X-Git-Url: https://www.fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=main.py;h=77b1b226f95c4d9a11dd6fe5990d575ae0795399;hb=b4c255babaae72d6a03b4c8e8e7e25f6ab0a19a0;hp=65922040a78fc60ada7bbf75c5373ace576ee4c6;hpb=a238fb780003e847d388861c41f0facdc5738dd0;p=mygpt.git diff --git a/main.py b/main.py index 6592204..77b1b22 100755 --- a/main.py +++ b/main.py @@ -111,8 +111,8 @@ for n in vars(args): ###################################################################### 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( @@ -120,13 +120,13 @@ def autoregression( 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 input in results.split(batch_size): for s in tqdm.tqdm(range(first, input.size(1)), desc = 'synth'): output = model(input) logits = output[:, s] @@ -157,6 +157,10 @@ import picoclvr class TaskPicoCLVR(Task): + def descr2tensor(self, descr): + t = [ [ self.token2id[u] for u in s ] for s in descr ] + return torch.tensor(t, device = self.device) + def __init__(self, batch_size, height, width, nb_colors = 5, device = torch.device('cpu')): @@ -193,10 +197,8 @@ class TaskPicoCLVR(Task): 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 = descr2tensor(self.train_descr) + self.test_input = descr2tensor(self.test_descr) def batches(self, split = 'train'): assert split in { 'train', 'test' } @@ -210,32 +212,21 @@ class TaskPicoCLVR(Task): 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 generate(self, primer_descr, model, nb_tokens): + results = autoregression( + model, self.batch_size, + 1, nb_tokens, primer = descr2tensor(primer_descr), + device = self.device + ) + return ' '.join([ self.id2token[t.item()] for t in results.flatten() ]) def produce_results(self, n_epoch, model, nb_tokens = None): if nb_tokens is None: nb_tokens = self.height * self.width + 3 - descr = [ ] + result_descr = [ ] nb_per_primer = 8 - for primer in [ + for primer_descr in [ 'red above green green top blue right of red ', 'there is red there is yellow there is blue ', 'red below yellow yellow below green green below blue red right yellow left green right blue left ', @@ -243,9 +234,10 @@ class TaskPicoCLVR(Task): ]: for k in range(nb_per_primer): - descr.append(self.generate(primer, model, nb_tokens)) + result_descr.append(self.generate(primer_descr, model, nb_tokens)) - img = [ picoclvr.descr2img(d, height = self.height, width = self.width) for d in descr ] + 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( @@ -255,13 +247,13 @@ class TaskPicoCLVR(Task): log_string(f'wrote {image_name}') 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}') ###################################################################### @@ -386,7 +378,7 @@ class TaskMNIST(Task): return 256 def produce_results(self, n_epoch, model, nb_samples = 64): - results = autoregression(model, nb_samples, 28 * 28, device = self.device) + 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) @@ -470,9 +462,9 @@ token_count = 0 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) +#log_string(f'train set perplexity {train_set_perplexity}') for k in range(nb_epochs_finished, nb_epochs): @@ -507,7 +499,7 @@ 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)