shared.py

import os
import subprocess
import re
import random
import matplotlib.pyplot as plt
import json
def get_gpu_memory_usage():
    """Returns a list of GPU memory usage in MB."""
    try:
        # Run nvidia-smi command and capture the output
        result = subprocess.run(['nvidia-smi', '--query-gpu=memory.used', '--format=csv,nounits,noheader'],
                                stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)

        # Check if the command was successful
        if result.returncode != 0:
            raise RuntimeError(f"nvidia-smi command failed with error: {result.stderr}")

        # Parse the output to get memory usage
        memory_usages = [int(x) for x in result.stdout.strip().split('\n')]
        return memory_usages
    except Exception as e:
        print(f"Error querying GPU memory usage: {e}")
        return []

def set_cuda_visible_device():
    """Sets the CUDA_VISIBLE_DEVICES environment variable to the GPU with the smallest memory usage."""
    memory_usages = get_gpu_memory_usage()

    if not memory_usages:
        print("No GPU memory usage data available.")
        return

    # Find the index of the GPU with the smallest memory usage
    min_memory_index = memory_usages.index(min(memory_usages))

    # Set the CUDA_VISIBLE_DEVICES environment variable
    os.environ["CUDA_VISIBLE_DEVICES"] = str(min_memory_index)
    print(f"Set CUDA_VISIBLE_DEVICES to GPU {min_memory_index} with {memory_usages[min_memory_index]} MB used.")

    return str(min_memory_index)

os.environ["ASN_ROOT_DIR"] = "/home/nickj/asn/second_order_lens"
os.chdir(os.environ["ASN_ROOT_DIR"])

import numpy as np
import torch
from PIL import Image
import os.path
import argparse
from pathlib import Path

from tqdm import tqdm
from utils.factory import create_model_and_transforms, get_tokenizer
from PIL import Image, ImageDraw

def get_model(model_name = "ViT-B/16", pretrained = "openai", device = "cuda:0"):
  torch.multiprocessing.set_sharing_strategy("file_system")
  model, _, preprocess = create_model_and_transforms(
      model_name, pretrained=pretrained, force_quick_gelu=True,
  )
  model.to(device)
  model.eval()
  context_length = model.context_length
  vocab_size = model.vocab_size

  return {
    "model": model,
    "model_name": model_name,
    "pretrained": pretrained,
    "preprocess": preprocess,
    "context_length": context_length,
    "vocab_size": vocab_size
  }

img_path = "/datasets/ilsvrc_2024-01-04_1913/val/n04398044/ILSVRC2012_val_00042447.JPEG"
# img_path = "./sample.jpeg"
def load_images(preprocess, image_folder = "/datasets/ilsvrc/current/val", count = 100, images_only = True):
  file_list = []

  for root, dirs, files in os.walk(image_folder):
      for file in files:
          file_list.append(os.path.join(root, file))

  if count > len(file_list):
    sampled_files = file_list
  else:
    sampled_files = random.sample(file_list, count)

  image_files = []

  for filename in sampled_files:
    image_files.append(preprocess(Image.open(filename)))
  if images_only:
    return image_files
  else:
    return image_files, sampled_files

def calc_neuron_potentials(model, attn_layers = (1, 2), include_layernorm = True):
  # Calculates the attention-shifting potential scores for every neuron to the attention heads defined by the given layers (relative to the MLP layer)

  embed_dim = model.visual.transformer.resblocks[0].attn.out_proj.in_features
  num_heads = model.visual.transformer.resblocks[0].attn.num_heads
  head_dim = embed_dim // num_heads
  layers = len(model.visual.transformer.resblocks)

  results = dict()

  for neuron_layer in tqdm(range(layers), desc = "Calculating attention shifting potentials"):
    neuron_projection = model.visual.transformer.resblocks[neuron_layer].state_dict()["mlp.c_proj.weight"]
    for l_attn in range(min(layers, neuron_layer + attn_layers[0]), min(layers, neuron_layer + attn_layers[1])):
      ln_vector = model.visual.transformer.resblocks[l_attn].ln_1.state_dict()["weight"]
      attn_matrix = model.visual.transformer.resblocks[l_attn].state_dict()["attn.in_proj_weight"]
      W_Q, W_K, W_V = (attn_matrix[:embed_dim].reshape(num_heads, head_dim, -1),
                       attn_matrix[embed_dim:2*embed_dim].reshape(num_heads, head_dim, -1),
                       attn_matrix[2*embed_dim:].reshape(num_heads, head_dim, -1))

