def create_and_check_causal_lm_model_as_decoder(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
encoder_hidden_states,
encoder_attention_mask,
):
config.add_cross_attention = True
model = TFRobertaForCausalLM(config=config)
inputs = {
"input_ids": input_ids,
"attention_mask": input_mask,
"token_type_ids": token_type_ids,
"encoder_hidden_states": encoder_hidden_states,
"encoder_attention_mask": encoder_attention_mask,
}
result = model(inputs)
inputs = [input_ids, input_mask]
result = model(inputs, token_type_ids=token_type_ids, encoder_hidden_states=encoder_hidden_states)
prediction_scores = result["logits"]
self.parent.assertListEqual(
list(prediction_scores.numpy().shape), [self.batch_size, self.seq_length, self.vocab_size]
)
def create_and_check_causal_lm_model_past_large_inputs(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
):
config.is_decoder = True
model = TFRobertaForCausalLM(config=config)
# special to `RobertaEmbeddings` in `Roberta`:
# - its `padding_idx` and its effect on `position_ids`
# (TFRobertaEmbeddings.create_position_ids_from_input_ids)
# - `1` here is `TFRobertaEmbeddings.padding_idx`
# avoid `padding_idx` in the past
input_ids = tf.where(input_ids == 1, 2, input_ids)
input_ids = input_ids[:1, :]
input_mask = input_mask[:1, :]
self.batch_size = 1
# first forward pass
outputs = model(input_ids, attention_mask=input_mask, use_cache=True)
past_key_values = outputs.past_key_values
# 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([input_mask, next_attn_mask], axis=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
output_hidden_states=True,
).hidden_states[0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
).hidden_states[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_roberta_for_causal_lm(self, config, input_ids,
token_type_ids, input_mask,
sequence_labels, token_labels,
choice_labels):
model = TFRobertaForCausalLM(config=config)
result = model([input_ids, input_mask, token_type_ids])
self.parent.assertEqual(
result.logits.shape,
(self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_causal_lm_model_past(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
):
config.is_decoder = True
model = TFRobertaForCausalLM(config=config)
# special to `RobertaEmbeddings` in `Roberta`:
# - its `padding_idx` and its effect on `position_ids`
# (TFRobertaEmbeddings.create_position_ids_from_input_ids)
# - `1` here is `TFRobertaEmbeddings.padding_idx`
input_ids = tf.where(input_ids == 1, 2, input_ids)
# 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)
past_key_values = outputs.past_key_values
# 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 attn_mask
next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
output_from_no_past = model(next_input_ids,
output_hidden_states=True).hidden_states[0]
output_from_past = model(next_tokens,
past_key_values=past_key_values,
output_hidden_states=True).hidden_states[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-6)
def create_and_check_causal_lm_model(
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
):
config.is_decoder = True
model = TFRobertaForCausalLM(config=config)
inputs = {
"input_ids": input_ids,
"attention_mask": input_mask,
"token_type_ids": token_type_ids,
}
prediction_scores = model(inputs)["logits"]
self.parent.assertListEqual(
list(prediction_scores.numpy().shape), [self.batch_size, self.seq_length, self.vocab_size]
)
def create_and_check_causal_lm_model_past_with_attn_mask(
self,
config,
input_ids,
token_type_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
):
config.is_decoder = True
model = TFRobertaForCausalLM(config=config)
# special to `RobertaEmbeddings` in `Roberta`:
# - its `padding_idx` and its effect on `position_ids`
# (TFRobertaEmbeddings.create_position_ids_from_input_ids)
# - `1` here is `TFRobertaEmbeddings.padding_idx`
# avoid `padding_idx` in the past
input_ids = tf.where(input_ids == 1, 2, input_ids)
# 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
outputs = 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)
past_key_values = outputs.past_key_values
# 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)
# avoid `padding_idx` in the past
input_ids = tf.where(input_ids == 1, 2, input_ids)
# append to next input_ids and
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,
)
output_from_no_past = model(
next_input_ids,
attention_mask=attn_mask,
output_hidden_states=True,
).hidden_states[0]
output_from_past = model(
next_tokens, past_key_values=past_key_values, attention_mask=attn_mask, output_hidden_states=True
).hidden_states[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-6)
