Spaces:
Running
Running
Upload 4 files
Browse files- utils/complete_model.py +276 -0
- utils/layer_mask.py +224 -0
- utils/modifiedGPT2.py +712 -0
- utils/processing.py +27 -0
utils/complete_model.py
ADDED
|
@@ -0,0 +1,276 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import torch
|
| 3 |
+
import torch.nn as nn
|
| 4 |
+
from transformers import AutoModel, GPT2Tokenizer
|
| 5 |
+
|
| 6 |
+
from utils.modifiedGPT2 import create_decoder
|
| 7 |
+
|
| 8 |
+
from utils.layer_mask import gaussian_layer_stack_pipeline
|
| 9 |
+
|
| 10 |
+
|
| 11 |
+
class DINOEncoder(nn.Module):
|
| 12 |
+
def __init__(self, model_id="facebook/dinov3-vits16-pretrain-lvd1689m", freeze=True):
|
| 13 |
+
super().__init__()
|
| 14 |
+
self.model = AutoModel.from_pretrained(model_id)
|
| 15 |
+
if freeze:
|
| 16 |
+
for p in self.model.parameters():
|
| 17 |
+
p.requires_grad = False
|
| 18 |
+
@torch.no_grad()
|
| 19 |
+
def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
|
| 20 |
+
"""
|
| 21 |
+
pixel_values: [B, C, H, W]
|
| 22 |
+
returns patches: [B, Np, Cenc]
|
| 23 |
+
"""
|
| 24 |
+
out = self.model(pixel_values=pixel_values)
|
| 25 |
+
tokens = out.last_hidden_state # [B, 1+Np, Cenc] (CLS + patches) for ViT-like
|
| 26 |
+
# Skip a few special tokens if your backbone adds them; adjust as needed.
|
| 27 |
+
patches = tokens[:, 5:, :] # [B, Np, Cenc]
|
| 28 |
+
return patches
|
| 29 |
+
|
| 30 |
+
class DinoUNet(nn.Module):
|
| 31 |
+
def __init__(self, model_name="facebook/dinov3-convnext-small-pretrain-lvd1689m", freeze=True):
|
| 32 |
+
super().__init__()
|
| 33 |
+
self.encoder = AutoModel.from_pretrained(model_name)
|
| 34 |
+
# NOTE: confirm channels of the chosen hidden state; 768 is common for small convnext/dinov3
|
| 35 |
+
self.channel_adapter = nn.Conv2d(768, 512, kernel_size=1)
|
| 36 |
+
self.decoder = nn.Sequential(
|
| 37 |
+
nn.Conv2d(512, 256, 3, padding=1), nn.ReLU(inplace=True),
|
| 38 |
+
nn.ConvTranspose2d(256, 128, 2, stride=2), nn.ReLU(inplace=True),
|
| 39 |
+
nn.ConvTranspose2d(128, 64, 2, stride=2), nn.ReLU(inplace=True),
|
| 40 |
+
nn.Conv2d(64, 1, 1)
|
| 41 |
+
)
|
| 42 |
+
if freeze:
|
| 43 |
+
for m in (self.encoder, self.channel_adapter, self.decoder):
|
| 44 |
+
for p in m.parameters():
|
| 45 |
+
p.requires_grad = False
|
| 46 |
+
|
| 47 |
+
@torch.no_grad()
|
| 48 |
+
def forward(self, x: torch.Tensor, num_layers: int) -> torch.Tensor:
|
| 49 |
+
"""
|
| 50 |
+
x: [B, C, H, W]; returns mask: [B, 1, H', W'] (your upsampling stack defines H',W')
|
| 51 |
+
"""
|
| 52 |
+
enc_feats = self.encoder(x, output_hidden_states=True, return_dict=True)
|
| 53 |
+
# take the last 4D feature map from hidden_states
|
| 54 |
+
feats = next(h for h in reversed(enc_feats.hidden_states) if isinstance(h, torch.Tensor) and h.ndim == 4)
|
| 55 |
+
feats = self.channel_adapter(feats)
|
| 56 |
+
pred = self.decoder(feats) # (B,1,h,w)
|
| 57 |
+
_, _, segmentation_mask = gaussian_layer_stack_pipeline(pred, n_layers = num_layers)
|
| 58 |
+
return segmentation_mask # [B, num_layers, h, w]
|
| 59 |
+
|
| 60 |
+
|
| 61 |
+
class LinearProjection(nn.Module):
|
| 62 |
+
def __init__(self, input_dim=384, output_dim=768, freeze=False):
|
| 63 |
+
super().__init__()
|
| 64 |
+
self.proj = nn.Linear(input_dim, output_dim)
|
| 65 |
+
if freeze:
|
| 66 |
+
for p in self.proj.parameters():
|
| 67 |
+
p.requires_grad = False
|
| 68 |
+
|
| 69 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
| 70 |
+
# x: [B, Np, input_dim] -> [B, Np, output_dim]
|
| 71 |
+
return self.proj(x)
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
class CustomModel(nn.Module):
|
| 75 |
+
def __init__(
|
| 76 |
+
self,
|
| 77 |
+
device: str = "cuda",
|
| 78 |
+
ENCODER_MODEL_PATH: str | None = "dino_encoder.pth",
|
| 79 |
+
SEGMENTER_MODEL_PATH: str | None = "dino_segmenter.pth",
|
| 80 |
+
DECODER_MODEL_PATH: str | None = "dino_decoder.pth",
|
| 81 |
+
LINEAR_PROJECTION_PATH: str | None = "linear_projection.pth",
|
| 82 |
+
freeze_encoder: bool = True,
|
| 83 |
+
freeze_segmenter: bool = True,
|
| 84 |
+
freeze_linear_projection: bool = False,
|
| 85 |
+
freeze_decoder: bool = False,
|
| 86 |
+
attention_implementation: str = "sdpa",
|
| 87 |
+
):
|
| 88 |
+
super().__init__()
|
| 89 |
+
self.device = torch.device(device)
|
| 90 |
+
|
| 91 |
+
# Encoder
|
| 92 |
+
self.encoder = DINOEncoder()
|
| 93 |
+
if ENCODER_MODEL_PATH and os.path.exists(ENCODER_MODEL_PATH):
|
| 94 |
+
self.encoder.load_state_dict(torch.load(ENCODER_MODEL_PATH, map_location="cpu"), strict=False)
|
| 95 |
+
print("Loaded encoder weights from", ENCODER_MODEL_PATH)
|
| 96 |
+
if freeze_encoder:
|
| 97 |
+
self.encoder.eval()
|
| 98 |
+
|
| 99 |
+
# Segmenter
|
| 100 |
+
self.segmenter = DinoUNet()
|
| 101 |
+
if SEGMENTER_MODEL_PATH and os.path.exists(SEGMENTER_MODEL_PATH):
|
| 102 |
+
self.segmenter.load_state_dict(torch.load(SEGMENTER_MODEL_PATH, map_location="cpu"), strict=False)
|
| 103 |
+
print("Loaded segmenter weights from", SEGMENTER_MODEL_PATH)
|
| 104 |
+
if freeze_segmenter:
|
| 105 |
+
self.segmenter.eval()
|
| 106 |
+
|
| 107 |
+
# Decoder (modified GPT-2)
|
| 108 |
+
self.decoder = create_decoder(attention=attention_implementation) # must expose .config.hidden_size & .config.num_hidden_layers
|
| 109 |
+
if DECODER_MODEL_PATH and os.path.exists(DECODER_MODEL_PATH):
|
| 110 |
+
self.decoder.load_state_dict(torch.load(DECODER_MODEL_PATH, map_location="cpu"), strict=False)
|
| 111 |
+
print("Loaded decoder weights from", DECODER_MODEL_PATH)
|
| 112 |
+
if freeze_decoder:
|
| 113 |
+
self.decoder.eval()
|
| 114 |
+
|
| 115 |
+
# Linear projection: DINO hidden -> GPT2 hidden
|
| 116 |
+
enc_h = self.encoder.model.config.hidden_size
|
| 117 |
+
dec_h = self.decoder.config.hidden_size
|
| 118 |
+
self.linear_projection = LinearProjection(input_dim=enc_h, output_dim=dec_h)
|
| 119 |
+
if LINEAR_PROJECTION_PATH and os.path.exists(LINEAR_PROJECTION_PATH):
|
| 120 |
+
self.linear_projection.load_state_dict(torch.load(LINEAR_PROJECTION_PATH, map_location="cpu"), strict=False)
|
| 121 |
+
print("Loaded linear projection weights from", LINEAR_PROJECTION_PATH)
|
| 122 |
+
if freeze_linear_projection:
|
| 123 |
+
self.linear_projection.eval()
|
| 124 |
+
|
| 125 |
+
# Tokenizer (pad token for GPT-2)
|
| 126 |
+
self.tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
|
| 127 |
+
if self.tokenizer.pad_token_id is None:
|
| 128 |
+
self.tokenizer.pad_token = self.tokenizer.eos_token
|
| 129 |
+
self.pad_token_id = self.tokenizer.pad_token_id # ✅ use ID, not string
|
| 130 |
+
|
| 131 |
+
self.num_layers = self.decoder.config.num_hidden_layers
|
| 132 |
+
|
| 133 |
+
# move everything once
|
| 134 |
+
self.to(self.device)
|
| 135 |
+
|
| 136 |
+
def forward(self, pixel_values: torch.Tensor, tgt_ids: torch.Tensor | None = None, **kwargs) -> dict:
|
| 137 |
+
"""
|
| 138 |
+
pixel_values: [B,C,H,W], float
|
| 139 |
+
