StyleSquirrel / model.py
Food Desert
Initial Space
2cd64de
"""Runtime model for the Style Tagger Space (nearest-neighbor over training set).
- Embeds query image with CLIP (use_fast=True for consistency).
- Projects with your trained StyleProjector.
- Looks up nearest neighbors in a retrieval bundle (FAISS if available).
- Tallies style tags from neighbors and returns {tag: score in [0,1]}.
Expected files next to app.py:
style_projector_and_topic_table.safetensors
style_topic_meta.json
style_vocab.json
retrieval_bundle/
├─ vectors.faiss (optional but recommended)
├─ tag_ids_concat.npy
├─ tag_offsets.npy
└─ image_ids.npy (optional, for debugging)
Notes:
- If FAISS index is missing but you also ship vectors.npy (optional), we fall back to NumPy top‑K.
- If neither is present, we return a stable dummy output so UI stays responsive.
"""
from __future__ import annotations
import os, json
from typing import Dict, List, Optional, Tuple
import numpy as np
from PIL import Image
import torch
import torch.nn as nn
from safetensors.torch import load_file as load_safetensors
from transformers import CLIPModel, CLIPProcessor
# Try FAISS (optional at runtime)
try:
import faiss # type: ignore
except Exception:
faiss = None
# -------------------
# Globals
# -------------------
_DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
_PROJECTOR: Optional[nn.Module] = None
_CLIP_MODEL: Optional[CLIPModel] = None
_CLIP_PROCESSOR: Optional[CLIPProcessor] = None
_VOCAB: Optional[List[str]] = None
# Retrieval bundle
_INDEX: Optional[object] = None # faiss index if available
_VEC_NP: Optional[np.ndarray] = None # fallback: raw vectors (optional file vectors.npy)
_TAGS_CONCAT: Optional[np.ndarray] = None
_TAGS_OFFSETS: Optional[np.ndarray] = None
_IMAGE_IDS: Optional[np.ndarray] = None # filenames aligned to index rows
_READY = False
# How many neighbors to tally by default
K_DEFAULT = int(os.getenv("STYLE_K", "30"))
# -------------------
# Model definition (matches training projector)
# -------------------
class StyleProjector(nn.Module):
def __init__(self, d_in: int, d_out: int, use_layer_norm: bool = True):
super().__init__()
blocks = [nn.Linear(d_in, 1024), nn.GELU()]
if use_layer_norm:
blocks.append(nn.LayerNorm(1024))
blocks += [nn.Dropout(0.0), nn.Linear(1024, d_out)]
self.net = nn.Sequential(*blocks)
def forward(self, x: torch.Tensor) -> torch.Tensor:
z = self.net(x)
return nn.functional.normalize(z, dim=-1)
# -------------------
# Helpers
# -------------------
def _embed_image(image: Image.Image) -> torch.Tensor:
"""CLIP-embed the (optionally cropped) image and L2-normalize to unit length."""
# Crop 1/8 top & bottom to match preprocessing, controlled via env var
if os.getenv("STYLE_CROP_TB", "1") not in ("0", "false", "False"):
w, h = image.size
dy = h // 8
if dy > 0:
image = image.crop((0, dy, w, h - dy))
inputs = _CLIP_PROCESSOR(images=[image.convert("RGB")], return_tensors="pt").to(_DEVICE)
with torch.no_grad():
feats = _CLIP_MODEL.get_image_features(**inputs)
feats = feats / feats.norm(dim=-1, keepdim=True)
return feats # [1, d_in]
def _load_retrieval_bundle(bundle_dir: str) -> None:
global _INDEX, _VEC_NP, _TAGS_CONCAT, _TAGS_OFFSETS, _IMAGE_IDS
# Required tag tables
_TAGS_CONCAT = np.load(os.path.join(bundle_dir, "tag_ids_concat.npy"))
_TAGS_OFFSETS = np.load(os.path.join(bundle_dir, "tag_offsets.npy"))
# Filenames (optional but recommended for diagnostics)
img_ids_path = os.path.join(bundle_dir, "image_ids.npy")
_IMAGE_IDS = np.load(img_ids_path) if os.path.exists(img_ids_path) else None
# Preferred: FAISS index
index_path = os.path.join(bundle_dir, "vectors.faiss")
if faiss is not None and os.path.exists(index_path):
_INDEX = faiss.read_index(index_path)
else:
_INDEX = None
# Optional fallback: raw vectors file if you ship it
vec_path = os.path.join(bundle_dir, "vectors.npy")
if os.path.exists(vec_path):
_VEC_NP = np.load(vec_path).astype(np.float32) # expected L2-normalized
else:
_VEC_NP = None
# -------------------
# Public API
# -------------------
def load():
"""Load CLIP, projector weights, vocab, and retrieval bundle if present.
This function is tolerant: it will not raise if files are missing; _READY will stay False.
