def _test_TFPegasus(self, size, large=False):
from transformers import PegasusTokenizer, TFPegasusModel
tokenizer = PegasusTokenizer.from_pretrained(size)
model = TFPegasusModel.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
}
# this comes from TFPegasusEncoder/Decoder like:
# self.embed_scale = tf.math.sqrt(float(config.d_model)) if config.scale_embedding else 1.0
# while this is mean to come from config tf tells us that those are model inputs
# this might be new in tensformers-2.4.2, we did not notice before that
extra_input = {
"tf_pegasus_model/model/decoder/mul/y:0":
np.array([32.], dtype=np.float32),
"tf_pegasus_model/model/encoder/mul/y:0":
np.array([32.], dtype=np.float32)
}
spec, input_dict = self.spec_and_pad(
input_dict, max_length=model.config.max_length)
outputs = ["last_hidden_state"]
self.run_test(model,
input_dict,
input_signature=spec,
outputs=outputs,
large=large,
extra_input=extra_input)
def check_decoder_model_past_large_inputs(self, config, inputs_dict):
model = TFPegasusModel(config=config).get_decoder()
input_ids = inputs_dict["input_ids"]
input_ids = input_ids[:1, :]
attention_mask = inputs_dict["attention_mask"][:1, :]
self.batch_size = 1
# first forward pass
outputs = model(input_ids, attention_mask=attention_mask, use_cache=True)
output, past_key_values = outputs.to_tuple()
past_key_values = past_key_values[1]
# create hypothetical next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_attn_mask = tf.cast(ids_tensor((self.batch_size, 3), 2), tf.int8)
# 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=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)