Commit
•
be4d0c2
1
Parent(s):
0791f60
Upload 5 files
Browse files- CNN_Architecture.ipynb +0 -0
- CNN_support.py +138 -0
- Data_preparation.ipynb +652 -0
- RNN_Architecture.ipynb +0 -0
- evaluation.py +99 -0
CNN_Architecture.ipynb
ADDED
The diff for this file is too large to render.
See raw diff
|
|
CNN_support.py
ADDED
@@ -0,0 +1,138 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
import librosa
|
4 |
+
import numpy as np
|
5 |
+
from scipy.io import wavfile
|
6 |
+
from sklearn.preprocessing import normalize
|
7 |
+
|
8 |
+
class SoundPreprocessing:
|
9 |
+
"""
|
10 |
+
Parameters
|
11 |
+
----------
|
12 |
+
|
13 |
+
sr (int): sampling rate
|
14 |
+
max_size (iterable): resulting shape of the tensor
|
15 |
+
n_fft (int): number related to FFT
|
16 |
+
n_mfcc (int): number of MFCC
|
17 |
+
|
18 |
+
"""
|
19 |
+
|
20 |
+
|
21 |
+
def __init__(self, *, sr, max_size, n_fft, n_mfcc = 60, hop_length = 512):
|
22 |
+
self.sr = sr
|
23 |
+
self.n_fft = n_fft
|
24 |
+
self.n_mfcc = n_mfcc
|
25 |
+
self.max_size = max_size
|
26 |
+
self.hop_length = hop_length
|
27 |
+
|
28 |
+
|
29 |
+
def padding(self, array, xx, yy):
|
30 |
+
"""
|
31 |
+
Parameters
|
32 |
+
----------
|
33 |
+
array: numpy array
|
34 |
+
xx: desired height
|
35 |
+
yy: desirex width
|
36 |
+
|
37 |
+
Returns: padded array
|
38 |
+
"""
|
39 |
+
self.array = array
|
40 |
+
self.xx = xx
|
41 |
+
self.yy = yy
|
42 |
+
|
43 |
+
h = array.shape[0]
|
44 |
+
w = array.shape[1]
|
45 |
+
a = max((xx - h) // 2,0)
|
46 |
+
aa = max(0,xx - a - h)
|
47 |
+
b = max(0,(yy - w) // 2)
|
48 |
+
bb = max(yy - b - w,0)
|
49 |
+
|
50 |
+
return np.pad(array, pad_width = ((a, aa), (b, bb)),
|
51 |
+
mode = "constant")
|
52 |
+
|
53 |
+
|
54 |
+
def generate_features(self, y_cut, sr, max_size, n_fft, n_mfcc, hop_length):
|
55 |
+
self.y_cut = y_cut
|
56 |
+
|
57 |
+
# Numeri -2 divisibili per 14
|
58 |
+
condition = np.arange(2, 1000)[np.where((np.arange(2, 1000) - 2)%14 == 0)]
|
59 |
+
|
60 |
+
global shape_changed
|
61 |
+
shape_changed = False
|
62 |
+
|
63 |
+
if max_size[0] not in condition:
|
64 |
+
# Get closest number to 'max_size' that respects 'condition'
|
65 |
+
new_max0 = sorted(condition, key = lambda v: abs(v - max_size[0]))[0]
|
66 |
+
shape_changed = True
|
67 |
+
max_size = (new_max0, max_size[1])
|
68 |
+
|
69 |
+
stft = self.padding(np.abs(librosa.stft(y = y_cut, n_fft = n_fft,
|
70 |
+
hop_length = 512)), max_size[0], max_size[1])
|
71 |
+
|
72 |
+
if max_size[0] < stft.shape[0]:
|
73 |
+
new_max0 = sorted(condition[condition >= stft.shape[0]],
|
74 |
+
key = lambda v: abs(v - stft.shape[0]))[0]
|
75 |
+
max_size = (new_max0, max_size[1])
|
76 |
+
shape_changed = True
|
77 |
+
|
78 |
+
stft = self.padding(np.abs(librosa.stft(y = y_cut, n_fft = n_fft,
|
79 |
+
hop_length = 512)), max_size[0], max_size[1])
|
80 |
+
|
81 |
+
MFCCs = self.padding(librosa.feature.mfcc(y = y_cut, n_fft = n_fft, sr = sr,
|
82 |
+
hop_length = hop_length, n_mfcc = n_mfcc),
|
83 |
+
max_size[0], max_size[1])
|
84 |
+
|
85 |
+
spec_centroid = librosa.feature.spectral_centroid(y = y_cut, sr = sr)
|
86 |
+
chroma_stft = librosa.feature.chroma_stft(y = y_cut, sr = sr)
|
87 |
+
spec_bw = librosa.feature.spectral_bandwidth(y = y_cut, sr = sr)
|
88 |
+
|
89 |
+
#Now the padding part
|
90 |
+
image = np.array([self.padding(normalize(spec_bw), 1, max_size[1])]).reshape(1, max_size[1])
|
91 |
+
image = np.append(image, self.padding(normalize(spec_centroid), 1, max_size[1]), axis = 0)
|
92 |
+
|
93 |
+
#repeat the padded spec_bw,spec_centroid and chroma stft until they are stft and MFCC-sized
|
94 |
+
for i in range( int((max_size[0]-2)/14) ):
|
95 |
+
image = np.append(image, self.padding(normalize(spec_bw), 1, max_size[1]), axis = 0)
|
96 |
+
image = np.append(image, self.padding(normalize(spec_centroid), 1, max_size[1]), axis = 0)