tgt_ids: [B,T], long (token IDs), padded with pad_token_id if any padding is present
|
| 140 |
+
"""
|
| 141 |
+
pixel_values = pixel_values.to(self.device, non_blocking=True)
|
| 142 |
+
|
| 143 |
+
# Visual path
|
| 144 |
+
patches = self.encoder(pixel_values) # [B,Np,Cenc]
|
| 145 |
+
projected_patches = self.linear_projection(patches) # [B,Np,n_embd]
|
| 146 |
+
|
| 147 |
+
# Segmentation path per layer
|
| 148 |
+
segmented_layers = self.segmenter(pixel_values, self.num_layers) # [B,n_layers,H,W] (per current decoder)
|
| 149 |
+
|
| 150 |
+
# Text path (optional teacher-forced training)
|
| 151 |
+
labels = None
|
| 152 |
+
if tgt_ids is not None:
|
| 153 |
+
if tgt_ids.dtype != torch.long:
|
| 154 |
+
tgt_ids = tgt_ids.long()
|
| 155 |
+
tgt_ids = tgt_ids.to(self.device, non_blocking=True) # [B,T]
|
| 156 |
+
text_embeds = self.decoder.transformer.wte(tgt_ids) # [B,T,n_embd]
|
| 157 |
+
inputs_embeds = torch.cat([projected_patches, text_embeds], dim=1) # [B,Np+T,n_embd]
|
| 158 |
+
|
| 159 |
+
# Labels: ignore prefix tokens (vision) and PADs in text
|
| 160 |
+
B, Np, _ = projected_patches.shape
|
| 161 |
+
labels_prefix = torch.full((B, Np), -100, device=self.device, dtype=torch.long)
|
| 162 |
+
text_labels = tgt_ids.clone()
|
| 163 |
+
text_labels[text_labels == self.pad_token_id] = -100 # ✅ compare to ID
|
| 164 |
+
labels = torch.cat([labels_prefix, text_labels], dim=1) # [B,Np+T]
|
| 165 |
+
else:
|
| 166 |
+
inputs_embeds = projected_patches
|
| 167 |
+
|
| 168 |
+
# Decoder forward
|
| 169 |
+
out = self.decoder(inputs_embeds=inputs_embeds, segmentation_mask=segmented_layers, labels=labels, **kwargs)
|
| 170 |
+
return out
|
| 171 |
+
|
| 172 |
+
@torch.inference_mode()
|
| 173 |
+
def generate(
|
| 174 |
+
self,
|
| 175 |
+
pixel_values: torch.Tensor,
|
| 176 |
+
max_new_tokens: int = 100,
|
| 177 |
+
output_attentions: bool = False,
|
| 178 |
+
) -> torch.Tensor:
|
| 179 |
+
"""
|
| 180 |
+
pixel_values: [B,C,H,W], float
|
| 181 |
+
returns generated_ids: [B, T]
|
| 182 |
+
"""
|
| 183 |
+
pixel_values = pixel_values.to(self.device, non_blocking=True)
|
| 184 |
+
|
| 185 |
+
# Visual path
|
| 186 |
+
patches = self.encoder(pixel_values) # [B,Np,Cenc]
|
| 187 |
+
projected_patches = self.linear_projection(patches) # [B,Np,n_embd]
|
| 188 |
+
|
| 189 |
+
# Segmentation path per layer
|
| 190 |
+
segmented_layers = self.segmenter(pixel_values, self.num_layers) # [B,n_layers,H,W] (per current decoder)
|
| 191 |
+
|
| 192 |
+
# Generate
|
| 193 |
+
output = self.decoder.generate(
|
| 194 |
+
inputs_embeds=projected_patches,
|
| 195 |
+
max_new_tokens=max_new_tokens,
|
| 196 |
+
do_sample=False,
|
| 197 |
+
repetition_penalty=1.2,
|
| 198 |
+
eos_token_id=self.tokenizer.eos_token_id,
|
| 199 |
+
pad_token_id=self.pad_token_id,
|
| 200 |
+
use_cache=True,
|
| 201 |
+
segmentation_mask=segmented_layers,
|
| 202 |
+
prefix_allowed_length=0,
|
| 203 |
+
plot_attention_mask=False,
|
| 204 |
+
plot_attention_mask_layer=[],
|
| 205 |
+
plot_attention_map=False,
|
| 206 |
+
plot_attention_map_layer=[],
|
| 207 |
+
plot_attention_map_generation=0,
|
| 208 |
+
output_attentions=output_attentions,
|
| 209 |
+
return_dict_in_generate=True,
|
| 210 |
+
)
|
| 211 |
+
# Remove prefix tokens (vision)
|
| 212 |
+
generated_ids = output.sequences#[:, projected_patches.shape[1]:] # [B,T]
|
| 213 |
+
generated_text = self.tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
|
| 214 |
+
return generated_ids, generated_text, output.attentions if output_attentions else None
|
| 215 |
+
|
| 216 |
+
def create_complete_model(device: str = "cuda", **kwargs) -> CustomModel:
|
| 217 |
+
model = CustomModel(device=device, **kwargs)
|
| 218 |
+
return model
|
| 219 |
+
|
| 220 |
+
def save_complete_model(model: CustomModel, save_path: str, device: str = "cuda") -> None:
|
| 221 |
+
# Ensure folder exists
|
| 222 |
+
os.makedirs(os.path.dirname(save_path) or ".", exist_ok=True)
|
| 223 |
+
|
| 224 |
+
# Save on CPU to keep checkpoint portable
|
| 225 |
+
orig_device = next(model.parameters()).device
|
| 226 |
+
model.to("cpu")
|
| 227 |
+
torch.save(model.state_dict(), save_path)
|
| 228 |
+
print(f"Saved complete model weights to {save_path}")
|
| 229 |
+
|
| 230 |
+
# Restore model device
|
| 231 |
+
model.to(device if isinstance(device, str) else orig_device)
|
| 232 |
+
|
| 233 |
+
def save_checkpoint(model: CustomModel, optimizer: torch.optim.Optimizer, save_path: str) -> None:
|
| 234 |
+
# Ensure folder exists
|
| 235 |
+
os.makedirs(os.path.dirname(save_path) or ".", exist_ok=True)
|
| 236 |
+
|
| 237 |
+
checkpoint = {
|
| 238 |
+
"model_state_dict": model.state_dict(),
|
| 239 |
+
"optimizer_state_dict": optimizer.state_dict(),
|
| 240 |
+
}
|
| 241 |
+
torch.save(checkpoint, save_path)
|
| 242 |
+
print(f"Saved checkpoint to {save_path}")
|
| 243 |
+
|
| 244 |
+
def load_complete_model(model: CustomModel, load_path: str, device: str = "cpu", strict: bool = True) -> CustomModel:
|
| 245 |
+
if not os.path.exists(load_path):
|
| 246 |
+
print(f"No weights found at {load_path}")
|
| 247 |
+
model.to(device)
|
| 248 |
+
return model
|
| 249 |
+
|
| 250 |
+
# Load to CPU first, then move to target device
|
| 251 |
+
state = torch.load(load_path, map_location="cpu")
|
| 252 |
+
missing, unexpected = model.load_state_dict(state, strict=strict)
|
| 253 |
+
if not strict:
|
| 254 |
+
if missing:
|
| 255 |
+
print(f"[load warning] Missing keys: {missing}")
|
| 256 |
+
if unexpected:
|
| 257 |
+
print(f"[load warning] Unexpected keys: {unexpected}")
|
| 258 |
+
|
| 259 |
+
model.to(device)
|
| 260 |
+
print(f"Loaded complete model weights from {load_path}")
|
| 261 |
+
return model
|
| 262 |
+
|
| 263 |
+
def load_checkpoint(model: CustomModel, optimizer: torch.optim.Optimizer, load_path: str, device: str = "cpu") -> tuple[CustomModel, torch.optim.Optimizer]:
|
| 264 |
+
if not os.path.exists(load_path):
|
| 265 |
+
print(f"No checkpoint found at {load_path}")
|
| 266 |
+
model.to(device)
|
| 267 |
+
return model, optimizer
|
| 268 |
+
|
| 269 |
+
# Load to CPU first, then move to target device
|
| 270 |
+
checkpoint = torch.load(load_path, map_location="cpu")
|
| 271 |
+
model.load_state_dict(checkpoint["model_state_dict"])
|
| 272 |
+
optimizer.load_state_dict(checkpoint["optimizer_state_dict"])
|
| 273 |
+
|
| 274 |
+
model.to(device)
|
| 275 |
+
print(f"Loaded checkpoint from {load_path}")
|
| 276 |
+
return model, optimizer
|
utils/layer_mask.py
ADDED
|
@@ -0,0 +1,224 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import torch.nn.functional as F
|
| 3 |
+
import math
|
| 4 |
+
import numpy as np
|
| 5 |
+
import matplotlib.pyplot as plt
|
| 6 |
+
|
| 7 |
+
@torch.no_grad()
|
| 8 |
+
def gaussian_layer_stack_pipeline(
|
| 9 |
+
x: torch.Tensor,
|
| 10 |
+
n_layers: int,
|
| 11 |
+
base_ksize: int = 3,
|
| 12 |
+
ksize_growth: int = 2,
|
| 13 |
+
sigma: float | None = None,
|
| 14 |
+
eps: float = 1e-8,
|
| 15 |
+
):
|
| 16 |
+
"""
|
| 17 |
