upload results

finetune/boolq/all_results.json

@@ -0,0 +1,16 @@
+{
+    "epoch": 10.0,
+    "eval_accuracy": 0.658367931842804,
+    "eval_f1": 0.7282728272827282,
+    "eval_loss": 0.9740044474601746,
+    "eval_mcc": 0.27328046278245693,
+    "eval_runtime": 1.2249,
+    "eval_samples": 723,
+    "eval_samples_per_second": 590.256,
+    "eval_steps_per_second": 74.292,
+    "train_loss": 0.3263397216796875,
+    "train_runtime": 70.9595,
+    "train_samples": 2072,
+    "train_samples_per_second": 291.998,
+    "train_steps_per_second": 4.651
+}

finetune/boolq/checkpoint-200/config.json

@@ -0,0 +1,41 @@
+{
+  "_name_or_path": "omarmomen/sf_babylm_1",
+  "architectures": [
+    "StructFormerModelForSequenceClassification"
+  ],
+  "auto_map": {
+    "AutoConfig": "structformer.StructFormerConfig",
+    "AutoModelForMaskedLM": "structformer.StructFormerModel",
+    "AutoModelForSequenceClassification": "structformer.StructFormerModelForSequenceClassification"
+  },
+  "conv_size": 9,
+  "dropatt": 0.1,
+  "dropout": 0.1,
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 512,
+  "id2label": {
+    "0": 0,
+    "1": 1
+  },
+  "initializer_range": 0.02,
+  "label2id": {
+    "0": 0,
+    "1": 1
+  },
+  "model_type": "structformer",
+  "n_parser_layers": 3,
+  "nhead": 8,
+  "nlayers": 8,
+  "ntokens": 16000,
+  "pad": 1,
+  "pos_emb": true,
+  "problem_type": "single_label_classification",
+  "relations": [
+    "head",
+    "child"
+  ],
+  "relative_bias": false,
+  "torch_dtype": "float32",
+  "transformers_version": "4.26.1",
+  "weight_act": "softmax"
+}

finetune/boolq/checkpoint-200/optimizer.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c3f651c5c9f9981bb14d59d7cb9fa2f4fd0096fc5c4dad0fb5f270c5d9d0205
+size 334487930

finetune/boolq/checkpoint-200/pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4f79d98ac75f5de6364c0f52da1493dc993efa95b33032b8444be736b945bf5
+size 167313902

finetune/boolq/checkpoint-200/rng_state.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:50aec6bc0ddef9c74c961b5729c90a128fd273f00030cf8bb80ad486887cc174
+size 14244

finetune/boolq/checkpoint-200/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b2de95c5fb4e38745233f92e6332f510dfd17896063583c072cd541a34cccb0
+size 1064

finetune/boolq/checkpoint-200/special_tokens_map.json

@@ -0,0 +1,15 @@
+{
+  "bos_token": "<s>",
+  "cls_token": "<s>",
+  "eos_token": "</s>",
+  "mask_token": {
+    "content": "<mask>",
+    "lstrip": true,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<pad>",
+  "sep_token": "</s>",
+  "unk_token": "<unk>"
+}

finetune/boolq/checkpoint-200/structformer.py

@@ -0,0 +1,867 @@
+# coding=utf-8
+# Copyright 2023 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""StructFormer and transformer model."""
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn import init
+from transformers import PretrainedConfig, PreTrainedModel
+from transformers.modeling_outputs import MaskedLMOutput, SequenceClassifierOutput
+
+def _get_activation_fn(activation):
+  """Get specified activation function."""
+  if activation == "relu":
+    return nn.ReLU()
+  elif activation == "gelu":
+    return nn.GELU()
+  elif activation == "leakyrelu":
+    return nn.LeakyReLU()
+
+  raise RuntimeError(
+      "activation should be relu/gelu, not {}".format(activation))
+
+
+class Conv1d(nn.Module):
+  """1D convolution layer."""
+
+  def __init__(self, hidden_size, kernel_size, dilation=1):
+    """Initialization.
+
+    Args:
+      hidden_size: dimension of input embeddings
+      kernel_size: convolution kernel size
+      dilation: the spacing between the kernel points
+    """
+    super(Conv1d, self).__init__()
+
+    if kernel_size % 2 == 0:
+      padding = (kernel_size // 2) * dilation
+      self.shift = True
+    else:
+      padding = ((kernel_size - 1) // 2) * dilation
+      self.shift = False
+    self.conv = nn.Conv1d(
+        hidden_size,
+        hidden_size,
+        kernel_size,
+        padding=padding,
+        dilation=dilation)
+
+  def forward(self, x):
+    """Compute convolution.
+
+    Args:
+      x: input embeddings
+    Returns:
+      conv_output: convolution results
+    """
+
+    if self.shift:
+      return self.conv(x.transpose(1, 2)).transpose(1, 2)[:, 1:]
+    else:
+      return self.conv(x.transpose(1, 2)).transpose(1, 2)
+
+
+class MultiheadAttention(nn.Module):
+  """Multi-head self-attention layer."""
+
+  def __init__(self,
+               embed_dim,
+               num_heads,
+               dropout=0.,
+               bias=True,
+               v_proj=True,
+               out_proj=True,
+               relative_bias=True):
+    """Initialization.
+
+    Args:
+      embed_dim: dimension of input embeddings
+      num_heads: number of self-attention heads
+      dropout: dropout rate
+      bias: bool, indicate whether include bias for linear transformations
+      v_proj: bool, indicate whether project inputs to new values
+      out_proj: bool, indicate whether project outputs to new values
+      relative_bias: bool, indicate whether use a relative position based
+        attention bias
+    """
+
+    super(MultiheadAttention, self).__init__()
+    self.embed_dim = embed_dim
+
+    self.num_heads = num_heads
+    self.drop = nn.Dropout(dropout)
+    self.head_dim = embed_dim // num_heads
+    assert self.head_dim * num_heads == self.embed_dim, ("embed_dim must be "
+                                                         "divisible by "
+                                                         "num_heads")
+
+    self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+    self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+    if v_proj:
+      self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+    else:
+      self.v_proj = nn.Identity()
+
+    if out_proj:
+      self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+    else:
+      self.out_proj = nn.Identity()
+
+    if relative_bias:
+      self.relative_bias = nn.Parameter(torch.zeros((self.num_heads, 512)))
+    else:
+      self.relative_bias = None
+
+    self._reset_parameters()
+
+  def _reset_parameters(self):
+    """Initialize attention parameters."""
+
+    init.xavier_uniform_(self.q_proj.weight)
+    init.constant_(self.q_proj.bias, 0.)
+
+    init.xavier_uniform_(self.k_proj.weight)
+    init.constant_(self.k_proj.bias, 0.)
+
+    if isinstance(self.v_proj, nn.Linear):
+      init.xavier_uniform_(self.v_proj.weight)
+      init.constant_(self.v_proj.bias, 0.)
+
+    if isinstance(self.out_proj, nn.Linear):
+      init.xavier_uniform_(self.out_proj.weight)
+      init.constant_(self.out_proj.bias, 0.)
+
+  def forward(self, query, key_padding_mask=None, attn_mask=None):
+    """Compute multi-head self-attention.
+
+    Args:
+      query: input embeddings
+      key_padding_mask: 3D mask that prevents attention to certain positions
+      attn_mask: 3D mask that rescale the attention weight at each position
+    Returns:
+      attn_output: self-attention output
+    """
+
+    length, bsz, embed_dim = query.size()
+    assert embed_dim == self.embed_dim
+
+    head_dim = embed_dim // self.num_heads
+    assert head_dim * self.num_heads == embed_dim, ("embed_dim must be "
+                                                    "divisible by num_heads")
+    scaling = float(head_dim)**-0.5
+
+    q = self.q_proj(query)
+    k = self.k_proj(query)
+    v = self.v_proj(query)
+
+    q = q * scaling
+
+    if attn_mask is not None:
+      assert list(attn_mask.size()) == [bsz * self.num_heads,
+                                        query.size(0), query.size(0)]
+
+    q = q.contiguous().view(length, bsz * self.num_heads,
+                            head_dim).transpose(0, 1)
+    k = k.contiguous().view(length, bsz * self.num_heads,
+                            head_dim).transpose(0, 1)
+    v = v.contiguous().view(length, bsz * self.num_heads,
+                            head_dim).transpose(0, 1)
+
+    attn_output_weights = torch.bmm(q, k.transpose(1, 2))
+    assert list(
+        attn_output_weights.size()) == [bsz * self.num_heads, length, length]
+
+    if self.relative_bias is not None:
+      pos = torch.arange(length, device=query.device)
+      relative_pos = torch.abs(pos[:, None] - pos[None, :]) + 256
+      relative_pos = relative_pos[None, :, :].expand(bsz * self.num_heads, -1,
+                                                     -1)
+
+      relative_bias = self.relative_bias.repeat_interleave(bsz, dim=0)
+      relative_bias = relative_bias[:, None, :].expand(-1, length, -1)
+      relative_bias = torch.gather(relative_bias, 2, relative_pos)
+      attn_output_weights = attn_output_weights + relative_bias
+
+    if key_padding_mask is not None:
+      attn_output_weights = attn_output_weights + key_padding_mask
+
+    if attn_mask is None:
+      attn_output_weights = torch.softmax(attn_output_weights, dim=-1)
+    else:
+      attn_output_weights = torch.sigmoid(attn_output_weights) * attn_mask
+
+    attn_output_weights = self.drop(attn_output_weights)
+
+    attn_output = torch.bmm(attn_output_weights, v)
+
+    assert list(attn_output.size()) == [bsz * self.num_heads, length, head_dim]
+    attn_output = attn_output.transpose(0, 1).contiguous().view(
+        length, bsz, embed_dim)
+    attn_output = self.out_proj(attn_output)
+
+    return attn_output
+
+
+class TransformerLayer(nn.Module):
+  """TransformerEncoderLayer is made up of self-attn and feedforward network."""
+
+  def __init__(self,
+               d_model,
+               nhead,
+               dim_feedforward=2048,
+               dropout=0.1,
+               dropatt=0.1,
+               activation="leakyrelu",
+               relative_bias=True):
+    """Initialization.
+
+    Args:
+      d_model: dimension of inputs
+      nhead: number of self-attention heads
+      dim_feedforward: dimension of hidden layer in feedforward layer
+      dropout: dropout rate
+      dropatt: drop attention rate
+      activation: activation function
+      relative_bias: bool, indicate whether use a relative position based
+        attention bias
+    """
+
+    super(TransformerLayer, self).__init__()
+    self.self_attn = MultiheadAttention(
+        d_model, nhead, dropout=dropatt, relative_bias=relative_bias)
+    # Implementation of Feedforward model
+    self.feedforward = nn.Sequential(
+        nn.LayerNorm(d_model), nn.Linear(d_model, dim_feedforward),
+        _get_activation_fn(activation), nn.Dropout(dropout),
+        nn.Linear(dim_feedforward, d_model))
+
+    self.norm = nn.LayerNorm(d_model)
+    self.dropout1 = nn.Dropout(dropout)
+    self.dropout2 = nn.Dropout(dropout)
+
+    self.nhead = nhead
+
+  def forward(self, src, attn_mask=None, key_padding_mask=None):
+    """Pass the input through the encoder layer.
+
+    Args:
+      src: the sequence to the encoder layer (required).
+      attn_mask: the mask for the src sequence (optional).
+      key_padding_mask: the mask for the src keys per batch (optional).
+    Returns:
+      src3: the output of transformer layer, share the same shape as src.
+    """
+    src2 = self.self_attn(
+        self.norm(src), attn_mask=attn_mask, key_padding_mask=key_padding_mask)
+    src2 = src + self.dropout1(src2)
+    src3 = self.feedforward(src2)
+    src3 = src2 + self.dropout2(src3)
+
+    return src3
+
+
+def cumprod(x, reverse=False, exclusive=False):
+  """cumulative product."""
+  if reverse:
+    x = x.flip([-1])
+
+  if exclusive:
+    x = F.pad(x[:, :, :-1], (1, 0), value=1)
+
+  cx = x.cumprod(-1)
+
+  if reverse:
+    cx = cx.flip([-1])
+  return cx
+
+def cumsum(x, reverse=False, exclusive=False):
+  """cumulative sum."""
+  bsz, _, length = x.size()
+  device = x.device
+  if reverse:
+    if exclusive:
+      w = torch.ones([bsz, length, length], device=device).tril(-1)
+    else:
+      w = torch.ones([bsz, length, length], device=device).tril(0)
+    cx = torch.bmm(x, w)
+  else:
+    if exclusive:
+      w = torch.ones([bsz, length, length], device=device).triu(1)
+    else:
+      w = torch.ones([bsz, length, length], device=device).triu(0)
+    cx = torch.bmm(x, w)
+  return cx
+
+def cummin(x, reverse=False, exclusive=False, max_value=1e9):
+  """cumulative min."""
+  if reverse:
+    if exclusive:
+      x = F.pad(x[:, :, 1:], (0, 1), value=max_value)
+    x = x.flip([-1]).cummin(-1)[0].flip([-1])
+  else:
+    if exclusive:
+      x = F.pad(x[:, :, :-1], (1, 0), value=max_value)
+    x = x.cummin(-1)[0]
+  return x
+
+class Transformer(nn.Module):
+  """Transformer model."""
+
+  def __init__(self,
+              hidden_size,
+              nlayers,
+              ntokens,
+              nhead=8,
+              dropout=0.1,
+              dropatt=0.1,
+              relative_bias=True,
+              pos_emb=False,
+              pad=0):
+    """Initialization.
+
+    Args:
+      hidden_size: dimension of inputs and hidden states
+      nlayers: number of layers
+      ntokens: number of output categories
+      nhead: number of self-attention heads
+      dropout: dropout rate
+      dropatt: drop attention rate
+      relative_bias: bool, indicate whether use a relative position based
+        attention bias
+      pos_emb: bool, indicate whether use a learnable positional embedding
+      pad: pad token index
+    """
+
+    super(Transformer, self).__init__()
+
+    self.drop = nn.Dropout(dropout)
+
+    self.emb = nn.Embedding(ntokens, hidden_size)
+    if pos_emb:
+      self.pos_emb = nn.Embedding(500, hidden_size)
+
+    self.layers = nn.ModuleList([
+        TransformerLayer(hidden_size, nhead, hidden_size * 4, dropout,
+                                dropatt=dropatt, relative_bias=relative_bias)
+        for _ in range(nlayers)])
+
+    self.norm = nn.LayerNorm(hidden_size)
+
+    self.output_layer = nn.Linear(hidden_size, ntokens)
+    self.output_layer.weight = self.emb.weight
+
+    self.init_weights()
+
+    self.nlayers = nlayers
+    self.nhead = nhead
+    self.ntokens = ntokens
+    self.hidden_size = hidden_size
+    self.pad = pad
+
+  def init_weights(self):
+    """Initialize token embedding and output bias."""
+    initrange = 0.1
+    self.emb.weight.data.uniform_(-initrange, initrange)
+    if hasattr(self, 'pos_emb'):
+      self.pos_emb.weight.data.uniform_(-initrange, initrange)
+    self.output_layer.bias.data.fill_(0)
+
+  def visibility(self, x, device):
+    """Mask pad tokens."""
+    visibility = (x != self.pad).float()
+    visibility = visibility[:, None, :].expand(-1, x.size(1), -1)
+    visibility = torch.repeat_interleave(visibility, self.nhead, dim=0)
+    return visibility.log()
+
+  def encode(self, x, pos):
+    """Standard transformer encode process."""
+    h = self.emb(x)
+    if hasattr(self, 'pos_emb'):
+      h = h + self.pos_emb(pos)
+    h_list = []
+    visibility = self.visibility(x, x.device)
+
+    for i in range(self.nlayers):
+      h_list.append(h)
+      h = self.layers[i](
+          h.transpose(0, 1), key_padding_mask=visibility).transpose(0, 1)
+
+    output = h
+    h_array = torch.stack(h_list, dim=2)
+
+    return output, h_array
+
+  def forward(self, x, pos):
+    """Pass the input through the encoder layer.
+
+    Args:
+      x: input tokens (required).
+      pos: position for each token (optional).
+    Returns:
+      output: probability distributions for missing tokens.
+      state_dict: parsing results and raw output
+    """
+
+    batch_size, length = x.size()
+
+    raw_output, _ = self.encode(x, pos)
+    raw_output = self.norm(raw_output)
+    raw_output = self.drop(raw_output)
+
+    output = self.output_layer(raw_output)
+    return output.view(batch_size * length, -1), {'raw_output': raw_output,}
+
+class StructFormer(Transformer):
+  """StructFormer model."""
+
+  def __init__(self,
+              hidden_size,
+              nlayers,
+              ntokens,
+              nhead=8,
+              dropout=0.1,
+              dropatt=0.1,
+              relative_bias=False,
+              pos_emb=False,
+              pad=0,
+              n_parser_layers=4,
+              conv_size=9,
+              relations=('head', 'child'),
+              weight_act='softmax'):
+    """Initialization.
+
+    Args:
+      hidden_size: dimension of inputs and hidden states
+      nlayers: number of layers
+      ntokens: number of output categories
+      nhead: number of self-attention heads
+      dropout: dropout rate
+      dropatt: drop attention rate
+      relative_bias: bool, indicate whether use a relative position based
+        attention bias
+      pos_emb: bool, indicate whether use a learnable positional embedding
+      pad: pad token index
+      n_parser_layers: number of parsing layers
+      conv_size: convolution kernel size for parser
+      relations: relations that are used to compute self attention
+      weight_act: relations distribution activation function
+    """
+
+    super(StructFormer, self).__init__(
+        hidden_size,
+        nlayers,
+        ntokens,
+        nhead=nhead,
+        dropout=dropout,
+        dropatt=dropatt,
+        relative_bias=relative_bias,
+        pos_emb=pos_emb,
+        pad=pad)
+
+    self.parser_layers = nn.ModuleList([
+        nn.Sequential(Conv1d(hidden_size, conv_size),
+                      nn.LayerNorm(hidden_size, elementwise_affine=False),
+                      nn.Tanh()) for i in range(n_parser_layers)])
+
+    self.distance_ff = nn.Sequential(
+        Conv1d(hidden_size, 2),
+        nn.LayerNorm(hidden_size, elementwise_affine=False), nn.Tanh(),
+        nn.Linear(hidden_size, 1))
+
+    self.height_ff = nn.Sequential(
+        nn.Linear(hidden_size, hidden_size),
+        nn.LayerNorm(hidden_size, elementwise_affine=False), nn.Tanh(),
+        nn.Linear(hidden_size, 1))
+
+    n_rel = len(relations)
+    self._rel_weight = nn.Parameter(torch.zeros((nlayers, nhead, n_rel)))
+    self._rel_weight.data.normal_(0, 0.1)
+
+    self._scaler = nn.Parameter(torch.zeros(2))
+
+    self.n_parse_layers = n_parser_layers
+    self.weight_act = weight_act
+    self.relations = relations
+
+  @property
+  def scaler(self):
+    return self._scaler.exp()
+
+  @property
+  def rel_weight(self):
+    if self.weight_act == 'sigmoid':
+      return torch.sigmoid(self._rel_weight)
+    elif self.weight_act == 'softmax':
+      return torch.softmax(self._rel_weight, dim=-1)
+
+  def parse(self, x, pos):
+    """Parse input sentence.
+
+    Args:
+      x: input tokens (required).
+      pos: position for each token (optional).
+    Returns:
+      distance: syntactic distance
+      height: syntactic height
+    """
+
+    mask = (x != self.pad)
+    mask_shifted = F.pad(mask[:, 1:], (0, 1), value=0)
+
+    h = self.emb(x)
+    for i in range(self.n_parse_layers):
+      h = h.masked_fill(~mask[:, :, None], 0)
+      h = self.parser_layers[i](h)
+
+    height = self.height_ff(h).squeeze(-1)
+    height.masked_fill_(~mask, -1e9)
+
+    distance = self.distance_ff(h).squeeze(-1)
+    distance.masked_fill_(~mask_shifted, 1e9)
+
+    # Calbrating the distance and height to the same level
+    length = distance.size(1)
+    height_max = height[:, None, :].expand(-1, length, -1)
+    height_max = torch.cummax(
+        height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9,
+        dim=-1)[0].triu(0)
+
+    margin_left = torch.relu(
+        F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max)
+    margin_right = torch.relu(distance[:, None, :] - height_max)
+    margin = torch.where(margin_left > margin_right, margin_right,
+                         margin_left).triu(0)
+
+    margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+    margin.masked_fill_(~margin_mask, 0)
+    margin = margin.max()
+
+    distance = distance - margin
+
+    return distance, height
+
+  def compute_block(self, distance, height):
+    """Compute constituents from distance and height."""
+
+    beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler[0]
+
+    gamma = torch.sigmoid(-beta_logits)
+    ones = torch.ones_like(gamma)
+
+    block_mask_left = cummin(
+        gamma.tril(-1) + ones.triu(0), reverse=True, max_value=1)
+    block_mask_left = block_mask_left - F.pad(
+        block_mask_left[:, :, :-1], (1, 0), value=0)
+    block_mask_left.tril_(0)
+
+    block_mask_right = cummin(
+        gamma.triu(0) + ones.tril(-1), exclusive=True, max_value=1)
+    block_mask_right = block_mask_right - F.pad(
+        block_mask_right[:, :, 1:], (0, 1), value=0)
+    block_mask_right.triu_(0)
+
+    block_p = block_mask_left[:, :, :, None] * block_mask_right[:, :, None, :]
+    block = cumsum(block_mask_left).tril(0) + cumsum(
+        block_mask_right, reverse=True).triu(1)
+
+    return block_p, block
+
+  def compute_head(self, height):
+    """Estimate head for each constituent."""
+
+    _, length = height.size()
+    head_logits = height * self.scaler[1]
+    index = torch.arange(length, device=height.device)
+
+    mask = (index[:, None, None] <= index[None, None, :]) * (
+        index[None, None, :] <= index[None, :, None])
+    head_logits = head_logits[:, None, None, :].repeat(1, length, length, 1)
+    head_logits.masked_fill_(~mask[None, :, :, :], -1e9)
+
+    head_p = torch.softmax(head_logits, dim=-1)
+
+    return head_p
+
+  def generate_mask(self, x, distance, height):
+    """Compute head and cibling distribution for each token."""
+
+    bsz, length = x.size()
+
+    eye = torch.eye(length, device=x.device, dtype=torch.bool)
+    eye = eye[None, :, :].expand((bsz, -1, -1))
+
+    block_p, block = self.compute_block(distance, height)
+    head_p = self.compute_head(height)
+    head = torch.einsum('blij,bijh->blh', block_p, head_p)
+    head = head.masked_fill(eye, 0)
+    child = head.transpose(1, 2)
+    cibling = torch.bmm(head, child).masked_fill(eye, 0)
+
+    rel_list = []
+    if 'head' in self.relations:
+      rel_list.append(head)
+    if 'child' in self.relations:
+      rel_list.append(child)
+    if 'cibling' in self.relations:
+      rel_list.append(cibling)
+
+    rel = torch.stack(rel_list, dim=1)
+
+    rel_weight = self.rel_weight
+
+    dep = torch.einsum('lhr,brij->lbhij', rel_weight, rel)
+    att_mask = dep.reshape(self.nlayers, bsz * self.nhead, length, length)
+
+    return att_mask, cibling, head, block
+
+  def encode(self, x, pos, att_mask):
+    """Structformer encoding process."""
+
+    visibility = self.visibility(x, x.device)
+    h = self.emb(x)
+    if hasattr(self, 'pos_emb'):
+      assert pos.max() < 500
+      h = h + self.pos_emb(pos)
+    for i in range(self.nlayers):
+      h = self.layers[i](
+          h.transpose(0, 1), attn_mask=att_mask[i],
+          key_padding_mask=visibility).transpose(0, 1)
+    return h
+
+  def forward(self, x, pos):
+    """Pass the input through the encoder layer.
+
+    Args:
+      x: input tokens (required).
+      pos: position for each token (optional).
+    Returns:
+      output: probability distributions for missing tokens.
+      state_dict: parsing results and raw output
+    """
+
+    batch_size, length = x.size()
+
+    distance, height = self.parse(x, pos)
+    att_mask, cibling, head, block = self.generate_mask(x, distance, height)
+
+    raw_output = self.encode(x, pos, att_mask)
+    raw_output = self.norm(raw_output)
+    raw_output = self.drop(raw_output)
+
+    output = self.output_layer(raw_output)
+
+    return output.view(batch_size * length, -1), \
+        {'raw_output': raw_output, 'distance': distance, 'height': height,
+        'cibling': cibling, 'head': head, 'block': block}
+
+
+##########################################
+# Clasication Head For BabyLM Evaluation Tasks
+##########################################
+class ClassificationHead(nn.Module):
+    """Head for sentence-level classification tasks."""
+    def __init__(self, config):
+        super(ClassificationHead, self).__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
+
+    def forward(self, features, **kwargs):
+        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
+        x = self.dropout(x)
+        x = self.dense(x)
+        x = torch.tanh(x)
+        x = self.dropout(x)
+        x = self.out_proj(x)
+        return x
+
+##########################################
+# HuggingFace Config
+##########################################
+class StructFormerConfig(PretrainedConfig):
+    model_type = "structformer"
+
+    def __init__(
+        self,
+        hidden_size=512,
+        nlayers=8,
+        ntokens=10_000,
+        nhead=8,
+        dropout=0.1,
+        dropatt=0.1,
+        relative_bias=False,
+        pos_emb=False,
+        pad=0,
+        n_parser_layers=4,
+        conv_size=9,
+        relations=('head', 'child'),
+        weight_act='softmax',
+        num_labels=1,
+        hidden_dropout_prob=0.1,
+        initializer_range=0.02,
+        **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.nlayers = nlayers
+        self.ntokens = ntokens
+        self.nhead = nhead
+        self.dropout = dropout
+        self.dropatt = dropatt
+        self.relative_bias = relative_bias
+        self.pos_emb = pos_emb
+        self.pad = pad
+        self.n_parser_layers = n_parser_layers
+        self.conv_size = conv_size
+        self.relations = relations
+        self.weight_act = weight_act
+        self.num_labels = num_labels
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.initializer_range=initializer_range
+        super().__init__(**kwargs)
+
+##########################################
+# HuggingFace Model
+##########################################
+class StructFormerModel(PreTrainedModel):
+    config_class = StructFormerConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = StructFormer(
+              hidden_size=config.hidden_size,
+              nlayers=config.nlayers,
+              ntokens=config.ntokens,
+              nhead=config.nhead,
+              dropout=config.dropout,
+              dropatt=config.dropatt,
+              relative_bias=config.relative_bias,
+              pos_emb=config.pos_emb,
+              pad=config.pad,
+              n_parser_layers=config.n_parser_layers,
+              conv_size=config.conv_size,
+              relations=config.relations,
+              weight_act=config.weight_act
+        )
+        self.config = config
+
+    def parse(self, input_ids, **kwargs):
+      x = input_ids
+      batch_size, length = x.size()
+      pos = kwargs['position_ids'] if 'position_ids' in kwargs.keys() else torch.arange(length, device=x.device).expand(batch_size, length)
+      
+      sf_output = self.model(x, pos)
+      
+      return sf_output[1]
+    
+    def forward(self, input_ids, labels=None, **kwargs):
+        x = input_ids
+        batch_size, length = x.size()
+        pos = kwargs['position_ids'] if 'position_ids' in kwargs.keys() else torch.arange(length, device=x.device).expand(batch_size, length)
+        
+        sf_output = self.model(x, pos)
+        
+        loss = None
+        if labels is not None:
+          loss_fct = nn.CrossEntropyLoss()
+          loss = loss_fct(sf_output[0], labels.reshape(-1))
+        
+        return MaskedLMOutput(
+            loss=loss, # shape: 1
+            logits=sf_output[0].view(batch_size, length, -1), # shape: (batch_size, length, ntokens)
+            hidden_states=None,
+            attentions=None
+        )
+
+class StructFormerModelForSequenceClassification(PreTrainedModel):
+    config_class = StructFormerConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = StructFormer(
+              hidden_size=config.hidden_size,
+              nlayers=config.nlayers,
+              ntokens=config.ntokens,
+              nhead=config.nhead,
+              dropout=config.dropout,
+              dropatt=config.dropatt,
+              relative_bias=config.relative_bias,
+              pos_emb=config.pos_emb,
+              pad=config.pad,
+              n_parser_layers=config.n_parser_layers,
+              conv_size=config.conv_size,
+              relations=config.relations,
+              weight_act=config.weight_act
+        )
+        self.config = config
+        self.num_labels = config.num_labels
+        self.model.classifier = ClassificationHead(config)
+    
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, nn.Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            if module.bias is not None:
+                module.bias.data.zero_()
+                module.weight.data.fill_(1.0)
+    
+    def forward(self, input_ids, labels=None, **kwargs):
+        x = input_ids
+        batch_size, length = x.size()
+        pos = kwargs['position_ids'] if 'position_ids' in kwargs.keys() else torch.arange(length, device=x.device).expand(batch_size, length)
+        
+        sf_output = self.model(x, pos)
+        
+        logits = self.model.classifier(sf_output[1]['raw_output'])
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = nn.MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = nn.CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = nn.BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=None,
+            attentions=None,
+        )

