def create_and_check_t5_model(self, config, input_ids, input_mask,
token_labels):
model = TFT5Model(config=config)
inputs = {
"input_ids": input_ids,
"decoder_input_ids": input_ids,
"decoder_attention_mask": input_mask,
}
result = model(inputs)
result = model(input_ids,
decoder_attention_mask=input_mask,
decoder_input_ids=input_ids)
decoder_output = result.last_hidden_state
decoder_past = result.decoder_past_key_values
encoder_output = result.encoder_last_hidden_state
self.parent.assertListEqual(
list(encoder_output.shape),
[self.batch_size, self.seq_length, self.hidden_size])
self.parent.assertListEqual(
list(decoder_output.shape),
[self.batch_size, self.seq_length, self.hidden_size])
self.parent.assertEqual(len(decoder_past), 2)
# decoder_past[0] should correspond to encoder output
self.parent.assertTrue(
tf.reduce_all(tf.math.equal(decoder_past[0][0], encoder_output)))
# There should be `num_layers` key value embeddings stored in decoder_past[1]
self.parent.assertEqual(len(decoder_past[1]), config.num_layers)
# There should be a self attn key, a self attn value, a cross attn key and a cross attn value stored in each decoder_past[1] tuple
self.parent.assertEqual(len(decoder_past[1][0]), 4)
def create_and_check_t5_decoder_model_past_large_inputs(
self, config, input_ids, decoder_input_ids, attention_mask
):
model = TFT5Model(config=config).get_decoder()
input_ids = input_ids[:1, :]
attention_mask = attention_mask[:1, :]
self.batch_size = 1
# first forward pass
outputs = model(input_ids, attention_mask=attention_mask, use_cache=True)
# create hypothetical next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_attn_mask = ids_tensor((self.batch_size, 3), 2)
# append to next input_ids and
next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
next_attention_mask = tf.concat([attention_mask, next_attn_mask], axis=-1)
output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)[0]
output_from_past = model(
next_tokens, attention_mask=next_attention_mask, past_key_values=outputs.past_key_values
)[0]
self.parent.assertEqual(next_tokens.shape[1], output_from_past.shape[1])
# select random slice
random_slice_idx = int(ids_tensor((1,), output_from_past.shape[-1]))
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx]
output_from_past_slice = output_from_past[:, :, random_slice_idx]
# test that outputs are equal for slice
tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-3)
def create_and_check_t5_decoder_model_past(self, config, input_ids, decoder_input_ids, attention_mask):
model = TFT5Model(config=config).get_decoder()
input_ids = input_ids[:1, :]
self.batch_size = 1
# first forward pass
outputs = model(input_ids, use_cache=True)
outputs_use_cache_conf = model(input_ids)
outputs_no_past = model(input_ids, use_cache=False)
self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
# create hypothetical next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
# append to next input_ids and
next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
output_from_no_past = model(next_input_ids)[0]
output_from_past = model(next_tokens, past_key_values=outputs.past_key_values)[0]
# select random slice
random_slice_idx = int(ids_tensor((1,), output_from_past.shape[-1]))
output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx]
output_from_past_slice = output_from_past[:, 0, random_slice_idx]
# test that outputs are equal for slice
tf.debugging.assert_near(output_from_past_slice, output_from_no_past_slice, rtol=1e-3)
def __init__(
self,
pretrained_model_name_or_path='t5-small',
reduce_output='sum',
trainable=True,
num_tokens=None,
**kwargs
):
super(T5Encoder, self).__init__()
try:
from transformers import TFT5Model
except ModuleNotFoundError:
logger.error(
' transformers is not installed. '
'In order to install all text feature dependencies run '
'pip install ludwig[text]'
)
sys.exit(-1)
self.transformer = TFT5Model.from_pretrained(