+
All-in-one GPU batch pipeline:
|
| 18 |
+
1) Per-sample min-max normalize to [0,1]
|
| 19 |
+
2) Resize to (32,32)
|
| 20 |
+
3) Apply L Gaussian blurs with increasing kernel size in a single
|
| 21 |
+
horizontal conv + single vertical conv using depthwise groups
|
| 22 |
+
(via a shared max kernel padded with zeros)
|
| 23 |
+
4) Renormalize each layer to [0,1]
|
| 24 |
+
5) Return stacked (B,L,32,32), flat (B,L,1024), tiled (B,L,1024,1024 view)
|
| 25 |
+
|
| 26 |
+
Args:
|
| 27 |
+
x: (B,H,W) or (B,1,H,W) tensor (any device/dtype)
|
| 28 |
+
n_layers: number of layers
|
| 29 |
+
base_ksize: starting odd kernel size (e.g., 3)
|
| 30 |
+
ksize_growth: increment per layer (e.g., 2) -> ensures odd sizes
|
| 31 |
+
sigma: if None, uses (ksize-1)/6 per layer; else fixed sigma for all
|
| 32 |
+
eps: small number for safe division
|
| 33 |
+
|
| 34 |
+
Returns:
|
| 35 |
+
stacked: (B, n_layers, 32, 32) float on x.device
|
| 36 |
+
flat: (B, n_layers, 1024)
|
| 37 |
+
tiled: (B, n_layers, 1024, 1024) (expand view; memory-cheap)
|
| 38 |
+
"""
|
| 39 |
+
assert n_layers >= 1, "n_layers must be >= 1"
|
| 40 |
+
|
| 41 |
+
# ---- Ensure 4D, 1 channel; cast to float (stay on same device) ----
|
| 42 |
+
if x.ndim == 3:
|
| 43 |
+
x = x.unsqueeze(1) # (B,1,H,W)
|
| 44 |
+
elif x.ndim != 4 or x.shape[1] not in (1,):
|
| 45 |
+
raise ValueError(f"Expected (B,H,W) or (B,1,H,W); got {tuple(x.shape)}")
|
| 46 |
+
x = x.float()
|
| 47 |
+
|
| 48 |
+
B, _, H, W = x.shape
|
| 49 |
+
|
| 50 |
+
# ---- Per-sample min-max normalize to [0,1] ----
|
| 51 |
+
xmin = x.amin(dim=(2, 3), keepdim=True)
|
| 52 |
+
xmax = x.amax(dim=(2, 3), keepdim=True)
|
| 53 |
+
denom = (xmax - xmin).clamp_min(eps)
|
| 54 |
+
x = (x - xmin) / denom # (B,1,H,W) in [0,1]
|
| 55 |
+
|
| 56 |
+
# ---- Resize to 32x32 on GPU ----
|
| 57 |
+
x = F.interpolate(x, size=(32, 32), mode="bilinear", align_corners=False) # (B,1,32,32)
|
| 58 |
+
|
| 59 |
+
# ---- Prepare per-layer kernel sizes (odd) ----
|
| 60 |
+
ksizes = []
|
| 61 |
+
for i in range(n_layers, 0, -1): # to keep your original ordering: L...1
|
| 62 |
+
k = base_ksize + i * ksize_growth
|
| 63 |
+
k = int(k)
|
| 64 |
+
if k % 2 == 0:
|
| 65 |
+
k += 1
|
| 66 |
+
k = max(k, 1)
|
| 67 |
+
ksizes.append(k)
|
| 68 |
+
|
| 69 |
+
Kmax = max(ksizes)
|
| 70 |
+
pad = Kmax // 2
|
| 71 |
+
|
| 72 |
+
# ---- Build per-layer 1D Gaussian vectors and embed into shared Kmax kernel ----
|
| 73 |
+
# We create horizontal weights of shape (L,1,1,Kmax) and vertical (L,1,Kmax,1)
|
| 74 |
+
device, dtype = x.device, x.dtype
|
| 75 |
+
weight_h = torch.zeros((n_layers, 1, 1, Kmax), device=device, dtype=dtype)
|
| 76 |
+
weight_v = torch.zeros((n_layers, 1, Kmax, 1), device=device, dtype=dtype)
|
| 77 |
+
|
| 78 |
+
for idx, k in enumerate(ksizes):
|
| 79 |
+
# choose sigma
|
| 80 |
+
sig = sigma if (sigma is not None and sigma > 0) else (k - 1) / 6.0
|
| 81 |
+
r = k // 2
|
| 82 |
+
xp = torch.arange(-r, r + 1, device=device, dtype=dtype)
|
| 83 |
+
g = torch.exp(-(xp * xp) / (2.0 * sig * sig))
|
| 84 |
+
g = g / g.sum() # (k,)
|
| 85 |
+
|
| 86 |
+
# center g into Kmax with zeros around
|
| 87 |
+
start = (Kmax - k) // 2
|
| 88 |
+
end = start + k
|
| 89 |
+
|
| 90 |
+
# horizontal row
|
| 91 |
+
weight_h[idx, 0, 0, start:end] = g # (1 x Kmax)
|
| 92 |
+
|
| 93 |
+
# vertical column
|
| 94 |
+
weight_v[idx, 0, start:end, 0] = g # (Kmax x 1)
|
| 95 |
+
|
| 96 |
+
# ---- Duplicate input across L channels (depthwise groups) ----
|
| 97 |
+
xL = x.expand(B, n_layers, 32, 32).contiguous() # (B,L,32,32)
|
| 98 |
+
|
| 99 |
+
# ---- Separable Gaussian blur with a single pass per axis (groups=L) ----
|
| 100 |
+
# Horizontal
|
| 101 |
+
xh = F.pad(xL, (pad, pad, 0, 0), mode="reflect")
|
| 102 |
+
xh = F.conv2d(xh, weight=weight_h, bias=None, stride=1, padding=0, groups=n_layers) # (B,L,32,32)
|
| 103 |
+
|
| 104 |
+
# Vertical
|
| 105 |
+
xv = F.pad(xh, (0, 0, pad, pad), mode="reflect")
|
| 106 |
+
yL = F.conv2d(xv, weight=weight_v, bias=None, stride=1, padding=0, groups=n_layers) # (B,L,32,32)
|
| 107 |
+
|
| 108 |
+
# ---- Renormalize each layer to [0,1] (per-sample, per-layer) ----
|
| 109 |
+
y_min = yL.amin(dim=(2, 3), keepdim=True)
|
| 110 |
+
y_max = yL.amax(dim=(2, 3), keepdim=True)
|
| 111 |
+
y_den = (y_max - y_min).clamp_min(eps)
|
| 112 |
+
stacked = (yL - y_min) / y_den # (B,L,32,32) in [0,1]
|
| 113 |
+
|
| 114 |
+
# ---- Flatten + tile (expand view; caution w/ later materialization) ----
|
| 115 |
+
flat = stacked.reshape(B, n_layers, 32 * 32) # (B,L,1024)
|
| 116 |
+
tiled = flat.unsqueeze(-2).expand(-1, -1, 32 * 32, -1) # (B,L,1024,1024) view
|
| 117 |
+
|
| 118 |
+
return stacked, flat, tiled
|
| 119 |
+
|
| 120 |
+
def plot_layers_any(
|
| 121 |
+
x,
|
| 122 |
+
*,
|
| 123 |
+
max_batches=None,
|
| 124 |
+
vlim=(0, 1),
|
| 125 |
+
one_indexed: bool = False,
|
| 126 |
+
max_cols: int = 6,
|
| 127 |
+
):
|
| 128 |
+
"""
|
| 129 |
+
Plot layers for each batch sample in separate figures.
|
| 130 |
+
|
| 131 |
+
Accepts:
|
| 132 |
+
- stacked: (B, L, H, W)
|
| 133 |
+
- flat: (B, L, HW)
|
| 134 |
+
- tiled: (B, L, HW, HW)
|
| 135 |
+
|
| 136 |
+
Behavior:
|
| 137 |
+
- Creates one figure PER BATCH (up to `max_batches`).
|
| 138 |
+
- At most `max_cols` layers per row (default 6).
|
| 139 |
+
- Column headers: 'Layer {i}' descending from n-1 -> 0 (or n -> 1 if one_indexed=True).
|
| 140 |
+
- Figure title per batch: 'Masks for input {i} out of {B}'.
|
| 141 |
+
|
| 142 |
+
Returns:
|
| 143 |
+
A list of (fig, axes) tuples, one per plotted batch.
|
| 144 |
+
"""
|
| 145 |
+
# ---- Normalize input to torch ----
|
| 146 |
+
if isinstance(x, np.ndarray):
|
| 147 |
+
x = torch.from_numpy(x)
|
| 148 |
+
if not isinstance(x, torch.Tensor):
|
| 149 |
+
raise TypeError(f"Expected torch.Tensor or np.ndarray, got {type(x)}")
|
| 150 |
+
|
| 151 |
+
if x.ndim not in (3, 4):
|
| 152 |
+
raise ValueError(f"Expected ndim 3 or 4, got shape {tuple(x.shape)}")
|
| 153 |
+
|
| 154 |
+
# ---- Convert to (B, L, H, W) 'stacked' ----
|
| 155 |
+
if x.ndim == 4:
|
| 156 |
+
B, L, A, B_ = x.shape
|
| 157 |
+
if A == B_:
|
| 158 |
+
# Could be stacked (H==W) or tiled (HW x HW). Heuristic: if A is a perfect square
|
| 159 |
+
# and reasonably large (e.g., 1024), treat as tiled and collapse to flat.
|
| 160 |
+
s = int(math.isqrt(A))
|
| 161 |
+
if s * s == A and A >= 64:
|
| 162 |
+
flat = x[..., 0, :].detach() # (B, L, HW)
|
| 163 |
+
H = W = s
|
| 164 |
+
stacked = flat.reshape(B, L, H, W)
|
| 165 |
+
else:
|
| 166 |
+
stacked = x.detach()
|
| 167 |
+
else:
|
| 168 |
+
stacked = x.detach()
|
| 169 |
+
else:
|
| 170 |
+
# x.ndim == 3 -> (B, L, HW)
|
| 171 |
+
B, L, HW = x.shape
|
| 172 |
+
s = int(math.isqrt(HW))
|
| 173 |
+
if s * s != HW:
|
| 174 |
+
if HW != 32 * 32:
|
| 175 |
+
raise ValueError(
|
| 176 |
+
f"Cannot infer square image size from HW={HW}. "
|
| 177 |
+
f"Provide stacked (B,L,H,W) or flat with square HW."