finetune/boolq/checkpoint-200/tokenizer_config.json

@@ -0,0 +1,65 @@
+{
+  "add_prefix_space": false,
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "cls_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "errors": "replace",
+  "mask_token": {
+    "__type": "AddedToken",
+    "content": "<mask>",
+    "lstrip": true,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "model_max_length": 512,
+  "name_or_path": "omarmomen/sf_babylm_1",
+  "pad_token": {
+    "__type": "AddedToken",
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "sep_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "special_tokens_map_file": null,
+  "tokenizer_class": "RobertaTokenizer",
+  "trim_offsets": true,
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

finetune/boolq/checkpoint-200/trainer_state.json

@@ -0,0 +1,27 @@
+{
+  "best_metric": 0.7282728272827282,
+  "best_model_checkpoint": "omarmomen/sf_babylm_1/finetune/boolq/checkpoint-200",
+  "epoch": 6.0606060606060606,
+  "global_step": 200,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 6.06,
+      "eval_accuracy": 0.658367931842804,
+      "eval_f1": 0.7282728272827282,
+      "eval_loss": 0.9740044474601746,
+      "eval_mcc": 0.27328046278245693,
+      "eval_runtime": 1.2669,
+      "eval_samples_per_second": 570.67,
+      "eval_steps_per_second": 71.827,
+      "step": 200
+    }
+  ],
+  "max_steps": 330,
+  "num_train_epochs": 10,
+  "total_flos": 321868603269120.0,
+  "trial_name": null,
+  "trial_params": null
+}

finetune/boolq/checkpoint-200/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:553f78e16a096fe90d36b0a7936e92787a1172a9f02cf4e4611e107248227ef4
+size 3960

finetune/boolq/config.json

@@ -0,0 +1,41 @@
+{
+  "_name_or_path": "omarmomen/sf_babylm_1",
+  "architectures": [
+    "StructFormerModelForSequenceClassification"
+  ],
+  "auto_map": {
+    "AutoConfig": "structformer.StructFormerConfig",
+    "AutoModelForMaskedLM": "structformer.StructFormerModel",
+    "AutoModelForSequenceClassification": "structformer.StructFormerModelForSequenceClassification"
+  },
+  "conv_size": 9,
+  "dropatt": 0.1,
+  "dropout": 0.1,
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 512,
+  "id2label": {
+    "0": 0,
+    "1": 1
+  },
+  "initializer_range": 0.02,
+  "label2id": {
+    "0": 0,
+    "1": 1
+  },
+  "model_type": "structformer",
+  "n_parser_layers": 3,
+  "nhead": 8,
+  "nlayers": 8,
+  "ntokens": 16000,
+  "pad": 1,
+  "pos_emb": true,
+  "problem_type": "single_label_classification",
+  "relations": [
+    "head",
+    "child"
+  ],
+  "relative_bias": false,
+  "torch_dtype": "float32",
+  "transformers_version": "4.26.1",
+  "weight_act": "softmax"
+}

finetune/boolq/eval_results.json

@@ -0,0 +1,11 @@
+{
+    "epoch": 10.0,
+    "eval_accuracy": 0.658367931842804,
+    "eval_f1": 0.7282728272827282,
+    "eval_loss": 0.9740044474601746,
+    "eval_mcc": 0.27328046278245693,
+    "eval_runtime": 1.2249,
+    "eval_samples": 723,
+    "eval_samples_per_second": 590.256,
+    "eval_steps_per_second": 74.292
+}

