Skip to content

Instantly share code, notes, and snippets.

@dienhoa

dienhoa/cryceleb.ipynb

Created May 7, 2023 21:19
Show Gist options
  • Save dienhoa/c391e69128ee5d5b35cc419092167741 to your computer and use it in GitHub Desktop.
Save dienhoa/c391e69128ee5d5b35cc419092167741 to your computer and use it in GitHub Desktop.
Download ZIP
Display the source blob
Display the rendered blob
Raw
cryceleb.ipynb
{
"cells": [
{
"cell_type": "code",
"execution_count": 1,
"id": "bfc2065e-6db9-4acf-92d8-215d37c02cdb",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\u001b[31mERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.\n",
"torchvision 0.12.0a0 requires torch==1.11.0a0+17540c5, but you have torch 2.0.0 which is incompatible.\n",
"torchtext 0.12.0a0 requires torch==1.11.0a0+17540c5, but you have torch 2.0.0 which is incompatible.\n",
"fastai 2.7.4 requires torch<1.12,>=1.7.0, but you have torch 2.0.0 which is incompatible.\u001b[0m\u001b[31m\n",
"\u001b[0m"
]
}
],
"source": [
"!pip install -qq speechbrain\n",
"!pip install -qq seaborn\n",
"!pip install -Uqq huggingface_hub"
]
},
{
"cell_type": "code",
"execution_count": 1,
"id": "93ac4ec0-0c4a-4e76-a8b6-0b7c5d24e985",
"metadata": {},
"outputs": [
{
"name": "stderr",
"output_type": "stream",
"text": [
"/opt/conda/lib/python3.8/site-packages/torchvision/io/image.py:13: UserWarning: Failed to load image Python extension: /opt/conda/lib/python3.8/site-packages/torchvision/image.so: undefined symbol: _ZN5torch3jit17parseSchemaOrNameERKNSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE\n",
" warn(f\"Failed to load image Python extension: {e}\")\n"
]
}
],
"source": [
"import speechbrain as sb\n",
"from speechbrain.pretrained import SpeakerRecognition, EncoderClassifier\n",
"from speechbrain.dataio.dataio import read_audio\n",
"from speechbrain.utils.metric_stats import EER\n",
"import pandas as pd\n",
"import numpy as np\n",
"import torch\n",
"import seaborn as sns\n",
"import matplotlib.pyplot as plt\n",
"from huggingface_hub import hf_hub_download\n",
"from tqdm.notebook import tqdm\n",
"import random"
]
},
{
"cell_type": "code",
"execution_count": 2,
"id": "c1349881-7dfd-4fa5-b1e2-b9e80c9cf40e",
"metadata": {},
"outputs": [
{
"data": {
"text/html": [
"<style type=\"text/css\">\n",
"#T_36771 caption {\n",
" font-size: 20px;\n",
"}\n",
"</style>\n",
"<table id=\"T_36771\">\n",
" <caption>metadata</caption>\n",
" <thead>\n",
" <tr>\n",
" <th class=\"blank level0\" >&nbsp;</th>\n",
" <th id=\"T_36771_level0_col0\" class=\"col_heading level0 col0\" >baby_id</th>\n",
" <th id=\"T_36771_level0_col1\" class=\"col_heading level0 col1\" >period</th>\n",
" <th id=\"T_36771_level0_col2\" class=\"col_heading level0 col2\" >duration</th>\n",
" <th id=\"T_36771_level0_col3\" class=\"col_heading level0 col3\" >split</th>\n",
" <th id=\"T_36771_level0_col4\" class=\"col_heading level0 col4\" >chronological_index</th>\n",
" <th id=\"T_36771_level0_col5\" class=\"col_heading level0 col5\" >file_name</th>\n",
