In this repo we experimented with two changes to the RoBERTa baseline:
Proposed in Press et al., 2022, ALiBi uses linear biases to the attention weights instead of positional encoding to represent positions of the tokens. Since masked LMs like RoBERTa both look ahead and back to determine the masked tokens, considerations must be made regarding how to distinguish them (ofirpress/attention_with_linear_biases#5). Unless otherwise noted, our implementation achieves this by shifting the linear biases ahead by 0.5 * slope, i.e. the constant bias (right) of Figure 3 in Press et al., 2022 becomes
0 -.5 -1.5 -2.5 -3.5
-1 0 -.5 -1.5 -2.5
-2 -1 0 -.5 -1.5
-3 -2 -1 0 -.5
-4 -3 -2 -1 0
Inspired by CLIP but actually goes back all the way to the origin of weight tying (Press and Wolf, 2017), CLAP head is the simplest possible prediction head for the missing token except the thermodynamic beta (inverse temperature):
fairseq/fairseq/models/roberta/model.py
Lines 527 to 543 in 8143446
embed_dim x embed_dim fully-connected layer, activation function (GELU), layer norm, and the output_dim trainable bias. On the other hand, we added the trainable thermodynamic beta and L2-normalize the embeddings before feeding them to the transformer and computing the inner products between them and the transformer output, scaled by beta.
At first we tested the changes with the WikiText-103 dataset with a GeForce RTX 3080 16 GB Laptop GPU, using the validation set MLM perplexity as the metric. We tested the baseline (learned positional encoding + RoBERTa prediction head), learned-clap (learned positional encoding + CLAP head), ALiBi (ALiBi + RoBERTa prediction head), and zero-clap (ALiBi + CLAP head), in addition to baseline but with sinusoidal positional encoding instead of learned positional encoding:
where solid lines are what's considered "canonical" setup and dotted lines are experiments with minor differences in setup. We tested the following and found them to be irrelevant:
- Whether we use attention dropout or not
- Whether we use symmetrical ALiBi (option 1 in ofirpress/attention_with_linear_biases#5) or asymmetrical ALiBi above
Whether we use zero vector or a separate learnable embedding for the mask embedding- Whether we L2-normalize the embeddings for the CLAP head or not
- Whether we scale the L2-normalized embeddings by
sqrt(embed_dim)(no_scale_embedding=False) or not
As we can see, the dotted lines are almost on top of the solid lines. Notably, sinusoidal positional encoding underperforms significantly compared to learned positional encoding.
As the next step, we scaled our experiments to train on the Pile for one epoch. About half of the examples in the Pile has sequence length > 1024, so we set sequence length to 2048. Even so, ~1/7 of the examples have sequence length > 2048 and had to be discarded. In the end, one epoch consists of 133082 updates and we employ cosine learning rate schedule while "overestimating" the number of training steps by 10%, as inspired by the Chinchilla paper. In addition to the validation MLM perplexity, we also fine-tuned the models on the GLUE benchmark. As in the original RoBERTa paper, we tested both the roberta.base with 125M parameters and roberta.large with 355M parameters. These experiments were performed on 8 x A100 40GB SXM4 GPUs, where the roberta.base experiments took ~3 days and roberta.large experiments took ~9 days. In the table below, PPL is the final validation MLM perplexity, STS-B is the best validation loss, and all the others are the best validation accuracies over 10 epochs of finetuning.
PPL↓ CoLA MNLI MRPC QNLI QQP RTE SST-2 STS-B↓
baseline 2.94 83.6 84.2 90 91.6 91.3 73.6 92.1 0.028
learned-clap 2.86 81.7 84.4 86.3 90.9 91.2 72.6 92.5 0.027
alibi 2.93 69.2 85.1 80.9 92 91.5 63.9 93.1 0.033
zero-clap 2.83 70.5 84.9 75.5 90.6 91.1 54.9 89.7 0.041
*Baseline but with sinusoidal positional encoding instead of learned positional encoding failed to converge.
PPL↓ CoLA MNLI MRPC QNLI QQP RTE SST-2 STS-B↓
baseline* 2.55 83.7 86.8 84.3 92.5 91.8 79.8 93.3 0.027
learned-clap 2.5 84.1 86.3 89.7 92.8 91.7 79.8 93.7 0.023
alibi 2.65 69.1 86.5 68.4 92.4 91.7 52.7 93.6 0.123
zero-clap 2.54 69.1 86.7 81.9 92.2 91.6 52.7 93.1 0.031
*Loss spiked somewhere between 24000-24500 updates and the model failed to recover. Loosely following the practice of 5.1 Training Instability in the PaLM paper, we solved the issue by restarting the training from the 20000 updates checkpoint with the PyTorch random seed changed from 1 to 2.
We found that ALiBi no longer helps lowering the validation MLM perplexity. Furthermore, ALiBi turned out to be harmful for several specific GLUE tasks (CoLA, MRPC, and RTE). CLAP head on its own, however, seems to be competitive and in fact outperforms the baseline with roberta.large.
This seems to be another case where models with lower perplexity do not necessarily yield higher accuracies for downstream tasks and architectural changes beneficial for models at smaller scales do not imply the same for models at larger scales (Tay et al., 2022). CLAP head, however, is simpler than the standard prediction head for MLMs, requires minimal changes, and may be worth trying especially at larger scales.
Final checkpoints for models trained on the Pile:
baseline
learned-clap
alibi
zero-clap
baseline
learned-clap
alibi
zero-clap
To load them, install this fork following the original instructions and download the GPT-2 fairseq dictionary:
wget -O gpt2_bpe/dict.txt https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/dict.txt
Then all of the checkpoints above except the zero-clap ones can load as follows:
$ python
Python 3.8.10 (default, Jun 22 2022, 20:18:18)
[GCC 9.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from fairseq.models.roberta import RobertaModel
>>> roberta = RobertaModel.from_pretrained('/checkpoint-dir', 'learned-clap-large.pt', '/dict-dir')
(...)
>>> roberta.fill_mask('The capital of China is <mask>.', topk=3)
[('The capital of China is Beijing.', 0.7009016871452332, ' Beijing'), ('The capital of China is Shanghai.', 0.23566904664039612, ' Shanghai'), ('The capital of China is Moscow.', 0.010170688852667809, ' Moscow')]
>>>
The zero-clap ones were trained without the last two madeupword's, so you need to delete them from dict.txt before loading, i.e.:
(...)
50009 0
50256 0
madeupword0000 0
madeupword0001 0
madeupword0002 0
$ python
Python 3.8.10 (default, Jun 22 2022, 20:18:18)
[GCC 9.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from fairseq.models.roberta import RobertaModel
>>> roberta = RobertaModel.from_pretrained('/checkpoint-dir', 'zero-clap-large.pt', '/dict-dir')
(...)
>>> roberta.fill_mask('The capital of China is <mask>.', topk=3)
[('The capital of China is Beijing.', 0.7051425576210022, ' Beijing'), ('The capital of China is Shanghai.', 0.21408841013908386, ' Shanghai'), ('The capital of China is Taiwan.', 0.007823833264410496, ' Taiwan')]
>>>
The rest of the original example usage should also just work. While these checkpoints have only been tested with this fork, the baseline ones should also work with the original fairseq repo with minimum changes to the state dict:
>>> path = '/checkpoint-dir/baseline-large.pt'
>>> with open(path, 'rb') as f:
... state = torch.load(f, map_location=torch.device("cpu"))
...
>>>
>>> del state['cfg']['task']['omit_mask']
(...)
>>> torch.save(state, '/checkpoint-dir/compatible.pt')