|
97 |
+
image = np.append(image, self.padding(normalize(chroma_stft), 12, max_size[1]), axis = 0)
|
98 |
+
|
99 |
+
image = np.dstack((image, np.abs(stft)))
|
100 |
+
image = np.dstack((image, MFCCs))
|
101 |
+
|
102 |
+
return image
|
103 |
+
|
104 |
+
|
105 |
+
def get_features(self, df, filepath):
|
106 |
+
self.df = df
|
107 |
+
self.filepath = filepath
|
108 |
+
|
109 |
+
# Get data for CNN
|
110 |
+
X = []
|
111 |
+
y = np.zeros(shape = (len(df), 1))
|
112 |
+
|
113 |
+
for i in df.index:
|
114 |
+
|
115 |
+
sr_i, aud = wavfile.read("{}\\{}".format(filepath, df.loc[i, "filename"]))
|
116 |
+
aud = aud.astype(np.float16)
|
117 |
+
|
118 |
+
X += [self.generate_features(y_cut = aud, sr = sr_i,
|
119 |
+
n_fft = self.n_fft,
|
120 |
+
n_mfcc = self.n_mfcc,
|
121 |
+
max_size = self.max_size,
|
122 |
+
hop_length = self.hop_length)]
|
123 |
+
|
124 |
+
y[i] = df.loc[i, "target"]
|
125 |
+
|
126 |
+
|
127 |
+
if shape_changed == True:
|
128 |
+
print(f"New max_size is {max_size}")
|
129 |
+
|
130 |
+
X = np.array(X)
|
131 |
+
|
132 |
+
return X, y
|
133 |
+
|
134 |
+
|
135 |
+
|
136 |
+
|
137 |
+
|
138 |
+
|
Data_preparation.ipynb
ADDED
@@ -0,0 +1,652 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "markdown",
|
5 |
+
"metadata": {
|
6 |
+
"id": "VNUnhmXWe9qz"
|
7 |
+
},
|
8 |
+
"source": [
|
9 |
+
"# Notebook for data preparation\n",
|
10 |
+
"\n",
|
11 |
+
"A.A. 2022-2023 - HUMAN DATA ANALYTICS\n",
|
12 |
+
"\n",
|
13 |
+
"Authors:\n",
|
14 |
+
"* Mattia Brocco\n",
|
15 |
+
"* Brenda Eloisa Tellez Juarez\n",
|
16 |
+
"\n",
|
17 |
+
"In the following notebook the pipeline for data import, preprocessing and storage (using `.parquet` format) is presented."
|
18 |
+
]
|
19 |
+
},
|
20 |
+
{
|
21 |
+
"cell_type": "code",
|
22 |
+
"execution_count": 1,
|
23 |
+
"metadata": {
|
24 |
+
"ExecuteTime": {
|
25 |
+
"end_time": "2023-02-12T22:43:39.436355Z",
|
26 |
+
"start_time": "2023-02-12T22:43:39.418449Z"
|
27 |
+
},
|
28 |
+
"colab": {
|
29 |
+
"base_uri": "https://localhost:8080/",
|
30 |
+
"height": 915
|
31 |
+
},
|
32 |
+
"id": "pz7MotpCfCUR",
|
33 |
+
"outputId": "fc916ed3-03d2-41ee-87db-237d79979cf0"
|
34 |
+
},
|
35 |
+
"outputs": [],
|
36 |
+
"source": [
|
37 |
+
"from google.colab import drive\n",
|
38 |
+
"drive.mount(\"/content/drive\")\n",
|
39 |
+
"\n",
|
40 |
+
"#%cd /content/drive/MyDrive/Environmental-sounds-UNIPD-2022"
|
41 |
+
]
|
42 |
+
},
|
43 |
+
{
|
44 |
+
"cell_type": "code",
|
45 |
+
"execution_count": 3,
|
46 |
+
"metadata": {
|
47 |
+
"id": "6YEmW9n_fOB8"
|
48 |
+
},
|
49 |
+
"outputs": [],
|
50 |
+
"source": [
|
51 |
+
"import os\n",
|
52 |
+
"import sys\n",
|
53 |
+
"import torch\n",
|
54 |
+
"import librosa\n",
|
55 |
+
"import matplotlib\n",
|
56 |
+
"import numpy as np\n",
|
57 |
+
"import pandas as pd\n",
|
58 |
+
"import seaborn as sns\n",
|
59 |
+
"import tensorflow as tf\n",
|
60 |
+
"from librosa import display\n",
|
61 |
+
"from scipy.io import wavfile\n",
|
62 |
+
"from tensorflow import keras\n",
|
63 |
+
"import IPython.display as ipd\n",
|
64 |
+
"import matplotlib.pyplot as plt\n",
|
65 |
+
"\n",
|
66 |
+
"from sklearn.metrics import confusion_matrix\n",
|
67 |
+
"from sklearn.metrics import classification_report\n",
|
68 |
+
"\n",
|
69 |
+
"import evaluation\n",
|
70 |
+
"import CNN_support as cnns\n",
|
71 |
+
"from gng import GrowingNeuralGas\n",
|
72 |
+
"\n",
|
73 |
+
"%load_ext autoreload\n",
|
74 |
+
"%autoreload 2"
|
75 |
+
]
|
76 |
+
},
|
77 |
+
{
|
78 |
+
"cell_type": "code",
|
79 |
+
"execution_count": 4,
|
80 |
+
"metadata": {
|
81 |
+
"colab": {
|
82 |
+
"base_uri": "https://localhost:8080/",
|
83 |
+
"height": 206
|
84 |
+
},
|
85 |
+
"execution": {
|
86 |
+
"iopub.execute_input": "2023-01-14T19:51:27.903698Z",
|
87 |
+
"iopub.status.busy": "2023-01-14T19:51:27.903426Z",