" <th id=\"T_36771_level0_col6\" class=\"col_heading level0 col6\" >file_id</th>\n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" <tr>\n",
" <th id=\"T_36771_level0_row0\" class=\"row_heading level0 row0\" >0</th>\n",
" <td id=\"T_36771_row0_col0\" class=\"data row0 col0\" >0694</td>\n",
" <td id=\"T_36771_row0_col1\" class=\"data row0 col1\" >B</td>\n",
" <td id=\"T_36771_row0_col2\" class=\"data row0 col2\" >1.320000</td>\n",
" <td id=\"T_36771_row0_col3\" class=\"data row0 col3\" >dev</td>\n",
" <td id=\"T_36771_row0_col4\" class=\"data row0 col4\" >000</td>\n",
" <td id=\"T_36771_row0_col5\" class=\"data row0 col5\" >audio/dev/0694/B/0694_B_000.wav</td>\n",
" <td id=\"T_36771_row0_col6\" class=\"data row0 col6\" >0694_B_000</td>\n",
" </tr>\n",
" <tr>\n",
" <th id=\"T_36771_level0_row1\" class=\"row_heading level0 row1\" >1</th>\n",
" <td id=\"T_36771_row1_col0\" class=\"data row1 col0\" >0694</td>\n",
" <td id=\"T_36771_row1_col1\" class=\"data row1 col1\" >B</td>\n",
" <td id=\"T_36771_row1_col2\" class=\"data row1 col2\" >0.940000</td>\n",
" <td id=\"T_36771_row1_col3\" class=\"data row1 col3\" >dev</td>\n",
" <td id=\"T_36771_row1_col4\" class=\"data row1 col4\" >001</td>\n",
" <td id=\"T_36771_row1_col5\" class=\"data row1 col5\" >audio/dev/0694/B/0694_B_001.wav</td>\n",
" <td id=\"T_36771_row1_col6\" class=\"data row1 col6\" >0694_B_001</td>\n",
" </tr>\n",
" <tr>\n",
" <th id=\"T_36771_level0_row2\" class=\"row_heading level0 row2\" >2</th>\n",
" <td id=\"T_36771_row2_col0\" class=\"data row2 col0\" >0694</td>\n",
" <td id=\"T_36771_row2_col1\" class=\"data row2 col1\" >B</td>\n",
" <td id=\"T_36771_row2_col2\" class=\"data row2 col2\" >0.880000</td>\n",
" <td id=\"T_36771_row2_col3\" class=\"data row2 col3\" >dev</td>\n",
" <td id=\"T_36771_row2_col4\" class=\"data row2 col4\" >002</td>\n",
" <td id=\"T_36771_row2_col5\" class=\"data row2 col5\" >audio/dev/0694/B/0694_B_002.wav</td>\n",
" <td id=\"T_36771_row2_col6\" class=\"data row2 col6\" >0694_B_002</td>\n",
" </tr>\n",
" <tr>\n",
" <th id=\"T_36771_level0_row3\" class=\"row_heading level0 row3\" >3</th>\n",
" <td id=\"T_36771_row3_col0\" class=\"data row3 col0\" >0694</td>\n",
" <td id=\"T_36771_row3_col1\" class=\"data row3 col1\" >B</td>\n",
" <td id=\"T_36771_row3_col2\" class=\"data row3 col2\" >1.130000</td>\n",
" <td id=\"T_36771_row3_col3\" class=\"data row3 col3\" >dev</td>\n",
" <td id=\"T_36771_row3_col4\" class=\"data row3 col4\" >003</td>\n",
" <td id=\"T_36771_row3_col5\" class=\"data row3 col5\" >audio/dev/0694/B/0694_B_003.wav</td>\n",
" <td id=\"T_36771_row3_col6\" class=\"data row3 col6\" >0694_B_003</td>\n",
" </tr>\n",
" <tr>\n",
" <th id=\"T_36771_level0_row4\" class=\"row_heading level0 row4\" >4</th>\n",
" <td id=\"T_36771_row4_col0\" class=\"data row4 col0\" >0694</td>\n",
" <td id=\"T_36771_row4_col1\" class=\"data row4 col1\" >B</td>\n",
" <td id=\"T_36771_row4_col2\" class=\"data row4 col2\" >1.180000</td>\n",
" <td id=\"T_36771_row4_col3\" class=\"data row4 col3\" >dev</td>\n",
" <td id=\"T_36771_row4_col4\" class=\"data row4 col4\" >004</td>\n",
