Spaces:
Build error
Build error
File size: 20,942 Bytes
46fcc2f 247dc37 46fcc2f 247dc37 46fcc2f 247dc37 5124a31 247dc37 41ed540 247dc37 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 |
import numpy as np
import hdbscan
from minisom import MiniSom
import pickle
from collections import Counter
import matplotlib.pyplot as plt
import phate
import imageio
from tqdm import tqdm
import io
import plotly.graph_objs as go
import plotly.subplots as sp
import umap
from sklearn.datasets import make_blobs
from sklearn.preprocessing import LabelEncoder
from sklearn.cluster import KMeans
from sklearn.semi_supervised import LabelSpreading
from moviepy.editor import *
class ClusterSOM:
def __init__(self):
self.hdbscan_model = None
self.som_models = {}
self.sigma_values = {}
self.mean_values = {}
self.cluster_mapping = {}
self.embedding = None
self.dim_red_op = None
def train(self, dataset, min_samples_per_cluster=100, n_clusters=None, som_size=(20, 20), sigma=1.0, learning_rate=0.5, num_iteration=200000, random_seed=42, n_neighbors=5, coverage=0.95):
"""
Train HDBSCAN and SOM models on the given dataset.
"""
# Train HDBSCAN model
print('Identifying clusters in the embedding ...')
self.hdbscan_model = hdbscan.HDBSCAN(min_cluster_size=min_samples_per_cluster)
self.hdbscan_model.fit(dataset)
# Calculate n_clusters if not provided
if n_clusters is None:
cluster_labels, counts = zip(*Counter(self.hdbscan_model.labels_).most_common())
cluster_labels = list(cluster_labels)
total_points = sum(counts)
covered_points = 0
n_clusters = 0
for count in counts:
covered_points += count
n_clusters += 1
if covered_points / total_points >= coverage:
break
# Train SOM models for the n_clusters most common clusters in the HDBSCAN model
cluster_labels, counts = zip(*Counter(self.hdbscan_model.labels_).most_common(n_clusters + 1))
cluster_labels = list(cluster_labels)
if -1 in cluster_labels:
cluster_labels.remove(-1)
else:
cluster_labels.pop()
for i, label in tqdm(enumerate(cluster_labels), total=len(cluster_labels), desc="Fitting 2D maps"):
if label == -1:
continue # Ignore noise
cluster_data = dataset[self.hdbscan_model.labels_ == label]
som = MiniSom(som_size[0], som_size[1], dataset.shape[1], sigma=sigma, learning_rate=learning_rate, random_seed=random_seed)
som.train_random(cluster_data, num_iteration)
self.som_models[i+1] = som
self.cluster_mapping[i+1] = label
# Compute sigma values
mean_cluster, sigma_cluster = self.compute_sigma_values(cluster_data, som_size, som, n_neighbors=n_neighbors)
self.sigma_values[i+1] = sigma_cluster
self.mean_values[i+1] = mean_cluster
def compute_sigma_values(self, cluster_data, som_size, som, n_neighbors=5):
som_weights = som.get_weights()
# Assign each datapoint to its nearest node
partitions = {idx: [] for idx in np.ndindex(som_size[0], som_size[1])}
for sample in cluster_data:
x, y = som.winner(sample)
partitions[(x, y)].append(sample)
# Compute the mean distance and std deviation of these partitions
mean_cluster = np.zeros(som_size)
sigma_cluster = np.zeros(som_size)
for idx in partitions:
if len(partitions[idx]) > 0:
partition_data = np.array(partitions[idx])
mean_distance = np.mean(np.linalg.norm(partition_data - som_weights[idx], axis=-1))
std_distance = np.std(np.linalg.norm(partition_data - som_weights[idx], axis=-1))
else:
mean_distance = 0
std_distance = 0
mean_cluster[idx] = mean_distance
sigma_cluster[idx] = std_distance
return mean_cluster, sigma_cluster
def train_label(self, labeled_data, labels):
"""
Train on labeled data to find centroids and compute distances to the labels.
"""
le = LabelEncoder()
encoded_labels = le.fit_transform(labels)
unique_labels = np.unique(encoded_labels)
# Use label spreading to propagate the labels
label_prop_model = LabelSpreading(kernel='knn', n_neighbors=5)
label_prop_model.fit(labeled_data, encoded_labels)
# Find the centroids for each label using KMeans
kmeans = KMeans(n_clusters=len(unique_labels), random_state=42)
kmeans.fit(labeled_data)
# Store the label centroids and label encodings
self.label_centroids = kmeans.cluster_centers_
self.label_encodings = le
def predict(self, data, sigma_factor=1.5):
"""
Predict the cluster and BMU SOM coordinate for each sample in the data if it's inside the sigma value.
