Spaces:
Sleeping
Sleeping
Jan Winkler
commited on
Commit
•
719aa88
1
Parent(s):
2a66be9
delete redundant code, update paths and restructuring (#6)
Browse files- README.md +1 -1
- python/notebooks/eda_jan.py +59 -171
README.md
CHANGED
@@ -64,6 +64,6 @@ now you shoold see the running docker containers.
|
|
64 |
1. `chmod +x format` so that the `format` file is executable
|
65 |
2. then simply use `./format` before adding your changes and all the files will be autoformatted
|
66 |
|
67 |
-
|
68 |
|
69 |
![](./docs/data_overview_yuri.jpg)
|
|
|
64 |
1. `chmod +x format` so that the `format` file is executable
|
65 |
2. then simply use `./format` before adding your changes and all the files will be autoformatted
|
66 |
|
67 |
+
## awesome data overview
|
68 |
|
69 |
![](./docs/data_overview_yuri.jpg)
|
python/notebooks/eda_jan.py
CHANGED
@@ -12,114 +12,6 @@
|
|
12 |
# name: python3
|
13 |
# ---
|
14 |
|
15 |
-
# %%
|
16 |
-
import os
|
17 |
-
import numpy as np
|
18 |
-
import librosa
|
19 |
-
import librosa.display
|
20 |
-
import matplotlib.pyplot as plt
|
21 |
-
from sklearn.cluster import KMeans
|
22 |
-
from sklearn.decomposition import PCA
|
23 |
-
from IPython.display import Audio, display
|
24 |
-
|
25 |
-
# %%
|
26 |
-
# Load the entire audio file
|
27 |
-
cwd = os.getcwd()
|
28 |
-
relative_path = "data/soundscape_data/PER_001_S01_20190116_100007Z.flac"
|
29 |
-
file_path = os.path.join(cwd, relative_path)
|
30 |
-
y, sr = librosa.load(file_path, sr=44100)
|
31 |
-
|
32 |
-
# %%
|
33 |
-
# split soundfile in to 10s chunks
|
34 |
-
window_size = 10 # window size in seconds
|
35 |
-
hop_size = 10 # hop size in seconds
|
36 |
-
|
37 |
-
# Convert window and hop size to samples
|
38 |
-
window_samples = int(window_size * sr)
|
39 |
-
hop_samples = int(hop_size * sr)
|
40 |
-
|
41 |
-
# Total number of windows
|
42 |
-
num_windows = (len(y) - window_samples) // hop_samples + 1
|
43 |
-
|
44 |
-
print(f"Total number of windows: {num_windows}")
|
45 |
-
|
46 |
-
|
47 |
-
# %%
|
48 |
-
# Define frequency bands (in Hz)
|
49 |
-
bands = {
|
50 |
-
"Sub-bass": (20, 60),
|
51 |
-
"Bass": (60, 250),
|
52 |
-
"Low Midrange": (250, 500),
|
53 |
-
"Midrange": (500, 2000),
|
54 |
-
"Upper Midrange": (2000, 4000),
|
55 |
-
"Presence": (4000, 6000),
|
56 |
-
"Brilliance": (6000, 20000),
|
57 |
-
}
|
58 |
-
|
59 |
-
# Initialize a list to hold the features
|
60 |
-
all_features = []
|
61 |
-
|
62 |
-
for i in range(num_windows):
|
63 |
-
start_sample = i * hop_samples
|
64 |
-
end_sample = start_sample + window_samples
|
65 |
-
y_window = y[start_sample:end_sample]
|
66 |
-
|
67 |
-
# Compute STFT
|
68 |
-
S = librosa.stft(y_window)
|
69 |
-
S_db = librosa.amplitude_to_db(np.abs(S))
|
70 |
-
|
71 |
-
# Compute features for each band
|
72 |
-
features = []
|
73 |
-
for band, (low_freq, high_freq) in bands.items():
|
74 |
-
low_bin = int(np.floor(low_freq * (S.shape[0] / sr)))
|
75 |
-
high_bin = int(np.ceil(high_freq * (S.shape[0] / sr)))
|
76 |
-
band_energy = np.mean(S_db[low_bin:high_bin, :], axis=0)
|
77 |
-
features.append(band_energy)
|
78 |
-
|
79 |
-
# Flatten the feature array and add to all_features
|
80 |
-
features_flat = np.concatenate(features)
|
81 |
-
all_features.append(features_flat)
|
82 |
-
|
83 |
-
# Convert to numpy array
|
84 |
-
all_features = np.array(all_features)
|
85 |
-
|
86 |
-
|
87 |
-
# %%
|
88 |
-