pretrained_model_name_or_path
)
self.reduce_output = reduce_output
self.reduce_sequence = SequenceReducer(reduce_mode=reduce_output)
self.transformer.trainable = trainable
self.transformer.resize_token_embeddings(num_tokens)
def create_and_check_t5_decoder_model_attention_mask_past(
self, config, input_ids, decoder_input_ids, attention_mask):
model = TFT5Model(config=config).get_decoder()
# create attention mask
half_seq_length = self.seq_length // 2
attn_mask_begin = tf.ones((self.batch_size, half_seq_length),
dtype=tf.int32)
attn_mask_end = tf.zeros(
(self.batch_size, self.seq_length - half_seq_length),
dtype=tf.int32)
attn_mask = tf.concat([attn_mask_begin, attn_mask_end], axis=1)
# first forward pass
_, past_key_value_states = model(input_ids,
attention_mask=attn_mask,
use_cache=True)
# create hypothetical next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
# change a random masked slice from input_ids
random_seq_idx_to_change = ids_tensor(
(1, ), half_seq_length).numpy() + 1
random_other_next_tokens = ids_tensor(
(self.batch_size, self.seq_length), config.vocab_size)
vector_condition = tf.range(
self.seq_length) == (self.seq_length - random_seq_idx_to_change)
condition = tf.transpose(
tf.broadcast_to(tf.expand_dims(vector_condition, -1),
(self.seq_length, self.batch_size)))
input_ids = tf.where(condition, random_other_next_tokens, input_ids)
# append to next input_ids and attn_mask
next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
attn_mask = tf.concat(
[attn_mask,
tf.ones((attn_mask.shape[0], 1), dtype=tf.int32)],
axis=1,
)
# get two different outputs
output_from_no_past = model(next_input_ids,
attention_mask=attn_mask)[0]
output_from_past = model(next_tokens,
past_key_value_states=past_key_value_states,
attention_mask=attn_mask)[0]
# select random slice
random_slice_idx = ids_tensor(
(1, ), output_from_past.shape[-1]).numpy().item()
output_from_no_past_slice = output_from_no_past[:, -1,
random_slice_idx]
output_from_past_slice = output_from_past[:, 0, random_slice_idx]
# test that outputs are equal for slice
tf.debugging.assert_near(output_from_past_slice,
output_from_no_past_slice,
rtol=1e-3)
def _test_TFT5Model(self, size, large=False):
from transformers import T5Tokenizer, TFT5Model
tokenizer = T5Tokenizer.from_pretrained(size)
model = TFT5Model.from_pretrained(size)
input_ids = \
tokenizer("Studies have been shown that owning a dog is good for you", return_tensors="tf").input_ids
decoder_input_ids = \
tokenizer("Studies show that", return_tensors="tf").input_ids
input_dict = {"input_ids": input_ids, "decoder_input_ids": decoder_input_ids}
spec, input_dict = self.spec_and_pad(input_dict)
outputs = ["last_hidden_state"]
self.run_test(model, input_dict,
input_signature=spec, outputs=outputs, large=large)
def create_and_check_t5_model(self, config, input_ids, input_mask, token_labels):
model = TFT5Model(config=config)
inputs = {
"inputs": input_ids,
"decoder_input_ids": input_ids,
"decoder_attention_mask": input_mask,
}
encoder_output, decoder_output = model(inputs)
encoder_output, decoder_output = model(
input_ids, decoder_attention_mask=input_mask, decoder_input_ids=input_ids
)
result = {
"encoder_output": encoder_output.numpy(),
"decoder_output": decoder_output.numpy(),
}
self.parent.assertListEqual(
list(result["encoder_output"].shape), [self.batch_size, self.seq_length, self.hidden_size]
)
self.parent.assertListEqual(
list(result["decoder_output"].shape), [self.batch_size, self.seq_length, self.hidden_size]
)
def test_model_from_pretrained(self):
for model_name in ["t5-small"]:
model = TFT5Model.from_pretrained(model_name)
self.assertIsNotNone(model)
def test_model_from_pretrained(self):
model = TFT5Model.from_pretrained("t5-small")
self.assertIsNotNone(model)
def test_model_from_pretrained(self):
for model_name in ["t5-small"]:
model = TFT5Model.from_pretrained(model_name, cache_dir=CACHE_DIR)
self.assertIsNotNone(model)