" <td id=\"T_36771_row4_col5\" class=\"data row4 col5\" >audio/dev/0694/B/0694_B_004.wav</td>\n",
" <td id=\"T_36771_row4_col6\" class=\"data row4 col6\" >0694_B_004</td>\n",
" </tr>\n",
" </tbody>\n",
"</table>\n"
],
"text/plain": [
"<pandas.io.formats.style.Styler at 0x7fe7a30fd9a0>"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"data": {
"text/html": [
"<style type=\"text/css\">\n",
"#T_e6b97 caption {\n",
" font-size: 20px;\n",
"}\n",
"</style>\n",
"<table id=\"T_e6b97\">\n",
" <caption>dev_pairs</caption>\n",
" <thead>\n",
" <tr>\n",
" <th class=\"blank level0\" >&nbsp;</th>\n",
" <th id=\"T_e6b97_level0_col0\" class=\"col_heading level0 col0\" >baby_id_B</th>\n",
" <th id=\"T_e6b97_level0_col1\" class=\"col_heading level0 col1\" >baby_id_D</th>\n",
" <th id=\"T_e6b97_level0_col2\" class=\"col_heading level0 col2\" >id</th>\n",
" <th id=\"T_e6b97_level0_col3\" class=\"col_heading level0 col3\" >label</th>\n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" <tr>\n",
" <th id=\"T_e6b97_level0_row0\" class=\"row_heading level0 row0\" >0</th>\n",
" <td id=\"T_e6b97_row0_col0\" class=\"data row0 col0\" >0133</td>\n",
" <td id=\"T_e6b97_row0_col1\" class=\"data row0 col1\" >0611</td>\n",
" <td id=\"T_e6b97_row0_col2\" class=\"data row0 col2\" >0133B_0611D</td>\n",
" <td id=\"T_e6b97_row0_col3\" class=\"data row0 col3\" >0</td>\n",
" </tr>\n",
" <tr>\n",
" <th id=\"T_e6b97_level0_row1\" class=\"row_heading level0 row1\" >1</th>\n",
" <td id=\"T_e6b97_row1_col0\" class=\"data row1 col0\" >0593</td>\n",
" <td id=\"T_e6b97_row1_col1\" class=\"data row1 col1\" >0584</td>\n",
" <td id=\"T_e6b97_row1_col2\" class=\"data row1 col2\" >0593B_0584D</td>\n",
" <td id=\"T_e6b97_row1_col3\" class=\"data row1 col3\" >0</td>\n",
" </tr>\n",
" <tr>\n",
" <th id=\"T_e6b97_level0_row2\" class=\"row_heading level0 row2\" >2</th>\n",
" <td id=\"T_e6b97_row2_col0\" class=\"data row2 col0\" >0094</td>\n",
" <td id=\"T_e6b97_row2_col1\" class=\"data row2 col1\" >0292</td>\n",
" <td id=\"T_e6b97_row2_col2\" class=\"data row2 col2\" >0094B_0292D</td>\n",
" <td id=\"T_e6b97_row2_col3\" class=\"data row2 col3\" >0</td>\n",
" </tr>\n",
" <tr>\n",
" <th id=\"T_e6b97_level0_row3\" class=\"row_heading level0 row3\" >3</th>\n",
" <td id=\"T_e6b97_row3_col0\" class=\"data row3 col0\" >0563</td>\n",
" <td id=\"T_e6b97_row3_col1\" class=\"data row3 col1\" >0094</td>\n",
" <td id=\"T_e6b97_row3_col2\" class=\"data row3 col2\" >0563B_0094D</td>\n",
" <td id=\"T_e6b97_row3_col3\" class=\"data row3 col3\" >0</td>\n",
" </tr>\n",
" <tr>\n",
" <th id=\"T_e6b97_level0_row4\" class=\"row_heading level0 row4\" >4</th>\n",
" <td id=\"T_e6b97_row4_col0\" class=\"data row4 col0\" >0122</td>\n",
" <td id=\"T_e6b97_row4_col1\" class=\"data row4 col1\" >0694</td>\n",
" <td id=\"T_e6b97_row4_col2\" class=\"data row4 col2\" >0122B_0694D</td>\n",
" <td id=\"T_e6b97_row4_col3\" class=\"data row4 col3\" >0</td>\n",
" </tr>\n",
" </tbody>\n",
"</table>\n"
],
"text/plain": [
"<pandas.io.formats.style.Styler at 0x7fe7f7a24e80>"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"data": {
