app.py

import functools
import io
import json
import logging
import math
import os
import pathlib
import random

import beartype
import einops.layers.torch
import gradio as gr
import numpy as np
import open_clip
import requests
import saev.nn
import torch
from jaxtyping import Float, jaxtyped
from PIL import Image, ImageDraw
from torch import Tensor
from torchvision.transforms import v2

log_format = "[%(asctime)s] [%(levelname)s] [%(name)s] %(message)s"
logging.basicConfig(level=logging.INFO, format=log_format)
logger = logging.getLogger("app.py")


####################
# Global Constants #
####################

DEBUG = True
"""Whether we are debugging."""

n_sae_latents = 3
"""Number of SAE latents to show."""

n_sae_examples = 4
"""Number of SAE examples per latent to show."""

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
"""Hardware accelerator, if any."""

vit_ckpt = "ViT-B-16/openai"
"""CLIP checkpoint."""

n_patches_per_img: int = 196
"""Number of patches per image in vit_ckpt."""

max_frequency = 1e-1
"""Maximum frequency. Any feature that fires more than this is ignored."""

CWD = pathlib.Path(__file__).parent

r2_url = "https://pub-289086e849214430853bc87bd8964988.r2.dev/"


logger.info("Set global constants.")


###########
# Helpers #
###########


@beartype.beartype
def get_cache_dir() -> str:
    """
    Get cache directory from environment variables, defaulting to the current working directory (.)

    Returns:
        A path to a cache directory (might not exist yet).
    """
    cache_dir = ""
    for var in ("HF_HOME", "HF_HUB_CACHE"):
        cache_dir = cache_dir or os.environ.get(var, "")
    return cache_dir or "."


@beartype.beartype
def load_model(fpath: str | pathlib.Path, *, device: str = "cpu") -> torch.nn.Module:
    """
    Loads a linear layer from disk.
    """
    with open(fpath, "rb") as fd:
        kwargs = json.loads(fd.readline().decode())
        buffer = io.BytesIO(fd.read())

    model = torch.nn.Linear(**kwargs)
    state_dict = torch.load(buffer, weights_only=True, map_location=device)
    model.load_state_dict(state_dict)
    model = model.to(device)
    return model


@beartype.beartype
@functools.lru_cache(maxsize=512)
def get_dataset_img(i: int) -> Image.Image:
    return Image.open(requests.get(r2_url + image_fpaths[i], stream=True).raw)


@beartype.beartype
def make_img(
    img: Image.Image,
    patches: Float[Tensor, " n_patches"],
    *,
    upper: int | None = None,
) -> Image.Image:
    # Resize to 256x256 and crop to 224x224
    resize_size_px = (512, 512)
    resize_w_px, resize_h_px = resize_size_px
    crop_size_px = (448, 448)
    crop_w_px, crop_h_px = crop_size_px
    crop_coords_px = (
        (resize_w_px - crop_w_px) // 2,
        (resize_h_px - crop_h_px) // 2,
        (resize_w_px + crop_w_px) // 2,
        (resize_h_px + crop_h_px) // 2,
    )

    img = img.resize(resize_size_px).crop(crop_coords_px)
    img = add_highlights(img, patches.numpy(), upper=upper, opacity=0.5)
    return img


##########
# Models #
##########


@jaxtyped(typechecker=beartype.beartype)
class SplitClip(torch.nn.Module):
    def __init__(self, *, n_end_layers: int):
        super().__init__()

        if vit_ckpt.startswith("hf-hub:"):
            clip, _ = open_clip.create_model_from_pretrained(
                vit_ckpt, cache_dir=get_cache_dir()
            )
        else:
            arch, ckpt = vit_ckpt.split("/")
            clip, _ = open_clip.create_model_from_pretrained(
                arch, pretrained=ckpt, cache_dir=get_cache_dir()
            )
        model = clip.visual
        model.proj = None
        model.output_tokens = True  # type: ignore
        self.vit = model.eval()
        assert not isinstance(self.vit, open_clip.timm_model.TimmModel)

        self.n_end_layers = n_end_layers

    @staticmethod
    def _expand_token(token, batch_size: int):
        return token.view(1, 1, -1).expand(batch_size, -1, -1)

    def forward_start(self, x: Float[Tensor, "batch channels width height"]):
        x = self.vit.conv1(x)  # shape = [*, width, grid, grid]
        x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
        x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]

