wassname/multifit
1
0
Fork 0
mirror of https://github.com/wassname/multifit.git synced 2026-09-09 11:27:26 +08:00
Code Issues Packages Projects Releases Wiki Activity

Extract params to an experiment data class

Browse Source 
You can run this as follows:

`python -m ulmfit.pretrain_lm --dir-path 'data/wiki/wikitext-2'  --qrnn=True train_lm --num_epochs=1`
This commit is contained in:
Piotr Czapla
2018-11-25 12:37:45 +01:00
parent eea9be09db
commit 80d4d4da29
2 changed files with 170 additions and 138 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/test_end_to_end.py
+8 -8
View File
@@ -50,17 +50,18 @@ def test_ulmfit_default_end_to_end():
test_data, wt2 = get_test_data()
lm_name = 'end-to-end-test-default'
cuda_id = 0
results = ulmfit.pretrain_lm.pretrain_lm(
exp = ulmfit.pretrain_lm.Experiment(
dir_path=wt2,
lang='en',
cuda_id=cuda_id,
qrnn=True,
subword=False,
max_vocab=1000,
bs=2,
num_epochs=1,
name=lm_name)
assert results['accuracy'] > 0.02
exp.train_lm(num_epochs=1)
assert exp.results['accuracy'] > 0.02
results = ulmfit.train_clas.new_train_clas(
data_dir=test_data,
@@ -82,7 +83,7 @@ def test_ulmfit_sentencepiece_end_to_end():
imdb, wt2 = get_test_data()
lm_name = 'end-to-end-test-spm'
cuda_id = 0
results = ulmfit.pretrain_lm.pretrain_lm(
exp = ulmfit.pretrain_lm.Experiment(
dir_path=wt2,
lang='en',
cuda_id=cuda_id,
@@ -90,11 +91,10 @@ def test_ulmfit_sentencepiece_end_to_end():
subword=True,
max_vocab=100,
bs=2,
num_epochs=1,
name=lm_name,
)
assert results['accuracy'] > 0.30
exp.train_lm(num_epochs=1)
assert exp.results['accuracy'] > 0.30
# NOTE: ds_pct is not available for sentencepiece -- tests are on the complete dataset
# sentencepiece for finetuning/classification is currently not implemented
ulmfit/pretrain_lm.py
+162 -130
View File
@@ -25,159 +25,191 @@ import fastai_contrib.data as contrib_data
# cupy needs to be installed for QRNN
def pretrain_lm(dir_path, lang='en', cuda_id=0, qrnn=True, subword=False, max_vocab=60000,
bs=70, bptt=70, name='wt-103', num_epochs=10, bidir=False, ds_pct=1.0):
"""
:param dir_path: The path to the directory of the file.
:param lang: the language unicode
:param cuda_id: The id of the GPU. Uses GPU 0 by default or no GPU when
run on CPU.
:param qrnn: Use a QRNN. Requires installing cupy.
:param subword: Use sub-word tokenization on the cleaned data.
:param max_vocab: The maximum size of the vocabulary.
:param bs: The batch size.
:param bptt: The back-propagation-through-time sequence length.
:param name: The name used for both the model and the vocabulary.
:param model_dir: The path to the directory where the models should be saved
:param bidir: whether the language model is bidirectional
"""
results = {}
# """
# :param dir_path: The path to the directory of the file.
# :param lang: the language unicode
# :param cuda_id: The id of the GPU. Uses GPU 0 by default or no GPU when
# run on CPU.
# :param qrnn: Use a QRNN. Requires installing cupy.
# :param subword: Use sub-word tokenization on the cleaned data.
# :param max_vocab: The maximum size of the vocabulary.
# :param bs: The batch size.
# :param bptt: The back-propagation-through-time sequence length.
# :param name: The name used for both the model and the vocabulary.
# :param model_dir: The path to the directory where the models should be saved
# :param bidir: whether the language model is bidirectional
# """
if not torch.cuda.is_available():
print('CUDA not available. Setting device=-1.')
cuda_id = -1
torch.cuda.set_device(cuda_id)
dir_path = Path(dir_path)
assert dir_path.exists()
model_dir = dir_path / 'models' # removed from params, as it is absolute models location in train_clas and here it is relative
model_dir.mkdir(exist_ok=True)
print('Batch size:', bs)
print('Max vocab:', max_vocab)
model_name = 'qrnn' if qrnn else 'lstm'
if qrnn:
print('Using QRNNs...')
