-
Notifications
You must be signed in to change notification settings - Fork 5
/
MNIST_generative_transformer.py
205 lines (189 loc) · 8.62 KB
/
MNIST_generative_transformer.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
## Import transformers
from transformers import get_linear_schedule_with_warmup
from transformers import BertTokenizer, BertForSequenceClassification
from transformers import BertModel, BertConfig
from transformers import GPT2LMHeadModel, GPT2Tokenizer, GPT2Config, GPT2Model
## import MNIST from torchvision package
from torchvision import datasets, transforms
## import torch
import os
from os.path import join
from tqdm import tqdm, trange
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.optim import AdamW, Adam
from torch.utils.data import Dataset, DataLoader
from torchvision.utils import make_grid, save_image
import matplotlib.pyplot as plt
#%%
from torchvision.datasets import MNIST
dataset = MNIST(root='/home/binxu/Datasets', download=True, transform=transforms.ToTensor())
#%%
# image to patch sequence by convolution
def img2patch(img, patch_size=2):
# img: (batch_size, channel, height, width)
# patch_size: int
# return: (batch_size, channel, height//patch_size, width//patch_size, patch_size**2)
batch_size, channel, height, width = img.shape
img = img.reshape(batch_size, channel, height//patch_size, patch_size, width//patch_size, patch_size)
img = img.permute(0, 2, 4, 1, 3, 5)
img = img.reshape(batch_size, height//patch_size, width//patch_size, channel * patch_size**2)
return img
def patch2seq(patches):
batch_size, height, width, channel = patches.shape
return patches.reshape(batch_size, height*width, channel)
def seq2img(patch_seq, patch_size=2):
# patch_seq: (batch_size, channel, height//patch_size, width//patch_size, patch_size**2)
# patch_size: int
# return: (batch_size, channel, height, width)
batch_size, HW, hidden = patch_seq.shape
height = width = int(math.sqrt(HW))
channel = hidden // patch_size**2
patch_seq = patch_seq.reshape(batch_size, height, width, channel, patch_size, patch_size)
patch_seq = patch_seq.permute(0, 3, 1, 4, 2, 5)
imgtsr = patch_seq.reshape(batch_size, channel, height*patch_size, width*patch_size)
return imgtsr
#%%
# GPT2Config()
dataloaders = DataLoader(dataset, batch_size=64, shuffle=True)
imgs, labels = next(iter(dataloaders))
print(patch2seq(img2patch(imgs, patch_size=4)).shape)
assert torch.allclose(seq2img(patch2seq(img2patch(imgs,)),), imgs)
assert torch.allclose(seq2img(patch2seq(img2patch(imgs, patch_size=4)), patch_size=4), imgs)
#%%
#%% Conditional model
patch_size = 4
config = GPT2Config(n_embd=128, n_layer=24, n_head=16, n_positions=256,
vocab_size=100, bos_token_id=101, eos_token_id=102,
add_cross_attention=True, )
model = GPT2Model(config).cuda()
patch_emb = nn.Linear(patch_size * patch_size, config.n_embd).cuda()
patch_readout = nn.Linear(config.n_embd, patch_size * patch_size).cuda()
digit_emb = nn.Embedding(10, config.n_embd).cuda()
optimizer = AdamW([*model.parameters(),
*patch_emb.parameters(),
*patch_readout.parameters(),
*digit_emb.parameters()], lr=5e-4)
#%%
def generate_img(prompt_digit, prompt_patch, model, digit_emb, patch_emb, patch_readout,
patch_size=4, pixel_size=28):
max_seq_len = (pixel_size // patch_size) ** 2 - 1
prompt_digit_emb = digit_emb(torch.tensor([*prompt_digit]).cuda())[:, None, :]
patch_seq = [prompt_patch]
input_patch_emb = patch_emb(prompt_patch)
with torch.no_grad():
for i in range(max_seq_len):
output = model(inputs_embeds=input_patch_emb,
encoder_hidden_states=prompt_digit_emb)
output_hiddens = output.last_hidden_state
next_patch = patch_readout(output_hiddens[:, -1:, :])
input_patch_emb = torch.cat([input_patch_emb, patch_emb(next_patch)], dim=1)
patch_seq.append(next_patch)
patch_seq_tsr = torch.cat(patch_seq, dim=1)
gen_imgs = seq2img(patch_seq_tsr, patch_size=patch_size)
return gen_imgs
#%%
# saveroot = r"D:\DL_Projects\Vision\pixel_GPT"
saveroot = r"/home/binxu/DL_Projects/patchGPT"
runname = "conditional"
os.makedirs(join(saveroot, runname), exist_ok=True)
batch_size = 512