|
88 |
+
"iopub.status.idle": "2023-01-14T19:51:27.930731Z",
|
89 |
+
"shell.execute_reply": "2023-01-14T19:51:27.929790Z",
|
90 |
+
"shell.execute_reply.started": "2023-01-14T19:51:27.903668Z"
|
91 |
+
},
|
92 |
+
"id": "ZjdASAl2emSc",
|
93 |
+
"outputId": "a209c1ff-299b-4e8d-c79a-911fc9fab8ca"
|
94 |
+
},
|
95 |
+
"outputs": [
|
96 |
+
{
|
97 |
+
"data": {
|
98 |
+
"text/html": [
|
99 |
+
"\n",
|
100 |
+
" <div id=\"df-ea413149-b901-4dd6-a8ae-0b417a0edf47\">\n",
|
101 |
+
" <div class=\"colab-df-container\">\n",
|
102 |
+
" <div>\n",
|
103 |
+
"<style scoped>\n",
|
104 |
+
" .dataframe tbody tr th:only-of-type {\n",
|
105 |
+
" vertical-align: middle;\n",
|
106 |
+
" }\n",
|
107 |
+
"\n",
|
108 |
+
" .dataframe tbody tr th {\n",
|
109 |
+
" vertical-align: top;\n",
|
110 |
+
" }\n",
|
111 |
+
"\n",
|
112 |
+
" .dataframe thead th {\n",
|
113 |
+
" text-align: right;\n",
|
114 |
+
" }\n",
|
115 |
+
"</style>\n",
|
116 |
+
"<table border=\"1\" class=\"dataframe\">\n",
|
117 |
+
" <thead>\n",
|
118 |
+
" <tr style=\"text-align: right;\">\n",
|
119 |
+
" <th></th>\n",
|
120 |
+
" <th>filename</th>\n",
|
121 |
+
" <th>fold</th>\n",
|
122 |
+
" <th>target</th>\n",
|
123 |
+
" <th>category</th>\n",
|
124 |
+
" <th>esc10</th>\n",
|
125 |
+
" <th>src_file</th>\n",
|
126 |
+
" <th>take</th>\n",
|
127 |
+
" </tr>\n",
|
128 |
+
" </thead>\n",
|
129 |
+
" <tbody>\n",
|
130 |
+
" <tr>\n",
|
131 |
+
" <th>0</th>\n",
|
132 |
+
" <td>1-100032-A-0.wav</td>\n",
|
133 |
+
" <td>1</td>\n",
|
134 |
+
" <td>0</td>\n",
|
135 |
+
" <td>dog</td>\n",
|
136 |
+
" <td>True</td>\n",
|
137 |
+
" <td>100032</td>\n",
|
138 |
+
" <td>A</td>\n",
|
139 |
+
" </tr>\n",
|
140 |
+
" <tr>\n",
|
141 |
+
" <th>1</th>\n",
|
142 |
+
" <td>1-100038-A-14.wav</td>\n",
|
143 |
+
" <td>1</td>\n",
|
144 |
+
" <td>14</td>\n",
|
145 |
+
" <td>chirping_birds</td>\n",
|
146 |
+
" <td>False</td>\n",
|
147 |
+
" <td>100038</td>\n",
|
148 |
+
" <td>A</td>\n",
|
149 |
+
" </tr>\n",
|
150 |
+
" <tr>\n",
|
151 |
+
" <th>2</th>\n",
|
152 |
+
" <td>1-100210-A-36.wav</td>\n",
|
153 |
+
" <td>1</td>\n",
|
154 |
+
" <td>36</td>\n",
|
155 |
+
" <td>vacuum_cleaner</td>\n",
|
156 |
+
" <td>False</td>\n",
|
157 |
+
" <td>100210</td>\n",
|
158 |
+
" <td>A</td>\n",
|
159 |
+
" </tr>\n",
|
160 |
+
" <tr>\n",
|
161 |
+
" <th>3</th>\n",
|
162 |
+
" <td>1-100210-B-36.wav</td>\n",
|
163 |
+
" <td>1</td>\n",
|
164 |
+
" <td>36</td>\n",
|
165 |
+
" <td>vacuum_cleaner</td>\n",
|
166 |
+
" <td>False</td>\n",
|
167 |
+
" <td>100210</td>\n",
|
168 |
+
" <td>B</td>\n",
|
169 |
+
" </tr>\n",
|
170 |
+
" <tr>\n",
|
171 |
+
" <th>4</th>\n",
|
172 |
+
" <td>1-101296-A-19.wav</td>\n",
|
173 |
+
" <td>1</td>\n",
|
174 |
+
" <td>19</td>\n",
|
175 |
+
" <td>thunderstorm</td>\n",
|
176 |
+
" <td>False</td>\n",
|
177 |
+
" <td>101296</td>\n",
|
178 |
+
" <td>A</td>\n",
|
179 |
+
" </tr>\n",
|
180 |
+
" </tbody>\n",
|
181 |
+
"</table>\n",
|
182 |
+
"</div>\n",
|
183 |
+
" <button class=\"colab-df-convert\" onclick=\"convertToInteractive('df-ea413149-b901-4dd6-a8ae-0b417a0edf47')\"\n",
|
184 |
+
" title=\"Convert this dataframe to an interactive table.\"\n",
|
185 |
+
" style=\"display:none;\">\n",
|
186 |
+
" \n",
|
187 |
+
" <svg xmlns=\"http://www.w3.org/2000/svg\" height=\"24px\"viewBox=\"0 0 24 24\"\n",
|
188 |
+
" width=\"24px\">\n",
|
189 |
+
" <path d=\"M0 0h24v24H0V0z\" fill=\"none\"/>\n",
|
190 |
+
" <path d=\"M18.56 5.44l.94 2.06.94-2.06 2.06-.94-2.06-.94-.94-2.06-.94 2.06-2.06.94zm-11 1L8.5 8.5l.94-2.06 2.06-.94-2.06-.94L8.5 2.5l-.94 2.06-2.06.94zm10 10l.94 2.06.94-2.06 2.06-.94-2.06-.94-.94-2.06-.94 2.06-2.06.94z\"/><path d=\"M17.41 7.96l-1.37-1.37c-.4-.4-.92-.59-1.43-.59-.52 0-1.04.2-1.43.59L10.3 9.45l-7.72 7.72c-.78.78-.78 2.05 0 2.83L4 21.41c.39.39.9.59 1.41.59.51 0 1.02-.2 1.41-.59l7.78-7.78 2.81-2.81c.8-.78.8-2.07 0-2.86zM5.41 20L4 18.59l7.72-7.72 1.47 1.35L5.41 20z\"/>\n",
