modeling_pcmlp.py

"""PC-MLP model for HuggingFace Hub with auto_map."""
import math
import torch
import torch.nn as nn
from transformers import PreTrainedModel
from .configuration_pcmlp import PCMLPConfig

class PCMLP(PreTrainedModel):
    config_class = PCMLPConfig

    def __init__(self, config):
        super().__init__(config)
        d = config.vocab_size
        h = config.hidden_dim
        scale = 1.0 / math.sqrt(h)

        self.tok = nn.Embedding(config.vocab_size, d)
        self.pos = nn.Embedding(config.seq_len, d)
        self.W1 = nn.Linear(d, h, bias=False)
        self.W2 = nn.Linear(h, h, bias=False)
        self.head = nn.Linear(h, config.num_classes)

        # match JAX init scale
        nn.init.normal_(self.tok.weight, std=0.05)
        nn.init.normal_(self.pos.weight, std=0.05)
        nn.init.normal_(self.W1.weight, std=scale)
        nn.init.normal_(self.W2.weight, std=scale)
        nn.init.normal_(self.head.weight, std=0.02)

    def forward(self, input_ids, labels=None):
        B, T = input_ids.shape
        pos_ids = torch.arange(T, device=input_ids.device)
        h = self.tok(input_ids) + self.pos(pos_ids)
        h = h.reshape(B, -1)[:, :self.W1.in_features]
        h0 = h
        h1 = torch.nn.functional.gelu(self.W1(h0))
        h2 = torch.nn.functional.gelu(self.W2(h1))
        logits = self.head(h2)

        loss = None
        if labels is not None:
            ce = torch.nn.functional.cross_entropy(logits, labels)
            local = ((h1.mean(1) - h0.mean(1)) ** 2).mean() + \
                    ((h2.mean(1) - h1.mean(1)) ** 2).mean()
            loss = ce + self.config.local_loss_weight * local
        return {"loss": loss, "logits": logits}