@dataclass
class Experiment:
dir_path: str
bidir: bool =False
bptt: int = 70
bs: int = 70
lang: str = 'en'
max_vocab: int = 60000
name: str = 'wt-103'
subword: bool = False
ds_pct: float = 1.0
qrnn: bool = True
cuda_id:int = 0
def __post_init__(self):
self.results = {}
trn_path = dir_path / f'{lang}.wiki.train.tokens'
val_path = dir_path / f'{lang}.wiki.valid.tokens'
tst_path = dir_path / f'{lang}.wiki.test.tokens'
for path_ in [trn_path, val_path, tst_path]:
assert path_.exists(), f'Error: {path_} does not exist.'
if not torch.cuda.is_available():
print('CUDA not available. Setting device=-1.')
cuda_id = -1
torch.cuda.set_device(self.cuda_id)
if subword:
# apply sentencepiece tokenization
trn_path = dir_path / f'{lang}.wiki.train.tokens'
val_path = dir_path / f'{lang}.wiki.valid.tokens'
self.dir_path = Path(self.dir_path)
assert self.dir_path.exists()
self.model_dir = self.dir_path / 'models' # removed from params, as it is absolute models location in train_clas and here it is relative
self.model_dir.mkdir(exist_ok=True)
print('Batch size:', self.bs)
print('Max vocab:', self.max_vocab)
read_file(trn_path, 'train')
read_file(val_path, 'valid')
if self.qrnn:
print('Using QRNNs...')
sp = get_sentencepiece(dir_path, trn_path, name, vocab_size=max_vocab)
@property
def model_name(self):
return 'qrnn' if self.qrnn else 'lstm'
lm_type = contrib_data.LanguageModelType.BiLM if bidir else contrib_data.LanguageModelType.FwdLM
def train_lm(self, num_epochs=10):
data_lm = self.load_data()
exe = Executor(self.create_lm_learner(data_lm), exp=self)
if num_epochs > 0:
exe.learn.fit_one_cycle(num_epochs, 5e-3, (0.8, 0.7), wd=1e-7)
exe.validate()
exe.save()
return exe
data_lm = TextLMDataBunch.from_csv(dir_path, 'train.csv', **sp, bs=bs, bptt=bptt, lm_type=lm_type)
itos = data_lm.train_ds.vocab.itos
stoi = data_lm.train_ds.vocab.stoi
else:
# read the already whitespace separated data without any preprocessing
trn_tok = read_whitespace_file(trn_path)
val_tok = read_whitespace_file(val_path)
if ds_pct < 1.0:
trn_tok = trn_tok[:max(20, int(len(trn_tok) * ds_pct))]
val_tok = val_tok[:max(20, int(len(val_tok) * ds_pct))]
print(f"Limiting data sets to {ds_pct*100}%, trn {len(trn_tok)}, val: {len(val_tok)}")
itos_fname = model_dir / f'itos_{name}.pkl'
if not itos_fname.exists():
# create the vocabulary
cnt = Counter(word for sent in trn_tok for word in sent)
itos = [o for o,c in cnt.most_common(n=max_vocab)]
itos.insert(1, PAD) #  set pad id to 1 to conform to fast.ai standard
assert UNK in itos, f'Unknown words are expected to have been replaced with {UNK} in the data.'
