A collection of VAT (Virtual Adversarial Training) methods, in PyTorch.
$ pip install vat-pytorch
The SMART paper proposes to find the noise that maximally perturbs the logits when added to the embedding layer, and to use a loss function to make sure that the perturbed logits are as close as possible to the predicted logits.
from vat_pytorch import SMARTLoss, inf_norm
loss = SMARTLoss(
model: nn.Module,
loss_fn: Callable,
loss_last_fn: Callable = None,
norm_fn: Callable = inf_norm,
num_steps: int = 1,
step_size: float = 1e-3,
epsilon: float = 1e-6,
noise_var: float = 1e-5
)The ALICE paper is analogous to the SMART paper, but adds an additional term to make sure that the perturbed logits are as close as possible to both the predicted logits and the ground truth labels.
from vat_pytorch import ALICELoss, inf_norm
loss = ALICEPPLoss(
model: nn.Module,
loss_fn: Callable,
loss_last_fn: Callable = None,
gold_loss_fn: Callable = None,
gold_loss_last_fn: Callable = None,
norm_fn: Callable = inf_norm,
alpha: float = 1,
num_steps: int = 1,
step_size: float = 1e-3,
epsilon: float = 1e-6,
noise_var: float = 1e-5
)The ALICE++ paper is analogous to the ALICE paper, but instead of adding noise to the embedding layer, it picks a random layer from the network at each iteration on which to add the noise.
from vat_pytorch import ALICEPPLoss, ALICEPPModule, inf_norm
loss = ALICEPPLoss(
model: ALICEPPModule,
loss_fn: Callable,
num_layers: int,
loss_last_fn: Callable = None,
gold_loss_fn: Callable = None,
gold_loss_last_fn: Callable = None,
norm_fn: Callable = inf_norm,
alpha: float = 1,
num_steps: int = 1,
step_size: float = 1e-3,
epsilon: float = 1e-6,
noise_var: float = 1e-5
)The first thing we have to do is extract the chunk of the model that we want to perturb adversarially. A generic example with Huggingface's RoBERTa for sequence classification is given.
import torch.nn as nn
from transformers import AutoModelForSequenceClassification
class ExtractedRoBERTa(nn.Module):
def __init__(self):
super().__init__()
model = AutoModelForSequenceClassification.from_pretrained('roberta-base')
self.roberta = model.roberta
self.layers = model.roberta.encoder.layer
self.classifier = model.classifier
self.attention_mask = None
self.num_layers = len(self.layers) - 1
def forward(self, hidden, with_hidden_states = False, start_layer = 0):
""" Forwards the hidden value from self.start_layer layer to the logits. """
hidden_states = [hidden]
for layer in self.layers[start_layer:]:
hidden = layer(hidden, attention_mask = self.attention_mask)[0]
hidden_states += [hidden]
logits = self.classifier(hidden)
return (logits, hidden_states) if with_hidden_states else logits
def get_embeddings(self, input_ids):
""" Computes first embedding layer given inputs_ids """
return self.roberta.embeddings(input_ids)
def set_attention_mask(self, attention_mask):
""" Sets the correct mask on all subsequent forward passes """
self.attention_mask = self.roberta.get_extended_attention_mask(
attention_mask,
input_shape = attention_mask.shape,
device = attention_mask.device
) # (b, 1, 1, s) The function set_attention_mask is used to fix the attention mask for all subsequent forward calls, this is necessary if we want to use a mask using any VAT loss. The parameter start_layer in the forward function is necessary only if we are using ALICEPPLoss since the loss function needs a way to change the start layer internally.