      for head_idx in range(num_heads):
        W_Q_h, W_K_h = W_Q[head_idx], W_K[head_idx]
        effects = []
        for i in range(neuron_projection.shape[1]):
          if include_layernorm:
            neuron_attn_effect = torch.norm(W_Q_h.T @ W_K_h @ (neuron_projection[:, i] * ln_vector))
          else:
            neuron_attn_effect = torch.norm(W_Q_h.T @ W_K_h @ neuron_projection[:, i])
          effects.append(neuron_attn_effect)

        results[(neuron_layer, l_attn, head_idx)] = torch.tensor(effects)
  return results

def calc_top_asns(shift_potentials, top_k = 10, per = "layer", layers_away = 1):
  num_layers = max([key[1] for key in shift_potentials.keys()]) + 1 # the last layer has no ASNs by definition
  num_heads = max([key[2] for key in shift_potentials.keys()])

  top_asns = []
  for layer in range(num_layers - layers_away):
    if per == "layer":
      potentials = []
      for head_idx in range(num_heads):
        potentials.append(shift_potentials[(layer, layer + layers_away, head_idx)])
      potentials = torch.max(torch.stack(potentials, dim = 0), dim = 0).values
      _, sorted_indices = torch.sort(potentials, descending = True)
      top_asns.append(sorted_indices[:top_k].tolist())
    elif per == "head":
      top_layer_asns = []
      for head_idx in range(num_heads):
        _, sorted_indices = torch.sort(shift_potentials[(layer, layer + layers_away, head_idx)], descending = True)
        top_layer_asns.append(sorted_indices[:top_k].tolist())
      top_asns.append(top_layer_asns)
    else:
      raise ValueError(f"Invalid per value: {per}")
  return top_asns

def aggregate_attn_map(attn_map, layer, head):
  num_tokens = attn_map.shape[-1]
  assert (num_tokens - 1) ** 0.5 % 1 == 0, "num_tokens - 1 is not a perfect square"

  num_patches = int((num_tokens - 1) ** 0.5)
  aggregate_scores = torch.sum(attn_map[:, layer, head, 1:, 1:], dim = 1).reshape((1, num_patches, num_patches))
  return aggregate_scores

def attn_map_cls_token(attn_map, layer, head):
  # Gets the attention map for the CLS token
  num_tokens = attn_map.shape[-1]
  assert (num_tokens - 1) ** 0.5 % 1 == 0, "num_tokens - 1 is not a perfect square"

  num_patches = int((num_tokens - 1) ** 0.5)
  attn_map_reshaped = attn_map[:, layer, head, 0, 1:].reshape((1, num_patches, num_patches))
  return attn_map_reshaped

def visualize_attn_shift(attn_map1, attn_map2, image, display=True, out=None, min_diff=None, max_diff=None):
  import matplotlib.pyplot as plt
  import numpy as np

  # Subtract attn_map1 from attn_map2
  diff_map = attn_map2 - attn_map1

  # Convert the image to RGBA
  image = image.convert("RGBA")
  overlay = Image.new("RGBA", image.size, (0, 0, 0, 0))
  draw = ImageDraw.Draw(overlay)

  # Calculate the size of each attention block
  block_size_x = image.size[0] / diff_map.shape[0]
  block_size_y = image.size[1] / diff_map.shape[1]

  # Create a colormap
  cmap = plt.get_cmap('coolwarm_r')  # 'cool' colormap for lighter to darker

  # Get the min and max values for scaling the colormap
  if max_diff is None:
    max_diff = diff_map.max()
  if min_diff is None:
    min_diff = diff_map.min()

  for i in range(diff_map.shape[0]):
    for j in range(diff_map.shape[1]):
      # Get the color from the colormap
      intensity = diff_map[i, j]
      normalized_intensity = (intensity - min_diff) / (max_diff - min_diff)  # Scale to [0, 1]
      rgba_color = cmap(1 - normalized_intensity)  # Invert the normalized intensity
      color = tuple(int(c * 255) for c in rgba_color[:3]) + (int(rgba_color[3] * 128),)

