oliverdk's picture
End of training
7e59788 verified
raw
history blame
4.8 kB
from typing import Optional, Tuple, Union
import torch
from torch.nn import BCEWithLogitsLoss
from transformers import PreTrainedModel, PreTrainedTokenizer
from transformers.tokenization_utils_base import PreTrainedTokenizerBase
from transformers.modeling_outputs import BaseModelOutputWithPast, SequenceClassifierOutputWithPast
from .sensor_loc_reg import SENSOR_LOC_REGISTRY
from .sensor_loc_finder import SensorLocFinder
class MeasurementPredictorMixin(PreTrainedModel):
def __init__(self, config):
super().__init__(config)
self.sensor_loc_type = config.sensor_loc_type
self.sensor_token = config.sensor_token
self.n_sensors = config.n_sensors
self.sensor_probes = torch.nn.ModuleList([
torch.nn.Linear(config.emb_dim, 1) for _ in range(config.n_sensors)
])
self.use_aggregated = config.use_aggregated
if config.use_aggregated:
self.aggregate_probe = torch.nn.Linear(config.emb_dim, 1)
self.sensors_weight = config.sensors_weight
self.aggregate_weight = config.aggregate_weight
self.get_sensor_locs: SensorLocFinder = None
def init_sensor_loc_finder(self, tokenizer: PreTrainedTokenizerBase):
self.get_sensor_locs = SENSOR_LOC_REGISTRY[self.sensor_loc_type](
tokenizer, sensor_token=self.sensor_token, n_sensors=self.n_sensors
)
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
attention_mask: Optional[torch.FloatTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
head_mask: Optional[torch.FloatTensor] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
labels: Optional[torch.LongTensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, SequenceClassifierOutputWithPast]:
r"""
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
`labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
base_model_output: BaseModelOutputWithPast = self.base_model(
input_ids,
past_key_values=past_key_values,
attention_mask=attention_mask,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
sensor_locs = self.get_sensor_locs(input_ids)
sensor_embs = base_model_output.last_hidden_state.gather(
1, sensor_locs.unsqueeze(-1).expand(-1, -1, self.config.emb_dim)
)
assert sensor_embs.shape == (input_ids.shape[0], self.n_sensors, self.config.emb_dim), f"{sensor_embs.shape} != {(input_ids.shape[0], self.n_sensors, self.config.emb_dim)}"
sensor_logits = torch.concat([self.sensor_probes[i](sensor_embs[:, i, :])
for i in range(self.n_sensors)], dim=-1)
logits = sensor_logits
if self.use_aggregated:
last_emb = base_model_output.last_hidden_state[:, -1, :]
aggregate_logits = self.aggregate_probe(last_emb)
logits = torch.concat([logits, aggregate_logits], dim=-1)
loss = None
if labels is not None:
loss_fct = BCEWithLogitsLoss()
sensor_loss = loss_fct(sensor_logits, labels[:, :self.n_sensors]) * self.sensors_weight
loss = sensor_loss
if self.use_aggregated: #TOOD: should be use aggregate
aggregate_loss = loss_fct(aggregate_logits, labels[:, -1:]) * self.aggregate_weight
loss += aggregate_loss
if not return_dict:
output = (logits, ) + base_model_output[1:]
return ((loss,) + output) if loss is not None else output
return SequenceClassifierOutputWithPast(
loss=loss,
logits=logits,
past_key_values=base_model_output.past_key_values,
hidden_states=base_model_output.hidden_states,
attentions=base_model_output.attentions,
)