File size: 16,309 Bytes
b7b35a5 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 | import numpy as np
import torch
from torch import nn
from torch.nn import functional as F
def get_patches_fast(images, cfg):
from einops import rearrange
batch_size, height, width, channels = images.shape
patch_size = cfg.patch_size ## n_patches = 8
patches = rearrange(images[:,:,:,:3], 'b (h p1) (w p2) c -> b (h w) (p1 p2 c)', p1 = patch_size, p2 = patch_size)
if channels > 3:
## History stacking in the channel dimension for observations only, not goal images.
patches = rearrange(images, 'b (h p1) (w p2) (c hs) -> b (h w hs) (p1 p2 c)', p1 = patch_size, p2 = patch_size, hs=cfg.policy.obs_stacking) ## Stack the history in the channel dimension
return patches
def calc_positional_embeddings(sequence_length, d):
result = torch.ones(sequence_length, d)
for i in range(sequence_length):
for j in range(d):
result[i][j] = np.sin(i / (10000 ** (j / d))) if j % 2 == 0 else np.cos(i / (10000 ** ((j - 1) / d)))
return result
class Head(nn.Module):
""" one head of self-attention """
def __init__(self, head_size, n_embd, dropout):
super().__init__()
self.key = nn.Linear(n_embd, head_size, bias=False)
self.query = nn.Linear(n_embd, head_size, bias=False)
self.value = nn.Linear(n_embd, head_size, bias=False)
self.dropout = nn.Dropout(dropout)
def forward(self, x, mask=None):
B,T,C = x.shape
# TODO:
## Provide the block masking logic for the attention head
k = self.key(x)
q = self.query(x)
wei = q @ k.transpose(-2,-1) * C**-0.5
wei = wei.masked_fill(mask == 0, float('-inf'))
wei = F.softmax(wei, dim=-1)
wei = self.dropout(wei)
v = self.value(x)
out = wei @ v
return out
class MultiHeadAttention(nn.Module):
def __init__(self, num_heads, head_size, n_embd, dropout):
super().__init__()
self.heads = nn.ModuleList([Head(head_size, n_embd=n_embd, dropout=dropout) for _ in range(num_heads)])
self.proj = nn.Linear(n_embd, n_embd)
self.dropout = nn.Dropout(dropout)
def forward(self, x, mask=None):
with torch.profiler.record_function("Self-Attention"):
out = torch.cat([h(x, mask) for h in self.heads], dim=-1)
out = self.dropout(self.proj(out))
return out
class FeedFoward(nn.Module):
def __init__(self, n_embd, dropout):
super().__init__()
self.net = nn.Sequential(
nn.Linear(n_embd, 4 * n_embd),
nn.ReLU(),
nn.Linear(4 * n_embd, n_embd),
nn.Dropout(dropout),
)
def forward(self, x):
return self.net(x)
class Block(nn.Module):
def __init__(self, n_embd, n_head, dropout):
super().__init__()
head_size = n_embd // n_head
self.sa = MultiHeadAttention(n_head, head_size, n_embd=n_embd, dropout=dropout)
self.ffwd = FeedFoward(n_embd, dropout)
self.ln1 = nn.LayerNorm(n_embd)
self.ln2 = nn.LayerNorm(n_embd)
def forward(self, x, mask=None):
x = x + self.sa(self.ln1(x), mask)
x = x + self.ffwd(self.ln2(x))
return x
class GRP(nn.Module):
def __init__(self, cfg, mlp_ratio=4):
super(GRP, self).__init__()
self._cfg = cfg
chars = cfg.dataset.chars_list
cfg.vocab_size = len(chars)
# TODO:
## Provide the logic for the GRP network
# 1) Patch embedding layer
self.patch_embedding = nn.Linear(cfg.patch_size * cfg.patch_size * 3, cfg.n_embd)
# 2) Learnable token embeddings for classification and goal image tokens
self.class_token = nn.Parameter(torch.zeros(1, 1, cfg.n_embd))
self.goal_token = nn.Parameter(torch.zeros(1, 1, cfg.n_embd))