# Reduce dimensionality with PCA
|
89 |
-
pca = PCA(n_components=2)
|
90 |
-
features_reduced = pca.fit_transform(all_features)
|
91 |
-
|
92 |
-
# Perform k-means clustering
|
93 |
-
kmeans = KMeans(n_clusters=5) # Example: 5 clusters
|
94 |
-
clusters = kmeans.fit_predict(features_reduced)
|
95 |
-
|
96 |
-
# Plot the clusters
|
97 |
-
plt.figure(figsize=(10, 6))
|
98 |
-
scatter = plt.scatter(
|
99 |
-
features_reduced[:, 0], features_reduced[:, 1], c=clusters, cmap="viridis"
|
100 |
-
)
|
101 |
-
plt.title("Clustered Frequency Band Features")
|
102 |
-
plt.xlabel("Principal Component 1")
|
103 |
-
plt.ylabel("Principal Component 2")
|
104 |
-
plt.colorbar(scatter, label="Cluster")
|
105 |
-
plt.show()
|
106 |
-
|
107 |
-
|
108 |
-
# %%
|
109 |
-
# Play the audio for a representative sample from each cluster
|
110 |
-
for cluster_label in np.unique(clusters):
|
111 |
-
# Find the first data point in the cluster
|
112 |
-
representative_index = np.where(clusters == cluster_label)[0][0]
|
113 |
-
|
114 |
-
# Use the original audio window at the representative index
|
115 |
-
start_sample = representative_index * hop_samples
|
116 |
-
end_sample = start_sample + window_samples
|
117 |
-
y_representative = y[start_sample:end_sample]
|
118 |
-
|
119 |
-
print(f"Cluster {cluster_label} representative audio:")
|
120 |
-
display(Audio(data=y_representative, rate=sr))
|
121 |
-
|
122 |
-
|
123 |
# %% [markdown]
|
124 |
# ## pipeline for all the files
|
125 |
|
@@ -140,12 +32,22 @@ from sklearn.ensemble import RandomForestClassifier
|
|
140 |
|
141 |
|
142 |
# %%
|
|
|
143 |
# Directory containing the audio files
|
144 |
# audio_dir = "data/soundscape_data"
|
145 |
audio_dir = (
|
146 |
-
"data/SoundMeters_Ingles_Primary-20240519T132658Z-002/SoundMeters_Ingles_Primary"
|
147 |
)
|
148 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
149 |
# Parameters for windowing
|
150 |
window_size = 10 # window size in seconds
|
151 |
hop_size = 10 # hop size in seconds
|
@@ -161,15 +63,11 @@ bands = {
|
|
161 |
"Brilliance": (6000, 20000),
|
162 |
}
|
163 |
|
164 |
-
# Directory to save features
|
165 |
-
features_dir = "features"
|
166 |
-
os.makedirs(features_dir, exist_ok=True)
|
167 |
-
|
168 |
# Iterate over each audio file in the directory
|
169 |
for filename in os.listdir(audio_dir):
|
170 |
if filename.endswith(".wav"):
|
171 |
file_path = os.path.join(audio_dir, filename)
|
172 |
-
y, sr = librosa.load(file_path, sr=
|
173 |
|
174 |
# Convert window and hop size to samples
|
175 |
window_samples = int(window_size * sr)
|
@@ -216,11 +114,9 @@ for filename in os.listdir(audio_dir):
|
|
216 |
|
217 |
|
218 |
# %%
|
219 |
-
#
|
220 |
-
features_dir = "features"
|
221 |
n_clusters = 5
|
222 |
|
223 |
-
|
224 |
# Load all features
|
225 |
all_features = []
|
226 |
for feature_file in os.listdir(features_dir):
|
@@ -239,21 +135,36 @@ features_pca = pca.fit_transform(all_features)
|
|
239 |
kmeans = KMeans(n_clusters=n_clusters) # Example: 5 clusters
|
240 |
clusters = kmeans.fit_predict(all_features)
|
241 |
|
242 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
243 |
plt.figure(figsize=(10, 6))
|
244 |
-
|
245 |
-
|
246 |
-
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
247 |
plt.title("PCA of Clustered Frequency Band Features")
|
248 |
plt.xlabel("Principal Component 1")
|
249 |
plt.ylabel("Principal Component 2")
|
250 |
-
plt.