|
| 178 |
+
)
|
| 179 |
+
s = 32
|
| 180 |
+
H = W = s
|
| 181 |
+
stacked = x.detach().reshape(B, L, H, W)
|
| 182 |
+
|
| 183 |
+
# Ensure float & CPU for plotting
|
| 184 |
+
stacked = stacked.to(torch.float32).cpu().numpy()
|
| 185 |
+
|
| 186 |
+
# ---- Batch selection ----
|
| 187 |
+
B, L, H, W = stacked.shape
|
| 188 |
+
plot_B = B if max_batches is None else max(1, min(B, int(max_batches)))
|
| 189 |
+
|
| 190 |
+
# ---- Layout params ----
|
| 191 |
+
cols = max(1, int(max_cols))
|
| 192 |
+
rows_needed = lambda L: (L + cols - 1) // cols
|
| 193 |
+
|
| 194 |
+
figs = []
|
| 195 |
+
for b in range(plot_B):
|
| 196 |
+
# number of rows for this batch
|
| 197 |
+
r = rows_needed(L)
|
| 198 |
+
fig, axes = plt.subplots(r, cols, figsize=(cols * 3, r * 3), squeeze=False)
|
| 199 |
+
fig.suptitle(f"Masks for input {b} out of {B}", fontsize=12, y=1.02)
|
| 200 |
+
|
| 201 |
+
for l in range(L):
|
| 202 |
+
rr = l // cols
|
| 203 |
+
cc = l % cols
|
| 204 |
+
ax = axes[rr, cc]
|
| 205 |
+
if vlim is None:
|
| 206 |
+
ax.imshow(stacked[b, l], cmap="gray")
|
| 207 |
+
else:
|
| 208 |
+
ax.imshow(stacked[b, l], cmap="gray", vmin=vlim[0], vmax=vlim[1])
|
| 209 |
+
ax.axis("off")
|
| 210 |
+
|
| 211 |
+
# Set column titles only on the first row of the grid
|
| 212 |
+
label_num = (l + 1) if one_indexed else l
|
| 213 |
+
ax.set_title(f"Layer {label_num}", fontsize=10)
|
| 214 |
+
|
| 215 |
+
# Hide any unused axes (when L is not a multiple of cols)
|
| 216 |
+
total_slots = r * cols
|
| 217 |
+
for empty_idx in range(L, total_slots):
|
| 218 |
+
rr = empty_idx // cols
|
| 219 |
+
cc = empty_idx % cols
|
| 220 |
+
axes[rr, cc].axis("off")
|
| 221 |
+
|
| 222 |
+
plt.tight_layout()
|
| 223 |
+
figs.append((fig, axes))
|
| 224 |
+
return figs
|
utils/modifiedGPT2.py
ADDED
|
@@ -0,0 +1,712 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from typing import Optional, Union
|
| 2 |
+
import torch
|
| 3 |
+
import torch.nn.functional as F
|
| 4 |
+
from transformers import GPT2LMHeadModel, GPT2Model, GPT2Config
|
| 5 |
+
from transformers.cache_utils import Cache, DynamicCache, EncoderDecoderCache
|
| 6 |
+
from transformers.masking_utils import create_causal_mask
|
| 7 |
+
from transformers.modeling_attn_mask_utils import _prepare_4d_attention_mask_for_sdpa
|
| 8 |
+
from transformers.modeling_outputs import (
|
| 9 |
+
BaseModelOutputWithPastAndCrossAttentions,
|
| 10 |
+
CausalLMOutputWithCrossAttentions,
|
| 11 |
+
)
|
| 12 |
+
from transformers.utils import (
|
| 13 |
+
logging,
|
| 14 |
+
)
|
| 15 |
+
|
| 16 |
+
from transformers.models.gpt2.modeling_gpt2 import GPT2Block, GPT2Attention, eager_attention_forward
|
| 17 |
+
from torch import nn
|
| 18 |
+
from typing import Callable
|
| 19 |
+
from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
|
| 20 |
+
|
| 21 |
+
import matplotlib.pyplot as plt
|
| 22 |
+
|
| 23 |
+
logger = logging.get_logger(__name__)
|
| 24 |
+
|
| 25 |
+
class GPT2AttentionModified(GPT2Attention):
|
| 26 |
+
def __init__(self, config, is_cross_attention=False, layer_idx=None):
|
| 27 |
+
super().__init__(config, is_cross_attention=is_cross_attention, layer_idx=layer_idx)
|
| 28 |
+
self.config = config
|
| 29 |
+
max_positions = 2048
|
| 30 |
+
self.register_buffer(
|
| 31 |
+
"bias",
|
| 32 |
+
torch.tril(torch.ones((max_positions, max_positions), dtype=torch.bool)).view(
|
| 33 |
+
1, 1, max_positions, max_positions
|
| 34 |
+
),
|
| 35 |
+
persistent=False,
|
| 36 |
+
)
|
| 37 |
+
|
| 38 |
+
def forward(
|
| 39 |
+
self,
|
| 40 |
+
hidden_states: Optional[tuple[torch.FloatTensor]],
|
| 41 |
+
past_key_values: Optional[Cache] = None,
|
| 42 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 43 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
| 44 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
| 45 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
| 46 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
| 47 |
+
output_attentions: Optional[bool] = False,
|
| 48 |
+
**kwargs,
|
| 49 |
+
) -> tuple[Union[torch.Tensor, tuple[torch.Tensor]], ...]:
|
| 50 |
+
is_cross_attention = encoder_hidden_states is not None
|
| 51 |
+
if past_key_values is not None:
|
| 52 |
+
if isinstance(past_key_values, EncoderDecoderCache):
|
| 53 |
+
is_updated = past_key_values.is_updated.get(self.layer_idx)
|
| 54 |
+
if is_cross_attention:
|
| 55 |
+
# after the first generated id, we can subsequently re-use all key/value_layer from cache
|
| 56 |
+
curr_past_key_value = past_key_values.cross_attention_cache
|
| 57 |
+
else:
|
| 58 |
+
curr_past_key_value = past_key_values.self_attention_cache
|
| 59 |
+
else:
|
| 60 |
+
curr_past_key_value = past_key_values
|
| 61 |
+
|
| 62 |
+
if is_cross_attention:
|
| 63 |
+
if not hasattr(self, "q_attn"):
|
| 64 |
+
raise ValueError(
|
| 65 |
+
"If class is used as cross attention, the weights `q_attn` have to be defined. "
|
| 66 |
+
"Please make sure to instantiate class with `GPT2Attention(..., is_cross_attention=True)`."
