device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
######################################################################
-
parser = argparse.ArgumentParser(description = 'My own GPT.')
parser.add_argument('--log_filename',
######################################################################
def autoregression(
- model,
- nb_samples, nb_tokens_to_generate, starting_input = None,
+ model, batch_size,
+ nb_samples, nb_tokens_to_generate, primer = None,
device = torch.device('cpu')
):
results = torch.zeros(
dtype = torch.int64, device = device
)
- if starting_input is None:
+ if primer is None:
first = 0
else:
- first = starting_input.size(1)
- results = torch.cat((starting_input, results), 1)
+ first = primer.size(1)
+ results = torch.cat((primer, results), 1)
- for input in results.split(args.batch_size):
- for s in tqdm.tqdm(range(first, input.size(1)), desc = 'synth'):
+ for input in results.split(batch_size):
+ for s in range(first, input.size(1)):
output = model(input)
logits = output[:, s]
if args.synthesis_sampling:
def vocabulary_size(self):
pass
- def produce_results(self, n_epoch, model, nb_tokens = 50):
+ def produce_results(self, n_epoch, model):
pass
######################################################################
class TaskPicoCLVR(Task):
+ # Make a tensor from a list of strings
+ def tensorize(self, descr):
+ token_descr = [ s.strip().split(' ') for s in descr ]
+ l = max([ len(s) for s in token_descr ])
+ #token_descr = [ [ '<nul>' ] * (l - len(s)) + s for s in token_descr ]
+ token_descr = [ s + [ '<nul>' ] * (l - len(s)) for s in token_descr ]
+ id_descr = [ [ self.token2id[u] for u in s ] for s in token_descr ]
+ return torch.tensor(id_descr, device = self.device)
+
+ def trim(self, x, token = '<nul>'):
+ n = self.token2id[token]
+ i = (1 - (F.pad(x, (1, 1), value = n) == n).min(0).values.long()).cumsum(0)
+ a, b = (i == 0).nonzero().max(), (i == i.max()).nonzero().min()
+ return x[:, a:b]
+
def __init__(self, batch_size,
height, width, nb_colors = 5,
device = torch.device('cpu')):
def generate_descr(nb):
- descr = picoclvr.generate(
+ return picoclvr.generate(
nb,
height = self.height, width = self.width,
nb_colors = nb_colors
)
- descr = [ s.strip().split(' ') for s in descr ]
- l = max([ len(s) for s in descr ])
- #descr = [ [ '<unk>' ] * (l - len(s)) + s for s in descr ]
- descr = [ s + [ '<unk>' ] * (l - len(s)) for s in descr ]
-
- return descr
-
self.height = height
self.width = width
self.batch_size = batch_size
self.device = device
nb = args.data_size if args.data_size > 0 else 250000
+ log_string(f'generating {nb} samples (can take some time)')
self.train_descr = generate_descr((nb * 4) // 5)
self.test_descr = generate_descr((nb * 1) // 5)
# Build the tokenizer
- tokens = set()
+ tokens = { '<nul>' }
for d in [ self.train_descr, self.test_descr ]:
for s in d:
- for t in s: tokens.add(t)
+ for t in s.strip().split(' '): tokens.add(t)
self.token2id = dict([ (t, n) for n, t in enumerate(tokens) ])
self.id2token = dict([ (n, t) for n, t in enumerate(tokens) ])
# Tokenize the train and test sets
- t = [ [ self.token2id[u] for u in s ] for s in self.train_descr ]
- self.train_input = torch.tensor(t, device = self.device)
- t = [ [ self.token2id[u] for u in s ] for s in self.test_descr ]
- self.test_input = torch.tensor(t, device = self.device)
+ self.train_input = self.tensorize(self.train_descr)
+ self.test_input = self.tensorize(self.test_descr)
def batches(self, split = 'train'):
assert split in { 'train', 'test' }
- if split == 'train':
- for batch in tqdm.tqdm(self.train_input.split(self.batch_size), desc = f'epoch-{split}'):
- yield batch
- else:
- for batch in tqdm.tqdm(self.test_input.split(self.batch_size), desc = f'epoch-{split}'):
- yield batch
+ input = self.train_input if split == 'train' else self.test_input
+ for batch in tqdm.tqdm(input.split(self.batch_size), desc = f'epoch-{split}'):
+ yield self.trim(batch)
def vocabulary_size(self):
return len(self.token2id)
- def generate(self, primer, model, nb_tokens):
- t_primer = primer.strip().split(' ')
- t_generated = [ ]
-
- for j in range(nb_tokens):
- t = [ [ self.token2id[u] for u in t_primer + t_generated ] ]
- input = torch.tensor(t, device = self.device)
- input = F.pad(input, (0, 1)) # Add the next token, the one to predict
- output = model(input)
- logits = output[0, -1]
- if args.synthesis_sampling:
- dist = torch.distributions.categorical.Categorical(logits = logits)
- t_next = dist.sample()
- else:
- t_next = logits.argmax()
- t_generated.append(self.id2token[t_next.item()])
-
- return ' '.join(t_primer + t_generated)
-
