--- /dev/null
+#!/usr/bin/env python
+
+# Any copyright is dedicated to the Public Domain.
+# https://creativecommons.org/publicdomain/zero/1.0/
+
+# Written by Francois Fleuret <francois@fleuret.org>
+
+import math, sys, argparse, time, tqdm, itertools
+
+import torch, torchtext, torchvision
+from torch import nn
+from torch.nn import functional as F
+
+######################################################################
+
+device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+
+######################################################################
+
+parser = argparse.ArgumentParser(description = 'My own GPT.')
+
+parser.add_argument('--log_filename',
+ type = str, default = 'train.log')
+
+parser.add_argument('--download',
+ type = bool, default = False)
+
+parser.add_argument('--seed',
+ type = int, default = 0)
+
+parser.add_argument('--nb_epochs',
+ type = int, default = 100)
+
+parser.add_argument('--batch_size',
+ type = int, default = 25)
+
+parser.add_argument('--data',
+ type = str, default = 'wiki103')
+
+parser.add_argument('--data_size',
+ type = int, default = -1)
+
+parser.add_argument('--optim',
+ type = str, default = 'adam')
+
+parser.add_argument('--learning_rate',
+ type = float, default = 1e-4)
+
+parser.add_argument('--dim_model',
+ type = int, default = 512)
+
+parser.add_argument('--dim_keys',
+ type = int, default = 64)
+
+parser.add_argument('--dim_hidden',
+ type = int, default = 2048)
+
+parser.add_argument('--nb_heads',
+ type = int, default = 8)
+
+parser.add_argument('--nb_blocks',
+ type = int, default = 12)
+
+parser.add_argument('--dropout',
+ type = float, default = 0.1)
+
+parser.add_argument('--synthesis_sampling',
+ type = bool, default = True)
+
+######################################################################
+
+args = parser.parse_args()
+
+log_file = open(args.log_filename, 'w')
+
+if args.seed >= 0:
+ torch.manual_seed(args.seed)
+
+######################################################################
+
+def log_string(s):
+ t = time.strftime('%Y%m%d-%H:%M:%S ', time.localtime())
+
+ if log_file is not None:
+ log_file.write(t + s + '\n')
+ log_file.flush()
+
+ print(t + s)
+ sys.stdout.flush()
+
+for n in vars(args):
+ log_string(f'args.{n} {getattr(args, n)}')
+
+##############################
+
+class Residual(nn.Module):
+ def __init__(self, *f):
+ super().__init__()
+ self.f = f[0] if len(f) == 1 else nn.Sequential(*f)
+
+ def forward(self, x):
+ return x + self.f(x)
+
+##############################
+
+class PositionalEncoding(nn.Module):
+ def __init__(self, len_max):
+ super().__init__()
+ self.len_max = len_max
+
+ # From Vaswani et al 2018
+ # PE_{t,2i} = sin(t/(L^{2i/D}))
+ # PE_{t,2i+1} = cos(t/(L^{2i/D}))
+ def forward(self, x):
+ t = torch.arange(x.size(1), dtype = x.dtype, device = x.device)[:, None]
+ j = torch.arange(x.size(2), dtype = x.dtype, device = x.device)[None, :]
+ k = j%2
+ return x + torch.sin(t / (self.len_max ** ((j - k) / x.size(2))) + math.pi/2 * k)[None, :, :]
+
+##############################
+
+class QKVAttention(nn.Module):
+ def __init__(self, dim_in, dim_qk, dim_v, nb_heads = 1, causal = False, attention_dropout = 0.0):
+ super().__init__()
+
+ def randw(*d):
+ return nn.Parameter(torch.empty(*d).normal_(0, 1 / math.sqrt(d[-1])))
+
+ self.wq = randw(nb_heads, dim_qk, dim_in)
+ self.wk = randw(nb_heads, dim_qk, dim_in)
+ self.wv = randw(nb_heads, dim_v, dim_in)
+ self.causal = causal
+ self.attention_dropout = attention_dropout
+
+ def forward(self, x):
+ q = torch.einsum('ntc,hdc->nhtd', x, self.wq)
+ k = torch.einsum('ntc,hdc->nhtd', x, self.wk)
+ v = torch.einsum('ntc,hdc->nhtd', x, self.wv)
+ r = math.sqrt(q.size(3))
+ a = torch.einsum('nhtd,nhsd->nhts', q, k).div(r)
+ if self.causal:
+ mask = torch.tril(q.new_ones(a.size(2), a.size(3)))[None, None, :, :] == 0
+ a = a.masked_fill(mask, float('-inf'))
+ a = a.softmax(dim = 3)
+ a = F.dropout(a, self.attention_dropout, self.training)
+ y = torch.einsum('nhts,nhsd->nhtd', a, v)
+ return y.permute(0, 2, 1, 3).flatten(2) # nhtd -> nt(hd)
+
+##############################
+
+class MyGPT(nn.Module):
