SAG-ViT / modeling_sagvit.py
shravvvv's picture
Added files
b99e299
import torch
from torch import nn
from transformers import AutoConfig, PretrainedConfig, AutoModel, PreTrainedModel
from transformers.models.auto import AutoConfig, CONFIG_MAPPING, MODEL_MAPPING
from transformers.utils import logging
from transformers.modeling_utils import ModelOutput
from huggingface_hub import PyTorchModelHubMixin
from torch_geometric.data import Batch
from model_components import EfficientNetV2FeatureExtractor, GATGNN, TransformerEncoder, MLPBlock
from graph_construction import build_graph_from_patches, build_graph_data_from_patches
###############################################################################
# SAG-ViT Model:
# This class combines:
# 1) CNN backbone to produce high-fidelity feature maps (Section 3.1),
# 2) Graph construction and GAT to refine local patch embeddings (Section 3.2 and 3.3),
# 3) A Transformer encoder to capture global relationships (Section 3.3),
# 4) A final MLP classifier.
###############################################################################
# Custom model registration
class SAGViTConfig(PretrainedConfig):
model_type = "sagvit"
def __init__(self, **kwargs):
super().__init__(**kwargs)
self.d_model = kwargs.get("d_model", 64)
self.dim_feedforward = kwargs.get("dim_feedforward", 64)
self.gcn_hidden = kwargs.get("gcn_hidden", 128)
self.gcn_out = kwargs.get("gcn_out", 64)
self.hidden_mlp_features = kwargs.get("hidden_mlp_features", 64)
self.in_channels = kwargs.get("in_channels", 2560)
self.nhead = kwargs.get("nhead", 4)
self.num_classes = kwargs.get("num_classes", 10)
self.num_layers = kwargs.get("num_layers", 2)
self.patch_size = kwargs.get("patch_size", (4, 4))
class SAGViTClassifier(PreTrainedModel):
"""
SAG-ViT: Scale-Aware Graph Attention Vision Transformer
This model integrates the following steps:
- Extract multi-scale features from images using a CNN backbone (EfficientNetv2 here).
- Partition the feature map into patches and build a graph where each node is a patch.
- Use a Graph Attention Network (GAT) to refine patch embeddings based on local spatial relationships.
- Utilize a Transformer encoder to model long-range dependencies and integrate multi-scale information.
- Finally, classify the resulting representation into desired classes.
Inputs:
- x (Tensor): Input images (B, 3, H, W)
Outputs:
- out (Tensor): Classification logits (B, num_classes)
"""
config_class = SAGViTConfig
def __init__(self, config):
super().__init__(config)
self.patch_size = config.patch_size
self.num_classes = config.num_classes
# CNN feature extractor (frozen pre-trained EfficientNetv2)
self.cnn = EfficientNetV2FeatureExtractor()
# Graph Attention Network to process patch embeddings
self.gcn = GATGNN(
in_channels=config.in_channels,
hidden_channels=config.gcn_hidden,
out_channels=config.gcn_out,
)
# Learnable positional embedding for Transformer input
self.positional_embedding = nn.Parameter(torch.randn(1, 1, config.d_model))
# Extra embedding token (similar to class token) to summarize global info
self.extra_embedding = nn.Parameter(torch.randn(1, config.d_model))
# Transformer encoder to capture long-range global dependencies
self.transformer_encoder = TransformerEncoder(
d_model=config.d_model,
nhead=config.nhead,
num_layers=config.num_layers,
dim_feedforward=config.dim_feedforward,
)
# MLP classification head
self.mlp = MLPBlock(config.d_model, config.hidden_mlp_features, config.num_classes)
def forward(self, x, **kwargs):
# Step 1: High-fidelity feature extraction from CNN
feature_map = self.cnn(x)
# Step 2: Build graphs from patches
G_global_batch, patches = build_graph_from_patches(feature_map, self.patch_size)
# Step 3: Convert to PyG Data format and batch
data_list = build_graph_data_from_patches(G_global_batch, patches)
device = x.device
batch = Batch.from_data_list(data_list).to(device)
# Step 4: GAT stage
x_gcn = self.gcn(batch)
# Step 5: Reshape GCN output back to (B, N, D)
# The number of patches per image is determined by patch size and feature map dimensions.
B = x.size(0)
D = x_gcn.size(-1)
# N is automatically inferred
# Thus x_gcn is (B, D) now. We need a sequence dimension for the Transformer.
# Let's treat each image-level embedding as one "patch token" plus an extra token:
patch_embeddings = x_gcn.unsqueeze(1) # (B, 1, D)
# Add positional embedding
patch_embeddings = patch_embeddings + self.positional_embedding # (B, 1, D)
# Add an extra learnable embedding (like a CLS token)
patch_embeddings = torch.cat([patch_embeddings, self.extra_embedding.unsqueeze(0).expand(B, -1, -1)], dim=1) # (B, 2, D)
# Step 6: Transformer encoder
x_trans = self.transformer_encoder(patch_embeddings)
# Step 7: Global pooling (here we just take the mean)
x_pooled = x_trans.mean(dim=1) # (B, D)
# Classification
logits = self.mlp(x_pooled)
return ModelOutput(logits=logits)
# Register custom model and config
CONFIG_MAPPING.register("sagvit", SAGViTConfig)
MODEL_MAPPING.register(SAGViTConfig, SAGViTClassifier)