|
| 67 |
+
)
|
| 68 |
+
query_states = self.q_attn(hidden_states)
|
| 69 |
+
attention_mask = encoder_attention_mask
|
| 70 |
+
|
| 71 |
+
# Try to get key/value states from cache if possible
|
| 72 |
+
if past_key_values is not None and is_updated:
|
| 73 |
+
key_states = curr_past_key_value.layers[self.layer_idx].keys
|
| 74 |
+
value_states = curr_past_key_value.layers[self.layer_idx].values
|
| 75 |
+
else:
|
| 76 |
+
key_states, value_states = self.c_attn(encoder_hidden_states).split(self.split_size, dim=2)
|
| 77 |
+
shape_kv = (*key_states.shape[:-1], -1, self.head_dim)
|
| 78 |
+
key_states = key_states.view(shape_kv).transpose(1, 2)
|
| 79 |
+
value_states = value_states.view(shape_kv).transpose(1, 2)
|
| 80 |
+
else:
|
| 81 |
+
query_states, key_states, value_states = self.c_attn(hidden_states).split(self.split_size, dim=2)
|
| 82 |
+
shape_kv = (*key_states.shape[:-1], -1, self.head_dim)
|
| 83 |
+
key_states = key_states.view(shape_kv).transpose(1, 2)
|
| 84 |
+
value_states = value_states.view(shape_kv).transpose(1, 2)
|
| 85 |
+
|
| 86 |
+
shape_q = (*query_states.shape[:-1], -1, self.head_dim)
|
| 87 |
+
query_states = query_states.view(shape_q).transpose(1, 2)
|
| 88 |
+
|
| 89 |
+
if (past_key_values is not None and not is_cross_attention) or (
|
| 90 |
+
past_key_values is not None and is_cross_attention and not is_updated
|
| 91 |
+
):
|
| 92 |
+
# save all key/value_layer to cache to be re-used for fast auto-regressive generation
|
| 93 |
+
cache_position = cache_position if not is_cross_attention else None
|
| 94 |
+
key_states, value_states = curr_past_key_value.update(
|
| 95 |
+
key_states, value_states, self.layer_idx, {"cache_position": cache_position}
|
| 96 |
+
)
|
| 97 |
+
# set flag that curr layer for cross-attn is already updated so we can re-use in subsequent calls
|
| 98 |
+
if is_cross_attention:
|
| 99 |
+
past_key_values.is_updated[self.layer_idx] = True
|
| 100 |
+
|
| 101 |
+
is_causal = attention_mask is None and query_states.shape[-2] > 1 and not is_cross_attention
|
| 102 |
+
|
| 103 |
+
using_eager = self.config._attn_implementation == "eager"
|
| 104 |
+
attention_interface: Callable = eager_attention_forward
|
| 105 |
+
if self.config._attn_implementation != "eager":
|
| 106 |
+
attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
|
| 107 |
+
|
| 108 |
+
if using_eager and self.reorder_and_upcast_attn:
|
| 109 |
+
attn_output, attn_weights = self._upcast_and_reordered_attn(
|
| 110 |
+
query_states, key_states, value_states, attention_mask, head_mask
|
| 111 |
+
)
|
| 112 |
+
else:
|
| 113 |
+
if getattr(self.config, "prefix_allowed_length", None) is not None:
|
| 114 |
+
temp = self
|
| 115 |
+
temp.is_cross_attention = True
|
| 116 |
+
attn_output, attn_weights = attention_interface(
|
| 117 |
+
self if getattr(self.config, "prefix_allowed_length", None) is None else temp,
|
| 118 |
+
query_states,
|
| 119 |
+
key_states,
|
| 120 |
+
value_states,
|
| 121 |
+
attention_mask,
|
| 122 |
+
head_mask=head_mask,
|
| 123 |
+
dropout=self.attn_dropout.p if self.training else 0.0,
|
| 124 |
+
is_causal=is_causal if getattr(self.config, "is_prefix", None) is None else False,
|
| 125 |
+
**kwargs,
|
| 126 |
+
)
|
| 127 |
+
if getattr(self.config, "plot_attention_map", False) and self.layer_idx in getattr(self.config, "plot_attention_map_layer", []):
|
| 128 |
+
# pick batch=0, head=0
|
| 129 |
+
attn_bh = attn_weights[0, 0] # [L,S]
|
| 130 |
+
L, S = attn_bh.shape
|
| 131 |
+
if L > 1:
|
| 132 |
+
if getattr(self.config, "plot_attention_map_generation", 0) == 0:
|
| 133 |
+
print(f"Plotting attention map for inputs on layer {self.layer_idx}")
|
| 134 |
+
# full 2D heatmap
|
| 135 |
+
data = attn_bh.detach().float().cpu().numpy() # [L,S]
|
| 136 |
+
plt.figure(figsize=(6,5))
|
| 137 |
+
plt.imshow(data, aspect="auto", cmap="hot", vmin=0, vmax=0.01)
|
| 138 |
+
plt.colorbar()
|
| 139 |
+
plt.xlabel("Keys (S)")
|
| 140 |
+
plt.ylabel("Queries (L)")
|
| 141 |
+
plt.title(f"Attention map (B0,H0) L={L}, S={S}")
|
| 142 |
+
plt.show()
|
| 143 |
+
else:
|
| 144 |
+
if getattr(self.config, "plot_attention_map_generation", 0) == S:
|
| 145 |
+
print(f"Plotting attention row map for token {S} generation on layer {self.layer_idx}")
|
| 146 |
+
# attn_bh expected shape: [..., S] for the selected (B0, H0) row
|
| 147 |
+
row = attn_bh[0].detach().float().cpu().numpy() # -> np.ndarray shape [S]
|
| 148 |
+
n = row.shape[0]
|
| 149 |
+
|
| 150 |
+
# ----- First 1024 as 32x32 -----
|
| 151 |
+
head_1024 = row[:min(1024, n)]
|
| 152 |
+
grid = head_1024.reshape(32, 32)
|
| 153 |
+
|
| 154 |
+
plt.figure(figsize=(6, 5))
|
| 155 |
+
plt.imshow(grid, aspect="auto", cmap="hot", vmin=0, vmax=0.01)
|
| 156 |
+
plt.yticks([])
|
| 157 |
+
plt.colorbar()
|
| 158 |
+
plt.xlabel("Keys (S) [indices 0..1023]")
|
| 159 |
+
plt.title(f"Attention row (B0,H0) L={self.layer_idx}, S={S} — first 1024")
|
| 160 |
+
plt.tight_layout()
|
| 161 |
+
plt.show()
|
| 162 |
+
|
| 163 |
+
# ----- Tail (>=1024) as a single-row heatmap -----
|
| 164 |
+
tail = row[1024:]
|
| 165 |
+
if tail.size > 0:
|
| 166 |
+
plt.figure(figsize=(10, 1.2))
|
| 167 |
+
# one-row heatmap
|
| 168 |
+
plt.imshow(tail[None, :], aspect="auto", cmap="hot", vmin=0, vmax=0.01)
|
| 169 |
+
plt.yticks([])
|
| 170 |
+
plt.colorbar()
|
| 171 |
+
plt.xlabel(f"Keys (S) [indices 1024..{n-1}]")
|
| 172 |
+
plt.title(f"Attention row tail (B0,H0) L={self.layer_idx}, S={S}")
|
| 173 |
+
plt.tight_layout()
|
| 174 |
+
plt.show()
|
| 175 |
+
|
| 176 |
+
attn_output = attn_output.reshape(*attn_output.shape[:-2], -1).contiguous()
|
| 177 |
+
attn_output = self.c_proj(attn_output)
|
| 178 |
+
attn_output = self.resid_dropout(attn_output)
|
| 179 |
+
|
| 180 |
+
return attn_output, attn_weights
|
| 181 |
+
|
| 182 |
+
class GPT2BlockModified(GPT2Block):
|
| 183 |
+
def __init__(self, config, layer_idx=None):
|
| 184 |
+
super().__init__(config=config)
|
| 185 |
+
self.attn = GPT2AttentionModified(config=config, layer_idx=layer_idx)
|
| 186 |
+
|
| 187 |
+
def forward(
|
| 188 |
+
self,
|
| 189 |
+
hidden_states: Optional[tuple[torch.FloatTensor]],
|
| 190 |
+
past_key_values: Optional[Cache] = None,
|
| 191 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 192 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
| 193 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
| 194 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
| 195 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
| 196 |
+
use_cache: Optional[bool] = False,
|
| 197 |
+
output_attentions: Optional[bool] = False,
|
| 198 |
+
**kwargs,
|
| 199 |
+
) -> Union[tuple[torch.Tensor], Optional[tuple[torch.Tensor, tuple[torch.FloatTensor, ...]]]]:
|
| 200 |
+
residual = hidden_states
|
| 201 |
+
hidden_states = self.ln_1(hidden_states)
|
| 202 |
+
attn_output, self_attn_weights = self.attn(
|
| 203 |
+
hidden_states,
|
| 204 |
+
past_key_values=past_key_values,
|
| 205 |
+
cache_position=cache_position,
|
| 206 |
+
attention_mask=attention_mask,
|
| 207 |
+
head_mask=head_mask,
|
| 208 |
+
use_cache=use_cache,
|
| 209 |
+
output_attentions=output_attentions,
|
| 210 |
+
**kwargs,
|
| 211 |
+
)
|
| 212 |
+
|
| 213 |
+
# residual connection
|
| 214 |
+
hidden_states = attn_output + residual
|
| 215 |
+
|
| 216 |
+
if encoder_hidden_states is not None:
|
| 217 |
+
# add one self-attention block for cross-attention
|
| 218 |
+
if not hasattr(self, "crossattention"):
|
| 219 |
+
raise ValueError(
|
| 220 |
+
f"If `encoder_hidden_states` are passed, {self} has to be instantiated with "
|
| 221 |
+
"cross-attention layers by setting `config.add_cross_attention=True`"
|
| 222 |
+
)
|
| 223 |
+
residual = hidden_states
|
| 224 |
+
hidden_states = self.ln_cross_attn(hidden_states)
|
| 225 |
+
cross_attn_output, cross_attn_weights = self.crossattention(
|
| 226 |
+
hidden_states,
|
| 227 |
+
past_key_values=past_key_values,
|
| 228 |
+
attention_mask=attention_mask,
|
| 229 |
+
head_mask=head_mask,
|
| 230 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 231 |
+
encoder_attention_mask=encoder_attention_mask,
|
| 232 |
+
output_attentions=output_attentions,
|
| 233 |
+
)
|
| 234 |
+
# residual connection
|
| 235 |
+
hidden_states = residual + cross_attn_output
|
| 236 |
+
|
| 237 |
+
residual = hidden_states
|
| 238 |
+
hidden_states = self.ln_2(hidden_states)
|
| 239 |
+
feed_forward_hidden_states = self.mlp(hidden_states)
|
| 240 |
+
# residual connection
|
| 241 |
+
hidden_states = residual + feed_forward_hidden_states
|
| 242 |
+
|
| 243 |
+
outputs = (hidden_states,)
|
| 244 |
+
if output_attentions:
|
| 245 |
+
outputs += (self_attn_weights,)
|
| 246 |
+
if encoder_hidden_states is not None:
|
| 247 |
+
outputs += (cross_attn_weights,)
|
| 248 |
+
|
| 249 |
+
return outputs
|
| 250 |
+
|
| 251 |
+
|
| 252 |
+
class GPT2ModelModified(GPT2Model):
|
| 253 |
+
def __init__(self, config):
|
| 254 |
+
super().__init__(config)
|
| 255 |
+
self.config = config
|
| 256 |
+
self.config_causal = config
|
| 257 |
+
self.config_causal._attn_implementation = "eager" # Ensure causal mask creation uses eager implementation
|
| 258 |
+
# TEMPORARY: override the transformer blocks to pass segmentation masks
|
| 259 |
+
self.h = nn.ModuleList([GPT2BlockModified(config, layer_idx=i) for i in range(config.num_hidden_layers)])
|
| 260 |
+
|
| 261 |
+
def forward(
|
| 262 |
+
self,
|
| 263 |
+
input_ids: Optional[torch.LongTensor] = None,
|
| 264 |
+
past_key_values: Optional[Union[tuple[tuple[torch.Tensor]], Cache]] = None,
|
| 265 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 266 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
| 267 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
| 268 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 269 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
| 270 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 271 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
| 272 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
| 273 |
+
use_cache: Optional[bool] = None,
|
| 274 |
+
output_attentions: Optional[bool] = None,
|
| 275 |
+
output_hidden_states: Optional[bool] = None,
|
| 276 |
+
return_dict: Optional[bool] = None,
|
| 277 |
+
segmentation_mask: Optional[torch.FloatTensor] = None,
|
| 278 |
+
**kwargs,
|
| 279 |
+
) -> Union[tuple, BaseModelOutputWithPastAndCrossAttentions]:
|
| 280 |
+
r"""
|
| 281 |
+
input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
|
| 282 |
+
`input_ids_length` = `sequence_length` if `past_key_values` is `None` else
|
| 283 |
+
`past_key_values.get_seq_length()` (`sequence_length` of input past key value states). Indices of input
|
| 284 |
+
sequence tokens in the vocabulary.