finetune/boolq/predict_results.txt

@@ -0,0 +1,724 @@
+index	prediction
+0	1
+1	1
+2	1
+3	1
+4	1
+5	1
+6	1
+7	1
+8	1
+9	0
+10	1
+11	1
+12	0
+13	0
+14	1
+15	1
+16	1
+17	1
+18	1
+19	1
+20	0
+21	1
+22	1
+23	0
+24	0
+25	0
+26	1
+27	1
+28	0
+29	1
+30	1
+31	1
+32	0
+33	1
+34	1
+35	1
+36	1
+37	1
+38	1
+39	1
+40	0
+41	0
+42	1
+43	1
+44	1
+45	1
+46	1
+47	1
+48	1
+49	1
+50	1
+51	1
+52	0
+53	1
+54	0
+55	1
+56	1
+57	0
+58	0
+59	0
+60	1
+61	0
+62	1
+63	0
+64	1
+65	1
+66	1
+67	0
+68	1
+69	1
+70	1
+71	0
+72	0
+73	1
+74	0
+75	0
+76	1
+77	0
+78	1
+79	0
+80	1
+81	0
+82	0
+83	1
+84	1
+85	1
+86	0
+87	0
+88	0
+89	1
+90	1
+91	0
+92	0
+93	1
+94	0
+95	1
+96	0
+97	1
+98	0
+99	1
+100	1
+101	0
+102	1
+103	0
+104	0
+105	1
+106	1
+107	0
+108	1
+109	1
+110	1
+111	1
+112	0
+113	0
+114	1
+115	0
+116	0
+117	1
+118	1
+119	1
+120	1
+121	1
+122	1
+123	0
+124	1
+125	1
+126	0
+127	1
+128	1
+129	1
+130	0
+131	0
+132	0
+133	1
+134	1
+135	1
+136	1
+137	0
+138	1
+139	0
+140	0
+141	1
+142	1
+143	1
+144	1
+145	0
+146	1
+147	0
+148	1
+149	1
+150	0
+151	1
+152	1
+153	1
+154	1
+155	0
+156	0
+157	1
+158	1
+159	1
+160	1
+161	1
+162	1
+163	0
+164	1
+165	0
+166	1
+167	0
+168	0
+169	0
+170	1
+171	0
+172	0
+173	1
+174	1
+175	1
+176	1
+177	1
+178	0
+179	1
+180	0
+181	1
+182	1
+183	1
+184	0
+185	1
+186	1
+187	0
+188	1
+189	1
+190	1
+191	0
+192	1
+193	0
+194	1
+195	0
+196	1
+197	1
+198	0
+199	1
+200	1
+201	1
+202	1
+203	0
+204	1
+205	1
+206	1
+207	1
+208	1
+209	1
+210	1
+211	1
+212	1
+213	1
+214	1
+215	1
+216	1
+217	1
+218	1
+219	0
+220	1
+221	1
+222	1
+223	0
+224	0
+225	0
+226	1
+227	1
+228	0
+229	1
+230	0
+231	0
+232	0
+233	0
+234	0
+235	1
+236	0
+237	1
+238	1
+239	1
+240	1
+241	0
+242	1
+243	1
+244	0
+245	0
+246	1
+247	0
+248	0
+249	1
+250	1
+251	0
+252	1
+253	1
+254	1
+255	0
+256	0
+257	0
+258	1
+259	0
+260	1
+261	1
+262	1
+263	1
+264	1
+265	1
+266	1
+267	0
+268	1
+269	1
+270	0
+271	1
+272	1
+273	1
+274	1
+275	1
+276	0
+277	1
+278	1
+279	0
+280	0
+281	1
+282	0
+283	0
+284	1
+285	0
+286	1
+287	1
+288	1
+289	0
+290	1
+291	0
+292	1
+293	1
+294	1
+295	1
+296	1
+297	0
+298	1
+299	0
+300	1
+301	0
+302	1
+303	0
+304	0
+305	0
+306	0
+307	0
+308	0
+309	0
+310	0
+311	1
+312	1
+313	1
+314	0
+315	1
+316	1
+317	1
+318	1
+319	1
+320	0
+321	1
+322	1
+323	1
+324	1
+325	0
+326	1
+327	1
+328	1
+329	1
+330	0
+331	1
+332	1
+333	1
+334	0
+335	0
+336	0
+337	1
+338	1
+339	1
+340	1
+341	0
+342	1
+343	0
+344	0
+345	1
+346	1
+347	1
+348	1
+349	0
+350	0
+351	0
+352	1
+353	1
+354	1
+355	1
+356	0
+357	1
+358	0
+359	1
+360	1
+361	1
+362	0
+363	1
+364	0
+365	1
+366	0
+367	1
+368	1
+369	1
+370	0
+371	1
+372	1
+373	1
+374	1
+375	1
+376	1
+377	1
+378	0
+379	0
+380	0
+381	1
+382	1
+383	0
+384	1
+385	0
+386	1
+387	1
+388	1
+389	0
+390	1
+391	0
+392	1
+393	1
+394	1
+395	0
+396	1
+397	1
+398	1
+399	1
+400	1
+401	0
+402	1
+403	1
+404	1
+405	1
+406	0
+407	0
+408	1
+409	1
+410	1
+411	0
+412	0
+413	1
+414	1
+415	0
+416	1
+417	0
+418	0
+419	1
+420	1
+421	0
+422	1
+423	0
+424	1
+425	0
+426	1
+427	1
+428	1
+429	1
+430	1
+431	0
+432	1
+433	1
+434	1
+435	0
+436	0
+437	1
+438	1
+439	1
+440	0
+441	0
+442	0
+443	1
+444	0
+445	1
+446	1
+447	0
+448	1
+449	1
+450	0
+451	1
+452	1
+453	0
+454	1
+455	0
+456	1
+457	0
+458	1
+459	1
+460	1
+461	1
+462	0
+463	1
+464	0
+465	0
+466	1
+467	1
+468	1
+469	0
+470	1
+471	1
+472	1
+473	1
+474	1
+475	1
+476	1
+477	0
+478	0
+479	1
+480	1
+481	1
+482	0
+483	1
+484	1
+485	1
+486	1
+487	1
+488	1
+489	0
+490	1
+491	1
+492	1
+493	0
+494	1
+495	1
+496	1
+497	0
+498	0
+499	1
+500	1
+501	1
+502	0
+503	1
+504	0
+505	0
+506	1
+507	0
+508	1
+509	1
+510	1
+511	1
+512	1
+513	1
+514	0
+515	0
+516	1
+517	1
+518	1
+519	1
+520	1
+521	1
+522	0
+523	0
+524	1
+525	1
+526	0
+527	1
+528	0
+529	0
+530	1
+531	1
+532	0
+533	1
+534	1
+535	1
+536	1
+537	1
+538	1
+539	0
+540	1
+541	1
+542	1
+543	0
+544	1
+545	1
+546	0
+547	1
+548	1
+549	1
+550	1
+551	0
+552	0
+553	0
+554	0
+555	0
+556	1
+557	0
+558	0
+559	1
+560	0
+561	1
+562	1
+563	1
+564	0
+565	0
+566	1
+567	1
+568	1
+569	1
+570	1
+571	1
+572	1
+573	1
+574	1
+575	1
+576	1
+577	1
+578	1
+579	1
+580	1
+581	0
+582	1
+583	1
+584	1
+585	1
+586	1
+587	1
+588	0
+589	0
+590	1
+591	1
+592	0
+593	1
+594	1
+595	1
+596	0
+597	0
+598	0
+599	1
+600	0
+601	0
+602	1
+603	1
+604	1
+605	1
+606	1
+607	1
+608	1
+609	1
+610	1
+611	1
+612	0
+613	1
+614	1
+615	0
+616	1
+617	1
+618	1
+619	1
+620	1
+621	1
+622	0
+623	1
+624	1
+625	0
+626	1
+627	1
+628	1
+629	0
+630	1
+631	1
+632	0
+633	1
+634	1
+635	0
+636	1
+637	1
+638	1
+639	0
+640	1
+641	0
+642	0
+643	0
+644	0
+645	1
+646	1
+647	0
+648	1
+649	1
+650	1
+651	1
+652	1
+653	1
+654	0
+655	1
+656	1
+657	0
+658	0
+659	1
+660	0
+661	0
+662	1
+663	1
+664	1
+665	1
+666	0
+667	1
+668	0
+669	1
+670	1
+671	1
+672	1
+673	1
+674	0
+675	1
+676	1
+677	1
+678	0
+679	0
+680	1
+681	1
+682	1
+683	1
+684	0
+685	1
+686	1
+687	0
+688	0
+689	1
+690	1
+691	0
+692	0
+693	1
+694	1
+695	1
+696	1
+697	1
+698	1
+699	0
+700	1
+701	0
+702	1
+703	1
+704	1
+705	1
+706	1
+707	1
+708	1
+709	1
+710	1
+711	0
+712	1
+713	0
+714	0
+715	1
+716	0
+717	0
+718	0
+719	0
+720	0
+721	1
+722	1

finetune/boolq/pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4f79d98ac75f5de6364c0f52da1493dc993efa95b33032b8444be736b945bf5
+size 167313902

finetune/boolq/special_tokens_map.json

@@ -0,0 +1,15 @@
+{
+  "bos_token": "<s>",
+  "cls_token": "<s>",
+  "eos_token": "</s>",
+  "mask_token": {
+    "content": "<mask>",
+    "lstrip": true,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<pad>",
+  "sep_token": "</s>",
+  "unk_token": "<unk>"
+}

finetune/boolq/structformer.py

@@ -0,0 +1,867 @@
+# coding=utf-8
+# Copyright 2023 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""StructFormer and transformer model."""
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn import init
+from transformers import PretrainedConfig, PreTrainedModel
+from transformers.modeling_outputs import MaskedLMOutput, SequenceClassifierOutput
+
+def _get_activation_fn(activation):
+  """Get specified activation function."""
+  if activation == "relu":
+    return nn.ReLU()
+  elif activation == "gelu":
+    return nn.GELU()
+  elif activation == "leakyrelu":
+    return nn.LeakyReLU()
+
+  raise RuntimeError(
+      "activation should be relu/gelu, not {}".format(activation))
+
+
+class Conv1d(nn.Module):
+  """1D convolution layer."""
+
+  def __init__(self, hidden_size, kernel_size, dilation=1):
+    """Initialization.
+
+    Args:
+      hidden_size: dimension of input embeddings
+      kernel_size: convolution kernel size
+      dilation: the spacing between the kernel points
+    """
+    super(Conv1d, self).__init__()
+
+    if kernel_size % 2 == 0:
+      padding = (kernel_size // 2) * dilation
+      self.shift = True
+    else:
+      padding = ((kernel_size - 1) // 2) * dilation
+      self.shift = False
+    self.conv = nn.Conv1d(
+        hidden_size,
+        hidden_size,
+        kernel_size,
+        padding=padding,
+        dilation=dilation)
+
+  def forward(self, x):
+    """Compute convolution.
+
+    Args:
+      x: input embeddings
+    Returns:
+      conv_output: convolution results
+    """
+
+    if self.shift:
+      return self.conv(x.transpose(1, 2)).transpose(1, 2)[:, 1:]
+    else:
+      return self.conv(x.transpose(1, 2)).transpose(1, 2)
+
+
+class MultiheadAttention(nn.Module):
+  """Multi-head self-attention layer."""
+
+  def __init__(self,
+               embed_dim,
+               num_heads,
+               dropout=0.,
+               bias=True,
+               v_proj=True,
+               out_proj=True,
+               relative_bias=True):
+    """Initialization.
+
+    Args:
+      embed_dim: dimension of input embeddings
+      num_heads: number of self-attention heads
+      dropout: dropout rate
+      bias: bool, indicate whether include bias for linear transformations
+      v_proj: bool, indicate whether project inputs to new values
+      out_proj: bool, indicate whether project outputs to new values
+      relative_bias: bool, indicate whether use a relative position based
+        attention bias
+    """
+
+    super(MultiheadAttention, self).__init__()
+    self.embed_dim = embed_dim
+
+    self.num_heads = num_heads
+    self.drop = nn.Dropout(dropout)
+    self.head_dim = embed_dim // num_heads
+    assert self.head_dim * num_heads == self.embed_dim, ("embed_dim must be "
+                                                         "divisible by "
+                                                         "num_heads")
+
+    self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+    self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+    if v_proj:
+      self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+    else:
+      self.v_proj = nn.Identity()
+
+    if out_proj:
+      self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+    else:
+      self.out_proj = nn.Identity()
+
+    if relative_bias:
+      self.relative_bias = nn.Parameter(torch.zeros((self.num_heads, 512)))
+    else:
+      self.relative_bias = None
+
+    self._reset_parameters()
+
+  def _reset_parameters(self):
+    """Initialize attention parameters."""
+
+    init.xavier_uniform_(self.q_proj.weight)
+    init.constant_(self.q_proj.bias, 0.)
+
+    init.xavier_uniform_(self.k_proj.weight)
+    init.constant_(self.k_proj.bias, 0.)
+
+    if isinstance(self.v_proj, nn.Linear):
+      init.xavier_uniform_(self.v_proj.weight)
+      init.constant_(self.v_proj.bias, 0.)
+
+    if isinstance(self.out_proj, nn.Linear):
+      init.xavier_uniform_(self.out_proj.weight)
+      init.constant_(self.out_proj.bias, 0.)
+
+  def forward(self, query, key_padding_mask=None, attn_mask=None):
+    """Compute multi-head self-attention.
+
+    Args:
+      query: input embeddings
+      key_padding_mask: 3D mask that prevents attention to certain positions
+      attn_mask: 3D mask that rescale the attention weight at each position
+    Returns:
+      attn_output: self-attention output
+    """
+
+    length, bsz, embed_dim = query.size()
+    assert embed_dim == self.embed_dim
+
+    head_dim = embed_dim // self.num_heads
+    assert head_dim * self.num_heads == embed_dim, ("embed_dim must be "
+                                                    "divisible by num_heads")
+    scaling = float(head_dim)**-0.5
+
+    q = self.q_proj(query)
+    k = self.k_proj(query)
+    v = self.v_proj(query)
+
+    q = q * scaling
+
+    if attn_mask is not None:
+      assert list(attn_mask.size()) == [bsz * self.num_heads,
+                                        query.size(0), query.size(0)]
+
+    q = q.contiguous().view(length, bsz * self.num_heads,
+                            head_dim).transpose(0, 1)
+    k = k.contiguous().view(length, bsz * self.num_heads,
+                            head_dim).transpose(0, 1)
+    v = v.contiguous().view(length, bsz * self.num_heads,
+                            head_dim).transpose(0, 1)
+
+    attn_output_weights = torch.bmm(q, k.transpose(1, 2))
+    assert list(
+        attn_output_weights.size()) == [bsz * self.num_heads, length, length]
+
+    if self.relative_bias is not None:
+      pos = torch.arange(length, device=query.device)
+      relative_pos = torch.abs(pos[:, None] - pos[None, :]) + 256
+      relative_pos = relative_pos[None, :, :].expand(bsz * self.num_heads, -1,
+                                                     -1)
+
+      relative_bias = self.relative_bias.repeat_interleave(bsz, dim=0)
+      relative_bias = relative_bias[:, None, :].expand(-1, length, -1)
+      relative_bias = torch.gather(relative_bias, 2, relative_pos)
+      attn_output_weights = attn_output_weights + relative_bias
+
+    if key_padding_mask is not None:
+      attn_output_weights = attn_output_weights + key_padding_mask
+
+    if attn_mask is None:
+      attn_output_weights = torch.softmax(attn_output_weights, dim=-1)
+    else:
+      attn_output_weights = torch.sigmoid(attn_output_weights) * attn_mask
+
+    attn_output_weights = self.drop(attn_output_weights)
+
+    attn_output = torch.bmm(attn_output_weights, v)
+
+    assert list(attn_output.size()) == [bsz * self.num_heads, length, head_dim]
+    attn_output = attn_output.transpose(0, 1).contiguous().view(
+        length, bsz, embed_dim)
+    attn_output = self.out_proj(attn_output)
+
+    return attn_output
+
+
+class TransformerLayer(nn.Module):
+  """TransformerEncoderLayer is made up of self-attn and feedforward network."""
+
+  def __init__(self,
+               d_model,
+               nhead,
+               dim_feedforward=2048,
+               dropout=0.1,
+               dropatt=0.1,
+               activation="leakyrelu",
+               relative_bias=True):
+    """Initialization.
+
+    Args:
+      d_model: dimension of inputs
+      nhead: number of self-attention heads
+      dim_feedforward: dimension of hidden layer in feedforward layer
+      dropout: dropout rate
+      dropatt: drop attention rate
+      activation: activation function
+      relative_bias: bool, indicate whether use a relative position based
+        attention bias
+    """
+
+    super(TransformerLayer, self).__init__()
+    self.self_attn = MultiheadAttention(
+        d_model, nhead, dropout=dropatt, relative_bias=relative_bias)
+    # Implementation of Feedforward model
+    self.feedforward = nn.Sequential(
+        nn.LayerNorm(d_model), nn.Linear(d_model, dim_feedforward),
+        _get_activation_fn(activation), nn.Dropout(dropout),
+        nn.Linear(dim_feedforward, d_model))
+
+    self.norm = nn.LayerNorm(d_model)
+    self.dropout1 = nn.Dropout(dropout)
+    self.dropout2 = nn.Dropout(dropout)
+
+    self.nhead = nhead
+
+  def forward(self, src, attn_mask=None, key_padding_mask=None):
+    """Pass the input through the encoder layer.
+
+    Args:
+      src: the sequence to the encoder layer (required).
+      attn_mask: the mask for the src sequence (optional).
+      key_padding_mask: the mask for the src keys per batch (optional).
+    Returns:
+      src3: the output of transformer layer, share the same shape as src.
+    """
+    src2 = self.self_attn(
+        self.norm(src), attn_mask=attn_mask, key_padding_mask=key_padding_mask)
+    src2 = src + self.dropout1(src2)
+    src3 = self.feedforward(src2)
+    src3 = src2 + self.dropout2(src3)
+
+    return src3
+
+
+def cumprod(x, reverse=False, exclusive=False):
+  """cumulative product."""
+  if reverse:
+    x = x.flip([-1])
+
+  if exclusive:
+    x = F.pad(x[:, :, :-1], (1, 0), value=1)
+
+  cx = x.cumprod(-1)
+
+  if reverse:
+    cx = cx.flip([-1])
+  return cx
+
+def cumsum(x, reverse=False, exclusive=False):
+  """cumulative sum."""
+  bsz, _, length = x.size()
+  device = x.device
+  if reverse:
+    if exclusive:
+      w = torch.ones([bsz, length, length], device=device).tril(-1)
+    else:
+      w = torch.ones([bsz, length, length], device=device).tril(0)
+    cx = torch.bmm(x, w)
+  else:
+    if exclusive:
+      w = torch.ones([bsz, length, length], device=device).triu(1)
+    else:
+      w = torch.ones([bsz, length, length], device=device).triu(0)
+    cx = torch.bmm(x, w)
+  return cx
+
+def cummin(x, reverse=False, exclusive=False, max_value=1e9):
+  """cumulative min."""
+  if reverse:
+    if exclusive:
+      x = F.pad(x[:, :, 1:], (0, 1), value=max_value)
+    x = x.flip([-1]).cummin(-1)[0].flip([-1])
+  else:
+    if exclusive:
+      x = F.pad(x[:, :, :-1], (1, 0), value=max_value)
+    x = x.cummin(-1)[0]
+  return x
+
+class Transformer(nn.Module):
+  """Transformer model."""
+
+  def __init__(self,
+              hidden_size,
+              nlayers,
+              ntokens,
+              nhead=8,
+              dropout=0.1,
+              dropatt=0.1,
+              relative_bias=True,
+              pos_emb=False,
+              pad=0):
+    """Initialization.
+
+    Args:
+      hidden_size: dimension of inputs and hidden states
+      nlayers: number of layers
+      ntokens: number of output categories
+      nhead: number of self-attention heads
+      dropout: dropout rate
+      dropatt: drop attention rate
