File size: 9,492 Bytes
d491737 0526506 d491737 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 |
import torch
import torch.nn as nn
import torch.nn.functional as F
from gist1.gpt import GPT
from gist1.vqvae import VQVAE
from utils.misc import load_params
class VQVAETransformer(nn.Module):
def __init__(self, args):
super().__init__()
self.vqvae = self.load_vqvae(args)
self.transformer = self.load_transformer(args)
# self.sos_token = self.get_sos_token(args)
self.pkeep = args['pkeep']
self.vqvae_vocab_size = args['vocab_size']
self.loc_vocab_size = args['loc_vocab_size']
self.block_size = args['block_size']
def load_vqvae(self, args):
# VQVAE_path = args['vqvae_checkpoint']
# VQVAE_cfg = args['vqvae_cfg']
# cfg = load_params(VQVAE_cfg)
# seed= cfg['seed']
# torch.manual_seed(seed)
num_hiddens = args['vqvae_num_hiddens']
num_residual_layers = args['vqvae_num_residual_layers']
num_residual_hiddens = args['vqvae_num_residual_hiddens']
num_embeddings = args['latent_dim']
latent_dim = args['vqvae_latent_dim']
commitment_cost = args['vqvae_commitment_cost']
decay = args['vqvae_decay']
model = VQVAE(num_hiddens, num_residual_layers, num_residual_hiddens,
num_embeddings, latent_dim, commitment_cost,
decay)
# model.load_state_dict(torch.load(VQVAE_path))
# model = model.eval()
# update args from vqvae cfg
args['vocab_size'] = num_embeddings
return model
def load_transformer(self, args):
# seed= args['seed']
# torch.manual_seed(seed)
latent_dim = args['latent_dim']
heads = args['heads']
N = args['N']
block_size = args['block_size']
vocab_size = args['vocab_size'] + args['loc_vocab_size']
model = GPT(vocab_size, latent_dim, N, heads, block_size)
return model
@torch.no_grad()
def encode_to_z(self, x):
quantized, indices = self.vqvae.encode(x)
indices = indices.view(quantized.shape[0], -1)
return quantized, indices
@ torch.no_grad()
def z_to_isovist(self, indices):
indices[indices > self.vqvae_vocab_size-1] = self.vqvae_vocab_size-1
embedding_dim = self.vqvae.vq.embedding_dim
ix_to_vectors = self.vqvae.vq.embedding(indices).reshape(indices.shape[0], -1, embedding_dim)
ix_to_vectors = ix_to_vectors.permute(0, 2, 1)
isovist = self.vqvae.decode(ix_to_vectors)
return isovist
def loc_to_indices(self, x):
starting_index = self.vqvae_vocab_size
indices = x.long() + starting_index
return indices
def indices_to_loc(self, indices):
starting_index = self.vqvae_vocab_size
locs = indices - starting_index
locs[locs < 0] = 0
locs[locs > (self.loc_vocab_size-1)] = self.loc_vocab_size-1
return locs
def seq_encode(self, locs, isovists):
# BSW
indices_seq = []
# indices_loc = []
for i in range(isovists.shape[1]): # iterate trought the sequence
loc = locs[:, i].unsqueeze(1) # BL
indices_seq.append(self.loc_to_indices(loc))
isovist = isovists[:, i, :].unsqueeze(1) # BCW
_, indices = self.encode_to_z(isovist)
indices_seq.append(indices)
indices = torch.cat(indices_seq, dim=1)
return indices
def forward(self, indices):
device = indices.device
# indices = self.seq_encode(locs, isovists)
if self.training and self.pkeep < 1.0:
mask = torch.bernoulli(self.pkeep*torch.ones(indices.shape, device=device))
mask = mask.round().to(dtype=torch.int64)
random_indices = torch.randint_like(indices, self.vqvae_vocab_size) # doesn't include sos token
new_indices = mask*indices + (1-mask)*random_indices
else:
new_indices = indices
target = indices[:, 1:]
logits = self.transformer(new_indices[:, :-1])
return logits, target
def top_k_logits(self, logits, k):
v, ix = torch.topk(logits, k)
out = logits.clone()