# 3) Token embedding table for text goals (if not using T5)
self.token_embedding_table = nn.Embedding(cfg.vocab_size, cfg.n_embd)
self.dropout = nn.Dropout(cfg.dropout)
# 4) Transformer encoder blocks
self.blocks = nn.ModuleList([Block(cfg.n_embd, cfg.n_head, cfg.dropout) for _ in range(cfg.n_blocks)])
self.ln_f = nn.LayerNorm(cfg.n_embd)
# 5) Classification MLP head
if cfg.action_space == "continuous":
self.action_head = nn.Sequential(
nn.Linear(cfg.n_embd, cfg.n_embd * mlp_ratio),
nn.ReLU(),
nn.Linear(cfg.n_embd * mlp_ratio, cfg.action_dim * cfg.policy.action_stacking)
)
elif cfg.action_space == "discrete":
self.action_head = nn.Sequential(
nn.Linear(cfg.n_embd, cfg.n_embd * mlp_ratio),
nn.ReLU(),
nn.Linear(cfg.n_embd * mlp_ratio, cfg.action_dim * cfg.policy.action_stacking * 14)
)
# Weight initialization
self.apply(self._init_weights)
# Initialize learnable tokens with small random values instead of zeros
def _init_weights(self, module):
if isinstance(module, nn.Linear):
torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
if module.bias is not None:
torch.nn.init.zeros_(module.bias)
elif isinstance(module, nn.Embedding):
torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
def forward(self, images, goals_txt, goal_imgs, targets=None, pose=None, mask_=True):
n, c, h, w = images.shape
obs_patches = get_patches_fast(images, self._cfg)
patches_g = get_patches_fast(goal_imgs, self._cfg)
if self._cfg.dataset.encode_with_t5:
goals_e = goals_txt
B, T, E = goals_txt.shape
else:
goals_e = self.token_embedding_table(goals_txt)
B, E = goals_txt.shape
T = self._cfg.max_block_size
# TODO:
## Provide the logic to produce the output and loss for the GRP
# Map the vector corresponding to each patch to the hidden size dimension
obs_tokens = self.patch_embedding(obs_patches) # (n, n_patches, n_embd)
goal_img_tokens = self.patch_embedding(patches_g) # (n, n_patches, n_embd)
# Adding classification and goal_img tokens to the tokens
cls_token = self.class_token.expand(B, -1, -1) # (batch, 1, n_embd)
goal_token = self.goal_token.expand(B, -1, -1) # (batch, 1, n_embd)
x = torch.cat([cls_token, obs_tokens, goal_token, goal_img_tokens, goals_e], dim=1) # (batch, total_tokens, n_embd)
# Adding positional embedding
pos_emb = calc_positional_embeddings(x.shape[1], self._cfg.n_embd).to(x.device)
x = x + pos_emb.unsqueeze(0)[:, :x.shape[1], :]
x = self.dropout(x)
# Compute blocked masks
att_mask = torch.ones((B, x.shape[1]), device=x.device)
if mask_:
total_patches = (self._cfg.image_shape[0] // self._cfg.patch_size) * (self._cfg.image_shape[1] // self._cfg.patch_size)
obs_start = 1
obs_end = obs_start + total_patches * self._cfg.policy.obs_stacking
goal_img_start = obs_end + 1
goal_img_end = goal_img_start + total_patches
goal_text_start = goal_img_end
goal_text_end = goal_text_start + T
assert x.shape[1] == goal_text_end
# Create attention mask
# Randomly mask Text or Image goal
rand_val = torch.rand(n, device=x.device)
mask_text = (rand_val < 0.33).unsqueeze(1) # (B, 1)
mask_image = (rand_val > 0.66).unsqueeze(1) # (B, 1)
# Apply masking (0 = ignore, 1 = attend)
att_mask[:, goal_text_start:goal_text_end].masked_fill_(mask_text, 0)
att_mask[:, goal_img_start:goal_img_end].masked_fill_(mask_image, 0)
block_mask = att_mask.unsqueeze(1) # (B, 1, T)