      # Draw the rectangle on the overlay with transparency
      draw.rectangle(
        [j * block_size_x, i * block_size_y, (j + 1) * block_size_x, (i + 1) * block_size_y],
        fill=color  # Add transparency to the color
      )

  # Composite the overlay with the original image
  combined = Image.alpha_composite(image, overlay)

  if display:
    # Display the result
    combined.show()

    # Show the color scale
    plt.figure(figsize=(6, 1))
    plt.imshow([np.linspace(min_diff, max_diff, 256)], cmap='coolwarm_r', aspect='auto')
    plt.gca().set_visible(False)
    plt.colorbar(orientation="horizontal")
    plt.show()

  if out is not None:
    combined.save(out)

  return combined

def visualize_attn_shift_binary(attn_map1, attn_map2, image, display=True, out=None):
  # Creates a visualization of the attention shift where green = positive, red = negative values.
  # This is useful when there are outliers in the difference map causing the middle values around 0 to be messed into one color
  # Subtract attn_map1 from attn_map2
  diff_map = attn_map2 - attn_map1

  # Normalize the difference map to range [0, 1] for visualization
  diff_map_normalized = (diff_map - diff_map.min()) / (diff_map.max() - diff_map.min())
  # Convert the image to RGBA
  image = image.convert("RGBA")
  overlay = Image.new("RGBA", image.size, (0, 0, 0, 0))
  draw = ImageDraw.Draw(overlay)

  # Calculate the size of each attention block
  block_size_x = image.size[0] / diff_map.shape[0]
  block_size_y = image.size[1] / diff_map.shape[1]

  for i in range(diff_map.shape[0]):
    for j in range(diff_map.shape[1]):
      # Calculate the color intensity based on the difference
      intensity = diff_map_normalized[i, j]
      alpha = int(255 * 0.5)  # Tone down the alpha to 50%
      if diff_map[i, j] > 0:
        color = (0, int(255 * intensity), 0, alpha)  # Green for positive
      else:
        color = (int(255 * (1 - intensity)), 0, 0, alpha)  # Red for negative

      # Draw the rectangle on the overlay
      draw.rectangle(
        [j * block_size_x, i * block_size_y, (j + 1) * block_size_x, (i + 1) * block_size_y],
        fill=color
      )

  # Composite the overlay with the original image
  combined = Image.alpha_composite(image, overlay)

  if display:
    # Display the result
    combined.show()

  if out is not None:
    combined.save(out)

  return combined

def is_outlier(mean, std, value):
  return value < mean - 2 * std or value > mean + 2 * std


def get_neuron_activations(images, prs_group, model, device = "cuda:0"):
  # Returns neuron activations in shape (num_images, num_layers, num_patches, num_neurons)
  random_neuron_acts = []
  for image in tqdm(images, desc="Processing images"):
    prs_group.reinit()
    image_input = image.unsqueeze(0).to(device)
    representation = model.encode_image(
      image_input, attn_method="head", normalize=False
    )
    prs_group.finalize()
    gelu_outs = prs_group.post_gelu_outputs()
    random_neuron_acts.append(gelu_outs)
  random_neuron_acts = torch.stack(random_neuron_acts, dim = 0)
  return random_neuron_acts

def normalize_array(arr):
  min_val = np.min(arr)
  max_val = np.max(arr)
  # Avoid division by zero if all values are the same
  if max_val - min_val == 0:
      return np.zeros_like(arr)
  normalized_arr = (arr - min_val) / (max_val - min_val)
  return normalized_arr

def np_l2(arr1, arr2):
  return np.linalg.norm(arr1 - arr2)

def best_class(classifier, representation):
  cs = torch.cosine_similarity(classifier, representation.permute(1, 0), dim = 0)
  return torch.argmax(cs).item(), cs[torch.argmax(cs).item()].item()

def load_group_attn_shifts(timestamp):
  # Load from Supp1B
  results_dir = "./results/supp1B"
  # dirs = [os.path.join(results_dir, d) for d in os.listdir(results_dir)
  #         if os.path.isdir(os.path.join(results_dir, d))]
  # latest_dir = max(dirs, key=os.path.getmtime)

  latest_dir = os.path.join(results_dir, timestamp)
  print(f"Using latest results directory: {latest_dir}")