+ def __init__(self,
+ vocabulary_size,
+ dim_model, dim_keys, dim_hidden,
+ nb_heads, nb_blocks, dropout = 0.):
+
+ super().__init__()
+
+ assert dim_model % nb_heads == 0
+
+ self.embedding = nn.Sequential(
+ nn.Embedding(vocabulary_size, dim_model),
+ nn.Dropout(dropout),
+ PositionalEncoding(len_max = 1e5),
+ )
+
+ trunk_blocks = [ ]
+
+ for _ in range(nb_blocks):
+ trunk_blocks += [
+ Residual(
+ nn.LayerNorm(dim_model),
+ QKVAttention(
+ dim_in = dim_model,
+ dim_qk = dim_keys, dim_v = dim_model // nb_heads,
+ nb_heads = nb_heads,
+ causal = True, attention_dropout = dropout
+ ),
+ nn.Linear(in_features = dim_model, out_features = dim_model),
+ ),
+ Residual(
+ nn.LayerNorm(dim_model),
+ nn.Linear(in_features = dim_model, out_features = dim_hidden),
+ nn.ReLU(),
+ nn.Linear(in_features = dim_hidden, out_features = dim_model),
+ nn.Dropout(dropout),
+ ),
+ ]
+
+ self.trunk = nn.Sequential(*trunk_blocks)
+
+ self.readout = nn.Linear(in_features = dim_model, out_features = vocabulary_size)
+
+ def forward(self, x):
+ x = self.embedding(x)
+ x = self.trunk(x)
+ x = self.readout(x)
+ return x
+
+######################################################################
+
+class Task:
+ def batches(self, split = 'train'):
+ pass
+
+ def vocabulary_size(self):
+ pass
+
+ def produce_results(self, n_epoch, model, nb_tokens = 50):
+ pass
+
+######################################################################
+
+import picoclvr
+
+class TaskPicoCLVR(Task):
+
+ def __init__(self, batch_size, height = 6, width = 8, 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 = [ 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 ]
+
+ tokens = set()
+ for s in descr:
+ 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:]
+
+ 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'):
+ yield batch
+ else:
+ for batch in tqdm.tqdm(self.test_input.split(self.batch_size), desc = 'epoch'):
+ yield batch
+
+ def vocabulary_size(self):
+ return len(self.token2id)
+
+ def produce_results(self, n_epoch, model, nb_tokens = 50):
+ img = [ ]
+ nb_per_primer = 8
+ for primer 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):
+ 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) ]
+
+ 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}')
+
+######################################################################
+
+class TaskWiki103(Task):
+
+ def __init__(self, batch_size, len_min = 10, len_max = 200, min_freq = 100,
+ device = torch.device('cpu')):
+
+ self.batch_size = batch_size
+ self.len_min = len_min
+ self.len_max = len_max
+ self.min_freq = min_freq
+ self.device = device
+
+ self.tokenizer = torchtext.data.get_tokenizer('basic_english')
+ train_iter = torchtext.datasets.WikiText103(split = 'train', root = './data/nlp/')
+
+ # Mostly for debug
+ if args.data_size > 0:
+ train_iter = itertools.islice(train_iter, args.data_size)
+
+ def yield_tokens():
+ for l in tqdm.tqdm(train_iter, desc = 'vocab'):
+ yield self.tokenizer(l)
+
+ self.vocab = torchtext.vocab.build_vocab_from_iterator(
+ yield_tokens(),
+ specials = [ '<unk>', '<non>' ],
+ min_freq = self.min_freq
+ )
+
+ self.vocab.set_default_index(self.vocab[ '<unk>' ])
+
+ 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 ])
+
+ def yield_batches(self, ds):
+ s = [ ]
+ for l in ds:
+ q = self.tokenizer(l)
+ if len(q) >= self.len_min and len(q) <= self.len_max:
+ s += [ q ]
+ if len(s) == self.batch_size:
+ yield self.tensorize(s)
+ s = [ ]
+
+ if len(s) > 0:
+ yield self.tensorize(s)
+
+ def batches(self, split = 'train'):
+ data_iter = torchtext.datasets.WikiText103(split = split, root = './data/nlp/')
+
+ # Mostly for debug
+ if args.data_size > 0:
+ data_iter = itertools.islice(data_iter, args.data_size)
+
+ return self.yield_batches(tqdm.tqdm(data_iter, desc = 'epoch'))
+
+ def vocabulary_size(self):
+ return len(self.vocab)
+
+ 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:
+ for primer in [
+ 'the cat is hunting a',
+ 'paris is the capital',
+ 'cars are convenient',
+ 'the difference between men and women is',
+ 'the object was blue all over and green all over it was',
