xla_lstm.py

import torch
import torch.nn as nn


class XlaLSTM(nn.Module):

  def __init__(self, input_sz, hidden_sz, batch_first=False, pad_value=0):
    super(XlaLSTM, self).__init__()
    self.input_sz = input_sz
    self.hidden_size = hidden_sz
    if batch_first:
      self.batch_dim, self.sequence_dim = 0, 1
    else:
      self.batch_dim, self.sequence_dim = 1, 0
    self.pad_value = pad_value
    self.weight_ih = nn.Parameter(torch.Tensor(input_sz, hidden_sz * 4))
    self.weight_hh = nn.Parameter(torch.Tensor(hidden_sz, hidden_sz * 4))
    self.bias = nn.Parameter(torch.Tensor(hidden_sz * 4))
    self.init_weights()

  def init_weights(self):
    for p in self.parameters():
      if p.data.ndimension() >= 2:
        nn.init.xavier_uniform_(p.data)
      else:
        nn.init.zeros_(p.data)

  def sequence_slice(self, t, embed_out, embed_in):
    ot = embed_out[:, t, :] if self.sequence_dim == 1 else embed_out[t, :, :]
    it = embed_in[:, t] if self.sequence_dim == 1 else embed_in[t, :]
    return ot, it

  def sizes(self, embed_out):
    size = embed_out.size()
    return size[self.batch_dim], size[self.sequence_dim]

  def forward(self, embed_out, embed_in, init_states=None):
    bs, seq_sz = self.sizes(embed_out)
    if init_states is None:
      h_t, c_t = (torch.zeros(bs, self.hidden_size, device=embed_out.device),
                  torch.zeros(bs, self.hidden_size, device=embed_out.device))
    else:
      h_t, c_t = init_states
    hstate = []
    HS = self.hidden_size
    for t in range(0, seq_sz):
      feat, iseq = self.sequence_slice(t, embed_out, embed_in)
      gates = feat @ self.weight_ih + h_t @ self.weight_hh + self.bias
      i_t, f_t, g_t, o_t = (
          torch.sigmoid(gates[:, :HS]),
          torch.sigmoid(gates[:, HS:HS * 2]),
          torch.tanh(gates[:, HS * 2:HS * 3]),
          torch.sigmoid(gates[:, HS * 3:]),
      )
      fwd = iseq.unsqueeze(1) != self.pad_value
      c_t = torch.where(fwd, f_t * c_t + i_t * g_t, c_t)
      h_t = torch.where(fwd, o_t * torch.tanh(c_t), h_t)
      hstate.append(h_t.unsqueeze(self.sequence_dim))
    hstate = torch.cat(hstate, dim=self.sequence_dim)
    return hstate, (h_t, c_t)


#### TEST
import collections
import csv
import nltk
import random
import torch.optim as optim
import torch_xla
import torch_xla.core.xla_model as xm
import torch_xla.debug.metrics as met
import torch_xla.utils.gcsfs as gcs


def maybe_int(v):
  try:
    return int(v)
  except ValueError:
    pass


def read_dataset(path):
  with gcs.generic_open(path, mode='r') as fd:
    reader = csv.DictReader(fd)
    vocab = {'PAD': 0}
    sentences, max_target = [], -1
    for fields in reader:
      target = maybe_int(fields['target'])
      if target is None:
        continue
      sentence = []
      for tok in nltk.word_tokenize(fields['sentence']):
        ltok = tok.lower()
        tid = vocab.get(ltok, None)
        if tid is None:
          tid = len(vocab)
          vocab[ltok] = tid
        sentence.append(tid)
      if sentence:
        sentences.append((sentence, target))
        max_target = max(max_target, target)

    return sentences, vocab, max_target


def split_and_pad(sentences, splits, pad_value=0):
  splits = sorted(splits)
  split_dict = collections.defaultdict(list)
  discarded = 0
  for sentence, target in sentences:
    lsent = len(sentence)
    for i in range(0, len(splits)):
      if lsent < splits[i]:
        break
    seqlen = splits[i]
    if lsent > seqlen:
      discarded += 1
      continue
    padded_sentence = sentence + [pad_value] * (seqlen - lsent)
    split_dict[seqlen].append((padded_sentence, target))