|
191 |
+
" </svg>\n",
|
192 |
+
" </button>\n",
|
193 |
+
" \n",
|
194 |
+
" <style>\n",
|
195 |
+
" .colab-df-container {\n",
|
196 |
+
" display:flex;\n",
|
197 |
+
" flex-wrap:wrap;\n",
|
198 |
+
" gap: 12px;\n",
|
199 |
+
" }\n",
|
200 |
+
"\n",
|
201 |
+
" .colab-df-convert {\n",
|
202 |
+
" background-color: #E8F0FE;\n",
|
203 |
+
" border: none;\n",
|
204 |
+
" border-radius: 50%;\n",
|
205 |
+
" cursor: pointer;\n",
|
206 |
+
" display: none;\n",
|
207 |
+
" fill: #1967D2;\n",
|
208 |
+
" height: 32px;\n",
|
209 |
+
" padding: 0 0 0 0;\n",
|
210 |
+
" width: 32px;\n",
|
211 |
+
" }\n",
|
212 |
+
"\n",
|
213 |
+
" .colab-df-convert:hover {\n",
|
214 |
+
" background-color: #E2EBFA;\n",
|
215 |
+
" box-shadow: 0px 1px 2px rgba(60, 64, 67, 0.3), 0px 1px 3px 1px rgba(60, 64, 67, 0.15);\n",
|
216 |
+
" fill: #174EA6;\n",
|
217 |
+
" }\n",
|
218 |
+
"\n",
|
219 |
+
" [theme=dark] .colab-df-convert {\n",
|
220 |
+
" background-color: #3B4455;\n",
|
221 |
+
" fill: #D2E3FC;\n",
|
222 |
+
" }\n",
|
223 |
+
"\n",
|
224 |
+
" [theme=dark] .colab-df-convert:hover {\n",
|
225 |
+
" background-color: #434B5C;\n",
|
226 |
+
" box-shadow: 0px 1px 3px 1px rgba(0, 0, 0, 0.15);\n",
|
227 |
+
" filter: drop-shadow(0px 1px 2px rgba(0, 0, 0, 0.3));\n",
|
228 |
+
" fill: #FFFFFF;\n",
|
229 |
+
" }\n",
|
230 |
+
" </style>\n",
|
231 |
+
"\n",
|
232 |
+
" <script>\n",
|
233 |
+
" const buttonEl =\n",
|
234 |
+
" document.querySelector('#df-ea413149-b901-4dd6-a8ae-0b417a0edf47 button.colab-df-convert');\n",
|
235 |
+
" buttonEl.style.display =\n",
|
236 |
+
" google.colab.kernel.accessAllowed ? 'block' : 'none';\n",
|
237 |
+
"\n",
|
238 |
+
" async function convertToInteractive(key) {\n",
|
239 |
+
" const element = document.querySelector('#df-ea413149-b901-4dd6-a8ae-0b417a0edf47');\n",
|
240 |
+
" const dataTable =\n",
|
241 |
+
" await google.colab.kernel.invokeFunction('convertToInteractive',\n",
|
242 |
+
" [key], {});\n",
|
243 |
+
" if (!dataTable) return;\n",
|
244 |
+
"\n",
|
245 |
+
" const docLinkHtml = 'Like what you see? Visit the ' +\n",
|
246 |
+
" '<a target=\"_blank\" href=https://colab.research.google.com/notebooks/data_table.ipynb>data table notebook</a>'\n",
|
247 |
+
" + ' to learn more about interactive tables.';\n",
|
248 |
+
" element.innerHTML = '';\n",
|
249 |
+
" dataTable['output_type'] = 'display_data';\n",
|
250 |
+
" await google.colab.output.renderOutput(dataTable, element);\n",
|
251 |
+
" const docLink = document.createElement('div');\n",
|
252 |
+
" docLink.innerHTML = docLinkHtml;\n",
|
253 |
+
" element.appendChild(docLink);\n",
|
254 |
+
" }\n",
|
255 |
+
" </script>\n",
|
256 |
+
" </div>\n",
|
257 |
+
" </div>\n",
|
258 |
+
" "
|
259 |
+
],
|
260 |
+
"text/plain": [
|
261 |
+
" filename fold target category esc10 src_file take\n",
|
262 |
+
"0 1-100032-A-0.wav 1 0 dog True 100032 A\n",
|
263 |
+
"1 1-100038-A-14.wav 1 14 chirping_birds False 100038 A\n",
|
264 |
+
"2 1-100210-A-36.wav 1 36 vacuum_cleaner False 100210 A\n",
|
265 |
+
"3 1-100210-B-36.wav 1 36 vacuum_cleaner False 100210 B\n",
|
266 |
+
"4 1-101296-A-19.wav 1 19 thunderstorm False 101296 A"
|
267 |
+
]
|
268 |
+
},
|
269 |
+
"execution_count": 4,
|
270 |
+
"metadata": {},
|
271 |
+
"output_type": "execute_result"
|
272 |
+
}
|
273 |
+
],
|
274 |
+
"source": [
|
275 |
+
"#reading the csv file\n",
|
276 |
+
"data = pd.read_csv('./data/meta/esc50.csv')\n",
|
277 |
+
"data.head()"
|
278 |
+
]
|
279 |
+
},
|
280 |
+
{
|
281 |
+
"cell_type": "markdown",
|
282 |
+
"metadata": {
|
283 |
+
"id": "EsFcOZlvqf-K"
|
284 |
+
},
|
285 |
+
"source": [
|
286 |
+
"### 2. Data import & preprocessing\n",
|
287 |
+
"With the aim of replicability, the whole pipeline is implemented with the use of `np.random.seed()`."