Also, predict the label and distance to the center of the label if labels are trained.
"""
results = []
for sample in data:
min_distance = float('inf')
nearest_cluster_idx = None
nearest_node = None
for i, som in self.som_models.items():
x, y = som.winner(sample)
node = som.get_weights()[x, y]
distance = np.linalg.norm(sample - node)
if distance < min_distance:
min_distance = distance
nearest_cluster_idx = i
nearest_node = (x, y)
# Check if the nearest node is within the sigma value
if min_distance <= self.mean_values[nearest_cluster_idx][nearest_node] * 1.5: # * self.sigma_values[nearest_cluster_idx][nearest_node] * sigma_factor:
if hasattr(self, 'label_centroids'):
# Predict the label and distance to the center of the label
label_idx = self.label_encodings.inverse_transform([nearest_cluster_idx - 1])[0]
label_distance = np.linalg.norm(sample - self.label_centroids[label_idx])
results.append((nearest_cluster_idx, nearest_node, label_idx, label_distance))
else:
results.append((nearest_cluster_idx, nearest_node))
else:
results.append((-1, None)) # Noise
return results
def plot_embedding(self, new_data=None, dim_reduction='umap', interactive=False):
"""
Plot the dataset and SOM grids for each cluster.
If new_data is provided, it will be used for plotting instead of the entire dataset.
"""
if self.hdbscan_model is None:
raise ValueError("HDBSCAN model not trained yet.")
if len(self.som_models) == 0:
raise ValueError("SOM models not trained yet.")
if dim_reduction not in ['phate', 'umap']:
raise ValueError("Invalid dimensionality reduction method. Use 'phate' or 'umap'.")
if self.dim_red_op is None or self.embedding is None:
n_components = 3
if dim_reduction == 'phate':
self.dim_red_op = phate.PHATE(n_components=n_components, random_state=42)
elif dim_reduction == 'umap':
self.dim_red_op = umap.UMAP(n_components=n_components, random_state=42)
self.embedding = self.dim_red_op.fit_transform(new_data)
if new_data is not None:
new_embedding = self.dim_red_op.transform(new_data)
else:
new_embedding = self.embedding
if interactive:
fig = sp.make_subplots(rows=1, cols=1, specs=[[{'type': 'scatter3d'}]])
else:
fig = plt.figure(figsize=(30, 30))
ax = fig.add_subplot(111, projection='3d')
colors = plt.cm.rainbow(np.linspace(0, 1, len(self.som_models) + 1))
for reindexed_label, som in self.som_models.items():
original_label = self.cluster_mapping[reindexed_label]
cluster_data = embedding[self.hdbscan_model.labels_ == original_label]
som_weights = som.get_weights()
som_embedding = dim_red_op.transform(som_weights.reshape(-1, dataset.shape[1])).reshape(som_weights.shape[0], som_weights.shape[1], n_components)
if interactive:
# Plot the original data points
fig.add_trace(
go.Scatter3d(
x=cluster_data[:, 0],
y=cluster_data[:, 1],
z=cluster_data[:, 2],
mode='markers',
marker=dict(color=colors[reindexed_label], size=1),
name=f"Cluster {reindexed_label}"
)
)
else:
# Plot the original data points
ax.scatter(cluster_data[:, 0], cluster_data[:, 1], cluster_data[:, 2], c=[colors[reindexed_label]], alpha=0.3, s=5, label=f"Cluster {reindexed_label}")
for x in range(som_embedding.shape[0]):
for y in range(som_embedding.shape[1]):
if interactive:
# Plot the SOM grid
fig.add_trace(
go.Scatter3d(
x=[som_embedding[x, y, 0]],
y=[som_embedding[x, y, 1]],
z=[som_embedding[x, y, 2]],
mode='markers+text',
marker=dict(color=colors[reindexed_label], size=3, symbol='circle'),
text=[f"{x},{y}"],
textposition="top center"
)
)
else:
# Plot the SOM grid
ax.plot([som_embedding[x, y, 0]], [som_embedding[x, y, 1]], [som_embedding[x, y, 2]], '+', markersize=8, mew=2, zorder=10, c=colors[reindexed_label])
for i in range(som_embedding.shape[0] - 1):
for j in range(som_embedding.shape[1] - 1):
if interactive:
# Plot the SOM connections
fig.add_trace(
go.Scatter3d(
x=np.append(som_embedding[i:i+2, j, 0], som_embedding[i, j:j+2, 0]),
y=np.append(som_embedding[i:i+2, j, 1], som_embedding[i, j:j+2, 1]),
z=np.append(som_embedding[i:i+2, j, 2], som_embedding[i, j:j+2, 2]),
mode='lines',
line=dict(color=colors[reindexed_label], width=2),
showlegend=False
)
)
else:
# Plot the SOM connections
ax.plot(som_embedding[i:i+2, j, 0], som_embedding[i:i+2, j, 1], som_embedding[i:i+2, j, 2], lw=1, c=colors[reindexed_label])
ax.plot(som_embedding[i, j:j+2, 0], som_embedding[i, j:j+2, 1], som_embedding[i, j:j+2, 2], lw=1, c=colors[reindexed_label])
if interactive:
# Plot noise
noise_data = embedding[self.hdbscan_model.labels_ == -1]
if len(noise_data) > 0:
fig.add_trace(
go.Scatter3d(
x=noise_data[:, 0],
y=noise_data[:, 1],
z=noise_data[:, 2],
mode='markers',
marker=dict(color="gray", size=1),
name="Noise"
)
)
fig.update_layout(scene=dict(xaxis_title='X', yaxis_title='Y', zaxis_title='Z'))
fig.show()
else:
# Plot noise
noise_data = embedding[self.hdbscan_model.labels_ == -1]
if len(noise_data) > 0:
ax.scatter(noise_data[:, 0], noise_data[:, 1], noise_data[:, 2], c="gray", label="Noise")
ax.legend()
plt.show()
def plot_label_heatmap(self):
"""
Plot a heatmap for each main cluster showing the best label for each coordinate in a single subplot layout.
"""
if not hasattr(self, 'label_centroids'):
raise ValueError("Labels not trained yet.")
n_labels = len(self.label_centroids)
label_colors = plt.cm.rainbow(np.linspace(0, 1, n_labels))
n_clusters = len(self.som_models)
# Create a subplot layout with a heatmap for each main cluster
n_rows = int(np.ceil(np.sqrt(n_clusters)))
n_cols = n_rows if n_rows * (n_rows - 1) < n_clusters else n_rows - 1
fig, axes = plt.subplots(n_rows, n_cols, figsize=(n_cols * 10, n_rows * 10), squeeze=False)
for i, (reindexed_label, som) in enumerate(self.som_models.items()):
som_weights = som.get_weights()
label_map = np.zeros(som_weights.shape[:2], dtype=int)
label_distance_map = np.full(som_weights.shape[:2], np.inf)
for label_idx, label_centroid in enumerate(self.label_centroids):
for x in range(som_weights.shape[0]):
for y in range(som_weights.shape[1]):
node = som_weights[x, y]
distance = np.linalg.norm(label_centroid - node)
if distance < label_distance_map[x, y]:
label_distance_map[x, y] = distance
label_map[x, y] = label_idx
row, col = i // n_cols, i % n_cols
ax = axes[row, col]
cmap = plt.cm.rainbow
cmap.set_under(color='white')
im = ax.imshow(label_map, cmap=cmap, origin='lower', interpolation='none', vmin=0.5)
ax.set_xticks(range(label_map.shape[1]))
ax.set_yticks(range(label_map.shape[0]))
ax.grid(True, linestyle='-', linewidth=0.5)
ax.set_title(f"Label Heatmap for Cluster {reindexed_label}")
# Add a colorbar for label colors
cbar_ax = fig.add_axes([0.92, 0.15, 0.02, 0.7])
cbar = fig.colorbar(im, cax=cbar_ax, ticks=range(n_labels))
cbar.ax.set_yticklabels(self.label_encodings.classes_)
# Adjust the layout to fit everything nicely
fig.subplots_adjust(wspace=0.5, hspace=0.5, right=0.9)
plt.show()
def plot_activation(self, data, filename='prediction_output', start=None, end=None):
"""
Generate a GIF visualization of the prediction output using the activation maps of individual SOMs.
"""
if len(self.som_models) == 0:
raise ValueError("SOM models not trained yet.")