"text/html": [
"<style type=\"text/css\">\n",
"#T_d5c4c caption {\n",
" font-size: 20px;\n",
"}\n",
"</style>\n",
"<table id=\"T_d5c4c\">\n",
" <caption>test_pairs</caption>\n",
" <thead>\n",
" <tr>\n",
" <th class=\"blank level0\" >&nbsp;</th>\n",
" <th id=\"T_d5c4c_level0_col0\" class=\"col_heading level0 col0\" >baby_id_B</th>\n",
" <th id=\"T_d5c4c_level0_col1\" class=\"col_heading level0 col1\" >baby_id_D</th>\n",
" <th id=\"T_d5c4c_level0_col2\" class=\"col_heading level0 col2\" >id</th>\n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" <tr>\n",
" <th id=\"T_d5c4c_level0_row0\" class=\"row_heading level0 row0\" >0</th>\n",
" <td id=\"T_d5c4c_row0_col0\" class=\"data row0 col0\" >anonymous027</td>\n",
" <td id=\"T_d5c4c_row0_col1\" class=\"data row0 col1\" >anonymous212</td>\n",
" <td id=\"T_d5c4c_row0_col2\" class=\"data row0 col2\" >anonymous027B_anonymous212D</td>\n",
" </tr>\n",
" <tr>\n",
" <th id=\"T_d5c4c_level0_row1\" class=\"row_heading level0 row1\" >1</th>\n",
" <td id=\"T_d5c4c_row1_col0\" class=\"data row1 col0\" >anonymous035</td>\n",
" <td id=\"T_d5c4c_row1_col1\" class=\"data row1 col1\" >anonymous225</td>\n",
" <td id=\"T_d5c4c_row1_col2\" class=\"data row1 col2\" >anonymous035B_anonymous225D</td>\n",
" </tr>\n",
" <tr>\n",
" <th id=\"T_d5c4c_level0_row2\" class=\"row_heading level0 row2\" >2</th>\n",
" <td id=\"T_d5c4c_row2_col0\" class=\"data row2 col0\" >anonymous029</td>\n",
" <td id=\"T_d5c4c_row2_col1\" class=\"data row2 col1\" >anonymous288</td>\n",
" <td id=\"T_d5c4c_row2_col2\" class=\"data row2 col2\" >anonymous029B_anonymous288D</td>\n",
" </tr>\n",
" <tr>\n",
" <th id=\"T_d5c4c_level0_row3\" class=\"row_heading level0 row3\" >3</th>\n",
" <td id=\"T_d5c4c_row3_col0\" class=\"data row3 col0\" >anonymous001</td>\n",
" <td id=\"T_d5c4c_row3_col1\" class=\"data row3 col1\" >anonymous204</td>\n",
" <td id=\"T_d5c4c_row3_col2\" class=\"data row3 col2\" >anonymous001B_anonymous204D</td>\n",
" </tr>\n",
" <tr>\n",
" <th id=\"T_d5c4c_level0_row4\" class=\"row_heading level0 row4\" >4</th>\n",
" <td id=\"T_d5c4c_row4_col0\" class=\"data row4 col0\" >anonymous075</td>\n",
" <td id=\"T_d5c4c_row4_col1\" class=\"data row4 col1\" >anonymous244</td>\n",
" <td id=\"T_d5c4c_row4_col2\" class=\"data row4 col2\" >anonymous075B_anonymous244D</td>\n",
" </tr>\n",
" </tbody>\n",
"</table>\n"
],
"text/plain": [
"<pandas.io.formats.style.Styler at 0x7fe7a30fd9a0>"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"data": {
"text/html": [
"<style type=\"text/css\">\n",
"#T_0c218 caption {\n",
" font-size: 20px;\n",
"}\n",
"</style>\n",
"<table id=\"T_0c218\">\n",
" <caption>sample_submission</caption>\n",
" <thead>\n",
" <tr>\n",
" <th class=\"blank level0\" >&nbsp;</th>\n",
" <th id=\"T_0c218_level0_col0\" class=\"col_heading level0 col0\" >id</th>\n",
" <th id=\"T_0c218_level0_col1\" class=\"col_heading level0 col1\" >score</th>\n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" <tr>\n",
" <th id=\"T_0c218_level0_row0\" class=\"row_heading level0 row0\" >0</th>\n",