|
288 |
+
]
|
289 |
+
},
|
290 |
+
{
|
291 |
+
"cell_type": "code",
|
292 |
+
"execution_count": 5,
|
293 |
+
"metadata": {
|
294 |
+
"colab": {
|
295 |
+
"base_uri": "https://localhost:8080/"
|
296 |
+
},
|
297 |
+
"id": "Q0aXZASmzZtM",
|
298 |
+
"outputId": "7556538e-41f2-4a0a-cb42-d1cca4d1d575"
|
299 |
+
},
|
300 |
+
"outputs": [
|
301 |
+
{
|
302 |
+
"name": "stderr",
|
303 |
+
"output_type": "stream",
|
304 |
+
"text": [
|
305 |
+
"/usr/local/lib/python3.8/dist-packages/librosa/core/pitch.py:153: UserWarning: Trying to estimate tuning from empty frequency set.\n",
|
306 |
+
" warnings.warn(\"Trying to estimate tuning from empty frequency set.\")\n",
|
307 |
+
"/usr/local/lib/python3.8/dist-packages/librosa/core/pitch.py:153: UserWarning: Trying to estimate tuning from empty frequency set.\n",
|
308 |
+
" warnings.warn(\"Trying to estimate tuning from empty frequency set.\")\n"
|
309 |
+
]
|
310 |
+
}
|
311 |
+
],
|
312 |
+
"source": [
|
313 |
+
"# DATA AUGMENTATION\n",
|
314 |
+
"\n",
|
315 |
+
"#np.random.seed(42)\n",
|
316 |
+
"#indexed_samples = np.random.choice(X.shape[0], size = 10000, replace = True)\n",
|
317 |
+
"np.random.seed(101)\n",
|
318 |
+
"randn_seeds = np.random.choice(len(data), size = len(data), replace = False)\n",
|
319 |
+
"\n",
|
320 |
+
"aug_iterations = 7\n",
|
321 |
+
"\n",
|
322 |
+
"new_X = []\n",
|
323 |
+
"#new_X2 = []\n",
|
324 |
+
"new_y = np.zeros(shape = (aug_iterations*len(randn_seeds), 1))\n",
|
325 |
+
"\n",
|
326 |
+
"input_length = 220500\n",
|
327 |
+
"row_count = 0\n",
|
328 |
+
"for i in data.index:\n",
|
329 |
+
"\n",
|
330 |
+
" sample, sr_sample = librosa.load(\"./data/audio/{}\".format(data.loc[i, \"filename\"]),\n",
|
331 |
+
" sr = 44100)\n",
|
332 |
+
" # Min-max scaler [0, 1]\n",
|
333 |
+
" sample = (sample - sample.min()) / (sample.max() - sample.min())\n",
|
334 |
+
"\n",
|
335 |
+
" if len(sample) > input_length:\n",
|
336 |
+
" sample = sample[:input_length]\n",
|
337 |
+
" else:\n",
|
338 |
+
" sample = np.pad(sample, (0, max(0, input_length - len(sample))), \"constant\")\n",
|
339 |
+
"\n",
|
340 |
+
" for n in range(aug_iterations):\n",
|
341 |
+
" \n",
|
342 |
+
" if n == 0:\n",
|
343 |
+
" # NOISE INJECTION\n",
|
344 |
+
" np.random.seed(randn_seeds[i])\n",
|
345 |
+
" noise = np.random.randn(len( sample ))\n",
|
346 |
+
" augmented_data = (sample + 0.005 * noise)\n",
|
347 |
+
"\n",
|
348 |
+
" elif n == 1:\n",
|
349 |
+
" # TIME SHIFT: right shift\n",
|
350 |
+
" augmented_data = np.roll(sample, 22050)\n",
|
351 |
+
"\n",
|
352 |
+
" elif n == 2:\n",
|
353 |
+
" # PITCH SHIFT: shift down by 3\n",
|
354 |
+
" augmented_data = librosa.effects.pitch_shift(y = sample, sr = sr_sample,\n",
|
355 |
+
" n_steps = 3)\n",
|
356 |
+
" elif n == 3:\n",
|
357 |
+
" # PITCH SHIFT: shift down by -3\n",
|
358 |
+
" augmented_data = librosa.effects.pitch_shift(y = sample, sr = sr_sample,\n",
|
359 |
+
" n_steps = -3)\n",
|
360 |
+
" elif n == 4:\n",
|
361 |
+
" # SPEED SHIFT: faster\n",
|
362 |
+
" augmented_data = librosa.effects.time_stretch(y = sample, rate = 1.25)\n",
|
363 |
+
" augmented_data = np.append(augmented_data,\n",
|
364 |
+
" np.zeros(shape = len(sample) - len(augmented_data)))\n",
|
365 |
+
" elif n == 5:\n",
|
366 |
+
" # SPEED SHIFT: slower (returns longer array)\n",
|
367 |
+
" augmented_data = librosa.effects.time_stretch(y = sample, rate = 0.8)\n",
|
368 |
+
" augmented_data = augmented_data[:len(sample)]\n",
|
369 |
+
"\n",
|
370 |
+
" else:\n",
|
371 |
+
" # KEEP NORMAL SAMPLE\n",
|
372 |
+
" augmented_data = sample\n",
|
373 |
+
"\n",
|
374 |
+
" new_instance = librosa.feature.mfcc(y = augmented_data, sr = sr_sample,\n",
|
375 |
+
" hop_length = 512, n_mfcc = 60)\n",
|
376 |
+
" \n",
|
377 |
+
" \"\"\"\n",
|
378 |
+
" For the CNN, the input is composed of three channels\n",
|
379 |
+
" stacked together as follows (commented lines).\n",
|
380 |
+
" \"\"\"\n",
|
381 |
+
" #new_MFCC = librosa.feature.mfcc(y = augmented_data, sr = sr_sample,\n",
|
382 |
+
" # hop_length = 512, n_mfcc = 60)\n",
|
383 |
+
" #new_chromagram = librosa.feature.chroma_stft(y = augmented_data, sr = sr_sample,\n",
|
384 |
+
" # hop_length = 512, win_length = 1024,\n",
|
385 |
+
" # n_chroma = 60)\n",
|
386 |
+
" #new_delta = librosa.feature.delta(new_MFCC)\n",
|
387 |
+
" \n",
|
388 |
+
" #new_instance = np.dstack((new_MFCC, new_chromagram, new_delta))\n",
|
389 |
+
"\n",
|
390 |
+
" \n",
|
391 |
+
" new_X += [new_instance]\n",
|
392 |
+
" #new_X2 += [new_instance2]\n",
|
393 |
+
" new_y[row_count] = data.loc[i, \"target\"]\n",
|
394 |
+
" \n",
|
395 |
+