|
251 |
plt.show()
|
252 |
|
253 |
-
# Save clustering results
|
254 |
-
clustering_results = {"clusters": clusters, "kmeans": kmeans, "pca": pca}
|
255 |
-
joblib.dump(clustering_results, "clustering_results.pkl")
|
256 |
|
|
|
257 |
# Plot the clusters
|
258 |
plt.figure(figsize=(10, 6))
|
259 |
for i in range(n_clusters):
|
@@ -264,21 +175,26 @@ plt.xlabel("Feature Index (Frequency Bands)")
|
|
264 |
plt.ylabel("Mean Feature Value (Energy in dB)")
|
265 |
plt.show()
|
266 |
|
|
|
267 |
# %%
|
268 |
-
#
|
269 |
-
|
270 |
-
|
271 |
-
|
272 |
-
)
|
273 |
-
|
274 |
-
|
|
|
|
|
|
|
275 |
|
|
|
276 |
# Parameters for windowing
|
277 |
window_size = 10 # window size in seconds
|
278 |
hop_size = 10 # hop size in seconds
|
279 |
|
280 |
# Load clustering results
|
281 |
-
clustering_results = joblib.load("clustering_results.pkl")
|
282 |
clusters = clustering_results["clusters"]
|
283 |
|
284 |
# Load all features
|
@@ -290,7 +206,7 @@ for feature_file in os.listdir(features_dir):
|
|
290 |
features, scaler = joblib.load(os.path.join(features_dir, feature_file))
|
291 |
filename = feature_file.replace("_features.npy", ".wav")
|
292 |
file_path = os.path.join(audio_dir, filename)
|
293 |
-
y, sr = librosa.load(file_path, sr=
|
294 |
|
295 |
# Convert window and hop size to samples
|
296 |
window_samples = int(window_size * sr)
|
@@ -307,7 +223,8 @@ for feature_file in os.listdir(features_dir):
|
|
307 |
# Flatten the list of all features
|
308 |
all_features = np.vstack(all_features)
|
309 |
|
310 |
-
|
|
|
311 |
for cluster_label in np.unique(clusters):
|
312 |
try:
|
313 |
# Find the first data point in the cluster
|
@@ -323,12 +240,17 @@ for cluster_label in np.unique(clusters):
|
|
323 |
print(f"Cluster {cluster_label} representative audio:")
|
324 |
display(Audio(data=y_representative, rate=sr))
|
325 |
|
|
|
|
|
|
|
|
|
|
|
326 |
except Exception as e:
|
327 |
print(f"Could not play audio for cluster {cluster_label}: {e}")
|
328 |
|
329 |
-
|
330 |
# %%
|
331 |
-
|
|
|
332 |
# Fit PCA
|
333 |
pca = PCA().fit(all_features_scaled)
|
334 |
|
@@ -367,40 +289,6 @@ kaiser_criterion = np.sum(eigenvalues > 1)
|
|
367 |
print(f"Number of components selected by Kaiser Criterion: {kaiser_criterion}")