# save vocabulary
print(f"Saving vocabulary as {itos_fname}")
results['itos_fname'] = itos_fname
with open(itos_fname, 'wb') as f:
pickle.dump(itos, f)
def create_lm_learner(self, data_lm):
# these hyperparameters are for training on ~100M tokens (e.g. WikiText-103)
# for training on smaller datasets, more dropout is necessary
if self.qrnn:
emb_sz, nh, nl = 400, 1550, 3
# dps = np.array([0.0, 0.0, 0.0, 0.0, 0.0])
dps = np.array([0.25, 0.1, 0.2, 0.02, 0.15])
drop_mult = 0.1
else:
print("Loading itos:", itos_fname)
itos = np.load(itos_fname)
vocab = Vocab(itos)
stoi = vocab.stoi
emb_sz, nh, nl = 400, 1150, 3
# emb_sz, nh, nl = 400, 1150, 3
dps = np.array([0.25, 0.1, 0.2, 0.02, 0.15])
drop_mult = 0.1
fastai.text.learner.default_dropout['language'] = dps
lm_learner = bilm_learner if self.bidir else language_model_learner
learn = lm_learner(data_lm, bptt=self.bptt, emb_sz=emb_sz, nh=nh, nl=nl, pad_token=1,
drop_mult=drop_mult, tie_weights=True, model_dir=self.model_dir.name,
bias=True, qrnn=self.qrnn, clip=0.12)
# compared to standard Adam, we set beta_1 to 0.8
learn.opt_fn = partial(optim.Adam, betas=(0.8, 0.99))
learn.true_wd = False
print("true_wd: ", learn.true_wd)
if self.bidir:
learn.metrics = [accuracy_fwd, accuracy_bwd]
else:
learn.metrics = [accuracy]
return learn
trn_ids = np.array([([stoi.get(w, stoi[UNK]) for w in s]) for s in trn_tok])
val_ids = np.array([([stoi.get(w, stoi[UNK]) for w in s]) for s in val_tok])
lm_type = contrib_data.LanguageModelType.BiLM if bidir else contrib_data.LanguageModelType.FwdLM
def load_data(self):
trn_path = self.dir_path / f'{self.lang}.wiki.train.tokens'
val_path = self.dir_path / f'{self.lang}.wiki.valid.tokens'
tst_path = self.dir_path / f'{self.lang}.wiki.test.tokens'
for path_ in [trn_path, val_path, tst_path]:
assert path_.exists(), f'Error: {path_} does not exist.'
if self.subword:
# apply sentencepiece tokenization
trn_path = self.dir_path / f'{self.lang}.wiki.train.tokens'
val_path = self.dir_path / f'{self.lang}.wiki.valid.tokens'
# data_lm = TextLMDataBunch.from_ids(dir_path, trn_ids, [], val_ids, [], len(itos))
data_lm = TextLMDataBunch.from_ids(path=dir_path, vocab=vocab, train_ids=trn_ids,
valid_ids=val_ids, bs=bs, bptt=bptt,
lm_type=lm_type
)
read_file(trn_path, 'train')
read_file(val_path, 'valid')
print('Size of vocabulary:', len(itos))
print('First 10 words in vocab:', ', '.join([itos[i] for i in range(10)]))
sp = get_sentencepiece(self.dir_path, trn_path, self.name, vocab_size=self.max_vocab)
# these hyperparameters are for training on ~100M tokens (e.g. WikiText-103)
# for training on smaller datasets, more dropout is necessary
if qrnn:
emb_sz, nh, nl = 400, 1550, 3
#dps = np.array([0.0, 0.0, 0.0, 0.0, 0.0])
dps = np.array([0.25, 0.1, 0.2, 0.02, 0.15])
drop_mult = 0.1
else:
emb_sz, nh, nl = 400, 1150, 3
# emb_sz, nh, nl = 400, 1150, 3
dps = np.array([0.25, 0.1, 0.2, 0.02, 0.15])
drop_mult = 0.1
lm_type = contrib_data.LanguageModelType.BiLM if self.bidir else contrib_data.LanguageModelType.FwdLM
fastai.text.learner.default_dropout['language'] = dps
data_lm = TextLMDataBunch.from_csv(self.dir_path, 'train.csv', **sp, bs=self.bs, bptt=self.bptt, lm_type=lm_type)
itos = data_lm.train_ds.vocab.itos
stoi = data_lm.train_ds.vocab.stoi
else:
# read the already whitespace separated data without any preprocessing
trn_tok = read_whitespace_file(trn_path)
val_tok = read_whitespace_file(val_path)
if self.ds_pct < 1.0:
trn_tok = trn_tok[:max(20, int(len(trn_tok) * self.ds_pct))]
val_tok = val_tok[:max(20, int(len(val_tok) * self.ds_pct))]
print(f"Limiting data sets to {self.ds_pct * 100}%, trn {len(trn_tok)}, val: {len(val_tok)}")
lm_learner = bilm_learner if bidir else language_model_learner
learn = lm_learner(data_lm, bptt=bptt, emb_sz=emb_sz, nh=nh, nl=nl, pad_token=1,
drop_mult=drop_mult, tie_weights=True, model_dir=model_dir.name,
bias=True, qrnn=qrnn, clip=0.12)
# compared to standard Adam, we set beta_1 to 0.8
learn.opt_fn = partial(optim.Adam, betas=(0.8, 0.99))
itos_fname = self.model_dir / f'itos_{self.name}.pkl'
if not itos_fname.exists():
# create the vocabulary
cnt = Counter(word for sent in trn_tok for word in sent)
itos = [o for o, c in cnt.most_common(n=self.max_vocab)]
itos.insert(1, PAD) #   set pad id to 1 to conform to fast.ai standard
assert UNK in itos, f'Unknown words are expected to have been replaced with {UNK} in the data.'