" row_count += 1\n",
|
396 |
+
" \n",
|
397 |
+
" \n",
|
398 |
+
"new_X = np.array(new_X)\n",
|
399 |
+
"#new_X2 = np.array(new_X2)"
|
400 |
+
]
|
401 |
+
},
|
402 |
+
{
|
403 |
+
"cell_type": "code",
|
404 |
+
"execution_count": 6,
|
405 |
+
"metadata": {
|
406 |
+
"colab": {
|
407 |
+
"base_uri": "https://localhost:8080/"
|
408 |
+
},
|
409 |
+
"id": "kXgmb61EKq2_",
|
410 |
+
"outputId": "c4af2309-c793-4c6b-a083-864b01c71a16"
|
411 |
+
},
|
412 |
+
"outputs": [
|
413 |
+
{
|
414 |
+
"data": {
|
415 |
+
"text/plain": [
|
416 |
+
"((14000, 60, 431, 3), (14000, 1))"
|
417 |
+
]
|
418 |
+
},
|
419 |
+
"execution_count": 6,
|
420 |
+
"metadata": {},
|
421 |
+
"output_type": "execute_result"
|
422 |
+
}
|
423 |
+
],
|
424 |
+
"source": [
|
425 |
+
"new_X.shape, new_y.shape"
|
426 |
+
]
|
427 |
+
},
|
428 |
+
{
|
429 |
+
"cell_type": "code",
|
430 |
+
"execution_count": 8,
|
431 |
+
"metadata": {
|
432 |
+
"id": "paUvHcNHmVfH"
|
433 |
+
},
|
434 |
+
"outputs": [],
|
435 |
+
"source": [
|
436 |
+
"# Reduce float precision in order to decrease the size of the files\n",
|
437 |
+
"new_X = new_X.astype(np.float32)"
|
438 |
+
]
|
439 |
+
},
|
440 |
+
{
|
441 |
+
"cell_type": "code",
|
442 |
+
"execution_count": 2,
|
443 |
+
"metadata": {
|
444 |
+
"ExecuteTime": {
|
445 |
+
"end_time": "2023-02-12T22:44:44.989798Z",
|
446 |
+
"start_time": "2023-02-12T22:44:44.984746Z"
|
447 |
+
},
|
448 |
+
"colab": {
|
449 |
+
"base_uri": "https://localhost:8080/",
|
450 |
+
"height": 811
|
451 |
+
},
|
452 |
+
"id": "CWxW80DewwYQ",
|
453 |
+
"outputId": "bebc360f-9f33-48f3-8106-62d0cc0c91ee"
|
454 |
+
},
|
455 |
+
"outputs": [],
|
456 |
+
"source": [
|
457 |
+
"def data_to_parquet(arr, name):\n",
|
458 |
+
" \"\"\"\n",
|
459 |
+
" Whether it is for the CNN or the RNN,\n",
|
460 |
+
" this function provides a flattening of all the \n",
|
461 |
+
" dimensions of the array except the first\n",
|
462 |
+
" (number of samples).\n",
|
463 |
+
" \n",
|
464 |
+
" When required, the files are then imported\n",
|
465 |
+
" via the 'pandas' library and prperly reshaped.\n",
|
466 |
+
" \"\"\"\n",
|
467 |
+
" if len(arr.shape) > 2:\n",
|
468 |
+
" arr2 = arr.reshape(arr.shape[0], -1)\n",
|
469 |
+
" arr2 = pd.DataFrame(arr2)\n",
|
470 |
+
" else:\n",
|
471 |
+
" arr2 = pd.DataFrame(arr)\n",
|
472 |
+
"\n",
|
473 |
+
" arr2.columns = [str(c) for c in arr2.columns]\n",
|
474 |
+
" arr2.to_parquet(os.getcwd() + f\"/data/{name}.parquet\")\n",
|
475 |
+
" \n",
|
476 |
+
"\n",
|
477 |
+
"data_to_parquet(new_X, \"X_CNN_60x431x3_7times\")\n",
|
478 |
+
"data_to_parquet(new_y, \"y_CNN_7times\")"
|
479 |
+
]
|
480 |
+
},
|
481 |
+
{
|
482 |
+
"cell_type": "markdown",
|
483 |
+
"metadata": {
|
484 |
+
"id": "lb16Lux3deLi"
|
485 |
+
},
|
486 |
+
"source": [
|
487 |
+
"```python\n",
|
488 |
+
"# Get data for RNN\n",
|
489 |
+
"X = []\n",
|
490 |
+
"y = np.zeros(shape = (len(data), 1))\n",
|
491 |
+
"\n",
|
492 |
+
"for i in data.index:\n",
|
493 |
+
" \n",
|
494 |
+
" sample, sr_sample = librosa.load(\"./data/audio/{}\".format(data.loc[i, \"filename\"]),\n",
|
495 |
+
" sr = 44100)\n",
|
496 |
+
" \n",
|
497 |
+
" MFCC = librosa.feature.mfcc(y = sample, sr = sr_sample,\n",
|
498 |
+
" hop_length = 512, n_mfcc = 60)\n",
|
499 |
+
" \n",
|
500 |
+
" #instance = MFCC.mean(axis = 0)\n",
|
501 |
+
" \n",
|
502 |
+
" X += [MFCC]\n",
|
503 |
+
" \n",
|
504 |
+
" y[i] = data.loc[i, \"target\"]\n",
|
505 |
+
" \n",
|
506 |
+
"X = np.array(X)\n",
|
507 |
+
"```"
|
508 |
+
]
|
509 |
+
},
|
510 |
+
{
|
511 |
+
"cell_type": "markdown",
|
512 |
+
"metadata": {
|
513 |
+
"id": "kNf0QXsLg8Yz"
|
514 |
+
},
|
515 |
+
"source": [
|
516 |
+
"### Adversarial attacks"
|
517 |
+
]
|
518 |
+
},
|
519 |
+
{
|
520 |
+
"cell_type": "code",
|
521 |
+
"execution_count": null,
|
522 |
+
"metadata": {
|
523 |
+
"execution": {
|
524 |
+
"iopub.execute_input": "2023-01-14T17:37:45.772463Z",
|
525 |
+
"iopub.status.busy": "2023-01-14T17:37:45.771814Z",
|
526 |
+
"iopub.status.idle": "2023-01-14T17:37:45.787426Z",
|
527 |
+
"shell.execute_reply": "2023-01-14T17:37:45.786380Z",
|
528 |
+
"shell.execute_reply.started": "2023-01-14T17:37:45.772366Z"
|
529 |
+
},
|
530 |
+
"id": "u8gNRa0xemS-"
|
531 |
+
},
|
532 |
+
"outputs": [],
|
533 |
+
"source": [
|
534 |
+
"# create an adversarial example\n",
|
535 |
+
"def create_adversarial_example(x2, y_new, model_bidirectional):\n",
|
536 |
+
" # convert the label to a one-hot encoded vector\n",
|
537 |
+
" y = tf.keras.utils.to_categorical(y_new, num_classes=50)\n",
|
538 |
+
"# compute the gradient of the loss with respect to the input\n",
|
539 |
+
" with tf.GradientTape() as tape:\n",
|
540 |
+
" tape.watch(x2)\n",
|
541 |
+
" logits = model_bidirectional(x2)\n",
|
542 |
+
" loss_value = tf.losses.categorical_crossentropy(y_new, logits)\n",
|
543 |
+
" grads = tape.gradient(loss_value, x2)\n",
|
544 |
+
"# create an adversarial example by adding the sign of the gradient to the input\n",
|
545 |
+
" epsilon = 0.01\n",
|
546 |
+
" x_adv = x2 + epsilon * tf.sign(grads)\n",
|
547 |
+
" x_adv = tf.clip_by_value(x_adv, 0, 1)\n",
|
548 |
+
" return x_adv"
|
549 |
+
]
|
550 |
+
},
|
551 |
+
{
|
552 |
+
"cell_type": "code",
|
553 |
+
"execution_count": null,
|
554 |
+
"metadata": {
|
555 |
+
"execution": {
|
556 |
+
"iopub.execute_input": "2023-01-14T17:34:00.160306Z",
|
557 |
+
"iopub.status.busy": "2023-01-14T17:34:00.159720Z",
|
558 |
+
"iopub.status.idle": "2023-01-14T17:34:00.166335Z",
|
559 |
+
"shell.execute_reply": "2023-01-14T17:34:00.165267Z",
|
560 |
+
"shell.execute_reply.started": "2023-01-14T17:34:00.160266Z"
|
561 |
+
},
|
562 |
+
"id": "fXKsE1PzemS_"
|
563 |
+
},
|
564 |
+
"outputs": [],
|
565 |
+
"source": [
|
566 |
+
"#def create_adversarial_example(x2, y_new, model_bidirectional):\n",
|
567 |
+
" # convert the label to a one-hot encoded vector\n",
|
568 |
+
" y = tf.keras.utils.to_categorical(y_new, num_classes=20)\n",
|
569 |
+
" # compute the gradient of the loss with respect to the input\n",
|
570 |
+
" logits = model_bidirectional(x2)\n",
|
571 |
+
" loss = tf.losses.categorical_crossentropy(y_new, logits)\n",
|
572 |
+
" grads, = tf.gradients(loss, x2)\n",
|
573 |
+
" # create an adversarial example by adding the sign of the gradient to the input\n",
|
574 |
+
" epsilon = 0.01\n",
|
575 |
+
" x_adv = x2 + epsilon * tf.sign(grads)\n",
|
576 |
+
" x_adv = tf.clip_by_value(x_adv, 0, 1)\n",
|
577 |
+
" return x_adv"
|
578 |
+
]
|
579 |
+
},
|
580 |
+
{
|
581 |
+
"cell_type": "code",
|
582 |
+
"execution_count": null,
|
583 |
+
"metadata": {
|
584 |
+
"execution": {
|
585 |
+
"iopub.execute_input": "2023-01-14T20:08:50.574052Z",
|
586 |
+
"iopub.status.busy": "2023-01-14T20:08:50.573757Z",
|
587 |
+
"iopub.status.idle": "2023-01-14T20:09:00.767976Z",
|
588 |
+
"shell.execute_reply": "2023-01-14T20:09:00.766358Z",
|
589 |
+
"shell.execute_reply.started": "2023-01-14T20:08:50.574022Z"
|
590 |
+
},
|
591 |
+
"id": "Tl_qP5S6emS_"
|
592 |
+
},
|
593 |
+
"outputs": [],
|
594 |
+
"source": [
|
595 |
+
"# create an adversarial example and test it with the model\n",
|
596 |
+
"x_adv = create_adversarial_example(x2, y_new, model_bidirectional)\n",
|
597 |
+
"y_pred_adv = model_bidirectional(x_adv).argmax() # get the predicted label\n",
|
598 |
+
"acc = (y_pred_adv == y_new).mean() # calculate the accuracy\n",
|
599 |
+
"print(f'Model accuracy on adversarial example: {acc:.2f}')"
|
600 |
+
]
|
601 |
+
},
|
602 |
+
{
|
603 |
+
"cell_type": "code",
|
604 |
+
"execution_count": null,
|
605 |
+
"metadata": {
|
606 |
+
"execution": {
|
607 |
+
"iopub.execute_input": "2023-01-14T19:56:28.078039Z",
|
608 |
+
"iopub.status.busy": "2023-01-14T19:56:28.074638Z",
|
609 |
+
"iopub.status.idle": "2023-01-14T19:56:39.922249Z",
|
610 |
+
"shell.execute_reply": "2023-01-14T19:56:39.920914Z",
|
611 |
+
"shell.execute_reply.started": "2023-01-14T19:56:28.077987Z"
|
612 |
+
},
|
613 |
+
"id": "bFH3lL8UemS_"
|
614 |
+
},
|
615 |
+
"outputs": [],
|
616 |
+
"source": [
|
617 |
+
"# test the adversarial example\n",
|
618 |
+
"x_adv = create_adversarial_example(x2, y_new, model_bidirectional)\n",
|
619 |
+
"logits_adv = model_bidirectional(x_adv)\n",
|
620 |
+
"y_pred_adv = np.argmax(logits_adv, axis=1)\n",
|
621 |
+
"accuracy = accuracy_score(y_new, y_pred_adv)\n",
|
622 |
+
"print('Accuracy on adversarial example:', accuracy)"
|
623 |
+
]
|
624 |
+
}
|
625 |
+
],
|
626 |
+
"metadata": {
|
627 |
+
"colab": {
|
628 |
+
"machine_shape": "hm",
|
629 |
+
"provenance": []
|
630 |
+
},
|
631 |
+
"gpuClass": "standard",
|
632 |
+
"kernelspec": {
|
633 |
+
"display_name": "Python 3 (ipykernel)",
|
634 |
+
"language": "python",
|
635 |
+
"name": "python3"
|
636 |
+
},
|
637 |
+
"language_info": {
|
638 |
+
"codemirror_mode": {
|
639 |
+
"name": "ipython",
|
640 |
+
"version": 3
|
641 |
+
},
|
642 |
+
"file_extension": ".py",
|
643 |
+
"mimetype": "text/x-python",
|
644 |
+
"name": "python",
|
645 |
+
"nbconvert_exporter": "python",
|
646 |
+
"pygments_lexer": "ipython3",
|
647 |
+
"version": "3.9.12"
|
648 |
+
}
|
649 |
+
},
|
650 |
+
"nbformat": 4,
|
651 |
+
"nbformat_minor": 1
|
652 |
+
}
|
RNN_Architecture.ipynb
ADDED
The diff for this file is too large to render.