|
| 285 |
+
|
| 286 |
+
If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
|
| 287 |
+
`input_ids`.
|
| 288 |
+
|
| 289 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
| 290 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
| 291 |
+
|
| 292 |
+
[What are input IDs?](../glossary#input-ids)
|
| 293 |
+
"""
|
| 294 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 295 |
+
output_hidden_states = (
|
| 296 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 297 |
+
)
|
| 298 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
| 299 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 300 |
+
|
| 301 |
+
if input_ids is not None and inputs_embeds is not None:
|
| 302 |
+
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
|
| 303 |
+
elif input_ids is not None:
|
| 304 |
+
self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
|
| 305 |
+
input_shape = input_ids.size()
|
| 306 |
+
input_ids = input_ids.view(-1, input_shape[-1])
|
| 307 |
+
batch_size = input_ids.shape[0]
|
| 308 |
+
elif inputs_embeds is not None:
|
| 309 |
+
input_shape = inputs_embeds.size()[:-1]
|
| 310 |
+
batch_size = inputs_embeds.shape[0]
|
| 311 |
+
else:
|
| 312 |
+
raise ValueError("You have to specify either input_ids or inputs_embeds")
|
| 313 |
+
|
| 314 |
+
device = input_ids.device if input_ids is not None else inputs_embeds.device
|
| 315 |
+
|
| 316 |
+
if token_type_ids is not None:
|
| 317 |
+
token_type_ids = token_type_ids.view(-1, input_shape[-1])
|
| 318 |
+
|
| 319 |
+
if self.gradient_checkpointing and self.training:
|
| 320 |
+
if use_cache:
|
| 321 |
+
logger.warning_once(
|
| 322 |
+
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
|
| 323 |
+
)
|
| 324 |
+
use_cache = False
|
| 325 |
+
|
| 326 |
+
# based on pattern from src/transformers/models/whisper/modeling_whisper.py::WhisperDecoder
|
| 327 |
+
if use_cache:
|
| 328 |
+
if past_key_values is None:
|
| 329 |
+
past_key_values = DynamicCache()
|
| 330 |
+
elif isinstance(past_key_values, tuple):
|
| 331 |
+
logger.warning_once(
|
| 332 |
+
"Passing a tuple of `past_key_values` is deprecated and will be removed in Transformers v4.53.0. "
|
| 333 |
+
"You should pass an instance of `Cache` instead, e.g. "
|
| 334 |
+
"`past_key_values=DynamicCache.from_legacy_cache(past_key_values)`."
|
| 335 |
+
)
|
| 336 |
+
past_key_values = DynamicCache.from_legacy_cache(past_key_values)
|
| 337 |
+
|
| 338 |
+
if self.config.add_cross_attention and not isinstance(past_key_values, EncoderDecoderCache):
|
| 339 |
+
past_key_values = EncoderDecoderCache(past_key_values, DynamicCache())
|
| 340 |
+
|
| 341 |
+
if inputs_embeds is None:
|
| 342 |
+
inputs_embeds = self.wte(input_ids)
|
| 343 |
+
|
| 344 |
+
if cache_position is None:
|
| 345 |
+
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
| 346 |
+
cache_position = torch.arange(
|
| 347 |
+
past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
|
| 348 |
+
)
|
| 349 |
+
if position_ids is None:
|
| 350 |
+
position_ids = cache_position.unsqueeze(0)
|
| 351 |
+
|
| 352 |
+
position_embeds = self.wpe(position_ids)
|
| 353 |
+
hidden_states = inputs_embeds + position_embeds.to(inputs_embeds.device)
|
| 354 |
+
|
| 355 |
+
# Attention mask.
|
| 356 |
+
# ._update_causal_mask() and ._prepare_4d_causal_attention_mask_with_cache_position() copied from LlamaModel
|
| 357 |
+
if attention_mask is not None and attention_mask.ndim < 4:
|
| 358 |
+
attention_mask = attention_mask.view(batch_size, -1)
|
| 359 |
+
|
| 360 |
+
causal_mask = create_causal_mask(
|
| 361 |
+
config=self.config_causal,
|
| 362 |
+
input_embeds=inputs_embeds,
|
| 363 |
+
attention_mask=attention_mask,
|
| 364 |
+
cache_position=cache_position,
|
| 365 |
+
past_key_values=past_key_values,
|
| 366 |
+
position_ids=position_ids,
|
| 367 |
+
)
|
| 368 |
+
|
| 369 |
+
# If a 2D or 3D attention mask is provided for the cross-attention
|
| 370 |
+
# we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
|
| 371 |
+
_use_sdpa = self._attn_implementation == "sdpa" and output_attentions is False and head_mask is None
|
| 372 |
+
if self.config.add_cross_attention and encoder_hidden_states is not None:
|
| 373 |
+
encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
|
| 374 |
+
encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
|
| 375 |
+
if encoder_attention_mask is None:
|
| 376 |
+
encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
|
| 377 |
+
if _use_sdpa:
|
| 378 |
+
encoder_attention_mask = _prepare_4d_attention_mask_for_sdpa(
|
| 379 |
+
mask=encoder_attention_mask, dtype=inputs_embeds.dtype, tgt_len=input_shape[-1]
|
| 380 |
+
)
|
| 381 |
+
elif self._attn_implementation != "flash_attention_2":
|
| 382 |
+
encoder_attention_mask = self.invert_attention_mask(encoder_attention_mask)
|
| 383 |
+
else:
|
| 384 |
+
encoder_attention_mask = None
|
| 385 |
+
|
| 386 |
+
# Prepare head mask if needed
|
| 387 |
+
# 1.0 in head_mask indicate we keep the head
|
| 388 |
+
# attention_probs has shape bsz x n_heads x N x N
|
| 389 |
+
# head_mask has shape n_layer x batch x n_heads x N x N
|
| 390 |
+
head_mask = self.get_head_mask(head_mask, self.config.n_layer)
|
| 391 |
+
|
| 392 |
+
if token_type_ids is not None:
|
| 393 |
+
token_type_embeds = self.wte(token_type_ids)
|
| 394 |
+
hidden_states = hidden_states + token_type_embeds
|
| 395 |
+
|
| 396 |
+
hidden_states = self.drop(hidden_states)
|
| 397 |
+
|
| 398 |
+
output_shape = (-1,) + input_shape[1:] + (hidden_states.size(-1),)
|
| 399 |
+
|
| 400 |
+
all_self_attentions = () if output_attentions else None
|
| 401 |
+
all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
|
| 402 |
+
all_hidden_states = () if output_hidden_states else None
|
| 403 |
+
for i, block in enumerate(self.h):
|
| 404 |
+
# Model parallel
|
| 405 |
+
if self.model_parallel:
|
| 406 |
+
torch.cuda.set_device(hidden_states.device)
|
| 407 |
+
if isinstance(head_mask, torch.Tensor):
|
| 408 |
+
head_mask = head_mask.to(hidden_states.device)
|
| 409 |
+
if output_hidden_states:
|
| 410 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
| 411 |
+
if segmentation_mask is not None and causal_mask is not None:
|
| 412 |
+
# Make a safe copy of the causal mask and ensure its spatial
|
| 413 |
+
# dimensions match the sequence length that the attention
|
| 414 |
+
# functions expect. This prevents off-by-one shape errors
|
| 415 |
+
# when using eager attention (torch.where requires same sizes).
|
| 416 |
+
causal_mask_modified = causal_mask.clone()
|
| 417 |
+
if getattr(self.config, "prefix_allowed_length", None) is not None:
|
| 418 |
+
causal_mask_modified[:, :, :, :self.config.prefix_allowed_length].zero_()
|
| 419 |
+
|
| 420 |
+
# Use the input sequence length to crop the causal mask if needed
|
| 421 |
+
seq_len = input_shape[-1]
|
| 422 |
+
if causal_mask_modified.shape[2] != seq_len or causal_mask_modified.shape[3] != seq_len:
|
| 423 |
+
causal_mask_modified = causal_mask_modified[:, :, :seq_len, :seq_len]
|
| 424 |
+
|
| 425 |
+
# Clip segmentation mask to fit into causal_mask_modified before adding.
|
| 426 |
+
_, _, M, N = segmentation_mask.shape
|
| 427 |
+
M = min(M, causal_mask_modified.shape[2])
|
| 428 |
+
N = min(N, causal_mask_modified.shape[3])
|
| 429 |
+
causal_mask_modified[:, :, :M, :N] += segmentation_mask[:, i, :M, :N].unsqueeze(1)
|
| 430 |
+
if getattr(self.config, "plot_attention_mask", False) and i in getattr(self.config, "plot_attention_mask_layer", [0]):
|
| 431 |
+
if segmentation_mask is not None and causal_mask is not None:
|
| 432 |
+
print(f"Block {i}: segmentation mask added to causal mask.")