  # Load metadata
  with open(os.path.join(latest_dir, "metadata.json"), "r") as f:
      metadata = json.load(f)

  # Load memory-mapped files
  attn_maps = np.memmap(os.path.join(latest_dir, "attention_maps.mmap"),
                      dtype=np.float32,
                      mode='r',
                      shape=tuple(metadata["attention_maps_shape"]))

  resblocks = np.memmap(os.path.join(latest_dir, "resblocks.mmap"),
                      dtype=np.float32,
                      mode='r',
                      shape=tuple(metadata["resblocks_shape"]))

  # Get file list from metadata
  file_list = metadata.get("file_list", [])

  # Get top_k values
  top_k_values = metadata.get("top_k_values", [0])

  return {
    "attn_maps": attn_maps,
    "resblocks": resblocks,
    "metadata": metadata,
    "file_list": file_list,
    "top_k_values": top_k_values,
    "num_layers": metadata.get("num_layers", 0),
    "num_images": metadata.get("num_images", 0),
    "num_heads": metadata.get("num_heads", 0)
  }

def load_individual_attn_shifts(timestamp, supp = "supp1D"):
  results_dir = f"./results/{supp}"
  # dirs = [os.path.join(results_dir, d) for d in os.listdir(results_dir)
  #         if os.path.isdir(os.path.join(results_dir, d))]
  # latest_dir = max(dirs, key=os.path.getmtime)

  latest_dir = os.path.join(results_dir, timestamp)
  print(f"Using latest results directory: {latest_dir}")

  # Load metadata
  with open(os.path.join(latest_dir, "metadata.json"), "r") as f:
      metadata = json.load(f)

  # Load memory-mapped files
  attn_maps = np.memmap(os.path.join(latest_dir, "attention_maps.mmap"),
                      dtype=np.float32,
                      mode='r',
                      shape=tuple(metadata["attention_maps_shape"]))

  baseline_attn_maps = np.memmap(os.path.join(latest_dir, "baseline_attention_maps.mmap"),
                      dtype=np.float32,
                      mode='r',
                      shape=tuple(metadata["baseline_attention_maps_shape"]))

  neuron_activations = np.memmap(os.path.join(latest_dir, "neuron_activations.mmap"),
                      dtype=np.float32,
                      mode='r',
                      shape=tuple(metadata["neuron_activations_shape"]))

  baseline_neuron_activations = np.memmap(os.path.join(latest_dir, "baseline_neuron_activations.mmap"),
                      dtype=np.float32,
                      mode='r',
                      shape=tuple(metadata["baseline_neuron_activations_shape"]))

  ablated_neurons = np.memmap(os.path.join(latest_dir, "ablated_neurons.mmap"),
                      dtype=np.float32,
                      mode='r',
                      shape=tuple(metadata["ablated_neurons_shape"]))

  # Get file list from metadata
  file_list = metadata.get("file_list", [])

  # Get k value
  k = metadata.get("k", 25)

  return {
    "attn_maps": attn_maps,
    "baseline_attn_maps": baseline_attn_maps,
    "neuron_activations": neuron_activations,
    "baseline_neuron_activations": baseline_neuron_activations,
    "ablated_neurons": ablated_neurons,
    "metadata": metadata,
    "file_list": file_list,
    "k": k,
    "num_layers": metadata.get("num_layers", 12),
    "num_images": metadata.get("num_images", 100),
    "model_name": metadata.get("model_name", "ViT-B-16"),
    "pretrained": metadata.get("pretrained", "openai")
  }

def find_register_neurons_cuda(model, preprocess, prs_group, register_norm_threshold = 30, highest_layer = -1, device = "cuda:0", processed_image_cnt = 500):
    num_layers = len(model.visual.transformer.resblocks)
    highest_layer = num_layers - 1 if highest_layer == -1 else highest_layer
    num_neurons = model.visual.transformer.resblocks[0].mlp.state_dict()["c_proj.weight"].shape[1]
    random_images = load_images(preprocess, count=processed_image_cnt)
    neuron_scores = torch.zeros((len(random_images), num_layers, num_neurons), device=device)
    alignment_scores = torch.zeros((len(random_images), num_layers, num_neurons), device=device)
    image_count = 0