learn.true_wd = False
print("true_wd: ", learn.true_wd)
# save vocabulary
print(f"Saving vocabulary as {itos_fname}")
self.results['itos_fname'] = itos_fname
with open(itos_fname, 'wb') as f:
pickle.dump(itos, f)
else:
print("Loading itos:", itos_fname)
itos = np.load(itos_fname)
vocab = Vocab(itos)
stoi = vocab.stoi
if bidir:
learn.metrics = [accuracy_fwd, accuracy_bwd]
else:
learn.metrics = [accuracy]
trn_ids = np.array([([stoi.get(w, stoi[UNK]) for w in s]) for s in trn_tok])
val_ids = np.array([([stoi.get(w, stoi[UNK]) for w in s]) for s in val_tok])
try:
learn.load(f'{model_name}_{name}')
print("Weights loaded")
except FileNotFoundError:
print("Starting from random weights")
pass
lm_type = contrib_data.LanguageModelType.BiLM if self.bidir else contrib_data.LanguageModelType.FwdLM
learn.fit_one_cycle(num_epochs, 5e-3, (0.8, 0.7), wd=1e-7)
# data_lm = TextLMDataBunch.from_ids(dir_path, trn_ids, [], val_ids, [], len(itos))
data_lm = TextLMDataBunch.from_ids(path=self.dir_path, vocab=vocab, train_ids=trn_ids,
valid_ids=val_ids, bs=self.bs, bptt=self.bptt,
lm_type=lm_type)
itos, stoi, trn_path = data_lm.vocab.itos, data_lm.vocab.stoi, data_lm.path
print('Size of vocabulary:', len(itos))
print('First 10 words in vocab:', ', '.join([itos[i] for i in range(10)]))
return data_lm
if not subword and max_vocab is None:
# only if we use the unpreprocessed version and the full vocabulary
# are the perplexity results comparable to previous work
print(f"Validating model performance with test tokens from: {trn_path}")
tst_tok = read_whitespace_file(trn_path)
tst_ids = np.array([([stoi.get(w, stoi[UNK]) for w in s]) for s in tst_tok])
logloss, perplexity = validate(learn.model, tst_ids, bptt)
print('Test logloss:', logloss.item(), 'perplexity:', perplexity.item())
print(f"Saving models at {learn.path / learn.model_dir}")
learn.save(f'{model_name}_{name}')
class Executor:
def __init__(self, learn, exp):
self.exp = exp
self.learn = learn
try:
self.learn.load(f'{self.exp.model_name}_{self.exp.name}')
print("Weights loaded")
except FileNotFoundError:
print("Starting from random weights")
pass
opt_state_path = learn.path / learn.model_dir / f'{model_name}3_{name}_state.pth'
print(f"Saving optimiser state at {opt_state_path}")
torch.save(learn.opt.opt.state_dict(), opt_state_path)
def validate(self):
if not self.exp.subword and self.exp.max_vocab is None:
raise NotImplementedError("figure out how to validate and save results")
# only if we use the unpreprocessed version and the full vocabulary
# are the perplexity results comparable to previous work
print(f"Validating model performance with test tokens from: {trn_path}")
tst_tok = read_whitespace_file(trn_path)
tst_ids = np.array([([stoi.get(w, stoi[UNK]) for w in s]) for s in tst_tok])
logloss, perplexity = validate(learn.model, tst_ids, self.exp.bptt)
print('Test logloss:', logloss.item(), 'perplexity:', perplexity.item())
def save(self):
print(f"Saving models at {self.learn.path / self.learn.model_dir}")
self.learn.save(f'{self.exp.model_name}_{self.exp.name}')
opt_state_path = self.learn.path / self.learn.model_dir / f'{self.exp.model_name}_{self.exp.name}_state.pth'
print(f"Saving optimiser state at {opt_state_path}")
torch.save(self.learn.opt.opt.state_dict(), opt_state_path)
self.exp.results['accuracy'] = self.learn.validate()[1] #TODO rewrite
results['accuracy'] = learn.validate()[1]
return results
if __name__ == '__main__':
fire.Fire(pretrain_lm)
fire.Fire(Experiment)