# New masking logic
# att_mask = torch.ones((B, x.shape[1], self._cfg.n_embd), device=x.device)
# if mask_:
# total_patches = (self._cfg.image_shape[0] // self._cfg.patch_size) * (self._cfg.image_shape[1] // self._cfg.patch_size)
# obs_start = 1
# obs_end = obs_start + total_patches * self._cfg.policy.obs_stacking
# goal_img_start = obs_end + 1
# goal_img_end = goal_img_start + total_patches
# goal_text_start = goal_img_end
# goal_text_end = goal_text_start + T
# assert x.shape[1] == goal_text_end
# # Create attention mask
# # Randomly mask Text or Image goal
# rand_val = torch.rand(n, device=x.device)
# mask_text = (rand_val < 0.33).unsqueeze(1).unsqueeze(2) # (B, 1, 1)
# mask_image = (rand_val > 0.66).unsqueeze(1).unsqueeze(2) # (B, 1, 1)
# # Apply masking (0 = ignore, 1 = attend)
# att_mask[:, goal_text_start:goal_text_end, :].masked_fill_(mask_text, 0)
# att_mask[:, goal_img_start:goal_img_end, :].masked_fill_(mask_image, 0)
# # Apply mask to the embedding dimension as well
# x = x * att_mask
# Pass the mask to transformer blocks
for block in self.blocks:
x = block(x, mask=block_mask)
x = self.ln_f(x)
if targets is not None:
if self._cfg.action_space == "continuous":
out = self.action_head(x[:, 0, :]) # (batch, action_dim * action_stacking)
loss = F.mse_loss(out, targets)
elif self._cfg.action_space == "discrete":
logits = self.action_head(x[:, 0, :]).view(B, -1, 14) # (B, action_dim * action_stacking, 14)
targets_clamped = torch.clamp(targets, -1, 1) # Ensure targets are within valid range
targets_bins = ((targets_clamped + 1) / 2 * 13).long() # Map targets from [-1, 1] to [0, 13]
loss = F.cross_entropy(logits.permute(0, 2, 1), targets_bins)
out = logits.argmax(dim=-1)
else:
if self._cfg.action_space == "continuous":
out = self.action_head(x[:, 0, :]) # (batch, action_dim * action_stacking)
loss = torch.tensor(0.0, device=out.device)
elif self._cfg.action_space == "discrete":
logits = self.action_head(x[:, 0, :]).view(B, -1, 14) # (B, action_dim * action_stacking, 14)
bin_idxs = logits.argmax(dim=-1).float()
# Convert bin indices back to continuous values in [-1, 1]
out = (bin_idxs / 13.0) * 2.0 - 1.0
loss = torch.tensor(0.0, device=out.device)
# print("No targets provided, loss set to 0.0")
# Compute output and loss
return (out, loss)
def resize_image(self, image):
"""
Docstring for resize_image
:param self: Description
:param image: Description
self._resize_state = lambda sf: cv2.resize(np.array(sf, dtype=np.float32), (cfg.image_shape[0], cfg.image_shape[1])) # resize state
"""
import cv2
import numpy as _np
img = _np.array(image, dtype=_np.float32)
img = cv2.resize(img, (self._cfg.image_shape[0], self._cfg.image_shape[1]))
return img
def normalize_state(self, image):
"""
Docstring for preprocess_state
:param self: Description
:param image: Description
self._encode_state = lambda af: ((af/(255.0)*2.0)-1.0) # encoder: take a float, output an integer
self._resize_state = lambda sf: cv2.resize(np.array(sf, dtype=np.float32), (cfg.image_shape[0], cfg.image_shape[1])) # resize state
"""
# img = _np.array(image, dtype=_np.float32)
# img = cv2.resize(img, (self._cfg.image_shape[0], self._cfg.image_shape[1]))
enc = ((image / 255.0) * 2.0) - 1.0
# t = _torch.tensor(enc, dtype=_torch.float32, device=self._cfg.device)
return enc
def preprocess_state(self, image):
img = self.resize_image(image)
img = self.normalize_state(img)
return img
def preprocess_goal_image(self, image):
return self.preprocess_state(image)
