type = str, default = 'train.log')
parser.add_argument('--download',
- type = bool, default = False)
+ action='store_true', default = False)
parser.add_argument('--seed',
type = int, default = 0)
type = float, default = 0.1)
parser.add_argument('--synthesis_sampling',
- type = bool, default = True)
+ action='store_true', default = True)
+
+parser.add_argument('--checkpoint_name',
+ type = str, default = 'checkpoint.pth')
+
+parser.add_argument('--picoclvr_many_colors',
+ action='store_true', default = False)
######################################################################
class TaskPicoCLVR(Task):
- def __init__(self, batch_size, height = 6, width = 8, device = torch.device('cpu')):
+ def __init__(self, batch_size,
+ height = 6, width = 8, many_colors = False,
+ device = torch.device('cpu')):
+
self.batch_size = batch_size
self.device = device
nb = args.data_size if args.data_size > 0 else 250000
- descr = picoclvr.generate(nb, height = height, width = width)
+ descr = picoclvr.generate(
+ nb,
+ height = height, width = width,
+ many_colors = many_colors
+ )
+
+ # self.test_descr = descr[:nb // 5]
+ # self.train_descr = descr[nb // 5:]
+
descr = [ s.strip().split(' ') for s in descr ]
l = max([ len(s) for s in descr ])
descr = [ s + [ '<unk>' ] * (l - len(s)) for s in 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 = 'epoch'):
+ 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 = 'epoch'):
+ for batch in tqdm.tqdm(self.test_input.split(self.batch_size), desc = f'epoch-{split}'):
yield 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)
+ output = model(input)
+ logits = output[0, -1]
+ if args.synthesis_sampling:
+ dist = torch.distributions.categorical.Categorical(logits = logits)
+ t = dist.sample()
+ else:
+ t = logits.argmax()
+ t_generated.append(self.id2token[t.item()])
+
+ return ' '.join(t_primer + t_generated)
+
def produce_results(self, n_epoch, model, nb_tokens = 50):
- img = [ ]
+ descr = [ ]
nb_per_primer = 8
for primer in [
]:
for k in range(nb_per_primer):
- 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)
- output = model(input)
- logits = output[0, -1]
- if args.synthesis_sampling:
- dist = torch.distributions.categorical.Categorical(logits = logits)
- t = dist.sample()
- else:
- t = logits.argmax()
- t_generated.append(self.id2token[t.item()])
-
- descr = [ ' '.join(t_primer + t_generated) ]
- img += [ picoclvr.descr2img(descr) ]
+ descr.append(self.generate(primer, model, nb_tokens))
+ img = [ picoclvr.descr2img(d) for d in descr ]
img = torch.cat(img, 0)
file_name = f'result_picoclvr_{n_epoch:04d}.png'
torchvision.utils.save_image(img / 255.,
file_name, nrow = nb_per_primer, pad_value = 0.8)
log_string(f'wrote {file_name}')
+ nb_missing = sum( [ x[2] for x in picoclvr.nb_missing_properties(descr) ] )
+ log_string(f'nb_missing {nb_missing / len(descr):.02f}')
+
######################################################################
class TaskWiki103(Task):
if args.data_size > 0:
data_iter = itertools.islice(data_iter, args.data_size)
- return self.yield_batches(tqdm.tqdm(data_iter, desc = 'epoch'))
+ return self.yield_batches(tqdm.tqdm(data_iter, desc = f'epoch-{split}'))
def vocabulary_size(self):
return len(self.vocab)
data_input = data_set.data.view(-1, 28 * 28).long()
if args.data_size >= 0:
data_input = data_input[:args.data_size]
- for batch in tqdm.tqdm(data_input.split(self.batch_size), desc = 'epoch'):
+ for batch in tqdm.tqdm(data_input.split(self.batch_size), desc = f'epoch-{split}'):
yield batch
def vocabulary_size(self):
elif args.data == 'mnist':
task = TaskMNIST(batch_size = args.batch_size, device = device)
elif args.data == 'picoclvr':
- task = TaskPicoCLVR(batch_size = args.batch_size, device = device)
+ task = TaskPicoCLVR(batch_size = args.batch_size, many_colors = args.picoclvr_many_colors, device = device)
else:
raise ValueError(f'Unknown dataset {args.data}.')
nb_heads = args.nb_heads, nb_blocks = args.nb_blocks, dropout = args.dropout
)
+model.to(device)
+
nb_parameters = sum(p.numel() for p in model.parameters())
log_string(f'nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)')
-model.to(device)
-
######################################################################
if args.optim == 'sgd':
else:
raise ValueError(f'Unknown optimizer {args.optim}.')
-for k in range(args.nb_epochs):
+######################################################################
+
+nb_epochs_finished = 0
+
+try:
+ checkpoint = torch.load(args.checkpoint_name, map_location = device)
+ nb_epochs_finished = checkpoint['nb_epochs_finished']
+ model.load_state_dict(checkpoint['model_state'])
+ optimizer.load_state_dict(checkpoint['optimizer_state'])
+ print(f'Checkpoint loaded with {nb_epochs_finished} epochs finished.')
+
+except FileNotFoundError:
+ print('Starting from scratch.')
+
+except:
+ print('Error when loading the checkpoint.')
+ exit(1)
+
+######################################################################
+
+for k in range(nb_epochs_finished, args.nb_epochs):
model.train()
task.produce_results(k, model)
+ checkpoint = {
+ 'nb_epochs_finished': k + 1,
+ 'model_state': model.state_dict(),
+ 'optimizer_state': optimizer.state_dict()
+ }
+
+ torch.save(checkpoint, args.checkpoint_name)
+
######################################################################
projects / mygpt.git / blobdiff