+      relative_bias: bool, indicate whether use a relative position based
+        attention bias
+      pos_emb: bool, indicate whether use a learnable positional embedding
+      pad: pad token index
+    """
+
+    super(Transformer, self).__init__()
+
+    self.drop = nn.Dropout(dropout)
+
+    self.emb = nn.Embedding(ntokens, hidden_size)
+    if pos_emb:
+      self.pos_emb = nn.Embedding(500, hidden_size)
+
+    self.layers = nn.ModuleList([
+        TransformerLayer(hidden_size, nhead, hidden_size * 4, dropout,
+                                dropatt=dropatt, relative_bias=relative_bias)
+        for _ in range(nlayers)])
+
+    self.norm = nn.LayerNorm(hidden_size)
+
+    self.output_layer = nn.Linear(hidden_size, ntokens)
+    self.output_layer.weight = self.emb.weight
+
+    self.init_weights()
+
+    self.nlayers = nlayers
+    self.nhead = nhead
+    self.ntokens = ntokens
+    self.hidden_size = hidden_size
+    self.pad = pad
+
+  def init_weights(self):
+    """Initialize token embedding and output bias."""
+    initrange = 0.1
+    self.emb.weight.data.uniform_(-initrange, initrange)
+    if hasattr(self, 'pos_emb'):
+      self.pos_emb.weight.data.uniform_(-initrange, initrange)
+    self.output_layer.bias.data.fill_(0)
+
+  def visibility(self, x, device):
+    """Mask pad tokens."""
+    visibility = (x != self.pad).float()
+    visibility = visibility[:, None, :].expand(-1, x.size(1), -1)
+    visibility = torch.repeat_interleave(visibility, self.nhead, dim=0)
+    return visibility.log()
+
+  def encode(self, x, pos):
+    """Standard transformer encode process."""
+    h = self.emb(x)
+    if hasattr(self, 'pos_emb'):
+      h = h + self.pos_emb(pos)
+    h_list = []
+    visibility = self.visibility(x, x.device)
+
+    for i in range(self.nlayers):
+      h_list.append(h)
+      h = self.layers[i](
+          h.transpose(0, 1), key_padding_mask=visibility).transpose(0, 1)
+
+    output = h
+    h_array = torch.stack(h_list, dim=2)
+
+    return output, h_array
+
+  def forward(self, x, pos):
+    """Pass the input through the encoder layer.
+
+    Args:
+      x: input tokens (required).
+      pos: position for each token (optional).
+    Returns:
+      output: probability distributions for missing tokens.
+      state_dict: parsing results and raw output
+    """
+
+    batch_size, length = x.size()
+
+    raw_output, _ = self.encode(x, pos)
+    raw_output = self.norm(raw_output)
+    raw_output = self.drop(raw_output)
+
+    output = self.output_layer(raw_output)
+    return output.view(batch_size * length, -1), {'raw_output': raw_output,}
+
+class StructFormer(Transformer):
+  """StructFormer model."""
+
+  def __init__(self,
+              hidden_size,
+              nlayers,
+              ntokens,
+              nhead=8,
+              dropout=0.1,
+              dropatt=0.1,
+              relative_bias=False,
+              pos_emb=False,
+              pad=0,
+              n_parser_layers=4,
+              conv_size=9,
+              relations=('head', 'child'),
+              weight_act='softmax'):
+    """Initialization.
+
+    Args:
+      hidden_size: dimension of inputs and hidden states
+      nlayers: number of layers
+      ntokens: number of output categories
+      nhead: number of self-attention heads
+      dropout: dropout rate
+      dropatt: drop attention rate
+      relative_bias: bool, indicate whether use a relative position based
+        attention bias
+      pos_emb: bool, indicate whether use a learnable positional embedding
+      pad: pad token index
+      n_parser_layers: number of parsing layers
+      conv_size: convolution kernel size for parser
+      relations: relations that are used to compute self attention
+      weight_act: relations distribution activation function
+    """
+
+    super(StructFormer, self).__init__(
+        hidden_size,
+        nlayers,
+        ntokens,
+        nhead=nhead,
+        dropout=dropout,
+        dropatt=dropatt,
+        relative_bias=relative_bias,
+        pos_emb=pos_emb,
+        pad=pad)
+
+    self.parser_layers = nn.ModuleList([
+        nn.Sequential(Conv1d(hidden_size, conv_size),
+                      nn.LayerNorm(hidden_size, elementwise_affine=False),
+                      nn.Tanh()) for i in range(n_parser_layers)])
+
+    self.distance_ff = nn.Sequential(
+        Conv1d(hidden_size, 2),
+        nn.LayerNorm(hidden_size, elementwise_affine=False), nn.Tanh(),
+        nn.Linear(hidden_size, 1))
+
+    self.height_ff = nn.Sequential(
+        nn.Linear(hidden_size, hidden_size),
+        nn.LayerNorm(hidden_size, elementwise_affine=False), nn.Tanh(),
+        nn.Linear(hidden_size, 1))
+
+    n_rel = len(relations)
+    self._rel_weight = nn.Parameter(torch.zeros((nlayers, nhead, n_rel)))
+    self._rel_weight.data.normal_(0, 0.1)
+
+    self._scaler = nn.Parameter(torch.zeros(2))
+
+    self.n_parse_layers = n_parser_layers
+    self.weight_act = weight_act
+    self.relations = relations
+
+  @property
+  def scaler(self):
+    return self._scaler.exp()
+
+  @property
+  def rel_weight(self):
+    if self.weight_act == 'sigmoid':
+      return torch.sigmoid(self._rel_weight)
+    elif self.weight_act == 'softmax':
+      return torch.softmax(self._rel_weight, dim=-1)
+
+  def parse(self, x, pos):
+    """Parse input sentence.
+
+    Args:
+      x: input tokens (required).
+      pos: position for each token (optional).
+    Returns:
+      distance: syntactic distance
+      height: syntactic height
+    """
+
+    mask = (x != self.pad)
+    mask_shifted = F.pad(mask[:, 1:], (0, 1), value=0)
+
+    h = self.emb(x)
+    for i in range(self.n_parse_layers):
+      h = h.masked_fill(~mask[:, :, None], 0)
+      h = self.parser_layers[i](h)
+
+    height = self.height_ff(h).squeeze(-1)
+    height.masked_fill_(~mask, -1e9)
+
+    distance = self.distance_ff(h).squeeze(-1)
+    distance.masked_fill_(~mask_shifted, 1e9)
+
+    # Calbrating the distance and height to the same level
+    length = distance.size(1)
+    height_max = height[:, None, :].expand(-1, length, -1)
+    height_max = torch.cummax(
+        height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9,
+        dim=-1)[0].triu(0)
+
+    margin_left = torch.relu(
+        F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max)
+    margin_right = torch.relu(distance[:, None, :] - height_max)
+    margin = torch.where(margin_left > margin_right, margin_right,
+                         margin_left).triu(0)
+
+    margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+    margin.masked_fill_(~margin_mask, 0)
+    margin = margin.max()
+
+    distance = distance - margin
+
+    return distance, height
+
+  def compute_block(self, distance, height):
+    """Compute constituents from distance and height."""
+
+    beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler[0]
+
+    gamma = torch.sigmoid(-beta_logits)
+    ones = torch.ones_like(gamma)
+
+    block_mask_left = cummin(
+        gamma.tril(-1) + ones.triu(0), reverse=True, max_value=1)
+    block_mask_left = block_mask_left - F.pad(
+        block_mask_left[:, :, :-1], (1, 0), value=0)
+    block_mask_left.tril_(0)
+
+    block_mask_right = cummin(
+        gamma.triu(0) + ones.tril(-1), exclusive=True, max_value=1)
+    block_mask_right = block_mask_right - F.pad(
+        block_mask_right[:, :, 1:], (0, 1), value=0)
+    block_mask_right.triu_(0)
+
+    block_p = block_mask_left[:, :, :, None] * block_mask_right[:, :, None, :]
+    block = cumsum(block_mask_left).tril(0) + cumsum(
+        block_mask_right, reverse=True).triu(1)
+
+    return block_p, block
+
+  def compute_head(self, height):
+    """Estimate head for each constituent."""
+
+    _, length = height.size()
+    head_logits = height * self.scaler[1]
+    index = torch.arange(length, device=height.device)
+
+    mask = (index[:, None, None] <= index[None, None, :]) * (
+        index[None, None, :] <= index[None, :, None])
+    head_logits = head_logits[:, None, None, :].repeat(1, length, length, 1)
+    head_logits.masked_fill_(~mask[None, :, :, :], -1e9)
+
+    head_p = torch.softmax(head_logits, dim=-1)
+
+    return head_p
+
+  def generate_mask(self, x, distance, height):
+    """Compute head and cibling distribution for each token."""
+
+    bsz, length = x.size()
+
+    eye = torch.eye(length, device=x.device, dtype=torch.bool)
+    eye = eye[None, :, :].expand((bsz, -1, -1))
+
+    block_p, block = self.compute_block(distance, height)
+    head_p = self.compute_head(height)
+    head = torch.einsum('blij,bijh->blh', block_p, head_p)
+    head = head.masked_fill(eye, 0)
+    child = head.transpose(1, 2)
+    cibling = torch.bmm(head, child).masked_fill(eye, 0)
+
+    rel_list = []
+    if 'head' in self.relations:
+      rel_list.append(head)
+    if 'child' in self.relations:
+      rel_list.append(child)
+    if 'cibling' in self.relations:
+      rel_list.append(cibling)
+
+    rel = torch.stack(rel_list, dim=1)
+
+    rel_weight = self.rel_weight
+
+    dep = torch.einsum('lhr,brij->lbhij', rel_weight, rel)
+    att_mask = dep.reshape(self.nlayers, bsz * self.nhead, length, length)
+
+    return att_mask, cibling, head, block
+
+  def encode(self, x, pos, att_mask):
+    """Structformer encoding process."""
+
+    visibility = self.visibility(x, x.device)
+    h = self.emb(x)
+    if hasattr(self, 'pos_emb'):
+      assert pos.max() < 500
+      h = h + self.pos_emb(pos)
+    for i in range(self.nlayers):
+      h = self.layers[i](
+          h.transpose(0, 1), attn_mask=att_mask[i],
+          key_padding_mask=visibility).transpose(0, 1)
+    return h
+
+  def forward(self, x, pos):
+    """Pass the input through the encoder layer.
+
+    Args:
+      x: input tokens (required).
+      pos: position for each token (optional).
+    Returns:
+      output: probability distributions for missing tokens.
+      state_dict: parsing results and raw output
+    """
+
+    batch_size, length = x.size()
+
+    distance, height = self.parse(x, pos)
+    att_mask, cibling, head, block = self.generate_mask(x, distance, height)
+
+    raw_output = self.encode(x, pos, att_mask)
+    raw_output = self.norm(raw_output)
+    raw_output = self.drop(raw_output)
+
+    output = self.output_layer(raw_output)
+
+    return output.view(batch_size * length, -1), \
+        {'raw_output': raw_output, 'distance': distance, 'height': height,
+        'cibling': cibling, 'head': head, 'block': block}
+
+
+##########################################
+# Clasication Head For BabyLM Evaluation Tasks
+##########################################
+class ClassificationHead(nn.Module):
+    """Head for sentence-level classification tasks."""
+    def __init__(self, config):
+        super(ClassificationHead, self).__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
+
+    def forward(self, features, **kwargs):
+        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
+        x = self.dropout(x)
+        x = self.dense(x)
+        x = torch.tanh(x)
+        x = self.dropout(x)
+        x = self.out_proj(x)
+        return x
+
+##########################################
+# HuggingFace Config
+##########################################
+class StructFormerConfig(PretrainedConfig):
+    model_type = "structformer"
+
+    def __init__(
+        self,
+        hidden_size=512,
+        nlayers=8,
+        ntokens=10_000,
+        nhead=8,
+        dropout=0.1,
+        dropatt=0.1,
+        relative_bias=False,
+        pos_emb=False,
+        pad=0,
+        n_parser_layers=4,
+        conv_size=9,
+        relations=('head', 'child'),
+        weight_act='softmax',
+        num_labels=1,
+        hidden_dropout_prob=0.1,
+        initializer_range=0.02,
+        **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.nlayers = nlayers
+        self.ntokens = ntokens
+        self.nhead = nhead
+        self.dropout = dropout
+        self.dropatt = dropatt
+        self.relative_bias = relative_bias
+        self.pos_emb = pos_emb
+        self.pad = pad
+        self.n_parser_layers = n_parser_layers
+        self.conv_size = conv_size
+        self.relations = relations
+        self.weight_act = weight_act
+        self.num_labels = num_labels
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.initializer_range=initializer_range
+        super().__init__(**kwargs)
+
+##########################################
+# HuggingFace Model
+##########################################
+class StructFormerModel(PreTrainedModel):
+    config_class = StructFormerConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = StructFormer(
+              hidden_size=config.hidden_size,
+              nlayers=config.nlayers,
+              ntokens=config.ntokens,
+              nhead=config.nhead,
+              dropout=config.dropout,
+              dropatt=config.dropatt,
+              relative_bias=config.relative_bias,
+              pos_emb=config.pos_emb,
+              pad=config.pad,
+              n_parser_layers=config.n_parser_layers,
+              conv_size=config.conv_size,
+              relations=config.relations,
+              weight_act=config.weight_act
+        )
+        self.config = config
+
+    def parse(self, input_ids, **kwargs):
+      x = input_ids
+      batch_size, length = x.size()
+      pos = kwargs['position_ids'] if 'position_ids' in kwargs.keys() else torch.arange(length, device=x.device).expand(batch_size, length)
+      
+      sf_output = self.model(x, pos)
+      
+      return sf_output[1]
+    
+    def forward(self, input_ids, labels=None, **kwargs):
+        x = input_ids
+        batch_size, length = x.size()
+        pos = kwargs['position_ids'] if 'position_ids' in kwargs.keys() else torch.arange(length, device=x.device).expand(batch_size, length)
+        
+        sf_output = self.model(x, pos)
+        
+        loss = None
+        if labels is not None:
+          loss_fct = nn.CrossEntropyLoss()
+          loss = loss_fct(sf_output[0], labels.reshape(-1))
+        
+        return MaskedLMOutput(
+            loss=loss, # shape: 1
+            logits=sf_output[0].view(batch_size, length, -1), # shape: (batch_size, length, ntokens)
+            hidden_states=None,
+            attentions=None
+        )
+
+class StructFormerModelForSequenceClassification(PreTrainedModel):
+    config_class = StructFormerConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = StructFormer(
+              hidden_size=config.hidden_size,
+              nlayers=config.nlayers,
+              ntokens=config.ntokens,
+              nhead=config.nhead,
+              dropout=config.dropout,
+              dropatt=config.dropatt,
+              relative_bias=config.relative_bias,
+              pos_emb=config.pos_emb,
+              pad=config.pad,
+              n_parser_layers=config.n_parser_layers,
+              conv_size=config.conv_size,
+              relations=config.relations,
+              weight_act=config.weight_act
+        )
+        self.config = config
+        self.num_labels = config.num_labels
+        self.model.classifier = ClassificationHead(config)
+    
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, nn.Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            if module.bias is not None:
+                module.bias.data.zero_()
+                module.weight.data.fill_(1.0)
+    
+    def forward(self, input_ids, labels=None, **kwargs):
+        x = input_ids
+        batch_size, length = x.size()
+        pos = kwargs['position_ids'] if 'position_ids' in kwargs.keys() else torch.arange(length, device=x.device).expand(batch_size, length)
+        
+        sf_output = self.model(x, pos)
+        
+        logits = self.model.classifier(sf_output[1]['raw_output'])
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = nn.MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = nn.CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = nn.BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=None,
+            attentions=None,
+        )