out[out < v[..., [-1]]] = -float("inf")
return out
def sample(self, x, steps, temp=1.0, top_k=100, seed=None, step_size=17, zeroing=False):
device = x.device
is_train = False
if self.transformer.training == True:
is_train = True
self.transformer.eval()
block_size = self.block_size
generator = None
if seed is not None:
generator = torch.Generator(device).manual_seed(seed)
for k in range(steps):
if x.size(1) < block_size:
x_cond = x
else:
remain = step_size - (x.size(1) % step_size)
x_cond = x[:, -(block_size-remain):] # crop context if needed
if zeroing:
x_cond = x_cond.clone()
x_cond[:, 0] = self.vqvae_vocab_size
logits = self.transformer(x_cond)
logits = logits[:, -1, :] / temp
if top_k is not None:
logits = self.top_k_logits(logits, top_k)
probs = F.softmax(logits, dim = -1)
ix = torch.multinomial(probs, num_samples=1, generator=generator)
x = torch.cat((x, ix), dim=1)
if is_train == True:
self.transformer.train()
return x
def get_loc(self, ploc, dir):
if dir == 0:
loc = ploc
elif dir == 1:
loc = (ploc[0]+1, ploc[1])
elif dir == 2:
loc = (ploc[0]+1, ploc[1]+1)
elif dir == 3:
loc = (ploc[0], ploc[1]+1)
elif dir == 4:
loc = (ploc[0]-1, ploc[1]+1)
elif dir == 5:
loc = (ploc[0]-1, ploc[1])
elif dir == 6:
loc = (ploc[0]-1, ploc[1]-1)
elif dir == 7:
loc = (ploc[0], ploc[1]-1)
elif dir == 8:
loc = (ploc[0]+1, ploc[1]-1)
else:
raise NameError('Direction unknown')
return loc
def init_loc(self, x, step_size):
device = x.device
loc_dict = {}
loc = None
cached_loc = None
if x.shape[1] > 1:
steps = x.shape[1] -1
for k in range(steps):
if k % step_size == 0:
dir = x[:,k].detach().item() - self.vqvae_vocab_size
if dir == 0:
loc = (0, 0) # init loc
else:
loc = self.get_loc(loc, dir) # getloc
loc_dict[loc] = torch.empty(1,0).long().to(device)
cached_loc = loc
else:
ix = x[:,[k]]
loc_dict[cached_loc] = torch.cat((loc_dict[cached_loc], ix), dim = 1)
return loc_dict, loc
def sample_memorized(self, x, steps, temp=1.0, top_k=100, seed=None, step_size=17, zeroing=False):
device = x.device
loc_dict, loc = self.init_loc(x, step_size)
is_train = False
if self.transformer.training == True:
is_train = True
self.transformer.eval()
block_size = self.block_size
generator = None
if seed is not None:
generator = torch.Generator(device).manual_seed(seed)
is_visited = False
cache_counter = 0
# loc = None
for k in range(steps):
# check directionality
if k % step_size == 0:
dir = x[:,-1].detach().item() - self.vqvae_vocab_size
if dir == 0:
is_visited = False
loc = (0, 0) # init loc
loc_dict[loc] = torch.empty(1,0).long().to(device)
else:
loc = self.get_loc(loc, dir) # getloc
if loc in loc_dict:
is_visited = True
cache_counter = 0
else:
is_visited = False
loc_dict[loc] = torch.empty(1,0).long().to(device)
if x.size(1) < block_size:
x_cond = x
else:
remain = step_size - (x.size(1) % step_size)
x_cond = x[:, -(block_size-remain):] # crop context if needed
if zeroing:
x_cond = x_cond.clone()
x_cond[:, 0] = self.vqvae_vocab_size
if is_visited == False:
logits = self.transformer(x_cond)
logits = logits[:, -1, :] / temp
if top_k is not None:
logits = self.top_k_logits(logits, top_k)
probs = F.softmax(logits, dim = -1)
ix = torch.multinomial(probs, num_samples=1, generator=generator)
# print('this shouldnt')
loc_dict[loc] = torch.cat((loc_dict[loc], ix), dim = 1)
else:
if cache_counter == 15: #reaching end of latent code
is_visited = False
ix = loc_dict[loc][:,[cache_counter]]
# print(ix)
cache_counter += 1
x = torch.cat((x, ix), dim=1)
if is_train == True:
self.transformer.train()
return x
|