def encode_text_goal(self, goal, tokenizer=None, text_model=None):
import numpy as _np
import torch as _torch
if self._cfg.dataset.encode_with_t5:
if tokenizer is None or text_model is None:
raise ValueError("tokenizer and text_model must be provided when using T5 encoding")
# TODO:
## Provide the logic converting text goal to T5 embedding tensor
device = text_model.device
tokens = tokenizer(goal, return_tensors="pt").input_ids.to(device)
with _torch.no_grad():
embedding = text_model.encoder(tokens).last_hidden_state
return embedding.to(self._cfg.device)
else:
pad = " " * self._cfg.max_block_size
goal_ = goal[:self._cfg.max_block_size] + pad[len(goal):self._cfg.max_block_size]
try:
stoi = {c: i for i, c in enumerate(self._cfg.dataset.chars_list)}
ids = [stoi.get(c, 0) for c in goal_]
except Exception:
ids = [0] * self._cfg.max_block_size
return _torch.tensor(_np.expand_dims(_np.array(ids, dtype=_np.int64), axis=0), dtype=_torch.long, device=self._cfg.device)
def process_text_embedding_for_buffer(self, goal, tokenizer=None, text_model=None):
"""
Process text goal embedding for storing in the circular buffer.
Returns a numpy array of shape (max_block_size, n_embd) without batch dimension.
"""
import numpy as _np
if tokenizer is None or text_model is None:
raise ValueError("tokenizer and text_model must be provided when using T5 encoding")
goal_ = _np.zeros((self._cfg.max_block_size, self._cfg.n_embd), dtype=_np.float32)
input_ids = tokenizer(goal, return_tensors="pt").input_ids
goal_t = text_model.encoder(input_ids).last_hidden_state.detach().cpu().numpy()
goal_[:len(goal_t[0]), :] = goal_t[0][:self._cfg.max_block_size]
return goal_
def decode_action(self, action_tensor):
"""
Docstring for decode_action
:param self: Description
:param action_tensor: Description
self._decode_action = lambda binN: (binN * action_std) + action_mean # Undo mapping to [-1, 1]
"""
import torch as _torch
## The action tensor is of shape (batch_size, action_dim * action_stacking) so we need to repeat the mean and std per action stacking
action_mean = _torch.tensor(np.repeat(self._cfg.dataset.action_mean, self._cfg.policy.action_stacking), dtype=action_tensor.dtype, device=action_tensor.device)
action_std = _torch.tensor(np.repeat(self._cfg.dataset.action_std, self._cfg.policy.action_stacking), dtype=action_tensor.dtype, device=action_tensor.device)
return (action_tensor * action_std) + action_mean
def encode_action(self, action_float):
"""
Docstring for encode_action
:param self: Description
:param action_float: Description
self._encode_action = lambda af: (af - action_mean)/(action_std) # encoder: take a float, output an integer
"""
import torch as _torch
action_mean = _torch.tensor(self._cfg.dataset.action_mean, dtype=action_float.dtype, device=action_float.device)
action_std = _torch.tensor(self._cfg.dataset.action_std, dtype=action_float.dtype, device=action_float.device)
return (action_float - action_mean) / action_std
@torch.no_grad()
def estimate_loss(model, dataset):
out = {}
model.eval()
for split in ['train', 'val']:
losses = torch.zeros(model._cfg.eval_iters)
for k in range(model._cfg.eval_iters):
X, x_pose, x_goal, x_goal_img, Y = dataset.get_batch_grp(split, model._cfg, model._cfg.batch_size)
logits, loss = model(X, x_goal, x_goal_img, Y, pose=x_pose)
losses[k] = loss.item()
out[split] = losses.mean()
model.train()
return out
|