finetune/boolq/tokenizer_config.json

@@ -0,0 +1,65 @@
+{
+  "add_prefix_space": false,
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "cls_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "errors": "replace",
+  "mask_token": {
+    "__type": "AddedToken",
+    "content": "<mask>",
+    "lstrip": true,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "model_max_length": 512,
+  "name_or_path": "omarmomen/sf_babylm_1",
+  "pad_token": {
+    "__type": "AddedToken",
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "sep_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "special_tokens_map_file": null,
+  "tokenizer_class": "RobertaTokenizer",
+  "trim_offsets": true,
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

finetune/boolq/train_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 10.0,
+    "train_loss": 0.3263397216796875,
+    "train_runtime": 70.9595,
+    "train_samples": 2072,
+    "train_samples_per_second": 291.998,
+    "train_steps_per_second": 4.651
+}

finetune/boolq/trainer_state.json

@@ -0,0 +1,36 @@
+{
+  "best_metric": 0.7282728272827282,
+  "best_model_checkpoint": "omarmomen/sf_babylm_1/finetune/boolq/checkpoint-200",
+  "epoch": 10.0,
+  "global_step": 330,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 6.06,
+      "eval_accuracy": 0.658367931842804,
+      "eval_f1": 0.7282728272827282,
+      "eval_loss": 0.9740044474601746,
+      "eval_mcc": 0.27328046278245693,
+      "eval_runtime": 1.2669,
+      "eval_samples_per_second": 570.67,
+      "eval_steps_per_second": 71.827,
+      "step": 200
+    },
+    {
+      "epoch": 10.0,
+      "step": 330,
+      "total_flos": 530980689469440.0,
+      "train_loss": 0.3263397216796875,
+      "train_runtime": 70.9595,
+      "train_samples_per_second": 291.998,
+      "train_steps_per_second": 4.651
+    }
+  ],
+  "max_steps": 330,
+  "num_train_epochs": 10,
+  "total_flos": 530980689469440.0,
+  "trial_name": null,
+  "trial_params": null
+}

finetune/boolq/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:553f78e16a096fe90d36b0a7936e92787a1172a9f02cf4e4611e107248227ef4
+size 3960

finetune/cola/all_results.json

@@ -0,0 +1,16 @@
+{
+    "epoch": 10.0,
+    "eval_accuracy": 0.7036309838294983,
+    "eval_f1": 0.8219339622641509,
+    "eval_loss": 0.6300716400146484,
+    "eval_mcc": 0.13076987235140602,
+    "eval_runtime": 1.7861,
+    "eval_samples": 1019,
+    "eval_samples_per_second": 570.517,
+    "eval_steps_per_second": 71.665,
+    "train_loss": 0.3184558808803558,
+    "train_runtime": 219.0841,
+    "train_samples": 8164,
+    "train_samples_per_second": 372.642,
+    "train_steps_per_second": 5.843
+}

finetune/cola/checkpoint-200/config.json

@@ -0,0 +1,41 @@
+{
+  "_name_or_path": "omarmomen/sf_babylm_1",
+  "architectures": [
+    "StructFormerModelForSequenceClassification"
+  ],
+  "auto_map": {
+    "AutoConfig": "structformer.StructFormerConfig",
+    "AutoModelForMaskedLM": "structformer.StructFormerModel",
+    "AutoModelForSequenceClassification": "structformer.StructFormerModelForSequenceClassification"
+  },
+  "conv_size": 9,
+  "dropatt": 0.1,
+  "dropout": 0.1,
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 512,
+  "id2label": {
+    "0": 0,
+    "1": 1
+  },
+  "initializer_range": 0.02,
+  "label2id": {
+    "0": 0,
+    "1": 1
+  },
+  "model_type": "structformer",
+  "n_parser_layers": 3,
+  "nhead": 8,
+  "nlayers": 8,
+  "ntokens": 16000,
+  "pad": 1,
+  "pos_emb": true,
+  "problem_type": "single_label_classification",
+  "relations": [
+    "head",
+    "child"
+  ],
+  "relative_bias": false,
+  "torch_dtype": "float32",
+  "transformers_version": "4.26.1",
+  "weight_act": "softmax"
+}

finetune/cola/checkpoint-200/optimizer.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:64479433f30cd6e1151d4bc028baf56e84cd5d91db9e9a98e4b4a26a4d9f2dd7
+size 334487930

finetune/cola/checkpoint-200/pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:782e41676fcfa55bd9f10569a44c812fc36f983ee47f32bd0cafdee7d732ee83
+size 167313902

finetune/cola/checkpoint-200/rng_state.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dcdca2997df87a0fc6c3a6ebc025ec0c13497f4dd53314621d65b6268c45356d
+size 14244

finetune/cola/checkpoint-200/scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:79d27acb67b7f4c7f14c70b0036e640a0cdc532e992a5af9432f781d3c7ad02e
+size 1064

finetune/cola/checkpoint-200/special_tokens_map.json

@@ -0,0 +1,15 @@
+{
+  "bos_token": "<s>",
+  "cls_token": "<s>",
+  "eos_token": "</s>",
+  "mask_token": {
+    "content": "<mask>",
+    "lstrip": true,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<pad>",
+  "sep_token": "</s>",
+  "unk_token": "<unk>"
+}