    for i in tqdm(range(len(random_images)), desc="Processing random images"):
        image = random_images[i].unsqueeze(0).to(device)
        prs_group.reinit()

        with torch.inference_mode():
            representation = model.encode_image(
                image, attn_method="head", normalize=False
            )
        prs_group.finalize()

        baseline_neuron_acts = prs_group.post_gelu_outputs().to(device)
        baseline_resblock_outputs = prs_group.resblock_outputs().to(device)

        # Calculate norm map using torch
        norm_map = torch.norm(baseline_resblock_outputs[-1], dim=1)
        filtered_norms = norm_map.clone()
        filtered_norms[filtered_norms < register_norm_threshold] = 0

        # Get register locations as a tensor
        register_locations = torch.where(filtered_norms > register_norm_threshold)[0]

        if len(register_locations) == 0:
            continue

        image_count += 1

        # Process all layers vectorized
        for layer in range(num_layers):
            # Get absolute activations for all neurons in this layer
            act_layer = torch.abs(baseline_neuron_acts[layer])  # Shape: [seq_len, num_neurons]

            # Check sparsity condition for all neurons at once
            sparse_neurons = torch.sum(act_layer < 0.5, dim=0) >= 0.5 * act_layer.shape[0]  # Shape: [num_neurons]

            # Skip computation if no neurons meet the condition
            if not torch.any(sparse_neurons):
                continue

            # Get values at register locations for all neurons simultaneously
            # This creates a tensor of shape [num_register_locations, num_neurons]
            register_values = act_layer[register_locations]

            # For neurons that pass sparsity condition, compute mean at register locations
            # First, compute mean for all neurons (this is fast)
            neuron_means = register_values.mean(dim=0)  # Shape: [num_neurons]

            # Then zero out means for neurons that don't pass sparsity condition
            neuron_means = neuron_means * sparse_neurons.float()

            # Store in score tensor
            neuron_scores[i, layer] = neuron_means

    # Rest of the code remains the same
    mean_neuron_scores = neuron_scores[:image_count].mean(dim=0)
    mean_alignment_scores = alignment_scores[:image_count].mean(dim=0)

    # Flatten and find top values
    flattened_scores = mean_neuron_scores.flatten()
    sorted_values, sorted_indices = torch.sort(flattened_scores, descending=True)

    flattened_alignment = mean_alignment_scores.flatten()
    sorted_alignment_values, sorted_alignment_indices = torch.sort(flattened_alignment, descending=True)

    # Convert indices to layer/neuron pairs
    top_indices = [(idx.item() // num_neurons, idx.item() % num_neurons) for idx in sorted_indices]
    top_alignment_indices = [(idx.item() // num_neurons, idx.item() % num_neurons) for idx in sorted_alignment_indices]

    register_norms = [
        (layer, neuron, sorted_values[i].item())
        for i, (layer, neuron) in enumerate(top_indices)
        if layer <= highest_layer
    ]

    best_alignment_scores = [
        (layer, neuron, sorted_alignment_values[i].item())
        for i, (layer, neuron) in enumerate(top_alignment_indices)
        if layer <= highest_layer
    ]

    return register_norms, best_alignment_scores

def find_register_neurons(model, preprocess, prs_group, register_norm_threshold = 30, highest_layer = -1, device = "cuda:0", processed_image_cnt = 500):
  num_layers = len(model.visual.transformer.resblocks)
  highest_layer = num_layers - 1 if highest_layer == -1 else highest_layer
  num_neurons = model.visual.transformer.resblocks[0].mlp.state_dict()["c_proj.weight"].shape[1]

  random_images = load_images(preprocess, count = processed_image_cnt)
  neuron_scores = torch.zeros((len(random_images), num_layers, num_neurons))
  for i in tqdm(range(len(random_images)), desc="Processing random images"):
    image = random_images[i].unsqueeze(0).to(device)

    prs_group.reinit()
    with torch.no_grad():
      representation = model.encode_image(
        image, attn_method="head", normalize=False
      )
    prs_group.finalize()

