TEST_FILES_IN_PATH = os.path.join(self.PREPROCESSED_DATA_OUT_DIR, "test/")
self.WORDS_VOCAB_FILE = os.path.join(self.OUT_DIR, str(self._text_col) + "_" + "vocab.tsv")
self.CHARS_VOCAB_FILE = os.path.join(self.OUT_DIR, str(self._text_col) + "_" + "chars_vocab.tsv")
self.ENTITY_VOCAB_FILE = os.path.join(self.OUT_DIR, str(self._entity_col) + "_vocab.tsv")
check_n_makedirs(self.OUT_DIR)
self.padded_shapes = (tf.TensorShape([None]), # sentence of unknown size
tf.TensorShape([None]),
tf.TensorShape([None])) # labels of unknown size
self.padding_values = (SpecialTokens.PAD_WORD,
SpecialTokens.PAD_CHAR, # sentence padded on the right with id_pad_word
SpecialTokens.PAD_TAG) # labels padded on the right with id_pad_tag
self._extract_vocab()