- def produce_results(self, n_epoch, model, nb_tokens = None):
- if nb_tokens is None:
- nb_tokens = self.height * self.width + 3
- descr = [ ]
+ def produce_results(self, n_epoch, model):
+ nb_tokens_to_generate = self.height * self.width + 3
+ result_descr = [ ]
nb_per_primer = 8
- for primer in [
+ for primer_descr in [
'red above green <sep> green top <sep> blue right of red <img>',
'there is red <sep> there is yellow <sep> there is blue <img>',
'red below yellow <sep> yellow below green <sep> green below blue <sep> red right <sep> yellow left <sep> green right <sep> blue left <img>',
'green bottom <sep> yellow bottom <sep> green left of blue <sep> yellow right of blue <sep> blue top <img>',
]:
- for k in range(nb_per_primer):
- descr.append(self.generate(primer, model, nb_tokens))
+ results = autoregression(
+ model,
+ self.batch_size,
+ nb_samples = nb_per_primer,
+ nb_tokens_to_generate = nb_tokens_to_generate,
+ primer = self.tensorize([ primer_descr ]).expand(nb_per_primer, -1),
+ device = self.device
+ )
- img = [ picoclvr.descr2img(d, height = self.height, width = self.width) for d in descr ]
- img = torch.cat(img, 0)
- image_name = f'result_picoclvr_{n_epoch:04d}.png'
- torchvision.utils.save_image(
- img / 255.,
- image_name, nrow = nb_per_primer, pad_value = 0.8
- )
- log_string(f'wrote {image_name}')
+ l = [ ' '.join([ self.id2token[t.item()] for t in r ]) for r in results ]
+ result_descr += l
np = picoclvr.nb_properties(
- descr,
+ result_descr,
height = self.height, width = self.width
)
nb_requested_properties, _, nb_missing_properties = zip(*np)
- log_string(f'nb_requested_properties {sum(nb_requested_properties) / len(descr):.02f} nb_missing_properties {sum(nb_missing_properties) / len(descr):.02f}')
+ log_string(f'nb_requested_properties {sum(nb_requested_properties) / len(result_descr):.02f} nb_missing_properties {sum(nb_missing_properties) / len(result_descr):.02f}')
+
+ img = [
+ picoclvr.descr2img(d, height = self.height, width = self.width)
+ for d in result_descr
+ ]
+
+ img = torch.cat(img, 0)
+ image_name = f'result_picoclvr_{n_epoch:04d}.png'
+ torchvision.utils.save_image(
+ img / 255.,
+ image_name, nrow = nb_per_primer, pad_value = 0.8
+ )
+ log_string(f'wrote {image_name}')
######################################################################
self.vocab = torchtext.vocab.build_vocab_from_iterator(
yield_tokens(),
- specials = [ '<unk>', '<non>' ],
+ specials = [ '<unk>', '<nul>' ],
min_freq = self.min_freq
)
self.vocab.set_default_index(self.vocab[ '<unk>' ])
+ # makes a tensor from a list of list of tokens
def tensorize(self, s):
a = max(len(x) for x in s)
- return torch.tensor([ self.vocab(x + [ '<non>' ] * (a - len(x))) for x in s ])
+ return torch.tensor([ self.vocab(x + [ '<nul>' ] * (a - len(x))) for x in s ])
def yield_batches(self, ds):
s = [ ]
def vocabulary_size(self):
return len(self.vocab)
- def produce_results(self, n_epoch, model, nb_tokens = 50):
+ def produce_results(self, n_epoch, model):
+ nb_tokens = 50
file_name = f'result_wiki103_{n_epoch:04d}.txt'
with open(file_name, 'w') as outfile:
else:
t_next = logits.argmax()
t_generated.append(self.vocab.lookup_token(t_next))
- if t_generated[-1] == '<non>': break
+ if t_generated[-1] == '<nul>': break
s = ' '.join(t_generated)
def vocabulary_size(self):
return 256
- def produce_results(self, n_epoch, model, nb_samples = 64):
- results = autoregression(model, nb_samples, 28 * 28, device = self.device)
+ def produce_results(self, n_epoch, model):
+ nb_samples = 64
+ results = autoregression(model, self.batch_size, nb_samples, 28 * 28, device = self.device)
image_name = f'result_mnist_{n_epoch:04d}.png'
torchvision.utils.save_image(1 - results.reshape(-1, 1, 28, 28) / 255.,
image_name, nrow = 16, pad_value = 0.8)
for input in task.batches(split = 'train'):
token_count += F.one_hot(input, num_classes = task.vocabulary_size()).sum((0, 1))
token_probas = token_count / token_count.sum()
-h = -torch.xlogy(token_probas, token_probas).sum()
-train_set_perplexity = math.exp(h)
-log_string(f'train set perplexity {train_set_perplexity}')
+entropy = -torch.xlogy(token_probas, token_probas).sum()
+train_set_perplexity = math.exp(entropy)
for k in range(nb_epochs_finished, nb_epochs):
train_perplexity = math.exp(min(100, acc_train_loss/nb_train_samples))
test_perplexity = math.exp(min(100, acc_test_loss/nb_test_samples))
- log_string(f'perplexity {k} train {train_perplexity} test {test_perplexity}')
+ log_string(f'perplexity {k} train_set {train_set_perplexity} train_prediction {train_perplexity} test_prediction {test_perplexity}')
task.produce_results(k, model)