        # class embeddings and positional embeddings
        x = torch.cat(
            [self._expand_token(self.vit.class_embedding, x.shape[0]).to(x.dtype), x],
            dim=1,
        )
        # shape = [*, grid ** 2 + 1, width]
        x = x + self.vit.positional_embedding.to(x.dtype)

        x = self.vit.patch_dropout(x)
        x = self.vit.ln_pre(x)
        for r in self.vit.transformer.resblocks[: -self.n_end_layers]:
            x = r(x)
        return x

    def forward_end(self, x: Float[Tensor, "batch n_patches dim"]):
        for r in self.vit.transformer.resblocks[-self.n_end_layers :]:
            x = r(x)

        x = self.vit.ln_post(x)
        pooled, _ = self.vit._global_pool(x)
        if self.vit.proj is not None:
            pooled = pooled @ self.vit.proj

        return pooled


# ViT
split_vit = SplitClip(n_end_layers=1)
split_vit = split_vit.to(device)
logger.info("Initialized CLIP ViT.")

# Linear classifier
clf_ckpt_fpath = CWD / "ckpts" / "clf.pt"
clf = load_model(clf_ckpt_fpath)
clf = clf.to(device).eval()
logger.info("Loaded linear classifier.")

# SAE
sae_ckpt_fpath = CWD / "ckpts" / "sae.pt"
sae = saev.nn.load(sae_ckpt_fpath.as_posix())
sae.to(device).eval()
logger.info("Loaded SAE.")


############
# Datasets #
############

human_transform = v2.Compose([
    v2.Resize((512, 512), interpolation=v2.InterpolationMode.NEAREST),
    v2.CenterCrop((448, 448)),
    v2.ToImage(),
    einops.layers.torch.Rearrange("channels width height -> width height channels"),
])

arch, ckpt = vit_ckpt.split("/")
_, vit_transform = open_clip.create_model_from_pretrained(
    arch, pretrained=ckpt, cache_dir=get_cache_dir()
)


with open(CWD / "data" / "image_fpaths.json") as fd:
    image_fpaths = json.load(fd)


with open(CWD / "data" / "image_labels.json") as fd:
    image_labels = json.load(fd)


logger.info("Loaded all datasets.")

#############
# Variables #
#############


@beartype.beartype
def load_tensor(path: str | pathlib.Path) -> Tensor:
    return torch.load(path, weights_only=True, map_location="cpu")


top_img_i = load_tensor(CWD / "data" / "top_img_i.pt")
top_values = load_tensor(CWD / "data" / "top_values_uint8.pt")
sparsity = load_tensor(CWD / "data" / "sparsity.pt")


mask = torch.ones((sae.cfg.d_sae), dtype=bool)
mask = mask & (sparsity < max_frequency)


#############
# Inference #
#############


@beartype.beartype
def get_image(image_i: int) -> list[Image.Image | int]:
    image = get_dataset_img(image_i)
    image = human_transform(image)
    return [Image.fromarray(image.numpy()), image_labels[image_i]]


@beartype.beartype
def get_random_class_image(cls: int) -> Image.Image:
    indices = [i for i, tgt in enumerate(image_labels) if tgt == cls]
    i = random.choice(indices)

    image = get_dataset_img(i)
    image = human_transform(image)
    return Image.fromarray(image.numpy())