" <td id=\"T_0c218_row0_col0\" class=\"data row0 col0\" >anonymous027B_anonymous212D</td>\n",
" <td id=\"T_0c218_row0_col1\" class=\"data row0 col1\" >0.548814</td>\n",
" </tr>\n",
" <tr>\n",
" <th id=\"T_0c218_level0_row1\" class=\"row_heading level0 row1\" >1</th>\n",
" <td id=\"T_0c218_row1_col0\" class=\"data row1 col0\" >anonymous035B_anonymous225D</td>\n",
" <td id=\"T_0c218_row1_col1\" class=\"data row1 col1\" >0.715189</td>\n",
" </tr>\n",
" <tr>\n",
" <th id=\"T_0c218_level0_row2\" class=\"row_heading level0 row2\" >2</th>\n",
" <td id=\"T_0c218_row2_col0\" class=\"data row2 col0\" >anonymous029B_anonymous288D</td>\n",
" <td id=\"T_0c218_row2_col1\" class=\"data row2 col1\" >0.602763</td>\n",
" </tr>\n",
" <tr>\n",
" <th id=\"T_0c218_level0_row3\" class=\"row_heading level0 row3\" >3</th>\n",
" <td id=\"T_0c218_row3_col0\" class=\"data row3 col0\" >anonymous001B_anonymous204D</td>\n",
" <td id=\"T_0c218_row3_col1\" class=\"data row3 col1\" >0.544883</td>\n",
" </tr>\n",
" <tr>\n",
" <th id=\"T_0c218_level0_row4\" class=\"row_heading level0 row4\" >4</th>\n",
" <td id=\"T_0c218_row4_col0\" class=\"data row4 col0\" >anonymous075B_anonymous244D</td>\n",
" <td id=\"T_0c218_row4_col1\" class=\"data row4 col1\" >0.423655</td>\n",
" </tr>\n",
" </tbody>\n",
"</table>\n"
],
"text/plain": [
"<pandas.io.formats.style.Styler at 0x7fe7f7a24e80>"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"# read metadata\n",
"metadata = pd.read_csv(f'metadata.csv', dtype={'baby_id':str, 'chronological_index':str})\n",
"dev_metadata = metadata.loc[metadata['split']=='dev'].copy()\n",
"# read sample submission\n",
"sample_submission = pd.read_csv(f\"sample_submission.csv\") # scores are unfiorm random\n",
"# read verification pairs\n",
"dev_pairs = pd.read_csv(f\"dev_pairs.csv\", dtype={'baby_id_B':str, 'baby_id_D':str})\n",
"test_pairs = pd.read_csv(f\"test_pairs.csv\")\n",
"\n",
"display(metadata.head().style.set_caption(f\"metadata\").set_table_styles([{'selector': 'caption','props': [('font-size', '20px')]}]))\n",
"display(dev_pairs.head().style.set_caption(f\"dev_pairs\").set_table_styles([{'selector': 'caption','props': [('font-size', '20px')]}]))\n",
"display(test_pairs.head().style.set_caption(f\"test_pairs\").set_table_styles([{'selector': 'caption','props': [('font-size', '20px')]}]))\n",
"display(sample_submission.head().style.set_caption(f\"sample_submission\").set_table_styles([{'selector': 'caption','props': [('font-size', '20px')]}]))"
]
},
{
"cell_type": "code",
"execution_count": 3,
"id": "9896df28-2a89-4a6b-97fb-f928553f6836",
"metadata": {},
"outputs": [],
"source": [
"encoder = SpeakerRecognition.from_hparams(\n",
" source=\"Ubenwa/ecapa-voxceleb-ft-cryceleb\",\n",
" savedir=f\"ecapa-voxceleb-ft-cryceleb\",\n",
" run_opts={\"device\":\"cuda\"} #comment out if no GPU available\n",
")"
]
},
{
"cell_type": "code",
"execution_count": 4,
"id": "61824528-fb66-4201-835f-9ef80d02d115",
"metadata": {},
"outputs": [],
"source": [
"def shuffle_group_and_concat(x, n=5):\n",
" concatenated_results = []\n",