|
| 433 |
+
plt.imshow(causal_mask_modified[0,0].detach().cpu(), aspect='auto', cmap='hot', vmin=-1, vmax=1)
|
| 434 |
+
plt.colorbar()
|
| 435 |
+
plt.title(f"Causal Mask with Segmentation (Block {i})")
|
| 436 |
+
plt.show()
|
| 437 |
+
else:
|
| 438 |
+
print(f"Block {i}: no segmentation mask applied.")
|
| 439 |
+
plt.imshow(causal_mask[0,0].detach().cpu(), aspect='auto', cmap='hot', vmin=-1, vmax=1)
|
| 440 |
+
plt.colorbar()
|
| 441 |
+
plt.title(f"Causal Mask (Block {i})")
|
| 442 |
+
plt.show()
|
| 443 |
+
|
| 444 |
+
|
| 445 |
+
outputs = block(
|
| 446 |
+
hidden_states,
|
| 447 |
+
past_key_values if not (self.gradient_checkpointing and self.training) else None,
|
| 448 |
+
cache_position,
|
| 449 |
+
causal_mask_modified if segmentation_mask is not None and causal_mask is not None else causal_mask,
|
| 450 |
+
head_mask[i],
|
| 451 |
+
encoder_hidden_states, # as a positional argument for gradient checkpointing
|
| 452 |
+
encoder_attention_mask=encoder_attention_mask,
|
| 453 |
+
use_cache=use_cache,
|
| 454 |
+
output_attentions=output_attentions,
|
| 455 |
+
**kwargs,
|
| 456 |
+
)
|
| 457 |
+
|
| 458 |
+
hidden_states = outputs[0]
|
| 459 |
+
|
| 460 |
+
if output_attentions:
|
| 461 |
+
all_self_attentions = all_self_attentions + (outputs[1],)
|
| 462 |
+
if self.config.add_cross_attention:
|
| 463 |
+
all_cross_attentions = all_cross_attentions + (outputs[2],)
|
| 464 |
+
|
| 465 |
+
# Model Parallel: If it's the last layer for that device, put things on the next device
|
| 466 |
+
if self.model_parallel:
|
| 467 |
+
for k, v in self.device_map.items():
|
| 468 |
+
if i == v[-1] and "cuda:" + str(k) != self.last_device:
|
| 469 |
+
hidden_states = hidden_states.to("cuda:" + str(k + 1))
|
| 470 |
+
|
| 471 |
+
hidden_states = self.ln_f(hidden_states)
|
| 472 |
+
|
| 473 |
+
hidden_states = hidden_states.view(output_shape)
|
| 474 |
+
# Add last hidden state
|
| 475 |
+
if output_hidden_states:
|
| 476 |
+
all_hidden_states = all_hidden_states + (hidden_states,)
|
| 477 |
+
|
| 478 |
+
past_key_values = past_key_values if use_cache else None
|
| 479 |
+
if not return_dict:
|
| 480 |
+
return tuple(
|
| 481 |
+
v
|
| 482 |
+
for v in [hidden_states, past_key_values, all_hidden_states, all_self_attentions, all_cross_attentions]
|
| 483 |
+
if v is not None
|
| 484 |
+
)
|
| 485 |
+
|
| 486 |
+
return BaseModelOutputWithPastAndCrossAttentions(
|
| 487 |
+
last_hidden_state=hidden_states,
|
| 488 |
+
past_key_values=past_key_values,
|
| 489 |
+
hidden_states=all_hidden_states,
|
| 490 |
+
attentions=all_self_attentions,
|
| 491 |
+
cross_attentions=all_cross_attentions,
|
| 492 |
+
)
|
| 493 |
+
|
| 494 |
+
class GPT2LMHeadModelModified(GPT2LMHeadModel):
|
| 495 |
+
def __init__(self, config):
|
| 496 |
+
super().__init__(config)
|
| 497 |
+
# replace the base transformer with our modified transformer implementation
|
| 498 |
+
self.transformer = GPT2ModelModified(config)
|
| 499 |
+
self.post_init()
|
| 500 |
+
|
| 501 |
+
def forward(
|
| 502 |
+
self,
|
| 503 |
+
input_ids: Optional[torch.LongTensor] = None,
|
| 504 |
+
past_key_values: Optional[tuple[tuple[torch.Tensor]]] = None,
|
| 505 |
+
cache_position: Optional[torch.LongTensor] = None,
|
| 506 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
| 507 |
+
token_type_ids: Optional[torch.LongTensor] = None,
|
| 508 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 509 |
+
head_mask: Optional[torch.FloatTensor] = None,
|
| 510 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 511 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
| 512 |
+
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
| 513 |
+
labels: Optional[torch.LongTensor] = None,
|
| 514 |
+
use_cache: Optional[bool] = None,
|
| 515 |
+
output_attentions: Optional[bool] = None,
|
| 516 |
+
output_hidden_states: Optional[bool] = None,
|
| 517 |
+
return_dict: Optional[bool] = None,
|
| 518 |
+
logits_to_keep: Union[int, torch.Tensor] = 0,
|
| 519 |
+
segmentation_mask: Optional[torch.FloatTensor] = None,
|
| 520 |
+
prefix_allowed_length: Optional[int] = None,
|
| 521 |
+
plot_attention_mask: Optional[bool] = False,
|
| 522 |
+
plot_attention_mask_layer: Optional[list[int]] = [0],
|
| 523 |
+
plot_attention_map: Optional[bool] = False,
|
| 524 |
+
plot_attention_map_layer: Optional[list[int]] = [0],
|
| 525 |
+
plot_attention_map_generation: Optional[int] = 0,
|
| 526 |
+
**kwargs,
|
| 527 |
+
) -> Union[tuple, CausalLMOutputWithCrossAttentions]:
|
| 528 |
+
r"""
|
| 529 |
+
input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
|
| 530 |
+
`input_ids_length` = `sequence_length` if `past_key_values` is `None` else
|
| 531 |
+
`past_key_values.get_seq_length()` (`sequence_length` of input past key value states). Indices of input
|
| 532 |
+
sequence tokens in the vocabulary.
|
| 533 |
+
|
| 534 |
+
If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
|
| 535 |
+
`input_ids`.
|
| 536 |
+
|
| 537 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
| 538 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
| 539 |
+
|
| 540 |
+
[What are input IDs?](../glossary#input-ids)
|
| 541 |
+
labels (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*):
|
| 542 |
+
Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
|
| 543 |
+
`labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
|
| 544 |
+
are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
|
| 545 |
+
"""
|
| 546 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 547 |
+
|
| 548 |
+
if prefix_allowed_length is not None:
|
| 549 |
+
self.config.prefix_allowed_length = prefix_allowed_length
|
| 550 |
+
|
| 551 |
+
if plot_attention_mask is not None:
|
| 552 |
+
self.config.plot_attention_mask = plot_attention_mask
|
| 553 |
+
if plot_attention_mask_layer is not None:
|
| 554 |
+
self.config.plot_attention_mask_layer = plot_attention_mask_layer
|
| 555 |
+
|
| 556 |
+
if plot_attention_map is not None:
|
| 557 |
+
if plot_attention_map_layer is not None:
|
| 558 |
+
self.config.plot_attention_map_layer = plot_attention_map_layer
|
| 559 |
+
if plot_attention_map_generation is not None:
|
| 560 |
+
self.config.plot_attention_map_generation = plot_attention_map_generation
|
| 561 |
+
self.config.plot_attention_map = plot_attention_map
|
| 562 |
+
|
| 563 |
+
transformer_outputs = self.transformer(
|
| 564 |
+
input_ids,
|
| 565 |
+
past_key_values=past_key_values,
|
| 566 |
+
attention_mask=attention_mask,
|
| 567 |
+
cache_position=cache_position,
|
| 568 |
+
token_type_ids=token_type_ids,
|
| 569 |
+
position_ids=position_ids,
|
| 570 |
+
head_mask=head_mask,
|
| 571 |
+
inputs_embeds=inputs_embeds,
|
| 572 |
+
encoder_hidden_states=encoder_hidden_states,
|
| 573 |
+
encoder_attention_mask=encoder_attention_mask,
|
| 574 |
+
use_cache=use_cache,
|
| 575 |
+
output_attentions=output_attentions,
|
| 576 |
+
output_hidden_states=output_hidden_states,
|
| 577 |
+
return_dict=return_dict,
|
| 578 |
+
segmentation_mask=segmentation_mask, #Added this parameter
|
| 579 |
+
**kwargs,
|
| 580 |
+
)
|
| 581 |
+
hidden_states = transformer_outputs[0]
|
| 582 |
+
|
| 583 |
+
# Set device for model parallelism
|
| 584 |
+
if self.model_parallel:
|
| 585 |
+
torch.cuda.set_device(self.transformer.first_device)
|
| 586 |
+
hidden_states = hidden_states.to(self.lm_head.weight.device)
|
| 587 |
+
|
| 588 |
+
slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
|
| 589 |
+
logits = self.lm_head(hidden_states[:, slice_indices, :])
|
| 590 |
+
|
| 591 |
+
loss = None
|
| 592 |
+
if labels is not None:
|
| 593 |
+
# Flatten the tokens
|
| 594 |
+
loss = self.loss_function(
|
| 595 |
+
logits,
|
| 596 |
+
labels,
|
| 597 |
+
vocab_size=self.config.vocab_size,
|
| 598 |
+
**kwargs,
|
| 599 |
+
)
|
| 600 |
+
|
| 601 |
+
if not return_dict:
|
| 602 |
+
output = (logits,) + transformer_outputs[1:]
|
| 603 |
+
return ((loss,) + output) if loss is not None else output
|
| 604 |
+
|
| 605 |
+
return CausalLMOutputWithCrossAttentions(
|
| 606 |
+
loss=loss,
|
| 607 |
+
logits=logits,
|
| 608 |
+
past_key_values=transformer_outputs.past_key_values,
|
| 609 |
+
hidden_states=transformer_outputs.hidden_states,
|
| 610 |
+
attentions=transformer_outputs.attentions,
|
| 611 |
+
cross_attentions=transformer_outputs.cross_attentions,
|
| 612 |
+
)
|
| 613 |
+
|
| 614 |
+
@torch.no_grad()
|
| 615 |
+
def expand_gpt2_positional_embeddings(
|
| 616 |
+
model: torch.nn.Module,
|
| 617 |
+
new_max_positions: int,
|
| 618 |
+
mode: str = "linear", # "linear" | "copy_last" | "zeros"
|
| 619 |
+
align_corners: bool = True, # for linear interpolation
|
| 620 |
+
):
|
| 621 |
+
"""
|
| 622 |
+
Expand GPT-2's learned positional embeddings (wpe) to `new_max_positions`.