if start is None:
start = 0
if end is None:
end = len(data)
images = []
for sample in tqdm(data[start:end], desc="Visualizing prediction output"):
prediction = self.predict([sample])[0]
# if prediction[0] == -1: # Noise
# continue
fig, axes = plt.subplots(1, len(self.som_models), figsize=(20, 5), sharex=True, sharey=True)
fig.suptitle(f"Activation map for SOM {prediction[0]}, node {prediction[1]}", fontsize=16)
for idx, (som_key, som) in enumerate(self.som_models.items()):
ax = axes[idx]
activation_map = np.zeros(som._weights.shape[:2])
for x in range(som._weights.shape[0]):
for y in range(som._weights.shape[1]):
activation_map[x, y] = np.linalg.norm(sample - som._weights[x, y])
winner = som.winner(sample) # Find the BMU for this SOM
activation_map[winner] = 0 # Set the BMU's value to 0 so it will be red in the colormap
if som_key == prediction[0]: # Active SOM
im_active = ax.imshow(activation_map, cmap='viridis', origin='lower', interpolation='none')
ax.plot(winner[1], winner[0], 'r+') # Mark the BMU with a red plus sign
ax.set_title(f"SOM {som_key}", color='blue', fontweight='bold')
if hasattr(self, 'label_centroids'):
label_idx = self.label_encodings.inverse_transform([som_key - 1])[0]
ax.set_xlabel(f"Label: {label_idx}", fontsize=12)
else: # Inactive SOM
im_inactive = ax.imshow(activation_map, cmap='gray', origin='lower', interpolation='none')
ax.set_title(f"SOM {som_key}")
ax.set_xticks(range(activation_map.shape[1]))
ax.set_yticks(range(activation_map.shape[0]))
ax.grid(True, linestyle='-', linewidth=0.5)
# Create a colorbar for each frame
fig.subplots_adjust(right=0.8)
cbar_ax = fig.add_axes([0.85, 0.15, 0.05, 0.7])
try:
fig.colorbar(im_active, cax=cbar_ax)
except:
pass
# Save the plot to a buffer
buf = io.BytesIO()
plt.savefig(buf, format='png')
buf.seek(0)
img = imageio.imread(buf)
images.append(img)
plt.close()
# Save the images as a GIF
imageio.mimsave(f"{filename}.gif", images, duration=500, loop=1)
# Load the gif
gif_file = f"{filename}.gif" # Replace with the path to your GIF file
clip = VideoFileClip(gif_file)
# Convert the gif to mp4
mp4_file = f"{filename}.mp4" # Replace with the desired output path
clip.write_videofile(mp4_file, codec='libx264')
# Close the clip to release resources
clip.close()
def save(self, file_path):
"""
Save the ClusterSOM model to a file.
"""
model_data = (self.hdbscan_model, self.som_models, self.mean_values, self.sigma_values, self.cluster_mapping)
if hasattr(self, 'label_centroids'):
model_data += (self.label_centroids, self.label_encodings)
with open(file_path, "wb") as f:
pickle.dump(model_data, f)
def load(self, file_path):
"""
Load a ClusterSOM model from a file.
"""
with open(file_path, "rb") as f:
model_data = pickle.load(f)
self.hdbscan_model, self.som_models, self.mean_values, self.sigma_values, self.cluster_mapping = model_data[:5]
if len(model_data) > 5:
self.label_centroids, self.label_encodings = model_data[5:]
def plot_activation_v2(self, data, slice_select):
"""
Generate a GIF visualization of the prediction output using the activation maps of individual SOMs.
"""
if len(self.som_models) == 0:
raise ValueError("SOM models not trained yet.")
try:
prediction = self.predict([data[int(slice_select)-1]])[0]
except:
prediction = self.predict([data[int(slice_select)-2]])[0]
fig, axes = plt.subplots(1, len(self.som_models), figsize=(20, 5), sharex=True, sharey=True)
fig.suptitle(f"Activation map for SOM {prediction[0]}, node {prediction[1]}", fontsize=16)
for idx, (som_key, som) in enumerate(self.som_models.items()):
ax = axes[idx]
activation_map = np.zeros(som._weights.shape[:2])
for x in range(som._weights.shape[0]):
for y in range(som._weights.shape[1]):
activation_map[x, y] = np.linalg.norm(data[int(slice_select)-1] - som._weights[x, y])
winner = som.winner(data[int(slice_select)-1]) # Find the BMU for this SOM
activation_map[winner] = 0 # Set the BMU's value to 0 so it will be red in the colormap
if som_key == prediction[0]: # Active SOM
im_active = ax.imshow(activation_map, cmap='viridis', origin='lower', interpolation='none')
ax.plot(winner[1], winner[0], 'r+') # Mark the BMU with a red plus sign
ax.set_title(f"SOM {som_key}", color='blue', fontweight='bold')
if hasattr(self, 'label_centroids'):
label_idx = self.label_encodings.inverse_transform([som_key - 1])[0]
ax.set_xlabel(f"Label: {label_idx}", fontsize=12)
else: # Inactive SOM
im_inactive = ax.imshow(activation_map, cmap='gray', origin='lower', interpolation='none')
ax.set_title(f"SOM {som_key}")
ax.set_xticks(range(activation_map.shape[1]))
ax.set_yticks(range(activation_map.shape[0]))
ax.grid(True, linestyle='-', linewidth=0.5)
plt.tight_layout()
return fig |