" for _ in range(n):\n",
" shuffled_values = x.values.copy()\n",
" random.shuffle(shuffled_values)\n",
" concatenated = np.concatenate(shuffled_values)\n",
" concatenated_results.append(concatenated)\n",
" return concatenated_results"
]
},
{
"cell_type": "code",
"execution_count": 5,
"id": "175d6912-30f2-4678-93e7-96caa032f89b",
"metadata": {},
"outputs": [],
"source": [
"def compute_cosine_similarity_score(row, cry_dict):\n",
" \"\"\" Average scores for all possible pairs \"\"\"\n",
" cos = torch.nn.CosineSimilarity(dim=-1)\n",
" encoded_cry_B = cry_dict[(row['baby_id_B'], 'B')]['cry_encoded']\n",
" encoded_cry_D = cry_dict[(row['baby_id_D'], 'D')]['cry_encoded']\n",
" \n",
" similarity_scores = []\n",
" for tensor_B in encoded_cry_B:\n",
" for tensor_D in encoded_cry_D:\n",
" similarity_score = cos(tensor_B, tensor_D)\n",
" similarity_scores.append(similarity_score.item())\n",
" \n",
" return sum(similarity_scores) / len(similarity_scores)"
]
},
{
"cell_type": "code",
"execution_count": 6,
"id": "b4358175-b830-44dd-8627-6ecf4aba7b4d",
"metadata": {},
"outputs": [
{
"data": {
"application/vnd.jupyter.widget-view+json": {
"model_id": "4c6d959ad6cc4cc3baeb852068f645cb",
"version_major": 2,
"version_minor": 0
},
"text/plain": [
" 0%| | 0/80 [00:00<?, ?it/s]"
]
},
"metadata": {},
"output_type": "display_data"
},
{
"name": "stderr",
"output_type": "stream",
"text": [
"/opt/conda/lib/python3.8/site-packages/torch/functional.py:641: UserWarning: stft with return_complex=False is deprecated. In a future pytorch release, stft will return complex tensors for all inputs, and return_complex=False will raise an error.\n",
"Note: you can still call torch.view_as_real on the complex output to recover the old return format. (Triggered internally at ../aten/src/ATen/native/SpectralOps.cpp:862.)\n",
" return _VF.stft(input, n_fft, hop_length, win_length, window, # type: ignore[attr-defined]\n"
]
}
],
"source": [
"dev_metadata = metadata.loc[metadata['split']=='dev'].copy()\n",
"dev_metadata['cry'] = dev_metadata.apply(lambda row: read_audio(row['file_name']).numpy(), axis=1)\n",
"grouped_data = dev_metadata.groupby(['baby_id', 'period'])['cry']\n",
"cry_dict = {}\n",
"for key, group in grouped_data:\n",
" cry_dict[key] = {'cry': shuffle_group_and_concat(group, 5)}\n",
" \n",
"for (baby_id, period), d in tqdm(cry_dict.items()):\n",
" cry_array = d['cry']\n",
" cry_encoded_list = []\n",
"\n",
" for row in cry_array:\n",
" encoded_row = encoder.encode_batch(torch.tensor(row), normalize=False)\n",
" cry_encoded_list.append(encoded_row)\n",
"\n",
" d['cry_encoded'] = cry_encoded_list\n",
"dev_pairs['score'] = dev_pairs.apply(lambda row: compute_cosine_similarity_score(row=row, cry_dict=cry_dict), axis=1)"
]
},
{
"cell_type": "code",
"execution_count": 7,
"id": "a1dc1403-aa80-44f6-a063-2e4134eec9d1",
"metadata": {},
"outputs": [
{
"data": {
"image/png": "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\n",
"text/plain": [
"<Figure size 432x288 with 1 Axes>"
]
},
"metadata": {
"needs_background": "light"
},
"output_type": "display_data"