|
| 623 |
+
|
| 624 |
+
Works with GPT2LMHeadModel or GPT2Model (HF). Updates model.config.n_positions (and n_ctx if present).
|
| 625 |
+
Does NOT mutate token embeddings; only position table + config.
|
| 626 |
+
|
| 627 |
+
Args:
|
| 628 |
+
model: HF GPT2LMHeadModel or GPT2Model (already loaded).
|
| 629 |
+
new_max_positions: int, desired max sequence length (e.g., 1536 or 2048).
|
| 630 |
+
mode: how to initialize new rows if expanding:
|
| 631 |
+
- "linear": 1D linear interpolation along position dim (recommended)
|
| 632 |
+
- "copy_last": copy the last learned vector into all new rows
|
| 633 |
+
- "zeros": initialize new rows to zero
|
| 634 |
+
align_corners: passed to F.interpolate for "linear" mode.
|
| 635 |
+
|
| 636 |
+
Returns:
|
| 637 |
+
model (same instance) with expanded wpe and updated config.
|
| 638 |
+
"""
|
| 639 |
+
# Locate the position embedding table.
|
| 640 |
+
# Support both:
|
| 641 |
+
# - GPT2LMHeadModel (has .transformer which is a GPT2Model with .wpe)
|
| 642 |
+
# - GPT2Model (exposes .wpe directly)
|
| 643 |
+
if hasattr(model, "transformer") and hasattr(model.transformer, "wpe"):
|
| 644 |
+
model_for_wpe = model.transformer
|
| 645 |
+
elif hasattr(model, "wpe"):
|
| 646 |
+
model_for_wpe = model
|
| 647 |
+
else:
|
| 648 |
+
raise ValueError("Model does not look like a GPT-2 family model with a position embedding 'wpe')")
|
| 649 |
+
|
| 650 |
+
wpe = model_for_wpe.wpe
|
| 651 |
+
|
| 652 |
+
old_n, d = wpe.weight.shape
|
| 653 |
+
if new_max_positions <= 0:
|
| 654 |
+
raise ValueError("new_max_positions must be positive")
|
| 655 |
+
if new_max_positions == old_n:
|
| 656 |
+
# Still update config for consistency
|
| 657 |
+
if hasattr(model.config, "n_positions"):
|
| 658 |
+
model.config.n_positions = new_max_positions
|
| 659 |
+
if hasattr(model.config, "n_ctx"):
|
| 660 |
+
model.config.n_ctx = new_max_positions
|
| 661 |
+
return model
|
| 662 |
+
|
| 663 |
+
device = wpe.weight.device
|
| 664 |
+
dtype = wpe.weight.dtype
|
| 665 |
+
|
| 666 |
+
if new_max_positions < old_n:
|
| 667 |
+
# Shrink (rare): just slice
|
| 668 |
+
new_weight = wpe.weight[:new_max_positions].clone()
|
| 669 |
+
else:
|
| 670 |
+
# Expand
|
| 671 |
+
if mode == "linear":
|
| 672 |
+
# Interpolate along position dimension.
|
| 673 |
+
# Treat embedding dim as channels: (1, d, old_n) -> (1, d, new_n) -> (new_n, d)
|
| 674 |
+
w = wpe.weight.transpose(0, 1).unsqueeze(0) # (1, d, old_n)
|
| 675 |
+
w_new = F.interpolate(w, size=new_max_positions, mode="linear", align_corners=align_corners)
|
| 676 |
+
new_weight = w_new.squeeze(0).transpose(0, 1).contiguous() # (new_n, d)
|
| 677 |
+
elif mode == "copy_last":
|
| 678 |
+
new_weight = torch.empty((new_max_positions, d), device=device, dtype=dtype)
|
| 679 |
+
new_weight[:old_n].copy_(wpe.weight)
|
| 680 |
+
new_weight[old_n:].copy_(wpe.weight[old_n - 1].expand(new_max_positions - old_n, d))
|
| 681 |
+
elif mode == "zeros":
|
| 682 |
+
new_weight = torch.zeros((new_max_positions, d), device=device, dtype=dtype)
|
| 683 |
+
new_weight[:old_n].copy_(wpe.weight)
|
| 684 |
+
else:
|
| 685 |
+
raise ValueError(f"Unknown mode '{mode}'")
|
| 686 |
+
|
| 687 |
+
# Replace embedding module on whichever object held the original table
|
| 688 |
+
new_wpe = torch.nn.Embedding(new_max_positions, d, device=device, dtype=dtype)
|
| 689 |
+
new_wpe.weight.copy_(new_weight)
|
| 690 |
+
|
| 691 |
+
# Keep requires_grad True (default). If you want to freeze, set .requires_grad_(False).
|
| 692 |
+
if hasattr(model, "transformer") and hasattr(model.transformer, "wpe"):
|
| 693 |
+
model.transformer.wpe = new_wpe
|
| 694 |
+
else:
|
| 695 |
+
model.wpe = new_wpe
|
| 696 |
+
|
| 697 |
+
# Update config fields used by HF
|
| 698 |
+
if hasattr(model.config, "n_positions"):
|
| 699 |
+
model.config.n_positions = new_max_positions
|
| 700 |
+
if hasattr(model.config, "n_ctx"):
|
| 701 |
+
model.config.n_ctx = new_max_positions
|
| 702 |
+
|
| 703 |
+
return model
|
| 704 |
+
|
| 705 |
+
def create_decoder(attention = "sdpa"):
|
| 706 |
+
config = GPT2Config.from_pretrained("gpt2")
|
| 707 |
+
config._attn_implementation = attention
|
| 708 |
+
new_max_positions = 2048
|
| 709 |
+
decoder = GPT2LMHeadModelModified.from_pretrained("gpt2", config=config)
|
| 710 |
+
decoder.config._attn_implementation = attention
|
| 711 |
+
decoder = expand_gpt2_positional_embeddings(decoder, new_max_positions=new_max_positions, mode="linear")
|
| 712 |
+
return decoder
|
utils/processing.py
ADDED
|
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torchvision.transforms as T
|
| 2 |
+
import fsspec
|
| 3 |
+
import io
|
| 4 |
+
from PIL import Image
|
| 5 |
+
|
| 6 |
+
def image_transform(img_size=512):
|
| 7 |
+
return T.Compose([
|
| 8 |
+
T.Resize((img_size, img_size), interpolation=T.InterpolationMode.BICUBIC),
|
| 9 |
+
T.ToTensor(),
|
| 10 |
+
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
|
| 11 |
+
])
|
| 12 |
+
|
| 13 |
+
def open_binary(path: str):
|
| 14 |
+
"""
|
| 15 |
+
Open any (local or gs://) file for binary reading.
|
| 16 |
+
Returns a file-like object (context manager).
|
| 17 |
+
"""
|
| 18 |
+
return fsspec.open(path, mode="rb").open()
|
| 19 |
+
|
| 20 |
+
def pil_from_path(path: str) -> Image.Image:
|
| 21 |
+
"""
|
| 22 |
+
Load an image from local or GCS; returns a PIL image in RGB.
|
| 23 |
+
"""
|
| 24 |
+
with open_binary(path) as f:
|
| 25 |
+
img_bytes = f.read()
|
| 26 |
+
im = Image.open(io.BytesIO(img_bytes)).convert("RGB")
|
| 27 |
+
return im
|