finetune/cola/checkpoint-200/structformer.py

@@ -0,0 +1,867 @@
+# coding=utf-8
+# Copyright 2023 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""StructFormer and transformer model."""
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn import init
+from transformers import PretrainedConfig, PreTrainedModel
+from transformers.modeling_outputs import MaskedLMOutput, SequenceClassifierOutput
+
+def _get_activation_fn(activation):
+  """Get specified activation function."""
+  if activation == "relu":
+    return nn.ReLU()
+  elif activation == "gelu":
+    return nn.GELU()
+  elif activation == "leakyrelu":
+    return nn.LeakyReLU()
+
+  raise RuntimeError(
+      "activation should be relu/gelu, not {}".format(activation))
+
+
+class Conv1d(nn.Module):
+  """1D convolution layer."""
+
+  def __init__(self, hidden_size, kernel_size, dilation=1):
+    """Initialization.
+
+    Args:
+      hidden_size: dimension of input embeddings
+      kernel_size: convolution kernel size
+      dilation: the spacing between the kernel points
+    """
+    super(Conv1d, self).__init__()
+
+    if kernel_size % 2 == 0:
+      padding = (kernel_size // 2) * dilation
+      self.shift = True
+    else:
+      padding = ((kernel_size - 1) // 2) * dilation
+      self.shift = False
+    self.conv = nn.Conv1d(
+        hidden_size,
+        hidden_size,
+        kernel_size,
+        padding=padding,
+        dilation=dilation)
+
+  def forward(self, x):
+    """Compute convolution.
+
+    Args:
+      x: input embeddings
+    Returns:
+      conv_output: convolution results
+    """
+
+    if self.shift:
+      return self.conv(x.transpose(1, 2)).transpose(1, 2)[:, 1:]
+    else:
+      return self.conv(x.transpose(1, 2)).transpose(1, 2)
+
+
+class MultiheadAttention(nn.Module):
+  """Multi-head self-attention layer."""
+
+  def __init__(self,
+               embed_dim,
+               num_heads,
+               dropout=0.,
+               bias=True,
+               v_proj=True,
+               out_proj=True,
+               relative_bias=True):
+    """Initialization.
+
+    Args:
+      embed_dim: dimension of input embeddings
+      num_heads: number of self-attention heads
+      dropout: dropout rate
+      bias: bool, indicate whether include bias for linear transformations
+      v_proj: bool, indicate whether project inputs to new values
+      out_proj: bool, indicate whether project outputs to new values
+      relative_bias: bool, indicate whether use a relative position based
+        attention bias
+    """
+
+    super(MultiheadAttention, self).__init__()
+    self.embed_dim = embed_dim
+
+    self.num_heads = num_heads
+    self.drop = nn.Dropout(dropout)
+    self.head_dim = embed_dim // num_heads
+    assert self.head_dim * num_heads == self.embed_dim, ("embed_dim must be "
+                                                         "divisible by "
+                                                         "num_heads")
+
+    self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+    self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+    if v_proj:
+      self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+    else:
+      self.v_proj = nn.Identity()
+
+    if out_proj:
+      self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+    else:
+      self.out_proj = nn.Identity()
+
+    if relative_bias:
+      self.relative_bias = nn.Parameter(torch.zeros((self.num_heads, 512)))
+    else:
+      self.relative_bias = None
+
+    self._reset_parameters()
+
+  def _reset_parameters(self):
+    """Initialize attention parameters."""
+
+    init.xavier_uniform_(self.q_proj.weight)
+    init.constant_(self.q_proj.bias, 0.)
+
+    init.xavier_uniform_(self.k_proj.weight)
+    init.constant_(self.k_proj.bias, 0.)
+
+    if isinstance(self.v_proj, nn.Linear):
+      init.xavier_uniform_(self.v_proj.weight)
+      init.constant_(self.v_proj.bias, 0.)
+
+    if isinstance(self.out_proj, nn.Linear):
+      init.xavier_uniform_(self.out_proj.weight)
+      init.constant_(self.out_proj.bias, 0.)
+
+  def forward(self, query, key_padding_mask=None, attn_mask=None):
+    """Compute multi-head self-attention.
+
+    Args:
+      query: input embeddings
+      key_padding_mask: 3D mask that prevents attention to certain positions
+      attn_mask: 3D mask that rescale the attention weight at each position
+    Returns:
+      attn_output: self-attention output
+    """
+
+    length, bsz, embed_dim = query.size()
+    assert embed_dim == self.embed_dim
+
+    head_dim = embed_dim // self.num_heads
+    assert head_dim * self.num_heads == embed_dim, ("embed_dim must be "
+                                                    "divisible by num_heads")
+    scaling = float(head_dim)**-0.5
+
+    q = self.q_proj(query)
+    k = self.k_proj(query)
+    v = self.v_proj(query)
+
+    q = q * scaling
+
+    if attn_mask is not None:
+      assert list(attn_mask.size()) == [bsz * self.num_heads,
+                                        query.size(0), query.size(0)]
+
+    q = q.contiguous().view(length, bsz * self.num_heads,
+                            head_dim).transpose(0, 1)
+    k = k.contiguous().view(length, bsz * self.num_heads,
+                            head_dim).transpose(0, 1)
+    v = v.contiguous().view(length, bsz * self.num_heads,
+                            head_dim).transpose(0, 1)
+
+    attn_output_weights = torch.bmm(q, k.transpose(1, 2))
+    assert list(
+        attn_output_weights.size()) == [bsz * self.num_heads, length, length]
+
+    if self.relative_bias is not None:
+      pos = torch.arange(length, device=query.device)
+      relative_pos = torch.abs(pos[:, None] - pos[None, :]) + 256
+      relative_pos = relative_pos[None, :, :].expand(bsz * self.num_heads, -1,
+                                                     -1)
+
+      relative_bias = self.relative_bias.repeat_interleave(bsz, dim=0)
+      relative_bias = relative_bias[:, None, :].expand(-1, length, -1)
+      relative_bias = torch.gather(relative_bias, 2, relative_pos)
+      attn_output_weights = attn_output_weights + relative_bias
+
+    if key_padding_mask is not None:
+      attn_output_weights = attn_output_weights + key_padding_mask
+
+    if attn_mask is None:
+      attn_output_weights = torch.softmax(attn_output_weights, dim=-1)
+    else:
+      attn_output_weights = torch.sigmoid(attn_output_weights) * attn_mask
+
+    attn_output_weights = self.drop(attn_output_weights)
+
+    attn_output = torch.bmm(attn_output_weights, v)
+
+    assert list(attn_output.size()) == [bsz * self.num_heads, length, head_dim]
+    attn_output = attn_output.transpose(0, 1).contiguous().view(
+        length, bsz, embed_dim)
+    attn_output = self.out_proj(attn_output)
+
+    return attn_output
+
+
+class TransformerLayer(nn.Module):
+  """TransformerEncoderLayer is made up of self-attn and feedforward network."""
+
+  def __init__(self,
+               d_model,
+               nhead,
+               dim_feedforward=2048,
+               dropout=0.1,
+               dropatt=0.1,
+               activation="leakyrelu",
+               relative_bias=True):
+    """Initialization.
+
+    Args:
+      d_model: dimension of inputs
+      nhead: number of self-attention heads
+      dim_feedforward: dimension of hidden layer in feedforward layer
+      dropout: dropout rate
+      dropatt: drop attention rate
+      activation: activation function
+      relative_bias: bool, indicate whether use a relative position based
+        attention bias
+    """
+
+    super(TransformerLayer, self).__init__()
+    self.self_attn = MultiheadAttention(
+        d_model, nhead, dropout=dropatt, relative_bias=relative_bias)
+    # Implementation of Feedforward model
+    self.feedforward = nn.Sequential(
+        nn.LayerNorm(d_model), nn.Linear(d_model, dim_feedforward),
+        _get_activation_fn(activation), nn.Dropout(dropout),
+        nn.Linear(dim_feedforward, d_model))
+
+    self.norm = nn.LayerNorm(d_model)
+    self.dropout1 = nn.Dropout(dropout)
+    self.dropout2 = nn.Dropout(dropout)
+
+    self.nhead = nhead
+
+  def forward(self, src, attn_mask=None, key_padding_mask=None):
+    """Pass the input through the encoder layer.
+
+    Args:
+      src: the sequence to the encoder layer (required).
+      attn_mask: the mask for the src sequence (optional).
+      key_padding_mask: the mask for the src keys per batch (optional).
+    Returns:
+      src3: the output of transformer layer, share the same shape as src.
+    """
+    src2 = self.self_attn(
+        self.norm(src), attn_mask=attn_mask, key_padding_mask=key_padding_mask)
+    src2 = src + self.dropout1(src2)
+    src3 = self.feedforward(src2)
+    src3 = src2 + self.dropout2(src3)
+
+    return src3
+
+
+def cumprod(x, reverse=False, exclusive=False):
+  """cumulative product."""
+  if reverse:
+    x = x.flip([-1])
+
+  if exclusive:
+    x = F.pad(x[:, :, :-1], (1, 0), value=1)
+
+  cx = x.cumprod(-1)
+
+  if reverse:
+    cx = cx.flip([-1])
+  return cx
+
+def cumsum(x, reverse=False, exclusive=False):
+  """cumulative sum."""
+  bsz, _, length = x.size()
+  device = x.device
+  if reverse:
+    if exclusive:
+      w = torch.ones([bsz, length, length], device=device).tril(-1)
+    else:
+      w = torch.ones([bsz, length, length], device=device).tril(0)
+    cx = torch.bmm(x, w)
+  else:
+    if exclusive:
+      w = torch.ones([bsz, length, length], device=device).triu(1)
+    else:
+      w = torch.ones([bsz, length, length], device=device).triu(0)
+    cx = torch.bmm(x, w)
+  return cx
+
+def cummin(x, reverse=False, exclusive=False, max_value=1e9):
+  """cumulative min."""
+  if reverse:
+    if exclusive:
+      x = F.pad(x[:, :, 1:], (0, 1), value=max_value)
+    x = x.flip([-1]).cummin(-1)[0].flip([-1])
+  else:
+    if exclusive:
+      x = F.pad(x[:, :, :-1], (1, 0), value=max_value)
+    x = x.cummin(-1)[0]
+  return x
+
+class Transformer(nn.Module):
+  """Transformer model."""
+
+  def __init__(self,
+              hidden_size,
+              nlayers,
+              ntokens,
+              nhead=8,
+              dropout=0.1,
+              dropatt=0.1,
+              relative_bias=True,
+              pos_emb=False,
+              pad=0):
+    """Initialization.
+
+    Args:
+      hidden_size: dimension of inputs and hidden states
+      nlayers: number of layers
+      ntokens: number of output categories
+      nhead: number of self-attention heads
+      dropout: dropout rate
+      dropatt: drop attention rate
+      relative_bias: bool, indicate whether use a relative position based
+        attention bias
+      pos_emb: bool, indicate whether use a learnable positional embedding
+      pad: pad token index
+    """
+
+    super(Transformer, self).__init__()
+
+    self.drop = nn.Dropout(dropout)
+
+    self.emb = nn.Embedding(ntokens, hidden_size)
+    if pos_emb:
+      self.pos_emb = nn.Embedding(500, hidden_size)
+
+    self.layers = nn.ModuleList([
+        TransformerLayer(hidden_size, nhead, hidden_size * 4, dropout,
+                                dropatt=dropatt, relative_bias=relative_bias)
+        for _ in range(nlayers)])
+
+    self.norm = nn.LayerNorm(hidden_size)
+
+    self.output_layer = nn.Linear(hidden_size, ntokens)
+    self.output_layer.weight = self.emb.weight
+
+    self.init_weights()
+
+    self.nlayers = nlayers
+    self.nhead = nhead
+    self.ntokens = ntokens
+    self.hidden_size = hidden_size
+    self.pad = pad
+
+  def init_weights(self):
+    """Initialize token embedding and output bias."""
+    initrange = 0.1
+    self.emb.weight.data.uniform_(-initrange, initrange)
+    if hasattr(self, 'pos_emb'):
+      self.pos_emb.weight.data.uniform_(-initrange, initrange)
+    self.output_layer.bias.data.fill_(0)
+
+  def visibility(self, x, device):
+    """Mask pad tokens."""
+    visibility = (x != self.pad).float()
+    visibility = visibility[:, None, :].expand(-1, x.size(1), -1)
+    visibility = torch.repeat_interleave(visibility, self.nhead, dim=0)
+    return visibility.log()
+
+  def encode(self, x, pos):
+    """Standard transformer encode process."""
+    h = self.emb(x)
+    if hasattr(self, 'pos_emb'):
+      h = h + self.pos_emb(pos)
+    h_list = []
+    visibility = self.visibility(x, x.device)
+
+    for i in range(self.nlayers):
+      h_list.append(h)
+      h = self.layers[i](
+          h.transpose(0, 1), key_padding_mask=visibility).transpose(0, 1)
+
+    output = h
+    h_array = torch.stack(h_list, dim=2)
+
+    return output, h_array
+
+  def forward(self, x, pos):
+    """Pass the input through the encoder layer.
+
+    Args:
+      x: input tokens (required).
+      pos: position for each token (optional).
+    Returns:
+      output: probability distributions for missing tokens.
+      state_dict: parsing results and raw output
+    """
+
+    batch_size, length = x.size()
+
+    raw_output, _ = self.encode(x, pos)
+    raw_output = self.norm(raw_output)
+    raw_output = self.drop(raw_output)
+
+    output = self.output_layer(raw_output)
+    return output.view(batch_size * length, -1), {'raw_output': raw_output,}
+
+class StructFormer(Transformer):
+  """StructFormer model."""
+
+  def __init__(self,
+              hidden_size,
+              nlayers,
+              ntokens,
+              nhead=8,
+              dropout=0.1,
+              dropatt=0.1,
+              relative_bias=False,
+              pos_emb=False,
+              pad=0,
+              n_parser_layers=4,
+              conv_size=9,
+              relations=('head', 'child'),
+              weight_act='softmax'):
+    """Initialization.
+
+    Args:
+      hidden_size: dimension of inputs and hidden states
+      nlayers: number of layers
+      ntokens: number of output categories
+      nhead: number of self-attention heads
+      dropout: dropout rate
+      dropatt: drop attention rate
+      relative_bias: bool, indicate whether use a relative position based
+        attention bias
+      pos_emb: bool, indicate whether use a learnable positional embedding
+      pad: pad token index
+      n_parser_layers: number of parsing layers
+      conv_size: convolution kernel size for parser
+      relations: relations that are used to compute self attention
+      weight_act: relations distribution activation function
+    """
+
+    super(StructFormer, self).__init__(
+        hidden_size,
+        nlayers,
+        ntokens,
+        nhead=nhead,
+        dropout=dropout,
+        dropatt=dropatt,
+        relative_bias=relative_bias,
+        pos_emb=pos_emb,
+        pad=pad)
+
+    self.parser_layers = nn.ModuleList([
+        nn.Sequential(Conv1d(hidden_size, conv_size),
+                      nn.LayerNorm(hidden_size, elementwise_affine=False),
+                      nn.Tanh()) for i in range(n_parser_layers)])
+
+    self.distance_ff = nn.Sequential(
+        Conv1d(hidden_size, 2),
+        nn.LayerNorm(hidden_size, elementwise_affine=False), nn.Tanh(),
+        nn.Linear(hidden_size, 1))
+
+    self.height_ff = nn.Sequential(
+        nn.Linear(hidden_size, hidden_size),
+        nn.LayerNorm(hidden_size, elementwise_affine=False), nn.Tanh(),
+        nn.Linear(hidden_size, 1))
+
+    n_rel = len(relations)
+    self._rel_weight = nn.Parameter(torch.zeros((nlayers, nhead, n_rel)))
+    self._rel_weight.data.normal_(0, 0.1)
+
+    self._scaler = nn.Parameter(torch.zeros(2))
+
+    self.n_parse_layers = n_parser_layers
+    self.weight_act = weight_act
+    self.relations = relations
+
+  @property
+  def scaler(self):
+    return self._scaler.exp()
+
+  @property
+  def rel_weight(self):
+    if self.weight_act == 'sigmoid':
+      return torch.sigmoid(self._rel_weight)
+    elif self.weight_act == 'softmax':
+      return torch.softmax(self._rel_weight, dim=-1)
+
+  def parse(self, x, pos):
+    """Parse input sentence.
+
+    Args:
+      x: input tokens (required).
+      pos: position for each token (optional).
+    Returns:
+      distance: syntactic distance
+      height: syntactic height
+    """
+
+    mask = (x != self.pad)
+    mask_shifted = F.pad(mask[:, 1:], (0, 1), value=0)
+
+    h = self.emb(x)
+    for i in range(self.n_parse_layers):
+      h = h.masked_fill(~mask[:, :, None], 0)
+      h = self.parser_layers[i](h)
+
+    height = self.height_ff(h).squeeze(-1)
+    height.masked_fill_(~mask, -1e9)
+
+    distance = self.distance_ff(h).squeeze(-1)
+    distance.masked_fill_(~mask_shifted, 1e9)
+
+    # Calbrating the distance and height to the same level
+    length = distance.size(1)
+    height_max = height[:, None, :].expand(-1, length, -1)
+    height_max = torch.cummax(
+        height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9,
+        dim=-1)[0].triu(0)
+
+    margin_left = torch.relu(
+        F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max)
+    margin_right = torch.relu(distance[:, None, :] - height_max)
+    margin = torch.where(margin_left > margin_right, margin_right,
+                         margin_left).triu(0)
+
+    margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+    margin.masked_fill_(~margin_mask, 0)
+    margin = margin.max()
+
+    distance = distance - margin
+
+    return distance, height
+
+  def compute_block(self, distance, height):
+    """Compute constituents from distance and height."""
+
+    beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler[0]
+
+    gamma = torch.sigmoid(-beta_logits)
+    ones = torch.ones_like(gamma)
+
+    block_mask_left = cummin(
+        gamma.tril(-1) + ones.triu(0), reverse=True, max_value=1)
+    block_mask_left = block_mask_left - F.pad(
+        block_mask_left[:, :, :-1], (1, 0), value=0)
+    block_mask_left.tril_(0)
+
+    block_mask_right = cummin(
+        gamma.triu(0) + ones.tril(-1), exclusive=True, max_value=1)
+    block_mask_right = block_mask_right - F.pad(
+        block_mask_right[:, :, 1:], (0, 1), value=0)
+    block_mask_right.triu_(0)
+
+    block_p = block_mask_left[:, :, :, None] * block_mask_right[:, :, None, :]
+    block = cumsum(block_mask_left).tril(0) + cumsum(
+        block_mask_right, reverse=True).triu(1)
+
+    return block_p, block
+
+  def compute_head(self, height):
+    """Estimate head for each constituent."""
+
+    _, length = height.size()
+    head_logits = height * self.scaler[1]
+    index = torch.arange(length, device=height.device)
+
+    mask = (index[:, None, None] <= index[None, None, :]) * (
+        index[None, None, :] <= index[None, :, None])
+    head_logits = head_logits[:, None, None, :].repeat(1, length, length, 1)
+    head_logits.masked_fill_(~mask[None, :, :, :], -1e9)
+
+    head_p = torch.softmax(head_logits, dim=-1)
+
+    return head_p
+
+  def generate_mask(self, x, distance, height):
+    """Compute head and cibling distribution for each token."""
+
+    bsz, length = x.size()
+
+    eye = torch.eye(length, device=x.device, dtype=torch.bool)
+    eye = eye[None, :, :].expand((bsz, -1, -1))
+
+    block_p, block = self.compute_block(distance, height)
+    head_p = self.compute_head(height)
+    head = torch.einsum('blij,bijh->blh', block_p, head_p)
+    head = head.masked_fill(eye, 0)
+    child = head.transpose(1, 2)
+    cibling = torch.bmm(head, child).masked_fill(eye, 0)
+
+    rel_list = []
+    if 'head' in self.relations:
+      rel_list.append(head)
+    if 'child' in self.relations:
+      rel_list.append(child)
+    if 'cibling' in self.relations:
+      rel_list.append(cibling)
+
+    rel = torch.stack(rel_list, dim=1)
+
+    rel_weight = self.rel_weight
+
+    dep = torch.einsum('lhr,brij->lbhij', rel_weight, rel)
+    att_mask = dep.reshape(self.nlayers, bsz * self.nhead, length, length)
+
+    return att_mask, cibling, head, block
+
+  def encode(self, x, pos, att_mask):
+    """Structformer encoding process."""
+
+    visibility = self.visibility(x, x.device)
+    h = self.emb(x)
+    if hasattr(self, 'pos_emb'):
+      assert pos.max() < 500
+      h = h + self.pos_emb(pos)
+    for i in range(self.nlayers):
+      h = self.layers[i](
+          h.transpose(0, 1), attn_mask=att_mask[i],
+          key_padding_mask=visibility).transpose(0, 1)
+    return h
+
+  def forward(self, x, pos):
+    """Pass the input through the encoder layer.
+
+    Args:
+      x: input tokens (required).
+      pos: position for each token (optional).
+    Returns:
+      output: probability distributions for missing tokens.
+      state_dict: parsing results and raw output
+    """
+
+    batch_size, length = x.size()
+
+    distance, height = self.parse(x, pos)
+    att_mask, cibling, head, block = self.generate_mask(x, distance, height)
+
+    raw_output = self.encode(x, pos, att_mask)
+    raw_output = self.norm(raw_output)
+    raw_output = self.drop(raw_output)
+
+    output = self.output_layer(raw_output)
+
+    return output.view(batch_size * length, -1), \
+        {'raw_output': raw_output, 'distance': distance, 'height': height,
+        'cibling': cibling, 'head': head, 'block': block}
+
+
+##########################################
+# Clasication Head For BabyLM Evaluation Tasks
+##########################################
+class ClassificationHead(nn.Module):
+    """Head for sentence-level classification tasks."""
+    def __init__(self, config):
+        super(ClassificationHead, self).__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
+
+    def forward(self, features, **kwargs):
+        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
+        x = self.dropout(x)
+        x = self.dense(x)
+        x = torch.tanh(x)
+        x = self.dropout(x)
+        x = self.out_proj(x)
+        return x
+
+##########################################
+# HuggingFace Config
+##########################################
+class StructFormerConfig(PretrainedConfig):
+    model_type = "structformer"
+
+    def __init__(
+        self,
+        hidden_size=512,
+        nlayers=8,
+        ntokens=10_000,
+        nhead=8,
+        dropout=0.1,
+        dropatt=0.1,
+        relative_bias=False,
+        pos_emb=False,
+        pad=0,
+        n_parser_layers=4,
+        conv_size=9,
+        relations=('head', 'child'),
+        weight_act='softmax',
+        num_labels=1,
+        hidden_dropout_prob=0.1,
+        initializer_range=0.02,
+        **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.nlayers = nlayers
+        self.ntokens = ntokens
+        self.nhead = nhead
+        self.dropout = dropout
+        self.dropatt = dropatt
+        self.relative_bias = relative_bias
+        self.pos_emb = pos_emb
+        self.pad = pad
+        self.n_parser_layers = n_parser_layers
+        self.conv_size = conv_size
+        self.relations = relations
+        self.weight_act = weight_act
+        self.num_labels = num_labels
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.initializer_range=initializer_range
+        super().__init__(**kwargs)
+
+##########################################
+# HuggingFace Model
+##########################################
+class StructFormerModel(PreTrainedModel):
+    config_class = StructFormerConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = StructFormer(
+              hidden_size=config.hidden_size,
+              nlayers=config.nlayers,
+              ntokens=config.ntokens,
+              nhead=config.nhead,
+              dropout=config.dropout,
+              dropatt=config.dropatt,
+              relative_bias=config.relative_bias,
+              pos_emb=config.pos_emb,
+              pad=config.pad,
+              n_parser_layers=config.n_parser_layers,
+              conv_size=config.conv_size,
+              relations=config.relations,
+              weight_act=config.weight_act
+        )
+        self.config = config
+
+    def parse(self, input_ids, **kwargs):
+      x = input_ids
+      batch_size, length = x.size()
+      pos = kwargs['position_ids'] if 'position_ids' in kwargs.keys() else torch.arange(length, device=x.device).expand(batch_size, length)
+      
+      sf_output = self.model(x, pos)
+      
+      return sf_output[1]
+    
+    def forward(self, input_ids, labels=None, **kwargs):
+        x = input_ids
+        batch_size, length = x.size()
+        pos = kwargs['position_ids'] if 'position_ids' in kwargs.keys() else torch.arange(length, device=x.device).expand(batch_size, length)
+        
+        sf_output = self.model(x, pos)
+        
+        loss = None
+        if labels is not None:
+          loss_fct = nn.CrossEntropyLoss()
+          loss = loss_fct(sf_output[0], labels.reshape(-1))
+        
+        return MaskedLMOutput(
+            loss=loss, # shape: 1
+            logits=sf_output[0].view(batch_size, length, -1), # shape: (batch_size, length, ntokens)
+            hidden_states=None,
+            attentions=None
+        )
+
+class StructFormerModelForSequenceClassification(PreTrainedModel):
+    config_class = StructFormerConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = StructFormer(
+              hidden_size=config.hidden_size,
+              nlayers=config.nlayers,
+              ntokens=config.ntokens,
+              nhead=config.nhead,
+              dropout=config.dropout,
+              dropatt=config.dropatt,
+              relative_bias=config.relative_bias,
+              pos_emb=config.pos_emb,
+              pad=config.pad,
+              n_parser_layers=config.n_parser_layers,
+              conv_size=config.conv_size,
+              relations=config.relations,
+              weight_act=config.weight_act
+        )
+        self.config = config
+        self.num_labels = config.num_labels
+        self.model.classifier = ClassificationHead(config)
+    
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, nn.Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            if module.bias is not None:
+                module.bias.data.zero_()
+                module.weight.data.fill_(1.0)
+    
+    def forward(self, input_ids, labels=None, **kwargs):
+        x = input_ids
+        batch_size, length = x.size()
+        pos = kwargs['position_ids'] if 'position_ids' in kwargs.keys() else torch.arange(length, device=x.device).expand(batch_size, length)
+        
+        sf_output = self.model(x, pos)
+        
+        logits = self.model.classifier(sf_output[1]['raw_output'])
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = nn.MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = nn.CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = nn.BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=None,
+            attentions=None,
+        )