dataloaders = DataLoader(dataset, batch_size=batch_size, shuffle=True)
for epoch in trange(1,50):
pbar = tqdm(dataloaders)
model.train()
for ibatch, (imgs, labels) in enumerate(pbar):
digit_hiddens = digit_emb(labels.cuda())[:, None, :]
patch_seq = patch2seq(img2patch(imgs.cuda(), patch_size=patch_size))
input_embeds = patch_emb(patch_seq)
output = model(inputs_embeds=input_embeds, encoder_hidden_states=digit_hiddens)
output_hiddens = output.last_hidden_state
output_patches = patch_readout(output_hiddens)
loss = F.mse_loss(output_patches[:, :-1, :], patch_seq[:, 1:, :])
optimizer.zero_grad()
loss.backward()
optimizer.step()
pbar.set_description(f"loss: {loss.item():.4f}")
# print(loss.item())
prompt_digit = range(10)
prompt_patch = torch.zeros(10, 1, patch_size * patch_size).cuda()
model.eval()
gen_imgs = generate_img(prompt_digit, prompt_patch, model, digit_emb, patch_emb, patch_readout,
patch_size=patch_size, pixel_size=28)
save_image(make_grid(gen_imgs, nrow=5), join(saveroot, runname, f'{epoch}_genimages.png'))
model.save_pretrained(join(saveroot, runname, "model"))
patch_emb.cpu().save(join(saveroot, runname, "patch_emb.pth"))
patch_readout.cpu().save(join(saveroot, runname, "patch_readout.pth"))
#%%
#%%
prompt_digit = range(10)
prompt_patch = torch.zeros(10, 1, patch_size * patch_size).cuda()
gen_imgs = generate_img(prompt_digit, prompt_patch, model, digit_emb, patch_emb, patch_readout,
patch_size=patch_size, pixel_size=28)
plt.figure()
plt.imshow(make_grid(gen_imgs, nrow=5).permute(1, 2, 0).cpu())
plt.show()
#%%
def generate_img_uncond(prompt_patch, model, patch_emb, patch_readout,
patch_size=4, pixel_size=28):
max_seq_len = (pixel_size // patch_size) ** 2 - 1
patch_seq = [prompt_patch]
input_patch_emb = patch_emb(prompt_patch)
with torch.no_grad():
for i in range(max_seq_len):
output = model(inputs_embeds=input_patch_emb,)
output_hiddens = output.last_hidden_state
next_patch = patch_readout(output_hiddens[:, -1:, :])
input_patch_emb = torch.cat([input_patch_emb, patch_emb(next_patch)], dim=1)
patch_seq.append(next_patch)
patch_seq_tsr = torch.cat(patch_seq, dim=1)
gen_imgs = seq2img(patch_seq_tsr, patch_size=patch_size)
return gen_imgs
#%%
patch_size = 4
config = GPT2Config(n_embd=128, n_layer=24, n_head=16, n_positions=256,
vocab_size=100, bos_token_id=101, eos_token_id=102, )
model = GPT2Model(config).cuda()
patch_emb = nn.Linear(patch_size * patch_size, config.n_embd).cuda()
patch_readout = nn.Linear(config.n_embd, patch_size * patch_size).cuda()
optimizer = AdamW([*model.parameters(),
*patch_emb.parameters(),
*patch_readout.parameters()], lr=5e-4)
saveroot = r"/home/binxu/DL_Projects/patchGPT"
runname = "unconditional"
os.makedirs(join(saveroot, runname), exist_ok=True)
#%%
batch_size = 512
dataloaders = DataLoader(dataset, batch_size=batch_size, shuffle=True)
for epoch in range(50):
pbar = tqdm(dataloaders)
model.train()
for ibatch, (imgs, labels) in enumerate(pbar):
patch_seq = patch2seq(img2patch(imgs.cuda(), patch_size=patch_size))
input_embeds = patch_emb(patch_seq)
output = model(inputs_embeds=input_embeds)
output_hiddens = output.last_hidden_state
output_patches = patch_readout(output_hiddens)
loss = F.mse_loss(output_patches[:, :-1, :], patch_seq[:, 1:, :])
optimizer.zero_grad()
loss.backward()
optimizer.step()
pbar.set_description(f"loss: {loss.item():.4f}")
# print(loss.item())
prompt_digit = range(10)
prompt_patch = torch.zeros(10, 1, patch_size * patch_size).cuda()
model.eval()
gen_imgs = generate_img_uncond(prompt_patch, model, patch_emb, patch_readout,
patch_size=patch_size, pixel_size=28)
save_image(make_grid(gen_imgs, nrow=5), join(saveroot, runname, f'{epoch}_genimages.png'))
model.save_pretrained(join(saveroot, runname, "model"))
patch_emb.cpu().save(join(saveroot, runname, "patch_emb.pth"))
patch_readout.cpu().save(join(saveroot, runname, "patch_readout.pth"))
#%%
make_grid(seq2img(output_patches, patch_size=patch_size))
# plt.imshow(make_grid(imgs[:16, :, :, :]).permute(1, 2, 0))
plt.figure()
plt.imshow(make_grid(seq2img(output_patches.detach().cpu(), patch_size=patch_size)).permute(1, 2, 0))
plt.show()