  return split_dict, discarded


def make_batches(split_dict, batch_size):
  batch_dict = dict()
  for seqlen, slist in split_dict.items():
    batches = []
    i = 0
    while i + batch_size <= len(slist):
      batches.append(slist[i:i + batch_size])
      i += batch_size
    if i < len(slist):
      batch = slist[i:]
      while len(batch) < batch_size:
        batch.append(slist[random.randint(0, len(slist) - 1)])
      batches.append(batch)
    batch_dict[seqlen] = batches
  return batch_dict


def to_one_hot(y, n_dims, dtype=torch.float32):
  scatter_dim = len(y.size())
  y_tensor = y.view(*y.size(), -1)
  zeros = torch.zeros(*y.size(), n_dims, dtype=dtype, device=y.device)
  return zeros.scatter(scatter_dim, y_tensor, 1)


def gen_tensors(batch_dict, target_dims, shuffle=True):
  slist = []
  for seqlen in sorted(batch_dict.keys()):
    slist += batch_dict[seqlen]
  if shuffle:
    random.shuffle(slist)
  tensors = []
  for bseq in slist:
    sentence_data, target_data = [], []
    for sentence, target in bseq:
      sentence_data.append(sentence)
      target_data.append(target)
    sentence_tensor = torch.tensor(sentence_data, dtype=torch.int64)
    target_tensor = torch.tensor(target_data, dtype=torch.int64)
    onehot_tensor = to_one_hot(target_tensor, target_dims)
    tensors.append((sentence_tensor, onehot_tensor))
  return tensors


class TestClassifier(nn.Module):

  def __init__(self,
               vocab_size,
               embedding_dim,
               hidden_dim,
               output_dim,
               padding_idx=None):
    super().__init__()
    self.embedding = nn.Embedding(
        vocab_size, embedding_dim, padding_idx=padding_idx)
    self.lstm = XlaLSTM(embedding_dim, hidden_dim, batch_first=True)
    self.fc = nn.Linear(hidden_dim, output_dim)
    self.act = nn.Sigmoid()

  def forward(self, sentence_tensor):
    embedded = self.embedding(sentence_tensor)
    output, (hidden, cell) = self.lstm(embedded, sentence_tensor)
    dense_outputs = self.fc(hidden)
    return self.act(dense_outputs)


def test_model(path,
               device,
               splits,
               batch_size,
               embed_size=8,
               per_target_hs_size=256,
               lr=0.01,
               momentum=None,
               epochs=1,
               log_interval=10):
  MIN_HS_SIZE = 128
  MAX_HS_SIZE = 1024 * 32

  sentences, vocab, max_target = read_dataset(path)
  split_dict, discarded = split_and_pad(sentences, splits)
  batch_dict = make_batches(split_dict, batch_size)
  train_tensors = gen_tensors(batch_dict, max_target + 1)

  hs_size = min(MAX_HS_SIZE,
                max((max_target + 1) * per_target_hs_size, MIN_HS_SIZE))

  model = TestClassifier(
      len(vocab), embed_size, hs_size, max_target + 1, padding_idx=0)
  model.to(device)
  model.train()

  optimizer = optim.SGD(model.parameters(), lr=lr, momentum=momentum)
  criterion = nn.MSELoss().to(device)

  for epoch in range(0, epochs):
    n = 0
    for sentence_tensor, target_tensor in train_tensors:
      sentence_tensor = sentence_tensor.to(device)
      target_tensor = target_tensor.to(device)

      optimizer.zero_grad()
      output = model(sentence_tensor)

      loss = criterion(output, target_tensor)
      loss.backward()

      xm.optimizer_step(optimizer, barrier=True)

      # print('OUTPUT:\n', torch_xla._XLAC._get_xla_tensors_text([output]))
      n += 1
      if n % log_interval == 0:
        print('[{}] Loss: {:.4f}'.format(epoch, loss.cpu().item()))


torch.manual_seed(11)
device = xm.xla_device()
cvs_path = 'gs://davide-stg1/lstm_test_data.csv'
nltk.download('punkt')

test_model(
    cvs_path,
    device, (8, 16, 32, 64),
    batch_size=16,
    embed_size=128,
    lr=0.01,
    momentum=0.9,
    epochs=10)

print(met.metrics_report())