|
368 |
|
369 |
|
370 |
-
# %%
|
371 |
-
# Method 4: Cross-Validation
|
372 |
-
# Evaluate a classifier with different numbers of principal components
|
373 |
-
|
374 |
-
## do not run if you dont have time, this takes forever.
|
375 |
-
# scores = []
|
376 |
-
# for n_components in range(1, len(explained_variance) + 1):
|
377 |
-
# pca = PCA(n_components=n_components)
|
378 |
-
# features_pca = pca.fit_transform(all_features_scaled)
|
379 |
-
# classifier = RandomForestClassifier() # Use your preferred model here
|
380 |
-
# score = np.mean(cross_val_score(classifier, features_pca, clusters, cv=n_clusters)) # Assuming `clusters` are your labels
|
381 |
-
# scores.append(score)
|
382 |
-
|
383 |
-
# # Plot cross-validation scores
|
384 |
-
# plt.figure(figsize=(10, 6))
|
385 |
-
# plt.plot(range(1, len(explained_variance) + 1), scores, marker='o')
|
386 |
-
# plt.xlabel('Number of Principal Components')
|
387 |
-
# plt.ylabel('Cross-Validation Score')
|
388 |
-
# plt.title('Cross-Validation Score vs. Number of Principal Components')
|
389 |
-
# plt.grid(True)
|
390 |
-
# plt.show()
|
391 |
-
|
392 |
-
# # Choosing the number of components that explain at least 95% of the variance
|
393 |
-
# n_components_variance = np.argmax(cumulative_explained_variance >= 0.95) + 1
|
394 |
-
# print(f"Number of components to retain 95% variance: {n_components_variance}")
|
395 |
-
|
396 |
-
# # Choose the optimal number of components based on your analysis
|
397 |
-
# optimal_n_components = n_components_variance # or based on the scree plot, cross-validation, etc.
|
398 |
-
# print(f"Optimal number of components: {optimal_n_components}")
|
399 |
-
|
400 |
-
# # Perform PCA with the selected number of components
|
401 |
-
# pca = PCA(n_components=optimal_n_components)
|
402 |
-
# features_pca = pca.fit_transform(all_features_scaled)
|
403 |
-
|
404 |
# %%
|
405 |
|
406 |
# %%
|
|
|
12 |
# name: python3
|
13 |
# ---
|
14 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
# %% [markdown]
|
16 |
# ## pipeline for all the files
|
17 |
|
|
|
32 |
|
33 |
|
34 |
# %%
|
35 |
+
## cockpit for directories
|
36 |
# Directory containing the audio files
|
37 |
# audio_dir = "data/soundscape_data"
|
38 |
audio_dir = (
|
39 |
+
"../data/SoundMeters_Ingles_Primary-20240519T132658Z-002/SoundMeters_Ingles_Primary"
|
40 |
)
|
41 |
|
42 |
+
# Directory to save features
|
43 |
+
features_dir = "../data/features"
|
44 |
+
os.makedirs(features_dir, exist_ok=True)
|
45 |
+
|
46 |
+
# Directory to save clusters information
|
47 |
+
clusters_dir = "../data/clusters"
|
48 |
+
os.makedirs(clusters_dir, exist_ok=True)
|
49 |
+
|
50 |
+
# %%
|
51 |
# Parameters for windowing
|
52 |
window_size = 10 # window size in seconds
|
53 |
hop_size = 10 # hop size in seconds
|
|
|
63 |
"Brilliance": (6000, 20000),
|
64 |
}
|
65 |
|
|
|
|
|
|
|
|
|
66 |
# Iterate over each audio file in the directory
|
67 |
for filename in os.listdir(audio_dir):
|
68 |
if filename.endswith(".wav"):
|
69 |
file_path = os.path.join(audio_dir, filename)
|
70 |
+
y, sr = librosa.load(file_path, sr=None)
|
71 |
|
72 |
# Convert window and hop size to samples
|
73 |
window_samples = int(window_size * sr)
|
|
|
114 |
|
115 |
|
116 |
# %%
|
117 |
+
# Number of clusters
|
|
|
118 |
n_clusters = 5
|
119 |
|
|
|
120 |
# Load all features
|
121 |
all_features = []
|
122 |
for feature_file in os.listdir(features_dir):
|
|
|
135 |
kmeans = KMeans(n_clusters=n_clusters) # Example: 5 clusters