import torch.nn as nn
import torch.nn.functional as F
from vat_pytorch import SMARTLoss, kl_loss, sym_kl_loss
class SMARTClassificationModel(nn.Module):
# b: batch_size, s: sequence_length, d: hidden_size , n: num_labels
def __init__(self, extracted_model, weight = 1.0):
super().__init__()
self.model = extracted_model
self.weight = weight
self.vat_loss = SMARTLoss(model = extracted_model, loss_fn = kl_loss, loss_last_fn = sym_kl_loss)
def forward(self, input_ids, attention_mask, labels):
""" input_ids: (b, s), attention_mask: (b, s), labels: (b,) """
# Get input embeddings
embeddings = self.model.get_embeddings(input_ids)
# Set mask and compute logits
self.model.set_attention_mask(attention_mask)
logits = self.model(embeddings)
# Compute CE loss
ce_loss = F.cross_entropy(logits.view(-1, 2), labels.view(-1))
# Compute VAT loss
vat_loss = self.vat_loss(embeddings, logits)
# Merge losses
loss = ce_loss + self.weight * vat_loss
return logits, lossimport torch.nn as nn
import torch.nn.functional as F
from vat_pytorch import ALICELoss, kl_loss
class ALICEClassificationModel(nn.Module):
# b: batch_size, s: sequence_length, d: hidden_size , n: num_labels
def __init__(self, extracted_model):
super().__init__()
self.model = extracted_model
self.vat_loss = ALICELoss(model = extracted_model, loss_fn = kl_loss)
def forward(self, input_ids, attention_mask, labels):
""" input_ids: (b, s), attention_mask: (b, s), labels: (b,) """
# Get input embeddings
embeddings = self.model.get_embeddings(input_ids)
# Set iteration specific data (e.g. attention mask)
self.model.set_attention_mask(attention_mask)
# Compute logits
logits = self.model(embeddings)
# Compute VAT loss
loss = self.vat_loss(embeddings, logits, labels)
return logits, lossimport torch.nn as nn
import torch.nn.functional as F
from vat_pytorch import ALICEPPLoss, kl_loss
class ALICEPPClassificationModel(nn.Module):
# b: batch_size, s: sequence_length, d: hidden_size , n: num_labels
def __init__(self, extracted_model):
super().__init__()
self.model = extracted_model
self.vat_loss = ALICEPPLoss(model = extracted_model, loss_fn = kl_loss, num_layers = self.model.num_layers)
def forward(self, input_ids, attention_mask, labels):
""" input_ids: (b, s), attention_mask: (b, s), labels: (b,) """
# Get input embeddings
embeddings = self.model.get_embeddings(input_ids)
# Set iteration specific data (e.g. attention mask)
self.model.set_attention_mask(attention_mask)
# Compute logits
logits, hidden_states = self.model(embeddings, with_hidden_states = True)
# Compute VAT loss
loss = self.vat_loss(hidden_states, logits, labels)
return logits, lossNote that extracted_model requires a function with the following signature forward(self, hidden: Tensor, *, start_layer: int) -> Tensor, the interface ALICEPPModule (from vat_pytorch import ALICEPPModule) can be used instead of the nn.Module class on the extracted model to make sure that the method is present.
Any of the above losses can be used as follows with the extracted model.
import torch
from transformers import AutoTokenizer
extracted_model = ExtractedRoBERTa()
tokenizer = AutoTokenizer.from_pretrained('roberta-base')
# Pick one:
model = SMARTClassificationModel(extracted_model)
model = ALICEClassificationModel(extracted_model)
model = ALICEPPClassificationModel(extracted_model)
# Compute inputs
text = ["This text belongs to class 1...", "This text belongs to class 0..."]
inputs = tokenizer(text, return_tensors='pt')
labels = torch.tensor([1, 0])
# Compute logits and loss
logits, loss = model(input_ids = inputs['input_ids'], attention_mask = inputs['attention_mask'], labels = labels)
# To finetune do this for many steps
loss.backward() @inproceedings{Jiang2020SMARTRA,
title={SMART: Robust and Efficient Fine-Tuning for Pre-trained Natural Language Models through Principled Regularized Optimization},
author={Haoming Jiang and Pengcheng He and Weizhu Chen and Xiaodong Liu and Jianfeng Gao and Tuo Zhao},
booktitle={ACL},
year={2020}
}@article{Pereira2020AdversarialTF,
title={Adversarial Training for Commonsense Inference},
author={Lis Kanashiro Pereira and Xiaodong Liu and Fei Cheng and Masayuki Asahara and Ichiro Kobayashi},
journal={ArXiv},
year={2020},
volume={abs/2005.08156}
}@inproceedings{Pereira2021ALICEAT,
title={ALICE++: Adversarial Training for Robust and Effective Temporal Reasoning},
author={Lis Kanashiro Pereira and Fei Cheng and Masayuki Asahara and Ichiro Kobayashi},
booktitle={PACLIC},
year={2021}
}