finetune/cola/checkpoint-200/tokenizer_config.json

@@ -0,0 +1,65 @@
+{
+  "add_prefix_space": false,
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "cls_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "errors": "replace",
+  "mask_token": {
+    "__type": "AddedToken",
+    "content": "<mask>",
+    "lstrip": true,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "model_max_length": 512,
+  "name_or_path": "omarmomen/sf_babylm_1",
+  "pad_token": {
+    "__type": "AddedToken",
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "sep_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "special_tokens_map_file": null,
+  "tokenizer_class": "RobertaTokenizer",
+  "trim_offsets": true,
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

finetune/cola/checkpoint-200/trainer_state.json

@@ -0,0 +1,27 @@
+{
+  "best_metric": 0.8219339622641509,
+  "best_model_checkpoint": "omarmomen/sf_babylm_1/finetune/cola/checkpoint-200",
+  "epoch": 1.5625,
+  "global_step": 200,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 1.56,
+      "eval_accuracy": 0.7036309838294983,
+      "eval_f1": 0.8219339622641509,
+      "eval_loss": 0.6300716400146484,
+      "eval_mcc": 0.13076987235140602,
+      "eval_runtime": 1.2993,
+      "eval_samples_per_second": 784.268,
+      "eval_steps_per_second": 98.515,
+      "step": 200
+    }
+  ],
+  "max_steps": 1280,
+  "num_train_epochs": 10,
+  "total_flos": 327301417273344.0,
+  "trial_name": null,
+  "trial_params": null
+}

finetune/cola/checkpoint-200/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c43ff5a8741c4285a95db833f1cab13594ae1c36859759e5123678abf6a7eba3
+size 3960

finetune/cola/config.json

@@ -0,0 +1,41 @@
+{
+  "_name_or_path": "omarmomen/sf_babylm_1",
+  "architectures": [
+    "StructFormerModelForSequenceClassification"
+  ],
+  "auto_map": {
+    "AutoConfig": "structformer.StructFormerConfig",
+    "AutoModelForMaskedLM": "structformer.StructFormerModel",
+    "AutoModelForSequenceClassification": "structformer.StructFormerModelForSequenceClassification"
+  },
+  "conv_size": 9,
+  "dropatt": 0.1,
+  "dropout": 0.1,
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 512,
+  "id2label": {
+    "0": 0,
+    "1": 1
+  },
+  "initializer_range": 0.02,
+  "label2id": {
+    "0": 0,
+    "1": 1
+  },
+  "model_type": "structformer",
+  "n_parser_layers": 3,
+  "nhead": 8,
+  "nlayers": 8,
+  "ntokens": 16000,
+  "pad": 1,
+  "pos_emb": true,
+  "problem_type": "single_label_classification",
+  "relations": [
+    "head",
+    "child"
+  ],
+  "relative_bias": false,
+  "torch_dtype": "float32",
+  "transformers_version": "4.26.1",
+  "weight_act": "softmax"
+}

finetune/cola/eval_results.json

@@ -0,0 +1,11 @@
+{
+    "epoch": 10.0,
+    "eval_accuracy": 0.7036309838294983,
+    "eval_f1": 0.8219339622641509,
+    "eval_loss": 0.6300716400146484,
+    "eval_mcc": 0.13076987235140602,
+    "eval_runtime": 1.7861,
+    "eval_samples": 1019,
+    "eval_samples_per_second": 570.517,
+    "eval_steps_per_second": 71.665
+}

finetune/cola/predict_results.txt

@@ -0,0 +1,1020 @@
+index	prediction
+0	1
+1	1
+2	1
+3	1
+4	1
+5	1
+6	1
+7	1
+8	1
+9	1
+10	0
+11	1
+12	1
+13	1
+14	1
+15	1
+16	1
+17	1
+18	1
+19	1
+20	1
+21	1
+22	1
+23	1
+24	1
+25	1
+26	1
+27	1
+28	1
+29	1
+30	1
+31	1
+32	1
+33	1
+34	1
+35	1
+36	1
+37	1
+38	1
+39	1
+40	1
+41	0
+42	1
+43	1
+44	1
+45	1
+46	1
+47	1
+48	1
+49	1
+50	1
+51	1
+52	1
+53	1
+54	1
+55	1
+56	1
+57	1
+58	1
+59	1
+60	1
+61	1
+62	1
+63	1
+64	1
+65	1
+66	1
+67	1
+68	1
+69	1
+70	1
+71	1
+72	1
+73	1
+74	1
+75	1
+76	1
+77	1
+78	1
+79	1
+80	0
+81	1
+82	1
+83	1
+84	1
+85	1
+86	1
+87	1
+88	1
+89	1
+90	1
+91	1
+92	1
+93	1
+94	1
+95	1
+96	1
+97	1
+98	1
+99	1
+100	1
+101	1
+102	1
+103	1
+104	1
+105	1
+106	1
+107	1
+108	1
+109	1
+110	1
+111	1
+112	1
+113	1
+114	1
+115	1
+116	1
+117	1
+118	1
+119	1
+120	1
+121	1
+122	1
+123	1
+124	1
+125	1
+126	1
+127	1
+128	1
+129	1
+130	1
+131	1
+132	1
+133	1
+134	1
+135	1
+136	1
+137	1
+138	1
+139	1
+140	1
+141	0
+142	1
+143	1
+144	1
+145	1
+146	1
+147	1
+148	1
+149	1
+150	1
+151	1
+152	1
+153	1
+154	1
+155	1
+156	1
+157	1
+158	1
+159	1
+160	1
+161	1
+162	1
+163	1
+164	1
+165	1
+166	0
+167	1
+168	0
+169	1
+170	1
+171	0
+172	1
+173	1
+174	1
+175	1
+176	1
+177	1
+178	1
+179	1
+180	1
+181	1
+182	1
+183	1
+184	1
+185	1
+186	1
+187	1
+188	1
+189	1
+190	1
+191	1
+192	1
+193	1
+194	1
+195	1
+196	1
+197	1
+198	1
+199	1
+200	1
+201	1
+202	1
+203	1
+204	1
+205	1
+206	1
+207	1
+208	1
+209	1
+210	1
+211	1
+212	1
+213	1
+214	1
+215	1
+216	1
+217	1
+218	1
+219	1
+220	1
+221	1
+222	1
+223	1
+224	1
+225	1
+226	1
+227	1
+228	1
+229	1
+230	1
+231	1
+232	0
+233	1
+234	1
+235	1
+236	1
+237	1
+238	1
+239	1
+240	1
+241	1
+242	0
+243	1
+244	1
+245	1
+246	1
+247	1
+248	0
+249	1
+250	1
+251	1
+252	1
+253	1
+254	1
+255	1
+256	1
+257	1
+258	1
+259	1
+260	1
+261	1
+262	1
+263	1
+264	1
+265	1
+266	1
+267	1
+268	1
+269	1
+270	1
+271	1
+272	0
+273	1
+274	1
+275	1
+276	1
+277	1
+278	1
+279	1
+280	1
+281	1
+282	1
+283	1
+284	1
+285	1
+286	0
+287	1
+288	1
+289	1
+290	1
+291	1
+292	1
+293	0
+294	1
+295	1
+296	0
+297	1
+298	1
+299	1
+300	0
+301	1
+302	1
+303	1
+304	1
+305	1
+306	1
+307	1
+308	1
+309	1
+310	1
+311	1
+312	1
+313	1
+314	1
+315	1
+316	1
+317	1
+318	1
+319	1
+320	1
+321	1
+322	1
+323	0
+324	1
+325	1
+326	1
+327	1
+328	1
+329	1
+330	1
+331	1
+332	0
+333	1
+334	1
+335	1
+336	1
+337	1
+338	1
+339	1
+340	1
+341	1
+342	1
+343	1
+344	1
+345	1
+346	1
+347	1
+348	1
+349	1
+350	1
+351	1
+352	1
+353	1
+354	1
+355	1
+356	1
+357	1
+358	1
+359	1
+360	1
+361	1
+362	1
+363	1
+364	1
+365	1
+366	1
+367	1
+368	1
+369	1
+370	1
+371	1
+372	1
+373	1
+374	1
+375	1
+376	1
+377	1
+378	1
+379	1
+380	1
+381	1
+382	1
+383	1
+384	1
+385	1
+386	1
+387	1
+388	1
+389	1
+390	1
+391	1
+392	0
+393	1
+394	1
+395	1
+396	1
+397	1
+398	1
+399	1
+400	1
+401	1
+402	1
+403	1
+404	1
+405	1
+406	1
+407	1
+408	1
+409	1
+410	1
+411	1
+412	1
+413	1
+414	1
+415	1
+416	1
+417	1
+418	1
+419	1
+420	1
+421	1
+422	1
+423	1
+424	1
+425	1
+426	1
+427	1
+428	1
+429	1
+430	1
+431	1
+432	1
+433	1
+434	1
+435	1
+436	1
+437	1
+438	1
+439	1
+440	1
+441	1
+442	1
+443	1
+444	1
+445	1
+446	1
+447	1
+448	1
+449	1
+450	1
+451	1
+452	1
+453	1
+454	1
+455	1
+456	1
+457	1
+458	1
+459	1
+460	1
+461	1
+462	1
+463	1
+464	1
+465	1
+466	1
+467	1
+468	1
+469	1
+470	1
+471	1
+472	1
+473	1
+474	1
+475	1
+476	1
+477	1
+478	1
+479	1
+480	1
+481	1
+482	1
+483	1
+484	1
+485	1
+486	1
+487	1
+488	1
+489	1
+490	1
+491	1
+492	1
+493	1
+494	1
+495	1
+496	1
+497	1
+498	1
+499	1
+500	1
+501	1
+502	1
+503	1
+504	1
+505	1
+506	1
+507	1
+508	0
+509	1
+510	1
+511	1
+512	1
+513	1
+514	1
+515	1
+516	1
+517	1
+518	1
+519	1
+520	1
+521	1
+522	1
+523	1
+524	1
+525	1
+526	1
+527	1
+528	1
+529	1
+530	1
+531	1
+532	1
+533	1
+534	1
+535	1
+536	1
+537	1
+538	1
+539	1
+540	1
+541	1
+542	1
+543	1
+544	1
+545	0
+546	1
+547	1
+548	1
+549	1
+550	1
+551	1
+552	1
+553	1
+554	1
+555	1
+556	1
+557	1
+558	1
+559	1
+560	1
+561	1
+562	1
+563	1
+564	1
+565	1
+566	1
+567	1
+568	1
+569	1
+570	1
+571	1
+572	1
+573	1
+574	1
+575	1
+576	1
+577	1
+578	1
+579	1
+580	1
+581	1
+582	1
+583	1
+584	1
+585	1
+586	1
+587	1
+588	1
+589	1
+590	1
+591	1
+592	1
+593	0
+594	1
+595	1
+596	1
+597	1
+598	1
+599	0
+600	1
+601	1
+602	1
+603	0
+604	0
+605	0
+606	1
+607	1
+608	1
+609	1
+610	1
+611	1
+612	1
+613	1
+614	1
+615	1
+616	1
+617	1
+618	1
+619	1
+620	1
+621	1
+622	1
+623	1
+624	1
+625	1
+626	1
+627	1
+628	1
+629	1
+630	1
+631	1
+632	1
+633	1
+634	1
+635	1
+636	1
+637	1
+638	1
+639	1
+640	1
+641	1
+642	1
+643	1
+644	1
+645	1
+646	1
+647	1
+648	1
+649	1
+650	1
+651	1
+652	1
+653	1
+654	1
+655	1
+656	1
+657	1
+658	1
+659	1
+660	1
+661	1
+662	1
+663	1
+664	1
+665	1
+666	1
+667	1
+668	1
+669	1
+670	1
+671	1
+672	1
+673	1
+674	1
+675	1
+676	1
+677	1
+678	1
+679	1
+680	1
+681	1
+682	1
+683	1
+684	1
+685	1
+686	1
+687	1
+688	1
+689	1
+690	1
+691	1
+692	1
+693	1
+694	1
+695	1
+696	1
+697	1
+698	1
+699	1
+700	1
+701	1
+702	1
+703	1
+704	1
+705	1
+706	1
+707	1
+708	1
+709	1
+710	1
+711	1
+712	1
+713	1
+714	1
+715	1
+716	1
+717	1
+718	1
+719	1
+720	1
+721	1
+722	1
+723	1
+724	1
+725	1
+726	1
+727	1
+728	1
+729	1
+730	1
+731	1
+732	1
+733	1
+734	1
+735	1
+736	1
+737	1
+738	1
+739	1
+740	1
+741	1
+742	1
+743	1
+744	1
+745	1
+746	1
+747	1
+748	1
+749	1
+750	1
+751	1
+752	1
+753	1
+754	1
+755	1
+756	1
+757	1
+758	1
+759	1
+760	1
+761	1
+762	1
+763	1
+764	1
+765	1
+766	1
+767	1
+768	1
+769	1
+770	1
+771	1
+772	1
+773	1
+774	1
+775	1
+776	1
+777	1
+778	1
+779	1
+780	1
+781	1
+782	1
+783	1
+784	1
+785	1
+786	1
+787	1
+788	1
+789	1
+790	1
+791	1
+792	1
+793	1
+794	0
+795	0
+796	1
+797	1
+798	1
+799	1
+800	1
+801	1
+802	1
+803	1
+804	1
+805	1
+806	1
+807	1
+808	1
+809	1
+810	1
+811	1
+812	1
+813	0
+814	1
+815	1
+816	1
+817	1
+818	1
+819	1
+820	1
+821	1
+822	1
+823	1
+824	1
+825	1
+826	1
+827	1
+828	1
+829	1
+830	1
+831	1
+832	1
+833	1
+834	1
+835	1
+836	1
+837	1
+838	1
+839	1
+840	1
+841	1
+842	1
+843	1
+844	1
+845	1
+846	1
+847	1
+848	1
+849	1
+850	1
+851	1
+852	1
+853	1
+854	1
+855	1
+856	1
+857	1
+858	1
+859	1
+860	1
+861	1
+862	1
+863	1
+864	1
+865	1
+866	1
+867	1
+868	1
+869	1
+870	1
+871	1
+872	1
+873	1
+874	1
+875	1
+876	1
+877	1
+878	1
+879	1
+880	1
+881	1
+882	1
+883	1
+884	1
+885	1
+886	1
+887	1
+888	1
+889	1
+890	1
+891	1
+892	1
+893	1
+894	1
+895	1
+896	1
+897	1
+898	1
+899	1
+900	1
+901	1
+902	0
+903	1
+904	1
+905	1
+906	1
+907	1
+908	1
+909	1
+910	1
+911	1
+912	1
+913	1
+914	1
+915	1
+916	1
+917	1
+918	1
+919	1
+920	1
+921	1
+922	1
+923	1
+924	1
+925	1
+926	1
+927	1
+928	1
+929	1
+930	1
+931	1
+932	1
+933	1
+934	1
+935	1
+936	1
+937	1
+938	1
+939	1
+940	1
+941	1
+942	1
+943	1
+944	1
+945	1
+946	1
+947	1
+948	1
+949	1
+950	1
+951	1
+952	1
+953	1
+954	1
+955	1
+956	1
+957	1
+958	1
+959	1
+960	1
+961	1
+962	1
+963	1
+964	1
+965	1
+966	1
+967	1
+968	1
+969	1
+970	1
+971	1
+972	1
+973	1
+974	1
+975	1
+976	1
+977	1
+978	0
+979	1
+980	1
+981	1
+982	1
+983	1
+984	1
+985	1
+986	1
+987	1
+988	1
+989	1
+990	1
+991	1
+992	1
+993	1
+994	1
+995	1
+996	0
+997	1
+998	1
+999	1
+1000	1
+1001	1
+1002	1
+1003	1
+1004	1
+1005	1
+1006	1
+1007	1
+1008	1
+1009	1
+1010	1
+1011	1
+1012	1
+1013	1
+1014	1
+1015	1
+1016	1
+1017	1
+1018	1