+ 'cherries are red and lemons are',
+ 'cherries are sweet and lemons are',
+ 'two plus three equals',
+ 'deep learning is',
+ ]:
+ t_primer = self.tokenizer(primer)
+ t_generated = [ ]
+
+ for j in range(nb_tokens):
+
+ input = self.tensorize([ t_primer + t_generated ]).to(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.vocab.lookup_token(t))
+ if t_generated[-1] == '<non>': break
+
+ s = ' '.join(t_generated)
+
+ outfile.write(f'<{primer}> {s}\n')
+
+ log_string(f'wrote {file_name}')
+
+######################################################################
+
+class TaskMNIST(Task):
+
+ def __init__(self, batch_size, device = torch.device('cpu')):
+ self.device = device
+ self.batch_size = batch_size
+
+ def batches(self, split = 'train'):
+ assert split in { 'train', 'test' }
+ data_set = torchvision.datasets.MNIST(
+ root = './data', train = (split == 'train'),
+ download = True
+ )
+ 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'):
+ yield batch
+
+ def vocabulary_size(self):
+ return 256
+
+ def produce_results(self, n_epoch, model, nb_samples = 64):
+ results = torch.zeros(nb_samples, 28 * 28, dtype = torch.int64, device = self.device)
+ for input in results.split(self.batch_size):
+ for s in tqdm.tqdm(range(input.size(1) - 1), desc = 'synth'):
+ output = model(input)
+ logits = output[:, s]
+ if args.synthesis_sampling:
+ dist = torch.distributions.categorical.Categorical(logits = logits)
+ t = dist.sample()
+ else:
+ t = logits.argmax(1)
+ input[:, s + 1] = t
+
+ 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)
+ log_string(f'wrote {image_name}')
+
+######################################################################
+
+def check_causality(model):
+ #m = model[1:]
+ input = torch.rand(1, 5, dim_model).requires_grad_()
+ output = m(input)
+ a = torch.zeros(output.size(1), input.size(1))
+ for k in range(output.size(1)):
+ for d in range(output.size(2)):
+ g, = torch.autograd.grad(output[0, k, d], input, retain_graph = True)
+ a[k] += g.squeeze(0).pow(2).sum(1)
+ print(a)
+
+######################################################################
+
+log_string(f'device {device}')
+
+if args.data == 'wiki103':
+ task = TaskWiki103(batch_size = args.batch_size, device = device)
+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)
+else:
+ raise ValueError(f'Unknown dataset {args.data}.')
+
+vocabulary_size = task.vocabulary_size()
+
+log_string(f'vocabulary_size {vocabulary_size}')
+
+##############################
+
+model = MyGPT(
+ vocabulary_size = vocabulary_size,
+ dim_model = args.dim_model, dim_keys = args.dim_keys, dim_hidden = args.dim_hidden,
+ nb_heads = args.nb_heads, nb_blocks = args.nb_blocks, dropout = args.dropout
+)
+
+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':
+ optimizer = torch.optim.SGD(model.parameters(), lr = args.learning_rate)
+elif args.optim == 'adam':
+ optimizer = torch.optim.Adam(model.parameters(), lr = args.learning_rate)
+elif args.optim == 'adamw':
+ optimizer = torch.optim.AdamW(model.parameters(), lr = args.learning_rate)
+else:
+ raise ValueError(f'Unknown optimizer {args.optim}.')
+
+for k in range(args.nb_epochs):
+
+ model.train()
+
+ nb_train_samples, acc_train_loss = 0, 0.0
+
+ for input in task.batches(split = 'train'):
+ input = input.to(device)
+ output = model(input)
+ loss = F.cross_entropy(output[:, :-1].transpose(1, 2), input[:, 1:])
+ acc_train_loss += loss.item() * input.size(0)
+ nb_train_samples += input.size(0)
+
+ optimizer.zero_grad()
+ loss.backward()
+ optimizer.step()
+
+ with torch.autograd.no_grad():
+
+ model.eval()
+
+ nb_test_samples, acc_test_loss = 0, 0.0
+
+ for input in task.batches(split = 'test'):
+ input = input.to(device)
+ output = model(input)
+ loss = F.cross_entropy(output[:, :-1].transpose(1, 2), input[:, 1:])
+ acc_test_loss += loss.item() * input.size(0)
+ nb_test_samples += input.size(0)
+
+ log_string(f'perplexity {k+1} train {math.exp(min(100, acc_train_loss/nb_train_samples))} test {math.exp(min(100, acc_test_loss/nb_test_samples))}')
+
+ task.produce_results(k, model)
+
+######################################################################