![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F16438831%2Fea0ee012f4482ab49855fc293197aed4%2Finter.JPG?generation=1754887244898651&alt=media\")
On average, each sequence has a length of 69, with considerable variability extending up to approximately 700.
\nTo standardize the input length, interpolation is applied to resample all sequences to a fixed length of 1024
\n
In the previous step, the sequence length was arbitrarily fixed at 1024. However, to generate a Mel Spectrogram, an appropriate sampling ratio is required. Since the competition dataset does not provide this sampling ratio, it can be set to one of the powers of two: 32, 64, 128 or 256
\nBecause the spectrogram contains frequency information up to half of the actual sampling rate:
\nAnalysis of the original signals reveals that the frequency band is relatively low. Therefore, by gradually increasing the sampling ratio and searching for an optimal value that sufficiently captures the frequency information in the original signal, it was confirmed that beyond a sampling ratio of 128, changes in the spectrogram information become negligible.
\n| Sampling Ratio | \nFrequency Coverage | \nEffect on Spectrogram | \nNotes | \n
|---|---|---|---|
| 32 | \nVery low | \nSignificant loss of high-frequency components | \nMay miss important details | \n
| 64 | \nLow | \nSome high-frequency info retained | \nStill somewhat limited | \n
| 128 | \nModerate | \nMost relevant frequency info captured | \nOptimal minimum threshold | \n
| 256 | \nHigh | \nMinimal change compared to 128 | \nSlightly more detail | \n
| 512 | \nVery high | \nExcessive empty space in spectrogram | \nDiminishing returns | \n
We utilize the librosa library to generate spectrograms. For each unique sequence_id, a single spectrogram is created from a sequence with a fixed length of 2048. The resulting spectrograms have dimensions of 128 (frequency bins) by 192 (time frames).
The core function used is:
\nmel_spec = librosa.feature.melspectrogram(\n y=x,\n sr=,\n hop_length=(x)//,\n n_fft=,\n n_mels=,\n fmin=,\n fmax=,\n win_length=\n)\ny represents the input time-series signal.sr denotes the sampling frequency; in this competition, EEG signals are sampled at 200 Hz.hop_length determines the number of samples between successive frames, effectively controlling the width of the output spectrogram (width = len(x)/hop_length).n_fft defines the window size for the Fourier transform, thus influencing the vertical resolution and overall quality of the spectrogram.n_mels specifies the number of Mel frequency bands, setting the spectrogram height (height = n_mels).fmin and fmax set the minimum and maximum frequency bounds captured by the spectrogram.win_length controls the length of the windowed signal segment used for each Fourier transform, affecting the horizontal resolution and spectral quality.win_lengthA shorter win_length relative to n_fft results in less tapering of the actual signal and more zero-padding. This causes subtle differences in spectral leakage patterns. To avoid the complexity of additional zero-padding, it is common practice to set win_length equal to n_fft.
hop_lengthhop_length (e.g., 256) leads to greater overlap between frames, producing smoother transitions along the time axis and better capturing rapid transient events (such as chirps). However, this increases the number of frames and computational cost.hop_length (e.g., 1024) reduces overlap, yielding fewer frames and lower time resolution, which may miss quick temporal changes.fmin and fmax)Setting fmin to 50 Hz excludes low-frequency noise such as mechanical rumble or microphone handling artifacts. Setting fmax to 14,000 Hz filters out very high-frequency noise sources (e.g., insect sounds or wind artifacts), thereby focusing the analysis on the most relevant frequency range.
n_mels)Importantly, if the number of Mel bands (n_mels) equals or exceeds the maximum frequency parameter (fmax), the frequency resolution becomes unrealistic, leading to distortion or loss of meaningful spectral information.
`import librosa
\ndef spectrum_from_imu(sequence, sr=CFG.sampling_rate):
\n = np.zeros((, , len(acc_list)),dtype='float32') \n = np.zeros((, , len(rot_list)), dtype='float32')\n\n i, col in enumerate(acc_list):\n = sequence[col].values\n \n = inter_signal(x)\n \n = librosa.feature.melspectrogram(y=x, sr=sr, hop_length=len(x)//,\n =, n_mels=, fmin=, fmax=, win_length=)\n \n = (mel_spec.shape[]//)* \n\n \n = librosa.power_to_db(mel_spec, ref=np.max).astype(np.float32)[:,:width]\n = (mel_spec_db+)/\n [:,:,i] = mel_spec_db\n\n i, col in enumerate(rot_list):\n = sequence[col].values\n \n = inter_signal(x)\n\n \n = librosa.feature.melspectrogram(y=x, sr=sr, hop_length=len(x)//, \n =, n_mels=, fmin=, fmax=, win_length=)\n\n = (mel_spec.shape[]//)* \n \n = librosa.power_to_db(mel_spec, ref=np.max).astype(np.float32)[:,:width]\n\n \n = (mel_spec_db+)/\n [:,:,i] = mel_spec_db\n\n img_acc, img_rot`\nThe IMU device records a total of seven features: acc_x, acc_y, acc_z, rot_x, rot_y, rot_z, and rot_w.
\nOur objective is to generate spectrograms for each of these seven features.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F16438831%2Fb8cb992e4ec6c046ed462e4f259b938c%2Fspec.JPG?generation=1754887883454088&alt=media\")
At the top, the four rotation features (rot_x, rot_y, rot_z, rot_w) are concatenated sequentially along the axis=1 dimension.