"""
global _PROJECTOR, _CLIP_MODEL, _CLIP_PROCESSOR, _VOCAB, _READY
try:
# ---- Metadata ----
with open("style_topic_meta.json", "r", encoding="utf-8") as f:
meta = json.load(f)
proj_dim = int(meta["proj_dim"])
use_ln = bool(meta.get("use_layer_norm", True))
# ---- CLIP (use_fast=True for consistency; requires torchvision) ----
model_id = os.getenv("CLIP_MODEL_ID", "openai/clip-vit-base-patch32")
_CLIP_MODEL = CLIPModel.from_pretrained(model_id).to(_DEVICE)
_CLIP_PROCESSOR = CLIPProcessor.from_pretrained(model_id, use_fast=True)
_CLIP_MODEL.eval()
# ---- Projector ----
d_in = int(_CLIP_MODEL.config.projection_dim)
_PROJECTOR = StyleProjector(d_in, proj_dim, use_layer_norm=use_ln).to(_DEVICE)
tensors = load_safetensors("style_projector_and_topic_table.safetensors")
with torch.no_grad():
_PROJECTOR.net[0].weight.copy_(tensors["projector.net.0.weight"])
_PROJECTOR.net[0].bias.copy_(tensors["projector.net.0.bias"])
if use_ln:
_PROJECTOR.net[2].weight.copy_(tensors["projector.net.ln.weight"])
_PROJECTOR.net[2].bias.copy_(tensors["projector.net.ln.bias"])
last = _PROJECTOR.net[4]
else:
last = _PROJECTOR.net[3]
last.weight.copy_(tensors["projector.net.last.weight"])
last.bias.copy_(tensors["projector.net.last.bias"])
_PROJECTOR.eval()
# ---- Vocab ----
with open("style_vocab.json", "r", encoding="utf-8") as f:
_VOCAB = json.load(f)
# ---- Retrieval bundle (optional but recommended) ----
bundle_dir = os.getenv("RETRIEVAL_DIR", os.path.join(os.getcwd(), "retrieval_bundle"))
if os.path.isdir(bundle_dir):
_load_retrieval_bundle(bundle_dir)
_READY = True
except FileNotFoundError as e:
print("model.load(): missing file (skeleton mode)", e)
_READY = False
def predict(image: Image.Image, k: Optional[int] = None) -> Tuple[Dict[str, float], List[Dict[str, object]], Dict[str, int]]:
"""Return (scores_norm, neighbors_detailed, counts_raw)
- scores_norm: {style_tag: score in [0,1]} (counts normalized by max count)
- neighbors_detailed: [{"filename": str, "similarity": float, "distance": float, "styles": [str, ...]}, ...]
- counts_raw: {style_tag: int} (exact tallies used to rank)
"""
global _READY, _INDEX, _VEC_NP, _TAGS_CONCAT, _TAGS_OFFSETS, _VOCAB, _IMAGE_IDS
if not _READY or _VOCAB is None:
fallback = _VOCAB[:10] if _VOCAB else ["watercolor","oil painting","pixel art","sketch","digital painting"]
return ({tag: 0.0 for tag in fallback}, [], {tag: 0 for tag in fallback})
# 1) Embed + project query
with torch.no_grad():
q_clip = _embed_image(image) # [1, d_in]
q_proj = _PROJECTOR(q_clip).cpu().numpy()[0] # [d_proj], L2-normed
# 2) Nearest neighbors (and similarities)
K = k or K_DEFAULT
if _INDEX is not None and faiss is not None:
q = q_proj[np.newaxis, :].astype(np.float32)
D, I = _INDEX.search(q, K) # D: sims, I: indices
nbrs = I[0]
sims = D[0]
elif _VEC_NP is not None:
sims_all = _VEC_NP @ q_proj # cosine via dot, [N]
if K < sims_all.shape[0]:
idx = np.argpartition(-sims_all, K)[:K]
order = np.argsort(-sims_all[idx])
nbrs = idx[order]
sims = sims_all[nbrs]
else:
order = np.argsort(-sims_all)
nbrs = order[:K]
sims = sims_all[nbrs]
else:
# No retrieval table present
fallback = {tag: 0.0 for tag in _VOCAB[:10]}
return (fallback, [], {k: 0 for k in fallback})
# Build neighbor detail list with filenames, sims, distances, and neighbor styles
neighbors: List[Dict[str, object]] = []
for i, s in zip(nbrs, sims):
# neighbor styles from ragged arrays
start, end = int(_TAGS_OFFSETS[i]), int(_TAGS_OFFSETS[i + 1])
tag_ids_i = _TAGS_CONCAT[start:end]
styles_i = [_VOCAB[int(tid)] for tid in tag_ids_i] if len(tag_ids_i) > 0 else []
# filename (if available)
fname = str(_IMAGE_IDS[i]) if _IMAGE_IDS is not None else str(int(i))
neighbors.append({
"filename": fname,
"similarity": float(s),
"distance": float(1.0 - s), # cosine distance since vectors are L2-normalized
"styles": styles_i,
})
# 3) Tally neighbor style tags → counts_raw
if len(nbrs) == 0:
fallback = {tag: 0.0 for tag in _VOCAB[:10]}
return (fallback, neighbors, {k: 0 for k in fallback})
# concatenate tags across neighbors and bincount
tag_ids_all = np.concatenate([
_TAGS_CONCAT[int(_TAGS_OFFSETS[i]):int(_TAGS_OFFSETS[i + 1])] for i in nbrs
]) if len(nbrs) else np.array([], dtype=np.int32)
counts = np.bincount(tag_ids_all, minlength=len(_VOCAB)) if tag_ids_all.size else np.zeros(len(_VOCAB), dtype=np.int64)
counts_raw: Dict[str, int] = { _VOCAB[i]: int(counts[i]) for i in np.nonzero(counts)[0] }
# 4) Normalize counts → scores_norm in [0,1]
maxc = float(counts.max()) if counts.size else 0.0
if maxc <= 0:
return ({tag: 0.0 for tag in _VOCAB[:10]}, neighbors, counts_raw)
scores = counts / maxc
scores_norm: Dict[str, float] = { _VOCAB[i]: float(scores[i]) for i in np.nonzero(counts)[0] }
return (scores_norm, neighbors, counts_raw)