See raw diff
|
|
evaluation.py
ADDED
@@ -0,0 +1,99 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import seaborn as sns
|
2 |
+
import tensorflow as tf
|
3 |
+
import matplotlib.pyplot as plt
|
4 |
+
|
5 |
+
def plot_loss(history, axis = None):
|
6 |
+
"""
|
7 |
+
Parameters
|
8 |
+
----------
|
9 |
+
|
10 |
+
history : 'tf.keras.callbacks.History' object
|
11 |
+
axis : 'matplotlib.pyplot.axis' object
|
12 |
+
"""
|
13 |
+
if axis is not None:
|
14 |
+
axis.plot(history.epoch, history.history["loss"],
|
15 |
+
label = "Train loss", color = "#191970")
|
16 |
+
axis.plot(history.epoch, history.history["val_loss"],
|
17 |
+
label = "Val loss", color = "#00CC33")
|
18 |
+
axis.set_title("Loss")
|
19 |
+
axis.legend()
|
20 |
+
else:
|
21 |
+
plt.plot(history.epoch, history.history["loss"],
|
22 |
+
label = "Train loss", color = "#191970")
|
23 |
+
plt.plot(history.epoch, history.history["val_loss"],
|
24 |
+
label = "Val loss", color = "#00CC33")
|
25 |
+
plt.title("Loss")
|
26 |
+
plt.legend()
|
27 |
+
|
28 |
+
|
29 |
+
def plot_accuracy(history, axis = None):
|
30 |
+
"""
|
31 |
+
Parameters
|
32 |
+
----------
|
33 |
+
|
34 |
+
history : 'tf.keras.callbacks.History' object
|
35 |
+
axis : 'matplotlib.pyplot.axis' object
|
36 |
+
"""
|
37 |
+
if axis is not None:
|
38 |
+
axis.plot(history.epoch, history.history["accuracy"],
|
39 |
+
label = "Train accuracy", color = "#191970")
|
40 |
+
axis.plot(history.epoch, history.history["val_accuracy"],
|
41 |
+
label = "Val accuracy", color = "#00CC33")
|
42 |
+
axis.set_ylim(0, 1.1)
|
43 |
+
axis.set_title("Accuracy")
|
44 |
+
axis.legend()
|
45 |
+
else:
|
46 |
+
plt.plot(history.epoch, history.history["accuracy"],
|
47 |
+
label = "Train accuracy", color = "#191970")
|
48 |
+
plt.plot(history.epoch, history.history["val_accuracy"],
|
49 |
+
label = "Val accuracy", color = "#00CC33")
|
50 |
+
plt.title("Accuracy")
|
51 |
+
plt.ylim(0, 1.1)
|
52 |
+
plt.legend()
|
53 |
+
|
54 |
+
|
55 |
+
def keras_model_memory_usage_in_bytes(model, *, batch_size: int):
|
56 |
+
"""
|
57 |
+
Return the estimated memory usage of a given Keras model in bytes.
|
58 |
+
This includes the model weights and layers, but excludes the dataset.
|
59 |
+
|
60 |
+
The model shapes are multipled by the batch size, but the weights are not.
|
61 |
+
|
62 |
+
Parameters
|
63 |
+
----------
|
64 |
+
model: A Keras model.
|
65 |
+
batch_size: The batch size you intend to run the model with. If you
|
66 |
+
have already specified the batch size in the model itself, then
|
67 |
+
pass `1` as the argument here.
|
68 |
+
|
69 |
+
Returns
|
70 |
+
-------
|
71 |
+
An estimate of the Keras model's memory usage in bytes.
|
72 |
+
|
73 |
+
"""
|
74 |
+
default_dtype = tf.keras.backend.floatx()
|
75 |
+
shapes_mem_count = 0
|
76 |
+
internal_model_mem_count = 0
|
77 |
+
for layer in model.layers:
|
78 |
+
if isinstance(layer, tf.keras.Model):
|
79 |
+
internal_model_mem_count += keras_model_memory_usage_in_bytes(layer,
|
80 |
+
batch_size = batch_size)
|
81 |
+
single_layer_mem = tf.as_dtype(layer.dtype or default_dtype).size
|
82 |
+
out_shape = layer.output_shape
|
83 |
+
if isinstance(out_shape, list):
|
84 |
+
out_shape = out_shape[0]
|
85 |
+
for s in out_shape:
|
86 |
+
if s is None:
|
87 |
+
continue
|
88 |
+
single_layer_mem *= s
|
89 |
+
shapes_mem_count += single_layer_mem
|
90 |
+
|
91 |
+
trainable_count = sum([tf.keras.backend.count_params(p)
|
92 |
+
for p in model.trainable_weights])
|
93 |
+
non_trainable_count = sum([tf.keras.backend.count_params(p)
|
94 |
+
for p in model.non_trainable_weights])
|
95 |
+
|
96 |
+
total_memory = (batch_size * shapes_mem_count + internal_model_mem_count\
|
97 |
+
+ trainable_count + non_trainable_count)
|
98 |
+
|
99 |
+
return total_memory
|