finetune/cola/pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:782e41676fcfa55bd9f10569a44c812fc36f983ee47f32bd0cafdee7d732ee83
+size 167313902

finetune/cola/special_tokens_map.json

@@ -0,0 +1,15 @@
+{
+  "bos_token": "<s>",
+  "cls_token": "<s>",
+  "eos_token": "</s>",
+  "mask_token": {
+    "content": "<mask>",
+    "lstrip": true,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<pad>",
+  "sep_token": "</s>",
+  "unk_token": "<unk>"
+}

finetune/cola/structformer.py

@@ -0,0 +1,867 @@
+# coding=utf-8
+# Copyright 2023 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""StructFormer and transformer model."""
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn import init
+from transformers import PretrainedConfig, PreTrainedModel
+from transformers.modeling_outputs import MaskedLMOutput, SequenceClassifierOutput
+
+def _get_activation_fn(activation):
+  """Get specified activation function."""
+  if activation == "relu":
+    return nn.ReLU()
+  elif activation == "gelu":
+    return nn.GELU()
+  elif activation == "leakyrelu":
+    return nn.LeakyReLU()
+
+  raise RuntimeError(
+      "activation should be relu/gelu, not {}".format(activation))
+
+
+class Conv1d(nn.Module):
+  """1D convolution layer."""
+
+  def __init__(self, hidden_size, kernel_size, dilation=1):
+    """Initialization.
+
+    Args:
+      hidden_size: dimension of input embeddings
+      kernel_size: convolution kernel size
+      dilation: the spacing between the kernel points
+    """
+    super(Conv1d, self).__init__()
+
+    if kernel_size % 2 == 0:
+      padding = (kernel_size // 2) * dilation
+      self.shift = True
+    else:
+      padding = ((kernel_size - 1) // 2) * dilation
+      self.shift = False
+    self.conv = nn.Conv1d(
+        hidden_size,
+        hidden_size,
+        kernel_size,
+        padding=padding,
+        dilation=dilation)
+
+  def forward(self, x):
+    """Compute convolution.
+
+    Args:
+      x: input embeddings
+    Returns:
+      conv_output: convolution results
+    """
+
+    if self.shift:
+      return self.conv(x.transpose(1, 2)).transpose(1, 2)[:, 1:]
+    else:
+      return self.conv(x.transpose(1, 2)).transpose(1, 2)
+
+
+class MultiheadAttention(nn.Module):
+  """Multi-head self-attention layer."""
+
+  def __init__(self,
+               embed_dim,
+               num_heads,
+               dropout=0.,
+               bias=True,
+               v_proj=True,
+               out_proj=True,
+               relative_bias=True):
+    """Initialization.
+
+    Args:
+      embed_dim: dimension of input embeddings
+      num_heads: number of self-attention heads
+      dropout: dropout rate
+      bias: bool, indicate whether include bias for linear transformations
+      v_proj: bool, indicate whether project inputs to new values
+      out_proj: bool, indicate whether project outputs to new values
+      relative_bias: bool, indicate whether use a relative position based
+        attention bias
+    """
+
+    super(MultiheadAttention, self).__init__()
+    self.embed_dim = embed_dim
+
+    self.num_heads = num_heads
+    self.drop = nn.Dropout(dropout)
+    self.head_dim = embed_dim // num_heads
+    assert self.head_dim * num_heads == self.embed_dim, ("embed_dim must be "
+                                                         "divisible by "
+                                                         "num_heads")
+
+    self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+    self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+    if v_proj:
+      self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+    else:
+      self.v_proj = nn.Identity()
+
+    if out_proj:
+      self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
+    else:
+      self.out_proj = nn.Identity()
+
+    if relative_bias:
+      self.relative_bias = nn.Parameter(torch.zeros((self.num_heads, 512)))
+    else:
+      self.relative_bias = None
+
+    self._reset_parameters()
+
+  def _reset_parameters(self):
+    """Initialize attention parameters."""
+
+    init.xavier_uniform_(self.q_proj.weight)
+    init.constant_(self.q_proj.bias, 0.)
+
+    init.xavier_uniform_(self.k_proj.weight)
+    init.constant_(self.k_proj.bias, 0.)
+
+    if isinstance(self.v_proj, nn.Linear):
+      init.xavier_uniform_(self.v_proj.weight)
+      init.constant_(self.v_proj.bias, 0.)
+
+    if isinstance(self.out_proj, nn.Linear):
+      init.xavier_uniform_(self.out_proj.weight)
+      init.constant_(self.out_proj.bias, 0.)
+
+  def forward(self, query, key_padding_mask=None, attn_mask=None):
+    """Compute multi-head self-attention.
+
+    Args:
+      query: input embeddings
+      key_padding_mask: 3D mask that prevents attention to certain positions
+      attn_mask: 3D mask that rescale the attention weight at each position
+    Returns:
+      attn_output: self-attention output
+    """
+
+    length, bsz, embed_dim = query.size()
+    assert embed_dim == self.embed_dim
+
+    head_dim = embed_dim // self.num_heads
+    assert head_dim * self.num_heads == embed_dim, ("embed_dim must be "
+                                                    "divisible by num_heads")
+    scaling = float(head_dim)**-0.5
+
+    q = self.q_proj(query)
+    k = self.k_proj(query)
+    v = self.v_proj(query)
+
+    q = q * scaling
+
+    if attn_mask is not None:
+      assert list(attn_mask.size()) == [bsz * self.num_heads,
+                                        query.size(0), query.size(0)]
+
+    q = q.contiguous().view(length, bsz * self.num_heads,
+                            head_dim).transpose(0, 1)
+    k = k.contiguous().view(length, bsz * self.num_heads,
+                            head_dim).transpose(0, 1)
+    v = v.contiguous().view(length, bsz * self.num_heads,
+                            head_dim).transpose(0, 1)
+
+    attn_output_weights = torch.bmm(q, k.transpose(1, 2))
+    assert list(
+        attn_output_weights.size()) == [bsz * self.num_heads, length, length]
+
+    if self.relative_bias is not None:
+      pos = torch.arange(length, device=query.device)
+      relative_pos = torch.abs(pos[:, None] - pos[None, :]) + 256
+      relative_pos = relative_pos[None, :, :].expand(bsz * self.num_heads, -1,
+                                                     -1)
+
+      relative_bias = self.relative_bias.repeat_interleave(bsz, dim=0)
+      relative_bias = relative_bias[:, None, :].expand(-1, length, -1)
+      relative_bias = torch.gather(relative_bias, 2, relative_pos)
+      attn_output_weights = attn_output_weights + relative_bias
+
+    if key_padding_mask is not None:
+      attn_output_weights = attn_output_weights + key_padding_mask
+
+    if attn_mask is None:
+      attn_output_weights = torch.softmax(attn_output_weights, dim=-1)
+    else:
+      attn_output_weights = torch.sigmoid(attn_output_weights) * attn_mask
+
+    attn_output_weights = self.drop(attn_output_weights)
+
+    attn_output = torch.bmm(attn_output_weights, v)
+
+    assert list(attn_output.size()) == [bsz * self.num_heads, length, head_dim]
+    attn_output = attn_output.transpose(0, 1).contiguous().view(
+        length, bsz, embed_dim)
+    attn_output = self.out_proj(attn_output)
+
+    return attn_output
+
+
+class TransformerLayer(nn.Module):
+  """TransformerEncoderLayer is made up of self-attn and feedforward network."""
+
+  def __init__(self,
+               d_model,
+               nhead,
+               dim_feedforward=2048,
+               dropout=0.1,
+               dropatt=0.1,
+               activation="leakyrelu",
+               relative_bias=True):
+    """Initialization.
+
+    Args:
+      d_model: dimension of inputs
+      nhead: number of self-attention heads
+      dim_feedforward: dimension of hidden layer in feedforward layer
+      dropout: dropout rate
+      dropatt: drop attention rate
+      activation: activation function
+      relative_bias: bool, indicate whether use a relative position based
+        attention bias
+    """
+
+    super(TransformerLayer, self).__init__()
+    self.self_attn = MultiheadAttention(
+        d_model, nhead, dropout=dropatt, relative_bias=relative_bias)
+    # Implementation of Feedforward model
+    self.feedforward = nn.Sequential(
+        nn.LayerNorm(d_model), nn.Linear(d_model, dim_feedforward),
+        _get_activation_fn(activation), nn.Dropout(dropout),
+        nn.Linear(dim_feedforward, d_model))
+
+    self.norm = nn.LayerNorm(d_model)
+    self.dropout1 = nn.Dropout(dropout)
+    self.dropout2 = nn.Dropout(dropout)
+
+    self.nhead = nhead
+
+  def forward(self, src, attn_mask=None, key_padding_mask=None):
+    """Pass the input through the encoder layer.
+
+    Args:
+      src: the sequence to the encoder layer (required).
+      attn_mask: the mask for the src sequence (optional).
+      key_padding_mask: the mask for the src keys per batch (optional).
+    Returns:
+      src3: the output of transformer layer, share the same shape as src.
+    """
+    src2 = self.self_attn(
+        self.norm(src), attn_mask=attn_mask, key_padding_mask=key_padding_mask)
+    src2 = src + self.dropout1(src2)
+    src3 = self.feedforward(src2)
+    src3 = src2 + self.dropout2(src3)
+
+    return src3
+
+
+def cumprod(x, reverse=False, exclusive=False):
+  """cumulative product."""
+  if reverse:
+    x = x.flip([-1])
+
+  if exclusive:
+    x = F.pad(x[:, :, :-1], (1, 0), value=1)
+
+  cx = x.cumprod(-1)
+
+  if reverse:
+    cx = cx.flip([-1])
+  return cx
+
+def cumsum(x, reverse=False, exclusive=False):
+  """cumulative sum."""
+  bsz, _, length = x.size()
+  device = x.device
+  if reverse:
+    if exclusive:
+      w = torch.ones([bsz, length, length], device=device).tril(-1)
+    else:
+      w = torch.ones([bsz, length, length], device=device).tril(0)
+    cx = torch.bmm(x, w)
+  else:
+    if exclusive:
+      w = torch.ones([bsz, length, length], device=device).triu(1)
+    else:
+      w = torch.ones([bsz, length, length], device=device).triu(0)
+    cx = torch.bmm(x, w)
+  return cx
+
+def cummin(x, reverse=False, exclusive=False, max_value=1e9):
+  """cumulative min."""
+  if reverse:
+    if exclusive:
+      x = F.pad(x[:, :, 1:], (0, 1), value=max_value)
+    x = x.flip([-1]).cummin(-1)[0].flip([-1])
+  else:
+    if exclusive:
+      x = F.pad(x[:, :, :-1], (1, 0), value=max_value)
+    x = x.cummin(-1)[0]
+  return x
+
+class Transformer(nn.Module):
+  """Transformer model."""
+
+  def __init__(self,
+              hidden_size,
+              nlayers,
+              ntokens,
+              nhead=8,
+              dropout=0.1,
+              dropatt=0.1,
+              relative_bias=True,
+              pos_emb=False,
+              pad=0):
+    """Initialization.
+
+    Args:
+      hidden_size: dimension of inputs and hidden states
+      nlayers: number of layers
+      ntokens: number of output categories
+      nhead: number of self-attention heads
+      dropout: dropout rate
+      dropatt: drop attention rate
+      relative_bias: bool, indicate whether use a relative position based
+        attention bias
+      pos_emb: bool, indicate whether use a learnable positional embedding
+      pad: pad token index
+    """
+
+    super(Transformer, self).__init__()
+
+    self.drop = nn.Dropout(dropout)
+
+    self.emb = nn.Embedding(ntokens, hidden_size)
+    if pos_emb:
+      self.pos_emb = nn.Embedding(500, hidden_size)
+
+    self.layers = nn.ModuleList([
+        TransformerLayer(hidden_size, nhead, hidden_size * 4, dropout,
+                                dropatt=dropatt, relative_bias=relative_bias)
+        for _ in range(nlayers)])
+
+    self.norm = nn.LayerNorm(hidden_size)
+
+    self.output_layer = nn.Linear(hidden_size, ntokens)
+    self.output_layer.weight = self.emb.weight
+
+    self.init_weights()
+
+    self.nlayers = nlayers
+    self.nhead = nhead
+    self.ntokens = ntokens
+    self.hidden_size = hidden_size
+    self.pad = pad
+
+  def init_weights(self):
+    """Initialize token embedding and output bias."""
+    initrange = 0.1
+    self.emb.weight.data.uniform_(-initrange, initrange)
+    if hasattr(self, 'pos_emb'):
+      self.pos_emb.weight.data.uniform_(-initrange, initrange)
+    self.output_layer.bias.data.fill_(0)
+
+  def visibility(self, x, device):
+    """Mask pad tokens."""
+    visibility = (x != self.pad).float()
+    visibility = visibility[:, None, :].expand(-1, x.size(1), -1)
+    visibility = torch.repeat_interleave(visibility, self.nhead, dim=0)
+    return visibility.log()
+
+  def encode(self, x, pos):
+    """Standard transformer encode process."""
+    h = self.emb(x)
+    if hasattr(self, 'pos_emb'):
+      h = h + self.pos_emb(pos)
+    h_list = []
+    visibility = self.visibility(x, x.device)
+
+    for i in range(self.nlayers):
+      h_list.append(h)
+      h = self.layers[i](
+          h.transpose(0, 1), key_padding_mask=visibility).transpose(0, 1)
+
+    output = h
+    h_array = torch.stack(h_list, dim=2)
+
+    return output, h_array
+
+  def forward(self, x, pos):
+    """Pass the input through the encoder layer.
+
+    Args:
+      x: input tokens (required).
+      pos: position for each token (optional).
+    Returns:
+      output: probability distributions for missing tokens.
+      state_dict: parsing results and raw output
+    """
+
+    batch_size, length = x.size()
+
+    raw_output, _ = self.encode(x, pos)
+    raw_output = self.norm(raw_output)
+    raw_output = self.drop(raw_output)
+
+    output = self.output_layer(raw_output)
+    return output.view(batch_size * length, -1), {'raw_output': raw_output,}
+
+class StructFormer(Transformer):
+  """StructFormer model."""
+
+  def __init__(self,
+              hidden_size,
+              nlayers,
+              ntokens,
+              nhead=8,
+              dropout=0.1,
+              dropatt=0.1,
+              relative_bias=False,
+              pos_emb=False,
+              pad=0,
+              n_parser_layers=4,
+              conv_size=9,
+              relations=('head', 'child'),
+              weight_act='softmax'):
+    """Initialization.
+
+    Args:
+      hidden_size: dimension of inputs and hidden states
+      nlayers: number of layers
+      ntokens: number of output categories
+      nhead: number of self-attention heads
+      dropout: dropout rate
+      dropatt: drop attention rate
+      relative_bias: bool, indicate whether use a relative position based
+        attention bias
+      pos_emb: bool, indicate whether use a learnable positional embedding
+      pad: pad token index
+      n_parser_layers: number of parsing layers
+      conv_size: convolution kernel size for parser
+      relations: relations that are used to compute self attention
+      weight_act: relations distribution activation function
+    """
+
+    super(StructFormer, self).__init__(
+        hidden_size,
+        nlayers,
+        ntokens,
+        nhead=nhead,
+        dropout=dropout,
+        dropatt=dropatt,
+        relative_bias=relative_bias,
+        pos_emb=pos_emb,
+        pad=pad)
+
+    self.parser_layers = nn.ModuleList([
+        nn.Sequential(Conv1d(hidden_size, conv_size),
+                      nn.LayerNorm(hidden_size, elementwise_affine=False),
+                      nn.Tanh()) for i in range(n_parser_layers)])
+
+    self.distance_ff = nn.Sequential(
+        Conv1d(hidden_size, 2),
+        nn.LayerNorm(hidden_size, elementwise_affine=False), nn.Tanh(),
+        nn.Linear(hidden_size, 1))
+
+    self.height_ff = nn.Sequential(
+        nn.Linear(hidden_size, hidden_size),
+        nn.LayerNorm(hidden_size, elementwise_affine=False), nn.Tanh(),
+        nn.Linear(hidden_size, 1))
+
+    n_rel = len(relations)
+    self._rel_weight = nn.Parameter(torch.zeros((nlayers, nhead, n_rel)))
+    self._rel_weight.data.normal_(0, 0.1)
+
+    self._scaler = nn.Parameter(torch.zeros(2))
+
+    self.n_parse_layers = n_parser_layers
+    self.weight_act = weight_act
+    self.relations = relations
+
+  @property
+  def scaler(self):
+    return self._scaler.exp()
+
+  @property
+  def rel_weight(self):
+    if self.weight_act == 'sigmoid':
+      return torch.sigmoid(self._rel_weight)
+    elif self.weight_act == 'softmax':
+      return torch.softmax(self._rel_weight, dim=-1)
+
+  def parse(self, x, pos):
+    """Parse input sentence.
+
+    Args:
+      x: input tokens (required).
+      pos: position for each token (optional).
+    Returns:
+      distance: syntactic distance
+      height: syntactic height
+    """
+
+    mask = (x != self.pad)
+    mask_shifted = F.pad(mask[:, 1:], (0, 1), value=0)
+
+    h = self.emb(x)
+    for i in range(self.n_parse_layers):
+      h = h.masked_fill(~mask[:, :, None], 0)
+      h = self.parser_layers[i](h)
+
+    height = self.height_ff(h).squeeze(-1)
+    height.masked_fill_(~mask, -1e9)
+
+    distance = self.distance_ff(h).squeeze(-1)
+    distance.masked_fill_(~mask_shifted, 1e9)
+
+    # Calbrating the distance and height to the same level
+    length = distance.size(1)
+    height_max = height[:, None, :].expand(-1, length, -1)
+    height_max = torch.cummax(
+        height_max.triu(0) - torch.ones_like(height_max).tril(-1) * 1e9,
+        dim=-1)[0].triu(0)
+
+    margin_left = torch.relu(
+        F.pad(distance[:, :-1, None], (0, 0, 1, 0), value=1e9) - height_max)
+    margin_right = torch.relu(distance[:, None, :] - height_max)
+    margin = torch.where(margin_left > margin_right, margin_right,
+                         margin_left).triu(0)
+
+    margin_mask = torch.stack([mask_shifted] + [mask] * (length - 1), dim=1)
+    margin.masked_fill_(~margin_mask, 0)
+    margin = margin.max()
+
+    distance = distance - margin
+
+    return distance, height
+
+  def compute_block(self, distance, height):
+    """Compute constituents from distance and height."""
+
+    beta_logits = (distance[:, None, :] - height[:, :, None]) * self.scaler[0]
+
+    gamma = torch.sigmoid(-beta_logits)
+    ones = torch.ones_like(gamma)
+
+    block_mask_left = cummin(
+        gamma.tril(-1) + ones.triu(0), reverse=True, max_value=1)
+    block_mask_left = block_mask_left - F.pad(
+        block_mask_left[:, :, :-1], (1, 0), value=0)
+    block_mask_left.tril_(0)
+
+    block_mask_right = cummin(
+        gamma.triu(0) + ones.tril(-1), exclusive=True, max_value=1)
+    block_mask_right = block_mask_right - F.pad(
+        block_mask_right[:, :, 1:], (0, 1), value=0)
+    block_mask_right.triu_(0)
+
+    block_p = block_mask_left[:, :, :, None] * block_mask_right[:, :, None, :]
+    block = cumsum(block_mask_left).tril(0) + cumsum(
+        block_mask_right, reverse=True).triu(1)
+
+    return block_p, block
+
+  def compute_head(self, height):
+    """Estimate head for each constituent."""
+
+    _, length = height.size()
+    head_logits = height * self.scaler[1]
+    index = torch.arange(length, device=height.device)
+
+    mask = (index[:, None, None] <= index[None, None, :]) * (
+        index[None, None, :] <= index[None, :, None])
+    head_logits = head_logits[:, None, None, :].repeat(1, length, length, 1)
+    head_logits.masked_fill_(~mask[None, :, :, :], -1e9)
+
+    head_p = torch.softmax(head_logits, dim=-1)
+
+    return head_p
+
+  def generate_mask(self, x, distance, height):
+    """Compute head and cibling distribution for each token."""
+
+    bsz, length = x.size()
+
+    eye = torch.eye(length, device=x.device, dtype=torch.bool)
+    eye = eye[None, :, :].expand((bsz, -1, -1))
+
+    block_p, block = self.compute_block(distance, height)
+    head_p = self.compute_head(height)
+    head = torch.einsum('blij,bijh->blh', block_p, head_p)
+    head = head.masked_fill(eye, 0)
+    child = head.transpose(1, 2)
+    cibling = torch.bmm(head, child).masked_fill(eye, 0)
+
+    rel_list = []
+    if 'head' in self.relations:
+      rel_list.append(head)
+    if 'child' in self.relations:
+      rel_list.append(child)
+    if 'cibling' in self.relations:
+      rel_list.append(cibling)
+
+    rel = torch.stack(rel_list, dim=1)
+
+    rel_weight = self.rel_weight
+
+    dep = torch.einsum('lhr,brij->lbhij', rel_weight, rel)
+    att_mask = dep.reshape(self.nlayers, bsz * self.nhead, length, length)
+
+    return att_mask, cibling, head, block
+
+  def encode(self, x, pos, att_mask):
+    """Structformer encoding process."""
+
+    visibility = self.visibility(x, x.device)
+    h = self.emb(x)
+    if hasattr(self, 'pos_emb'):
+      assert pos.max() < 500
+      h = h + self.pos_emb(pos)
+    for i in range(self.nlayers):
+      h = self.layers[i](
+          h.transpose(0, 1), attn_mask=att_mask[i],
+          key_padding_mask=visibility).transpose(0, 1)
+    return h
+
+  def forward(self, x, pos):
+    """Pass the input through the encoder layer.
+
+    Args:
+      x: input tokens (required).
+      pos: position for each token (optional).
+    Returns:
+      output: probability distributions for missing tokens.
+      state_dict: parsing results and raw output
+    """
+
+    batch_size, length = x.size()
+
+    distance, height = self.parse(x, pos)
+    att_mask, cibling, head, block = self.generate_mask(x, distance, height)
+
+    raw_output = self.encode(x, pos, att_mask)
+    raw_output = self.norm(raw_output)
+    raw_output = self.drop(raw_output)
+
+    output = self.output_layer(raw_output)
+
+    return output.view(batch_size * length, -1), \
+        {'raw_output': raw_output, 'distance': distance, 'height': height,
+        'cibling': cibling, 'head': head, 'block': block}
+
+
+##########################################
+# Clasication Head For BabyLM Evaluation Tasks
+##########################################
+class ClassificationHead(nn.Module):
+    """Head for sentence-level classification tasks."""
+    def __init__(self, config):
+        super(ClassificationHead, self).__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
+
+    def forward(self, features, **kwargs):
+        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
+        x = self.dropout(x)
+        x = self.dense(x)
+        x = torch.tanh(x)
+        x = self.dropout(x)
+        x = self.out_proj(x)
+        return x
+
+##########################################
+# HuggingFace Config
+##########################################
+class StructFormerConfig(PretrainedConfig):
+    model_type = "structformer"
+
+    def __init__(
+        self,
+        hidden_size=512,
+        nlayers=8,
+        ntokens=10_000,
+        nhead=8,
+        dropout=0.1,
+        dropatt=0.1,
+        relative_bias=False,
+        pos_emb=False,
+        pad=0,
+        n_parser_layers=4,
+        conv_size=9,
+        relations=('head', 'child'),
+        weight_act='softmax',
+        num_labels=1,
+        hidden_dropout_prob=0.1,
+        initializer_range=0.02,
+        **kwargs,
+    ):
+        self.hidden_size = hidden_size
+        self.nlayers = nlayers
+        self.ntokens = ntokens
+        self.nhead = nhead
+        self.dropout = dropout
+        self.dropatt = dropatt
+        self.relative_bias = relative_bias
+        self.pos_emb = pos_emb
+        self.pad = pad
+        self.n_parser_layers = n_parser_layers
+        self.conv_size = conv_size
+        self.relations = relations
+        self.weight_act = weight_act
+        self.num_labels = num_labels
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.initializer_range=initializer_range
+        super().__init__(**kwargs)
+
+##########################################
+# HuggingFace Model
+##########################################
+class StructFormerModel(PreTrainedModel):
+    config_class = StructFormerConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = StructFormer(
+              hidden_size=config.hidden_size,
+              nlayers=config.nlayers,
+              ntokens=config.ntokens,
+              nhead=config.nhead,
+              dropout=config.dropout,
+              dropatt=config.dropatt,
+              relative_bias=config.relative_bias,
+              pos_emb=config.pos_emb,
+              pad=config.pad,
+              n_parser_layers=config.n_parser_layers,
+              conv_size=config.conv_size,
+              relations=config.relations,
+              weight_act=config.weight_act
+        )
+        self.config = config
+
+    def parse(self, input_ids, **kwargs):
+      x = input_ids
+      batch_size, length = x.size()
+      pos = kwargs['position_ids'] if 'position_ids' in kwargs.keys() else torch.arange(length, device=x.device).expand(batch_size, length)
+      
+      sf_output = self.model(x, pos)
+      
+      return sf_output[1]
+    
+    def forward(self, input_ids, labels=None, **kwargs):
+        x = input_ids
+        batch_size, length = x.size()
+        pos = kwargs['position_ids'] if 'position_ids' in kwargs.keys() else torch.arange(length, device=x.device).expand(batch_size, length)
+        
+        sf_output = self.model(x, pos)
+        
+        loss = None
+        if labels is not None:
+          loss_fct = nn.CrossEntropyLoss()
+          loss = loss_fct(sf_output[0], labels.reshape(-1))
+        
+        return MaskedLMOutput(
+            loss=loss, # shape: 1
+            logits=sf_output[0].view(batch_size, length, -1), # shape: (batch_size, length, ntokens)
+            hidden_states=None,
+            attentions=None
+        )
+
+class StructFormerModelForSequenceClassification(PreTrainedModel):
+    config_class = StructFormerConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = StructFormer(
+              hidden_size=config.hidden_size,
+              nlayers=config.nlayers,
+              ntokens=config.ntokens,
+              nhead=config.nhead,
+              dropout=config.dropout,
+              dropatt=config.dropatt,
+              relative_bias=config.relative_bias,
+              pos_emb=config.pos_emb,
+              pad=config.pad,
+              n_parser_layers=config.n_parser_layers,
+              conv_size=config.conv_size,
+              relations=config.relations,
+              weight_act=config.weight_act
+        )
+        self.config = config
+        self.num_labels = config.num_labels
+        self.model.classifier = ClassificationHead(config)
+    
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, nn.Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            if module.bias is not None:
+                module.bias.data.zero_()
+                module.weight.data.fill_(1.0)
+    
+    def forward(self, input_ids, labels=None, **kwargs):
+        x = input_ids
+        batch_size, length = x.size()
+        pos = kwargs['position_ids'] if 'position_ids' in kwargs.keys() else torch.arange(length, device=x.device).expand(batch_size, length)
+        
+        sf_output = self.model(x, pos)
+        
+        logits = self.model.classifier(sf_output[1]['raw_output'])
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = nn.MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = nn.CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = nn.BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=None,
+            attentions=None,
+        )

finetune/cola/tokenizer_config.json

@@ -0,0 +1,65 @@
+{
+  "add_prefix_space": false,
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "cls_token": {
+    "__type": "AddedToken",
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "errors": "replace",
+  "mask_token": {
+    "__type": "AddedToken",
+    "content": "<mask>",
+    "lstrip": true,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "model_max_length": 512,
+  "name_or_path": "omarmomen/sf_babylm_1",
+  "pad_token": {
+    "__type": "AddedToken",
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "sep_token": {
+    "__type": "AddedToken",
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "special_tokens_map_file": null,
+  "tokenizer_class": "RobertaTokenizer",
+  "trim_offsets": true,
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

finetune/cola/train_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 10.0,
+    "train_loss": 0.3184558808803558,
+    "train_runtime": 219.0841,
+    "train_samples": 8164,
+    "train_samples_per_second": 372.642,
+    "train_steps_per_second": 5.843
+}

finetune/cola/trainer_state.json

@@ -0,0 +1,103 @@
+{
+  "best_metric": 0.8219339622641509,
+  "best_model_checkpoint": "omarmomen/sf_babylm_1/finetune/cola/checkpoint-200",
+  "epoch": 10.0,
+  "global_step": 1280,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 1.56,
+      "eval_accuracy": 0.7036309838294983,
+      "eval_f1": 0.8219339622641509,
+      "eval_loss": 0.6300716400146484,
+      "eval_mcc": 0.13076987235140602,
+      "eval_runtime": 1.2993,
+      "eval_samples_per_second": 784.268,
+      "eval_steps_per_second": 98.515,
+      "step": 200
+    },
+    {
+      "epoch": 3.12,
+      "eval_accuracy": 0.6997056007385254,
+      "eval_f1": 0.8048469387755102,
+      "eval_loss": 0.6442329287528992,
+      "eval_mcc": 0.19657323093149748,
+      "eval_runtime": 1.4161,
+      "eval_samples_per_second": 719.591,
+      "eval_steps_per_second": 90.39,
+      "step": 400
+    },
+    {
+      "epoch": 3.91,
+      "learning_rate": 3.0468750000000002e-05,
+      "loss": 0.511,
+      "step": 500
+    },
+    {
+      "epoch": 4.69,
+      "eval_accuracy": 0.708537757396698,
+      "eval_f1": 0.8109484404837684,
+      "eval_loss": 0.8354039192199707,
+      "eval_mcc": 0.22128351240839433,
+      "eval_runtime": 1.3442,
+      "eval_samples_per_second": 758.05,
+      "eval_steps_per_second": 95.221,
+      "step": 600
+    },
+    {
+      "epoch": 6.25,
+      "eval_accuracy": 0.7144259214401245,
+      "eval_f1": 0.8138195777351248,
+      "eval_loss": 1.0478105545043945,
+      "eval_mcc": 0.2432614865937686,
+      "eval_runtime": 1.3221,
+      "eval_samples_per_second": 770.731,
+      "eval_steps_per_second": 96.814,
+      "step": 800
+    },
+    {
+      "epoch": 7.81,
+      "learning_rate": 1.09375e-05,
+      "loss": 0.2308,
+      "step": 1000
+    },
+    {
+      "epoch": 7.81,
+      "eval_accuracy": 0.7095191478729248,
+      "eval_f1": 0.8097686375321337,
+      "eval_loss": 1.1782832145690918,
+      "eval_mcc": 0.23273105838579872,
+      "eval_runtime": 1.3212,
+      "eval_samples_per_second": 771.289,
+      "eval_steps_per_second": 96.884,
+      "step": 1000
+    },
+    {
+      "epoch": 9.38,
+      "eval_accuracy": 0.7144259214401245,
+      "eval_f1": 0.8135810377962844,
+      "eval_loss": 1.2869043350219727,
+      "eval_mcc": 0.24432326576634322,
+      "eval_runtime": 1.3121,
+      "eval_samples_per_second": 776.633,
+      "eval_steps_per_second": 97.555,
+      "step": 1200
+    },
+    {
+      "epoch": 10.0,
+      "step": 1280,
+      "total_flos": 2092145921249280.0,
+      "train_loss": 0.3184558808803558,
+      "train_runtime": 219.0841,
+      "train_samples_per_second": 372.642,
+      "train_steps_per_second": 5.843
+    }
+  ],
+  "max_steps": 1280,
+  "num_train_epochs": 10,
+  "total_flos": 2092145921249280.0,
+  "trial_name": null,
+  "trial_params": null
+}

finetune/cola/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c43ff5a8741c4285a95db833f1cab13594ae1c36859759e5123678abf6a7eba3
+size 3960