@torch.inference_mode
def get_sae_examples(
    image_i: int, patches: list[int]
) -> list[None | Image.Image | int]:
    """
    Given a particular cell, returns some highlighted images showing what feature fires most on this cell.
    """
    if not patches:
        return [None] * 12 + [-1] * 3

    logger.info("Getting SAE examples for patches %s.", patches)

    img = get_dataset_img(image_i)
    x = vit_transform(img)[None, ...].to(device)
    x_BPD = split_vit.forward_start(x)
    # Need to add 1 to account for [CLS] token.
    vit_acts_MD = x_BPD[0, [p + 1 for p in patches]].to(device)

    _, f_x_MS, _ = sae(vit_acts_MD)
    f_x_S = f_x_MS.sum(axis=0)

    latents = torch.argsort(f_x_S, descending=True).cpu()
    latents = latents[mask[latents]][:n_sae_latents].tolist()

    images = []
    for latent in latents:
        img_patch_pairs, seen_i_im = [], set()
        for i_im, values_p in zip(top_img_i[latent].tolist(), top_values[latent]):
            if i_im in seen_i_im:
                continue

            example_img = get_dataset_img(i_im)
            img_patch_pairs.append((example_img, values_p))
            seen_i_im.add(i_im)

        # How to scale values.
        upper = None
        if top_values[latent].numel() > 0:
            upper = top_values[latent].max().item()

        latent_images = [
            make_img(img, patches.to(float), upper=upper)
            for img, patches in img_patch_pairs[:n_sae_examples]
        ]

        while len(latent_images) < n_sae_examples:
            latent_images += [None]

        images.extend(latent_images)

    return images + latents


@torch.inference_mode
def get_pred_dist(i: int) -> dict[int, float]:
    img = get_dataset_img(i)
    x = vit_transform(img)[None, ...].to(device)
    x_BPD = split_vit.forward_start(x)
    x_BD = split_vit.forward_end(x_BPD)

    logits_BC = clf(x_BD)

    probs = torch.nn.functional.softmax(logits_BC[0], dim=0).cpu().tolist()
    return {i: prob for i, prob in enumerate(probs)}


@torch.inference_mode
def get_modified_dist(
    image_i: int,
    patches: list[int],
    latent1: int,
    latent2: int,
    latent3: int,
    value1: float,
    value2: float,
    value3: float,
) -> dict[int, float]:
    img = get_dataset_img(image_i)
    x = vit_transform(img)[None, ...].to(device)
    x_BPD = split_vit.forward_start(x)

    cls_B1D, x_BPD = x_BPD[:, :1, :], x_BPD[:, 1:, :]

    x_hat_BPD, f_x_BPS, _ = sae(x_BPD)
    err_BPD = x_BPD - x_hat_BPD

    values = torch.tensor(
        [
            unscaled(value, top_values[latent].max().item())
            for value, latent in [
                (value1, latent1),
                (value2, latent2),
                (value3, latent3),
            ]
        ],
        device=device,
    )
    patches = torch.tensor(patches, device=device)
    latents = torch.tensor([latent1, latent2, latent3], device=device)
    f_x_BPS[:, patches[:, None], latents[None, :]] = values

    # Reproduce the SAE forward pass after f_x
    modified_x_hat_BPD = (
        einops.einsum(
            f_x_BPS,
            sae.W_dec,
            "batch patches d_sae, d_sae d_vit -> batch patches d_vit",
        )
        + sae.b_dec
    )

    modified_BPD = torch.cat([cls_B1D, err_BPD + modified_x_hat_BPD], axis=1)

    modified_BD = split_vit.forward_end(modified_BPD)
    logits_BC = clf(modified_BD)

    probs = torch.nn.functional.softmax(logits_BC[0], dim=0).cpu().tolist()
    return {i: prob for i, prob in enumerate(probs)}


@beartype.beartype
def unscaled(x: float | int, max_obs: float | int) -> float:
    """Scale from [-20, 20] to [20 * -max_obs, 20 * max_obs]."""
    return map_range(x, (-20.0, 20.0), (-20.0 * max_obs, 20.0 * max_obs))


