Create bert-dep.py (#77)

Python version of notebooks/BertDependencies.ipynb

examples/bert-dep.py

@@ -0,0 +1,126 @@
+#!/usr/bin/env python3
+
+import torchtext
+import torch
+
+from torch import nn
+from torch_struct import DependencyCRF
+from torch_struct.data import SubTokenizedField, ConllXDataset, TokenBucket
+from torchtext.data import RawField, BucketIterator
+
+from pytorch_transformers import BertModel, BertTokenizer, AdamW, WarmupLinearSchedule
+
+config = {'bert': 'bert-base-cased', 'H': 768, 'dropout': 0.2}
+
+# parse conll dependency data
+model_class, tokenizer_class, pretrained_weights = BertModel, BertTokenizer, config['bert']
+tokenizer = tokenizer_class.from_pretrained(pretrained_weights)
+
+def batch_num(nums):
+    lengths = torch.tensor([len(n) for n in nums]).long()
+    n = lengths.max()
+    out = torch.zeros(len(nums), n).long()
+    for b, n in enumerate(nums):
+        out[b, :len(n)] = torch.tensor(n)
+    return out, lengths
+
+HEAD = RawField(preprocessing=lambda x: [int(i) for i in x],
+        postprocessing=batch_num)
+HEAD.is_target = True
+WORD = SubTokenizedField(tokenizer)
+
+def len_filt(x): return 5 < len(x.word[0]) < 40
+
+train = ConllXDataset('wsj.train.conllx', (('word', WORD), ('head', HEAD)),
+        filter_pred=len_filt)
+train_iter = TokenBucket(train, 750)
+val = ConllXDataset('wsj.dev.conllx', (('word', WORD), ('head', HEAD)),
+        filter_pred=len_filt)
+val_iter = BucketIterator(val, batch_size=20, device='cuda:0')
+
+# make bert model to compute potentials
+H = config['H']
+class Model(nn.Module):
+    def __init__(self, hidden):
+        super().__init__()
+        self.base_model = model_class.from_pretrained(pretrained_weights)
+        self.linear = nn.Linear(H, H)
+        self.bilinear = nn.Linear(H, H)
+        self.root = nn.Parameter(torch.rand(H))
+        self.dropout = nn.Dropout(config['dropout'])
+
+    def forward(self, words, mapper):
+        out = self.dropout(self.base_model(words)[0])
+        out = torch.matmul(mapper.float().cuda().transpose(1, 2), out)
+        final1 = torch.matmul(out, self.linear.weight)
+        final2 = torch.einsum('bnh,hg,bmg->bnm', out, self.bilinear.weight, final1)
+        root_score = torch.matmul(out, self.root)
+        final2 = final2[:, 1:-1, 1:-1]
+        N = final2.shape[1]
+        final2[:, torch.arange(N), torch.arange(N)] += root_score[:, 1:-1]
+        return final2
+
+model = Model(H)
+model.cuda()
+
+# validation and train loops
+def validate(val_iter):
+    incorrect_edges = 0
+    total_edges = 0
+    model.eval()
+    for i, ex in enumerate(val_iter):
+        words, mapper, _ = ex.word
+        label, lengths = ex.head
+        batch, _ = label.shape
+
+        final = model(words.cuda(), mapper)
+        for b in range(batch):
+            final[b, lengths[b]-1:, :] = 0
+            final[b, :, lengths[b]-1:] = 0
+        dist = DependencyCRF(final, lengths=lengths)
+        gold = dist.struct.to_parts(label, lengths=lengths).type_as(dist.argmax)
+        incorrect_edges += (dist.argmax[:, :].cpu() - gold[:, :].cpu()).abs().sum() / 2.0
+        total_edges += gold.sum()
+
+    print(total_edges, incorrect_edges)
+    model.train()
+
+def train(train_iter, val_iter, model):
+    opt = AdamW(model.parameters(), lr=1e-4, eps=1e-8)
+    scheduler = WarmupLinearSchedule(opt, warmup_steps=20, t_total=2500)
+    model.train()
+    losses = []
+    for i, ex in enumerate(train_iter):
+        opt.zero_grad()
+        words, mapper, _ = ex.word
+        label, lengths = ex.head
+        batch, _ = label.shape
+
+        # Model
+        final = model(words.cuda(), mapper)
+        for b in range(batch):
+            final[b, lengths[b]-1:, :] = 0
+            final[b, :, lengths[b]-1:] = 0
+
+        if not lengths.max() <= final.shape[1] + 1:
+            print("fail")
+            continue
+        dist = DependencyCRF(final, lengths=lengths)
+
+        labels = dist.struct.to_parts(label, lengths=lengths).type_as(final)
+        log_prob = dist.log_prob(labels)
+
+        loss = log_prob.sum()
+        (-loss).backward()
+        torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
+
+        opt.step()
+        scheduler.step()
+        losses.append(loss.detach())
+        if i % 50 == 1:
+            print(-torch.tensor(losses).mean(), words.shape)
+            losses = []
+        if i % 600 == 500:
+            validate(val_iter)
+
+train(train_iter, val_iter, model)