def create_and_check_gptj_model_past(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
model = TFGPTJModel(config=config)
# first forward pass
outputs = model(input_ids, token_type_ids=token_type_ids, use_cache=True)
outputs_use_cache_conf = model(input_ids, token_type_ids=token_type_ids)
outputs_no_past = model(input_ids, token_type_ids=token_type_ids, use_cache=False)
self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
output, past = outputs.to_tuple()
# create hypothetical next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
next_token_types = ids_tensor([self.batch_size, 1], self.type_vocab_size)
# append to next input_ids and token_type_ids
next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
next_token_type_ids = tf.concat([token_type_ids, next_token_types], axis=-1)
output_from_no_past = model(next_input_ids, token_type_ids=next_token_type_ids)["last_hidden_state"]
output_from_past = model(next_tokens, token_type_ids=next_token_types, past=past)["last_hidden_state"]
# select random slice
random_slice_idx = int(ids_tensor((1,), shape_list(output_from_past)[-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_gptj_model_attention_mask_past(
self, config, input_ids, input_mask, head_mask, token_type_ids,
*args):
model = TFGPTJModel(config=config)
# 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
output, past = model(input_ids, attention_mask=attn_mask).to_tuple()
# 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((shape_list(attn_mask)[0], 1), dtype=tf.int32)
],
axis=1)
# get two different outputs
output_from_no_past = model(
next_input_ids, attention_mask=attn_mask)["last_hidden_state"]
output_from_past = model(next_tokens,
past=past,
attention_mask=attn_mask)["last_hidden_state"]
# select random slice
random_slice_idx = int(
ids_tensor((1, ),
shape_list(output_from_past)[-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-12)
def create_and_check_gptj_model_past_large_inputs(self, config, input_ids,
input_mask, head_mask,
token_type_ids, *args):
model = TFGPTJModel(config=config)
input_ids = input_ids[:1, :]
input_mask = input_mask[:1, :]
token_type_ids = token_type_ids[:1, :]
self.batch_size = 1
# first forward pass
outputs = model(input_ids,
attention_mask=input_mask,
token_type_ids=token_type_ids,
use_cache=True)
output, past = outputs.to_tuple()
# 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)
next_token_types = ids_tensor((self.batch_size, 3),
self.type_vocab_size)
# append to next input_ids and token_type_ids
next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
next_attention_mask = tf.concat([input_mask, next_attn_mask], axis=-1)
next_token_type_ids = tf.concat([token_type_ids, next_token_types],
axis=-1)
output_from_no_past = model(
next_input_ids,
token_type_ids=next_token_type_ids,
attention_mask=next_attention_mask)["last_hidden_state"]
output_from_past = model(next_tokens,
token_type_ids=next_token_types,
attention_mask=next_attention_mask,
past=past)["last_hidden_state"]
self.parent.assertTrue(
output_from_past.shape[1] == next_tokens.shape[1])
# select random slice
random_slice_idx = int(
ids_tensor((1, ),
shape_list(output_from_past)[-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_gptj_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
model = TFGPTJModel(config=config)
inputs = {
"input_ids": input_ids,
"attention_mask": input_mask,
"token_type_ids": token_type_ids,
}
result = model(inputs)
inputs = [input_ids, None, input_mask] # None is the input for 'past'
result = model(inputs)
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def test_model_from_pretrained(self):
model = TFGPTJModel.from_pretrained("EleutherAI/gpt-j-6B",
from_pt=True)
self.assertIsNotNone(model)