README.md

---
license: cc-by-sa-4.0
language:
- en
tags:
- music
- spectrogram
size_categories:
- 10K<n<100K
---

# Google/MusicCapsをスペクトログラムにしたデータ。

## Dataset information
<table>
  <thead>
    <td>画像</td>
    <td>caption</td>
    <td>data_idx</td>
    <td>number</td>
  </thead>
  <tbody>
    <tr>
      <td>1025px × 216px</td>
      <td>音楽の説明</td>
      <td>どのデータから生成されたデータか</td>
      <td>5秒ずつ区切ったデータのうち、何番目か</td>
    </tr>
  </tbody>
</table>

## How this dataset was made

* コード：https://colab.research.google.com/drive/13m792FEoXszj72viZuBtusYRUL1z6Cu2?usp=sharing
* 参考にしたKaggle Notebook : https://www.kaggle.com/code/osanseviero/musiccaps-explorer

```python
from PIL import Image
import IPython.display
import cv2

# 1. wavファイルを解析
y, sr = librosa.load("wavファイルなど")

# 2. フーリエ変換を適用して周波数成分を取得
D = librosa.amplitude_to_db(np.abs(librosa.stft(y)), ref=np.max) # librosaを用いてデータを作る
image = Image.fromarray(np.uint8(D), mode='L')  # 'L'は1チャンネルのグレースケールモードを指定します
image.save('spectrogram_{}.png')
```

## Recover music(wave form) from sprctrogram
```python
im = Image.open("pngファイル")
db_ud = np.uint8(np.array(im))
amp = librosa.db_to_amplitude(db_ud)
print(amp.shape)
# (1025, 861)は20秒のwavファイルをスペクトログラムにした場合
# (1025, 431)は10秒のwavファイルをスペクトログラムにした場合
# (1025, 216)は5秒のwavファイルをスペクトログラムにした場合

y_inv = librosa.griffinlim(amp*200)
display(IPython.display.Audio(y_inv, rate=sr))
```

## Example : How to use this
* <font color="red">Subset <b>data 1300-1600</b> and <b>data 3400-3600</b> are not working now, so please get subset_name_list</n>
  those were removed first</font>.
### 1 : get information about this dataset:
```python
# Extract dataset's information using huggingface API
import requests
headers = {"Authorization": f"Bearer {API token}"}
API_URL = "https://datasets-server.huggingface.co/info?dataset=mb23%2FGraySpectrogram"
def query():
    response = requests.get(API_URL, headers=headers)
    return response.json()
data = query()

# Make subset name list.
subset__name_list = list()
for dic in data["failed"]:
    subset_name_list.append(dic["config"])
    # print(dic["config"])
subset_name_list = sorted(subset_list, key=natural_keys)


remove_list = [
    "data 1300-1600",
    "data 3400-3600"
]
for remove_dataset in remove_list:
  if remove_dataset in subset_list:
      subset_list.remove(remove_dataset)
  else:
      pass
subset_list

'''
return subset name list. for example, 
['data 0-200',
 'data 200-600',
 'data 600-1000',
 'data 1000-1300',
 'data 1600-2000',
 'data 2000-2200',
 'data 2200-2400',
 'data 2400-2600',
 'data 2600-2800',
 'data 3000-3200',
 'data 3200-3400',
 'data 3600-3800',
 'data 3800-4000',
 'data 4000-4200',
 'data 4200-4400',
 'data 4400-4600',
 'data 4600-4800',
 'data 4800-5000',
 'data 5000-5200',
 'data 5200-5520']
'''
```

### 2 : load dataset:
* 
```python
data = load_dataset("mb23/GraySpectrogram", subset_name_list[0])
for subset in subset_name_list:
    # Confirm subset_list doesn't include "remove_list" datasets in the above cell.
    print(subset)
    new_ds = load_dataset("mb23/GraySpectrogram", subset)
    new_dataset_train = datasets.concatenate_datasets([data["train"], new_ds["train"]])
    new_dataset_test = datasets.concatenate_datasets([data["test"], new_ds["test"]])

    # take place of data[split]
    data["train"] = new_dataset_train
    data["test"] = new_dataset_test

data
```


### 3 : load dataset and change to dataloader:
* You can use the code below:
* <font color="red">...but (;・∀・)I don't know whether this code works efficiently, because I haven't tried this code so far</color>
```python
import datasets
from datasets import load_dataset, DatasetDict
from torchvision import transforms
from torch.utils.data import DataLoader
# BATCH_SIZE = ??? 
# IMAGE_SIZE = ???
# TRAIN_SIZE = ??? # the number of training data
# TEST_SIZE = ??? # the number of test data

def load_datasets():

    # Define data transforms
    data_transforms = [
        transforms.Resize((IMG_SIZE, IMG_SIZE)),
        transforms.ToTensor(), # Scales data into [0,1]
        transforms.Lambda(lambda t: (t * 2) - 1) # Scale between [-1, 1]
    ]
    data_transform = transforms.Compose(data_transforms)

    data = load_dataset("mb23/GraySpectrogram", subset_list[0])
    for subset in subset_name_list:
        # Confirm subset_list doesn't include "remove_list" datasets in the above cell.
        print(subset)
        new_ds = load_dataset("mb23/GraySpectrogram", subset)
        new_dataset_train = datasets.concatenate_datasets([data["train"], new_ds["train"]])
        new_dataset_test = datasets.concatenate_datasets([data["test"], new_ds["test"]])

        # take place of data[split]
        data["train"] = new_dataset_train
        data["test"] = new_dataset_test

    # memo:
    # 特徴量上手く抽出する方法が...わからん。これは力づく。
    # 本当はload_dataset()の時点で抽出したかったけど、無理そう
    # リポジトリ作り直してpush_to_hub()したほうがいいかもしれない。

    new_dataset = dict()
    new_dataset["train"] = Dataset.from_dict({
        "image" : data["train"]["image"],
        "caption" : data["train"]["caption"]
    })
    
    new_dataset["test"] = Dataset.from_dict({
        "image" : data["test"]["image"],
        "caption" : data["test"]["caption"]
    })
    data = datasets.DatasetDict(new_dataset)
    train = data["train"]
    test = data["test"]

    for idx in range(len(train["image"])):
        train["image"][idx] = data_transform(train["image"][idx])
        test["image"][idx] = data_transform(test["image"][idx])
    
    train = Dataset.from_dict(train)
    train = train.with_format("torch") # リスト型回避
    test = Dataset.from_dict(train)
    test = test.with_format("torch") # リスト型回避

    # or
    train_loader = DataLoader(train, batch_size=BATCH_SIZE, shuffle=True, drop_last=True)
    test_loader = DataLoader(test, batch_size=BATCH_SIZE, shuffle=True, drop_last=True)
    return train_loader, test_loader

```
* then try this?
```
train_loader, test_loader = load_datasets()
```