def __init__(self, hparams):
assert Path(hparams.data_dir).exists()
self.output_dir = Path(hparams.output_dir)
self.output_dir.mkdir(exist_ok=True)
save_dir = self.output_dir.joinpath("student")
hparams.model_name_or_path = str(save_dir) # Tell lightning we are training the student
teacher = AutoModelForSeq2SeqLM.from_pretrained(hparams.teacher).eval()
use_task_specific_params(teacher, hparams.task) # We copy good generation parameters to student by default
student, e_layer_ids, d_layer_ids = create_student_by_copying_alternating_layers(
teacher, e=hparams.student_encoder_layers, d=hparams.student_decoder_layers, save_path=save_dir
)
if hparams.length_penalty != -1:
student.config.length_penalty = hparams.length_penalty
super().__init__(hparams, model=student, config=student.config)
model_type = student.config.model_type
self.e_layer_ids, self.d_layer_ids = e_layer_ids, d_layer_ids # type: List[int], List[int]
if model_type == "t5":
teacher_encoder_layers = len(teacher.get_encoder().block)
teacher_decoder_layers = len(teacher.get_decoder().block)
else:
teacher_encoder_layers = teacher.config.encoder_layers
teacher_decoder_layers = teacher.config.decoder_layers
self.different_encoder = hparams.student_encoder_layers != teacher_encoder_layers
self.different_decoder = hparams.student_decoder_layers != teacher_decoder_layers
self.teacher = teacher
freeze_params(self.teacher)
if not self.different_encoder: # To save RAM, delete teacher encoder and freeze student encoder.
try:
del self.teacher.model.encoder
except AttributeError: # T5
del self.teacher.encoder
# Intermediate supervision: Decide which layers to supervise
if hparams.supervise_forward:
self.e_matches = get_layers_to_supervise(n_student=len(self.e_layer_ids), n_teacher=teacher_encoder_layers)
self.d_matches = get_layers_to_supervise(n_student=len(self.d_layer_ids), n_teacher=teacher_decoder_layers)
else: # student layer should emulate hidden states of the teacher layer it was copied from
self.e_matches = self.e_layer_ids
self.d_matches = self.d_layer_ids
self.ce_loss_fct = nn.KLDivLoss(reduction="batchmean")
self.temperature = 2.0
self.alpha_mlm = hparams.alpha_mlm
self.alpha_ce = hparams.alpha_ce
self.alpha_hid = hparams.alpha_hid
gc.collect()
torch.cuda.empty_cache()
def __init__(self, hparams):
assert Path(hparams.data_dir).exists()
self.output_dir = Path(hparams.output_dir)
self.output_dir.mkdir(exist_ok=True)
save_dir = self.output_dir.joinpath("student")
hparams.model_name_or_path = str(
save_dir) # Tell lightning we are training the student
teacher = AutoModelForSeq2SeqLM.from_pretrained(hparams.teacher).eval()
use_task_specific_params(
teacher, hparams.task
) # We copy good generation parameters to student by default
if hparams.student is not None:
student = AutoModelForSeq2SeqLM.from_pretrained(hparams.student)
use_task_specific_params(student, hparams.task)
e_layer_ids, d_layer_ids = None, None
else:
student, e_layer_ids, d_layer_ids = create_student_by_copying_alternating_layers(
teacher,
e=hparams.student_encoder_layers,
d=hparams.student_decoder_layers,
save_path=save_dir)
if hparams.length_penalty != -1:
student.config.length_penalty = hparams.length_penalty
hparams.tokenizer_name = hparams.teacher # Use teacher's tokenizer
super().__init__(hparams, model=student, config=student.config)
assert (
student.config.model_type == teacher.config.model_type
), f"teacher, student model types should be the same, got {student.config.model_type} != {teacher.config.model_type}"
if student.config.model_type == "t5":
student_encoder_layers = len(student.get_encoder().block)
student_decoder_layers = len(student.get_decoder().block)
teacher_encoder_layers = len(teacher.get_encoder().block)
teacher_decoder_layers = len(teacher.get_decoder().block)
else:
student_encoder_layers = student.config.encoder_layers
student_decoder_layers = student.config.decoder_layers
teacher_encoder_layers = teacher.config.encoder_layers
teacher_decoder_layers = teacher.config.decoder_layers
self.different_base_models = not (hparams.student is None or
hparams.teacher == hparams.student)
self.do_calc_hidden_loss = (
not self.different_base_models) and hparams.alpha_hid > 0
self.different_encoder = self.different_base_models or (
student_encoder_layers != teacher_encoder_layers)
# self.different_encoder determines whether we need to run the teacher encoder
self.teacher = teacher
freeze_params(self.teacher)
if not self.different_encoder: # To save RAM, delete teacher encoder and freeze student encoder.
try:
del self.teacher.model.encoder
except AttributeError: # T5
del self.teacher.encoder
if e_layer_ids is None:
e_layer_ids = list(range(student_encoder_layers))
if d_layer_ids is None:
d_layer_ids = list(range(student_decoder_layers))
self.e_layer_ids, self.d_layer_ids = e_layer_ids, d_layer_ids # type: List[int], List[int]
if self.do_calc_hidden_loss: # Intermediate supervision: Decide which layers to supervise
if hparams.supervise_forward:
self.e_matches = get_layers_to_supervise(
n_student=len(self.e_layer_ids),
n_teacher=teacher_encoder_layers)
self.d_matches = get_layers_to_supervise(
n_student=len(self.d_layer_ids),
n_teacher=teacher_decoder_layers)
else: # student layer should emulate hidden states of the teacher layer it was copied from
self.e_matches = self.e_layer_ids
self.d_matches = self.d_layer_ids
else:
self.e_matches = None
self.d_matches = None
self.ce_loss_fct = nn.KLDivLoss(reduction="batchmean")
self.temperature = 2.0
self.alpha_mlm = hparams.alpha_mlm
self.alpha_ce = hparams.alpha_ce
self.alpha_hid = hparams.alpha_hid
gc.collect()
torch.cuda.empty_cache()