@beartype.beartype
def map_range(
    x: float | int,
    domain: tuple[float | int, float | int],
    range: tuple[float | int, float | int],
):
    a, b = domain
    c, d = range
    if not (a <= x <= b):
        raise ValueError(f"x={x:.3f} must be in {[a, b]}.")
    return c + (x - a) * (d - c) / (b - a)


@jaxtyped(typechecker=beartype.beartype)
def add_highlights(
    img: Image.Image,
    patches: Float[np.ndarray, " n_patches"],
    *,
    upper: int | None = None,
    opacity: float = 0.9,
) -> Image.Image:
    if not len(patches):
        return img

    iw_np, ih_np = int(math.sqrt(len(patches))), int(math.sqrt(len(patches)))
    iw_px, ih_px = img.size
    pw_px, ph_px = iw_px // iw_np, ih_px // ih_np
    assert iw_np * ih_np == len(patches)

    # Create a transparent overlay
    overlay = Image.new("RGBA", img.size, (0, 0, 0, 0))
    draw = ImageDraw.Draw(overlay)

    # Using semi-transparent red (255, 0, 0, alpha)
    for p, val in enumerate(patches):
        assert upper is not None
        val /= upper + 1e-9
        x_np, y_np = p % iw_np, p // ih_np
        draw.rectangle(
            [
                (x_np * pw_px, y_np * ph_px),
                (x_np * pw_px + pw_px, y_np * ph_px + ph_px),
            ],
            fill=(int(val * 256), 0, 0, int(opacity * val * 256)),
        )

    # Composite the original image and the overlay
    return Image.alpha_composite(img.convert("RGBA"), overlay)


#############
# Interface #
#############


with gr.Blocks() as demo:
    image_number = gr.Number(label="Test Example", precision=0)
    class_number = gr.Number(label="Test Class", precision=0)
    input_image = gr.Image(label="Input Image")
    get_input_image_btn = gr.Button(value="Get Input Image")
    get_input_image_btn.click(
        get_image,
        inputs=[image_number],
        outputs=[input_image, class_number],
        api_name="get-image",
    )
    get_random_class_image_btn = gr.Button(value="Get Random Class Image")
    get_input_image_btn.click(
        get_random_class_image,
        inputs=[image_number],
        outputs=[input_image],
        api_name="get-random-class-image",
    )

    patch_numbers = gr.CheckboxGroup(
        label="Image Patch", choices=list(range(n_patches_per_img))
    )
    top_latent_numbers = gr.CheckboxGroup(label="Top Latents")
    top_latent_numbers = [
        gr.Number(label=f"Top Latents #{j + 1}", precision=0)
        for j in range(n_sae_latents)
    ]
    sae_example_images = [
        gr.Image(label=f"Latent #{j}, Example #{i + 1}")
        for i in range(n_sae_examples)
        for j in range(n_sae_latents)
    ]
    get_sae_examples_btn = gr.Button(value="Get SAE Examples")
    get_sae_examples_btn.click(
        get_sae_examples,
        inputs=[image_number, patch_numbers],
        outputs=sae_example_images + top_latent_numbers,
        api_name="get-sae-examples",
        concurrency_limit=16,
    )

    pred_dist = gr.Label(label="Pred. Dist.")
    get_pred_dist_btn = gr.Button(value="Get Pred. Distribution")
    get_pred_dist_btn.click(
        get_pred_dist,
        inputs=[image_number],
        outputs=[pred_dist],
        api_name="get-preds",
    )

    latent_numbers = [gr.Number(label=f"Latent {i + 1}", precision=0) for i in range(3)]
    value_sliders = [
        gr.Slider(label=f"Value {i + 1}", minimum=-10, maximum=10) for i in range(3)
    ]
    get_modified_dist_btn = gr.Button(value="Get Modified Label")
    get_modified_dist_btn.click(
        get_modified_dist,
        inputs=[image_number, patch_numbers] + latent_numbers + value_sliders,
        outputs=[pred_dist],
        api_name="get-modified",
        concurrency_limit=16,
    )


if __name__ == "__main__":
    demo.launch()