}
],
"source": [
"def compute_eer_and_plot_verification_scores(pairs_df):\n",
" ''' pairs_df must have 'score' and 'label' columns'''\n",
" positive_scores = pairs_df.loc[pairs_df['label']==1]['score'].values\n",
" negative_scores = pairs_df.loc[pairs_df['label']==0]['score'].values\n",
" eer, threshold = EER(torch.tensor(positive_scores), torch.tensor(negative_scores))\n",
" ax = sns.histplot(pairs_df, x='score', hue='label', stat='percent', common_norm=False)\n",
" ax.set_title(f'EER={round(eer, 4)} - Thresh={round(threshold, 4)}')\n",
" plt.axvline(x=[threshold], color='red', ls='--');\n",
" return eer, threshold\n",
"eer, threshold = compute_eer_and_plot_verification_scores(pairs_df=dev_pairs)"
]
},
{
"cell_type": "markdown",
"id": "6d371a14-22b0-4e71-9edf-b7a2b76567c5",
"metadata": {},
"source": [
"## Test"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "e1ad25d3-7757-4842-a61e-275c7c0ca222",
"metadata": {},
"outputs": [
{
"data": {
"application/vnd.jupyter.widget-view+json": {
"model_id": "833ed79f6b2e48d799c190cee88b0dc2",
"version_major": 2,
"version_minor": 0
},
"text/plain": [
" 0%| | 0/320 [00:00<?, ?it/s]"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"test_metadata = metadata.loc[metadata['split']=='test'].copy()\n",
"test_metadata['cry'] = test_metadata.apply(lambda row: read_audio(row['file_name']).numpy(), axis=1)\n",
"grouped_data = test_metadata.groupby(['baby_id', 'period'])['cry']\n",
"cry_dict_test = {}\n",
"for key, group in grouped_data:\n",
" cry_dict_test[key] = {'cry': shuffle_group_and_concat(group, 5)}\n",
"for (baby_id, period), d in tqdm(cry_dict_test.items()):\n",
" cry_array = d['cry']\n",
" cry_encoded_list = []\n",
"\n",
" for row in cry_array:\n",
" encoded_row = encoder.encode_batch(torch.tensor(row), normalize=False)\n",
" cry_encoded_list.append(encoded_row)\n",
"\n",
" d['cry_encoded'] = cry_encoded_list\n",
"test_pairs['score'] = test_pairs.apply(lambda row: compute_cosine_similarity_score(row=row, cry_dict=cry_dict_test), axis=1)"
]
},
{
"cell_type": "code",
"execution_count": 90,
"id": "206bb9f2-528a-4fa2-9a54-08328c605d0c",
"metadata": {},
"outputs": [
{
"data": {
"text/html": [
"<div>\n",
"<style scoped>\n",
" .dataframe tbody tr th:only-of-type {\n",
" vertical-align: middle;\n",
" }\n",
"\n",
" .dataframe tbody tr th {\n",
" vertical-align: top;\n",
" }\n",
"\n",
" .dataframe thead th {\n",
" text-align: right;\n",
" }\n",
"</style>\n",
"<table border=\"1\" class=\"dataframe\">\n",
" <thead>\n",
" <tr style=\"text-align: right;\">\n",
" <th></th>\n",
" <th>baby_id_B</th>\n",
" <th>baby_id_D</th>\n",
" <th>id</th>\n",
" <th>score</th>\n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" <tr>\n",
" <th>0</th>\n",
" <td>anonymous027</td>\n",
" <td>anonymous212</td>\n",
" <td>anonymous027B_anonymous212D</td>\n",
" <td>-0.149320</td>\n",
" </tr>\n",
" <tr>\n",
" <th>1</th>\n",
" <td>anonymous035</td>\n",
" <td>anonymous225</td>\n",
" <td>anonymous035B_anonymous225D</td>\n",
" <td>0.058973</td>\n",
" </tr>\n",
" <tr>\n",
" <th>2</th>\n",
