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('--no_checkpoint',
+ action='store_true', default = False)
+
+parser.add_argument('--checkpoint_name',
+ type = str, default = 'checkpoint.pth')
+
+##############################
+# picoclvr options
+
+parser.add_argument('--picoclvr_many_colors',
+ action='store_true', default = False)
+
+parser.add_argument('--picoclvr_height',
+ type = int, default = 12)
+
+parser.add_argument('--picoclvr_width',
+ type = int, default = 16)
######################################################################
class TaskPicoCLVR(Task):
def __init__(self, batch_size,
- height = 6, width = 8, many_colors = False,
+ height, width, many_colors = False,
device = torch.device('cpu')):
+ def generate_descr(nb):
+ descr = picoclvr.generate(
+ nb,
+ height = self.height, width = self.width,
+ many_colors = many_colors
+ )
+
+ 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 ]
+
+ 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
- descr = picoclvr.generate(
- nb,
- height = height, width = width,
- many_colors = many_colors
- )
-
- 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 ]
+ self.train_descr = generate_descr((nb * 4) // 5)
+ self.test_descr = generate_descr((nb * 1) // 5)
tokens = set()
- for s in descr:
- for t in s: tokens.add(t)
+ for d in [ self.train_descr, self.test_descr ]:
+ for s in d:
+ for t in s: 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) ])
- t = [ [ self.token2id[u] for u in s ] for s in descr ]
- data_input = torch.tensor(t, device = self.device)
-
- self.test_input = data_input[:nb // 5]
- self.train_input = data_input[nb // 5:]
+ 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)
def batches(self, split = 'train'):
assert split in { 'train', 'test' }
def vocabulary_size(self):
return len(self.token2id)
- def produce_results(self, n_epoch, model, nb_tokens = 50):
- img = [ ]
+ 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 = None):
+ if nb_tokens is None:
+ nb_tokens = self.height * self.width + 3
+ 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, height = self.height, width = self.width) 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)
+ 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,
+ height = self.height, width = self.width
+ )
+ ] )
+
+ log_string(f'nb_missing {nb_missing / len(descr):.02f}')
+
######################################################################
class TaskWiki103(Task):
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,
+ height = args.picoclvr_height,
+ width = args.picoclvr_width,
+ 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
+
+if args.no_checkpoint:
+ log_string(f'Not trying to load checkpoint.')
+
+else:
+ 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'])
+ log_string(f'Checkpoint loaded with {nb_epochs_finished} epochs finished.')
+
+ except FileNotFoundError:
+ log_string('Starting from scratch.')
+
+ except:
+ log_string('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