    # Gather neuron activations and resblock outputs
    baseline_neuron_acts = prs_group.post_gelu_outputs().cpu().numpy()
    baseline_resblock_outputs = prs_group.resblock_outputs().cpu().numpy()

    # Calculate norms of the last resblock outputs. Only consider patches of the activation maps that correspond with registers
    norms = np.linalg.norm(baseline_resblock_outputs[-1], axis=1)
    norms[norms < register_norm_threshold] = 0
    register_locations = np.where(norms > register_norm_threshold)[0]

    # register_neurons = []
    for layer in range(num_layers):
      for neuron in range(num_neurons):
        neuron_map = baseline_neuron_acts[layer, :, neuron]
        mask = np.zeros_like(neuron_map, dtype=bool)
        mask[register_locations] = True
        neuron_map[~mask] = 0
        if np.any(neuron_map < 0):
          continue
        # dist = np.linalg.norm(normalize_array(norms) - normalize_array(neuron_map))
        # register_neurons.append((layer, neuron, dist.item(), neuron_map[register_locations].mean()))

        neuron_scores[i, layer, neuron] = torch.tensor(neuron_map[register_locations].mean())
  mean_neuron_scores = neuron_scores.mean(dim=0)
  # Flatten the 2D tensor to find global top values
  flattened_scores = mean_neuron_scores.flatten()
  sorted_values, sorted_indices = torch.sort(flattened_scores, descending=True)

  # Convert flat indices back to 2D coordinates (layer, neuron)
  top_indices = [(idx.item() // num_neurons, idx.item() % num_neurons) for idx in sorted_indices]

  return [(layer, neuron, sorted_values[i].item()) for i, (layer, neuron) in enumerate(top_indices) if layer <= highest_layer]


def plot_attn_maps(attn_maps, image_idx):

  num_layers, num_heads, patch_height, patch_width = attn_maps.shape
  print(f"Shape of image_shifts: {attn_maps.shape}")

  # Create a grid of plots for all layers and heads
  fig, axes = plt.subplots(num_layers, num_heads, figsize=(2*num_heads, 2*num_layers))
  fig.suptitle(f'Attention Shift Maps for Image #{image_idx}', fontsize=16)

  # Import the correct module for make_axes_locatable
  from mpl_toolkits.axes_grid1 import make_axes_locatable

  # Plot each layer-head combination
  for layer in range(num_layers):
      # Determine min and max for this layer for consistent colorbar scaling within the layer
      layer_vmin = attn_maps[layer].min().item()
      layer_vmax = attn_maps[layer].max().item()

      for head in range(num_heads):
          # Get the current axis (handle both 2D and 1D cases)
          if num_layers == 1 and num_heads == 1:
              ax = axes
          elif num_layers == 1:
              ax = axes[head]
          elif num_heads == 1:
              ax = axes[layer]
          else:
              ax = axes[layer, head]

          # Plot the attention shift map with layer-specific normalization
          im = ax.imshow(attn_maps[layer, head], cmap='viridis', vmin=layer_vmin, vmax=layer_vmax)

          # Remove ticks for cleaner appearance
          ax.set_xticks([])
          ax.set_yticks([])

          # Add layer and head labels only on the edges
          if head == 0:
              ax.set_ylabel(f'Layer {layer}')
          if layer == num_layers-1:
              ax.set_xlabel(f'Head {head}')

          # Add a colorbar for each layer (only once per row)
          if head == num_heads-1:
              # Create a colorbar that's properly sized relative to the plot
              divider = make_axes_locatable(ax)
              cax = divider.append_axes("right", size="5%", pad=0.05)
              plt.colorbar(im, cax=cax)

  # Adjust layout to make room for the colorbars
  plt.tight_layout()
  return plt

def calculate_iou(output, target):
  intersection = output * (output == target)
  area_inter = intersection.sum().item()
  area_pred = output.sum().item()
  area_target = target.sum().item()
  union = area_pred + area_target - area_inter
  iou = area_inter / union
  return area_inter, union, iou

def calculate_pixel_accuracy(output, target):
  correct = output * (output == target)
  correct = correct.sum().item()
  total = target.sum().item()
  return correct, total, correct / total