|
136 |
clusters = kmeans.fit_predict(all_features)
|
137 |
|
138 |
+
# Save clustering results
|
139 |
+
clustering_results = {"clusters": clusters, "kmeans": kmeans, "pca": pca}
|
140 |
+
joblib.dump(clustering_results, os.path.join(clusters_dir, "clustering_results.pkl"))
|
141 |
+
|
142 |
+
|
143 |
+
# %%
|
144 |
+
# Plot the PCA-reduced features with cluster labels using a legend
|
145 |
plt.figure(figsize=(10, 6))
|
146 |
+
|
147 |
+
# Define a colormap
|
148 |
+
colors = plt.cm.tab10(np.arange(kmeans.n_clusters))
|
149 |
+
|
150 |
+
for cluster_label in np.unique(clusters):
|
151 |
+
cluster_points = features_pca[clusters == cluster_label]
|
152 |
+
plt.scatter(
|
153 |
+
cluster_points[:, 0],
|
154 |
+
cluster_points[:, 1],
|
155 |
+
s=50,
|
156 |
+
color=colors[cluster_label],
|
157 |
+
label=f"Cluster {cluster_label}",
|
158 |
+
)
|
159 |
+
|
160 |
plt.title("PCA of Clustered Frequency Band Features")
|
161 |
plt.xlabel("Principal Component 1")
|
162 |
plt.ylabel("Principal Component 2")
|
163 |
+
plt.legend()
|
164 |
plt.show()
|
165 |
|
|
|
|
|
|
|
166 |
|
167 |
+
# %%
|
168 |
# Plot the clusters
|
169 |
plt.figure(figsize=(10, 6))
|
170 |
for i in range(n_clusters):
|
|
|
175 |
plt.ylabel("Mean Feature Value (Energy in dB)")
|
176 |
plt.show()
|
177 |
|
178 |
+
|
179 |
# %%
|
180 |
+
# Function to plot the spectrogram
|
181 |
+
def plot_spectrogram(y, sr, title):
|
182 |
+
S = librosa.stft(y)
|
183 |
+
S_db = librosa.amplitude_to_db(np.abs(S), ref=np.max)
|
184 |
+
plt.figure(figsize=(10, 6))
|
185 |
+
librosa.display.specshow(S_db, sr=sr, x_axis="time", y_axis="log")
|
186 |
+
plt.colorbar(format="%+2.0f dB")
|
187 |
+
plt.title(title)
|
188 |
+
plt.show()
|
189 |
+
|
190 |
|
191 |
+
# %%
|
192 |
# Parameters for windowing
|
193 |
window_size = 10 # window size in seconds
|
194 |
hop_size = 10 # hop size in seconds
|
195 |
|
196 |
# Load clustering results
|
197 |
+
clustering_results = joblib.load(os.path.join(clusters_dir, "clustering_results.pkl"))
|
198 |
clusters = clustering_results["clusters"]
|
199 |
|
200 |
# Load all features
|
|
|
206 |
features, scaler = joblib.load(os.path.join(features_dir, feature_file))
|
207 |
filename = feature_file.replace("_features.npy", ".wav")
|
208 |
file_path = os.path.join(audio_dir, filename)
|
209 |
+
y, sr = librosa.load(file_path, sr=None)
|
210 |
|
211 |
# Convert window and hop size to samples
|
212 |
window_samples = int(window_size * sr)
|
|
|
223 |
# Flatten the list of all features
|
224 |
all_features = np.vstack(all_features)
|
225 |
|
226 |
+
|
227 |
+
# %%
|
228 |
for cluster_label in np.unique(clusters):
|
229 |
try:
|
230 |
# Find the first data point in the cluster
|
|
|
240 |
print(f"Cluster {cluster_label} representative audio:")
|
241 |
display(Audio(data=y_representative, rate=sr))
|
242 |
|
243 |
+
# Plot the spectrogram
|
244 |
+
plot_spectrogram(
|
245 |
+
y_representative, sr, f"Spectrogram for Cluster {cluster_label}"
|
246 |
+
)
|
247 |
+
|
248 |
except Exception as e:
|
249 |
print(f"Could not play audio for cluster {cluster_label}: {e}")
|
250 |
|
|
|
251 |
# %%
|
252 |
+
scaler = StandardScaler()
|
253 |
+
all_features_scaled = scaler.fit_transform(all_features)
|
254 |
# Fit PCA
|
255 |
pca = PCA().fit(all_features_scaled)
|
256 |
|
|
|
289 |
print(f"Number of components selected by Kaiser Criterion: {kaiser_criterion}")
|
290 |
|
291 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
292 |
# %%
|
293 |
|
294 |
# %%
|