" <td>anonymous029</td>\n",
" <td>anonymous288</td>\n",
" <td>anonymous029B_anonymous288D</td>\n",
" <td>0.010437</td>\n",
" </tr>\n",
" <tr>\n",
" <th>3</th>\n",
" <td>anonymous001</td>\n",
" <td>anonymous204</td>\n",
" <td>anonymous001B_anonymous204D</td>\n",
" <td>-0.087753</td>\n",
" </tr>\n",
" <tr>\n",
" <th>4</th>\n",
" <td>anonymous075</td>\n",
" <td>anonymous244</td>\n",
" <td>anonymous075B_anonymous244D</td>\n",
" <td>0.040881</td>\n",
" </tr>\n",
" </tbody>\n",
"</table>\n",
"</div>"
],
"text/plain": [
" baby_id_B baby_id_D id score\n",
"0 anonymous027 anonymous212 anonymous027B_anonymous212D -0.149320\n",
"1 anonymous035 anonymous225 anonymous035B_anonymous225D 0.058973\n",
"2 anonymous029 anonymous288 anonymous029B_anonymous288D 0.010437\n",
"3 anonymous001 anonymous204 anonymous001B_anonymous204D -0.087753\n",
"4 anonymous075 anonymous244 anonymous075B_anonymous244D 0.040881"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"display(test_pairs.head())"
]
},
{
"cell_type": "code",
"execution_count": 91,
"id": "23404c1f-a4ee-46cc-9a13-65aa2fd057db",
"metadata": {},
"outputs": [
{
"data": {
"text/html": [
"<div>\n",
"<style scoped>\n",
" .dataframe tbody tr th:only-of-type {\n",
" vertical-align: middle;\n",
" }\n",
"\n",
" .dataframe tbody tr th {\n",
" vertical-align: top;\n",
" }\n",
"\n",
" .dataframe thead th {\n",
" text-align: right;\n",
" }\n",
"</style>\n",
"<table border=\"1\" class=\"dataframe\">\n",
" <thead>\n",
" <tr style=\"text-align: right;\">\n",
" <th></th>\n",
" <th>id</th>\n",
" <th>score</th>\n",
" </tr>\n",
" </thead>\n",
" <tbody>\n",
" <tr>\n",
" <th>0</th>\n",
" <td>anonymous027B_anonymous212D</td>\n",
" <td>-0.149320</td>\n",
" </tr>\n",
" <tr>\n",
" <th>1</th>\n",
" <td>anonymous035B_anonymous225D</td>\n",
" <td>0.058973</td>\n",
" </tr>\n",
" <tr>\n",
" <th>2</th>\n",
" <td>anonymous029B_anonymous288D</td>\n",
" <td>0.010437</td>\n",
" </tr>\n",
" <tr>\n",
" <th>3</th>\n",
" <td>anonymous001B_anonymous204D</td>\n",
" <td>-0.087753</td>\n",
" </tr>\n",
" <tr>\n",
" <th>4</th>\n",
" <td>anonymous075B_anonymous244D</td>\n",
" <td>0.040881</td>\n",
" </tr>\n",
" </tbody>\n",
"</table>\n",
"</div>"
],
"text/plain": [
" id score\n",
"0 anonymous027B_anonymous212D -0.149320\n",
"1 anonymous035B_anonymous225D 0.058973\n",
"2 anonymous029B_anonymous288D 0.010437\n",
"3 anonymous001B_anonymous204D -0.087753\n",
"4 anonymous075B_anonymous244D 0.040881"
]
},
"metadata": {},
"output_type": "display_data"
}
],
"source": [
"#submission must match the 'sample_submission.csv' format exactly\n",
"my_submission= test_pairs[['id', 'score']]\n",
"my_submission.to_csv('my_submission.csv', index=False)\n",
"display(my_submission.head())"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "330d31a0-469b-4cfc-bbb3-c00e6c7970f2",
"metadata": {},
"outputs": [],
"source": []
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3 (ipykernel)",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.8.12"
}
},
"nbformat": 4,
"nbformat_minor": 5
}
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment