def parse_file_full_embeddings_tapas(fname, outfilename):
model_name = 'google/tapas-base'
tokenizer = TapasTokenizer.from_pretrained(model_name)
config = TapasConfig('google-base-finetuned-wikisql-supervised')
model = TapasForQuestionAnswering.from_pretrained(model_name, config=config).to(device)
final_dict = {}
with open(fname) as f:
data = list(f)
print("Num Examples: {}".format(len(data)))
for i, line in enumerate(data):
#print(line)
result = json.loads(line)
tbl_id = result['table_id']
table_string = ' '.join(_tbl(result))
table_list = table_string.split(SPLIT_WORD)
table_list_filtered = [token for token in table_list if token != '']
dict_index = {key : [] for key in table_list_filtered}
table = pd.DataFrame(dict_index)
query = [' '.join(result['question']['words'])]
inputs = tokenizer(table=table, queries=query)
out = model(inputs)[0].tolist()
final_dict[tbl_id] = out
if i % 200 == 0:
print("Num Examples Done: {}".format(i))
with open(outfilename, 'w') as outfile:
json.dump(final_dict, outfile)
def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, tapas_config_file, pytorch_dump_path):
# Initialise PyTorch model
#config = TapasConfig.from_json_file(tapas_config_file)
config = TapasConfig(task="SQA")
print("Building PyTorch model from configuration: {}".format(str(config)))
#model = TapasForMaskedLM(config)
model = TapasForQuestionAnswering(config)
# Load weights from tf checkpoint
load_tf_weights_in_tapas(model, config, tf_checkpoint_path)
# Save pytorch-model
print("Save PyTorch model to {}".format(pytorch_dump_path))
torch.save(model.state_dict(), pytorch_dump_path)
def get_config(self):
return TapasConfig(
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
intermediate_size=self.intermediate_size,
hidden_act=self.hidden_act,
hidden_dropout_prob=self.hidden_dropout_prob,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
type_vocab_sizes=self.type_vocab_sizes,
initializer_range=self.initializer_range,
positive_weight=self.positive_weight,
num_aggregation_labels=self.num_aggregation_labels,
num_labels=self.num_labels,
aggregation_loss_importance=self.aggregation_loss_importance,
use_answer_as_supervision=self.use_answer_as_supervision,
answer_loss_importance=self.answer_loss_importance,
use_normalized_answer_loss=self.use_normalized_answer_loss,
huber_loss_delta=self.huber_loss_delta,
temperature=self.temperature,
agg_temperature=self.agg_temperature,
use_gumbel_for_cells=self.use_gumbel_for_cells,
use_gumbel_for_agg=self.use_gumbel_for_agg,
average_approximation_function=self.average_approximation_function,
cell_selection_preference=self.cell_selection_preference,
answer_loss_cutoff=self.answer_loss_cutoff,
max_num_rows=self.max_num_rows,
max_num_columns=self.max_num_columns,
average_logits_per_cell=self.average_logits_per_cell,
select_one_column=self.select_one_column,
allow_empty_column_selection=self.allow_empty_column_selection,
init_cell_selection_weights_to_zero=self.
init_cell_selection_weights_to_zero,
reset_position_index_per_cell=self.reset_position_index_per_cell,
disable_per_token_loss=self.disable_per_token_loss,
)
def prepare_config_and_inputs(self):
input_ids = ids_tensor([self.batch_size, self.seq_length],
self.vocab_size).to(torch_device)
input_mask = None
if self.use_input_mask:
input_mask = random_attention_mask(
[self.batch_size, self.seq_length]).to(torch_device)
token_type_ids = []
for type_vocab_size in self.type_vocab_sizes:
token_type_ids.append(
ids_tensor(shape=[self.batch_size, self.seq_length],
vocab_size=type_vocab_size))
token_type_ids = torch.stack(token_type_ids, dim=2).to(torch_device)
sequence_labels = None
token_labels = None
labels = None
numeric_values = None
numeric_values_scale = None
float_answer = None
aggregation_labels = None
if self.use_labels:
sequence_labels = ids_tensor(
[self.batch_size],
self.type_sequence_label_size).to(torch_device)
token_labels = ids_tensor([self.batch_size, self.seq_length],
self.num_labels).to(torch_device)
labels = ids_tensor([self.batch_size, self.seq_length],
vocab_size=2).to(torch_device)
numeric_values = floats_tensor([self.batch_size,
self.seq_length]).to(torch_device)
numeric_values_scale = floats_tensor(
[self.batch_size, self.seq_length]).to(torch_device)
float_answer = floats_tensor([self.batch_size]).to(torch_device)
aggregation_labels = ids_tensor(
[self.batch_size],
self.num_aggregation_labels).to(torch_device)
config = TapasConfig(
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
num_attention_heads=self.num_attention_heads,
intermediate_size=self.intermediate_size,
hidden_act=self.hidden_act,
hidden_dropout_prob=self.hidden_dropout_prob,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
type_vocab_sizes=self.type_vocab_sizes,
initializer_range=self.initializer_range,
positive_weight=self.positive_weight,
num_aggregation_labels=self.num_aggregation_labels,
num_labels=self.num_labels,
aggregation_loss_importance=self.aggregation_loss_importance,
use_answer_as_supervision=self.use_answer_as_supervision,
answer_loss_importance=self.answer_loss_importance,
use_normalized_answer_loss=self.use_normalized_answer_loss,
huber_loss_delta=self.huber_loss_delta,
temperature=self.temperature,
agg_temperature=self.agg_temperature,
use_gumbel_for_cells=self.use_gumbel_for_cells,
use_gumbel_for_agg=self.use_gumbel_for_agg,
average_approximation_function=self.average_approximation_function,
cell_selection_preference=self.cell_selection_preference,
answer_loss_cutoff=self.answer_loss_cutoff,
max_num_rows=self.max_num_rows,
max_num_columns=self.max_num_columns,
average_logits_per_cell=self.average_logits_per_cell,
select_one_column=self.select_one_column,
allow_empty_column_selection=self.allow_empty_column_selection,
init_cell_selection_weights_to_zero=self.
init_cell_selection_weights_to_zero,
reset_position_index_per_cell=self.reset_position_index_per_cell,
disable_per_token_loss=self.disable_per_token_loss,
)
return (
config,
input_ids,
input_mask,
token_type_ids,
sequence_labels,
token_labels,
labels,
numeric_values,
numeric_values_scale,
float_answer,
aggregation_labels,
)
def convert_tf_checkpoint_to_pytorch(task, reset_position_index_per_cell,
tf_checkpoint_path, tapas_config_file,
pytorch_dump_path):
# Initialise PyTorch model.
# If you want to convert a checkpoint that uses absolute position embeddings, make sure to set reset_position_index_per_cell of
# TapasConfig to False.
# initialize configuration from json file
config = TapasConfig.from_json_file(tapas_config_file)
# set absolute/relative position embeddings parameter
config.reset_position_index_per_cell = reset_position_index_per_cell
# set remaining parameters of TapasConfig as well as the model based on the task
if task == "SQA":
model = TapasForQuestionAnswering(config=config)
elif task == "WTQ":
# run_task_main.py hparams
config.num_aggregation_labels = 4
config.use_answer_as_supervision = True
# hparam_utils.py hparams
config.answer_loss_cutoff = 0.664694
config.cell_selection_preference = 0.207951
config.huber_loss_delta = 0.121194
config.init_cell_selection_weights_to_zero = True
config.select_one_column = True
config.allow_empty_column_selection = False
config.temperature = 0.0352513
model = TapasForQuestionAnswering(config=config)
elif task == "WIKISQL_SUPERVISED":
# run_task_main.py hparams
config.num_aggregation_labels = 4
config.use_answer_as_supervision = False
# hparam_utils.py hparams
config.answer_loss_cutoff = 36.4519
config.cell_selection_preference = 0.903421
config.huber_loss_delta = 222.088
config.init_cell_selection_weights_to_zero = True
config.select_one_column = True
config.allow_empty_column_selection = True
config.temperature = 0.763141
model = TapasForQuestionAnswering(config=config)
elif task == "TABFACT":
model = TapasForSequenceClassification(config=config)
elif task == "MLM":
model = TapasForMaskedLM(config=config)
elif task == "INTERMEDIATE_PRETRAINING":
model = TapasModel(config=config)
else:
raise ValueError(f"Task {task} not supported.")
print(f"Building PyTorch model from configuration: {config}")
# Load weights from tf checkpoint
load_tf_weights_in_tapas(model, config, tf_checkpoint_path)
# Save pytorch-model (weights and configuration)
print(f"Save PyTorch model to {pytorch_dump_path}")
model.save_pretrained(pytorch_dump_path)
# Save tokenizer files
print(f"Save tokenizer files to {pytorch_dump_path}")
tokenizer = TapasTokenizer(vocab_file=tf_checkpoint_path[:-10] +
"vocab.txt",
model_max_length=512)
tokenizer.save_pretrained(pytorch_dump_path)
print("Used relative position embeddings:",
model.config.reset_position_index_per_cell)