\nAt the bottom, the three acceleration features (acc_x, acc_y, acc_z) are concatenated sequentially along the axis=1 dimension as well.
Subsequently, since the input to the image classification model requires three channels, the identical spectrogram image is concatenated three times along the channel axis (axis=-1).
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F16438831%2Fbb4292eebb5e4f1027bc913b8f34a590%2Fspec2.JPG?generation=1754887915460130&alt=media\")
On average, each sequence has a length of 69, with considerable variability extending up to approximately 700.
\nTo standardize the input length, interpolation is applied to resample all sequences to a fixed length of 1024
\n
In the previous step, the sequence length was arbitrarily fixed at 1024. However, to generate a Mel Spectrogram, an appropriate sampling ratio is required. Since the competition dataset does not provide this sampling ratio, it can be set to one of the powers of two: 32, 64, 128 or 256
\nBecause the spectrogram contains frequency information up to half of the actual sampling rate:
\nAnalysis of the original signals reveals that the frequency band is relatively low. Therefore, by gradually increasing the sampling ratio and searching for an optimal value that sufficiently captures the frequency information in the original signal, it was confirmed that beyond a sampling ratio of 128, changes in the spectrogram information become negligible.
\n| Sampling Ratio | \nFrequency Coverage | \nEffect on Spectrogram | \nNotes | \n
|---|---|---|---|
| 32 | \nVery low | \nSignificant loss of high-frequency components | \nMay miss important details | \n
| 64 | \nLow | \nSome high-frequency info retained | \nStill somewhat limited | \n
| 128 | \nModerate | \nMost relevant frequency info captured | \nOptimal minimum threshold | \n
| 256 | \nHigh | \nMinimal change compared to 128 | \nSlightly more detail | \n
| 512 | \nVery high | \nExcessive empty space in spectrogram | \nDiminishing returns | \n
We utilize the librosa library to generate spectrograms. For each unique sequence_id, a single spectrogram is created from a sequence with a fixed length of 2048. The resulting spectrograms have dimensions of 128 (frequency bins) by 192 (time frames).
The core function used is:
\nmel_spec = librosa.feature.melspectrogram(\n y=x,\n sr=,\n hop_length=(x)//,\n n_fft=,\n n_mels=,\n fmin=,\n fmax=,\n win_length=\n)\ny represents the input time-series signal.sr denotes the sampling frequency; in this competition, EEG signals are sampled at 200 Hz.hop_length determines the number of samples between successive frames, effectively controlling the width of the output spectrogram (width = len(x)/hop_length).n_fft defines the window size for the Fourier transform, thus influencing the vertical resolution and overall quality of the spectrogram.n_mels specifies the number of Mel frequency bands, setting the spectrogram height (height = n_mels).fmin and fmax set the minimum and maximum frequency bounds captured by the spectrogram.win_length controls the length of the windowed signal segment used for each Fourier transform, affecting the horizontal resolution and spectral quality.win_lengthA shorter win_length relative to n_fft results in less tapering of the actual signal and more zero-padding. This causes subtle differences in spectral leakage patterns. To avoid the complexity of additional zero-padding, it is common practice to set win_length equal to n_fft.
hop_lengthhop_length (e.g., 256) leads to greater overlap between frames, producing smoother transitions along the time axis and better capturing rapid transient events (such as chirps). However, this increases the number of frames and computational cost.hop_length (e.g., 1024) reduces overlap, yielding fewer frames and lower time resolution, which may miss quick temporal changes.fmin and fmax)Setting fmin to 50 Hz excludes low-frequency noise such as mechanical rumble or microphone handling artifacts. Setting fmax to 14,000 Hz filters out very high-frequency noise sources (e.g., insect sounds or wind artifacts), thereby focusing the analysis on the most relevant frequency range.
n_mels)Importantly, if the number of Mel bands (n_mels) equals or exceeds the maximum frequency parameter (fmax), the frequency resolution becomes unrealistic, leading to distortion or loss of meaningful spectral information.
`import librosa
\ndef spectrum_from_imu(sequence, sr=CFG.sampling_rate):
\n = np.zeros((, , len(acc_list)),dtype='float32') \n = np.zeros((, , len(rot_list)), dtype='float32')\n\n i, col in enumerate(acc_list):\n = sequence[col].values\n \n = inter_signal(x)\n \n = librosa.feature.melspectrogram(y=x, sr=sr, hop_length=len(x)//,\n =, n_mels=, fmin=, fmax=, win_length=)\n \n = (mel_spec.shape[]//)* \n\n \n = librosa.power_to_db(mel_spec, ref=np.max).astype(np.float32)[:,:width]\n = (mel_spec_db+)/\n [:,:,i] = mel_spec_db\n\n i, col in enumerate(rot_list):\n = sequence[col].values\n \n = inter_signal(x)\n\n \n = librosa.feature.melspectrogram(y=x, sr=sr, hop_length=len(x)//, \n =, n_mels=, fmin=, fmax=, win_length=)\n\n = (mel_spec.shape[]//)* \n \n = librosa.power_to_db(mel_spec, ref=np.max).astype(np.float32)[:,:width]\n\n \n = (mel_spec_db+)/\n [:,:,i] = mel_spec_db\n\n img_acc, img_rot`\nThe IMU device records a total of seven features: acc_x, acc_y, acc_z, rot_x, rot_y, rot_z, and rot_w.
\nOur objective is to generate spectrograms for each of these seven features.
\n
At the top, the four rotation features (rot_x, rot_y, rot_z, rot_w) are concatenated sequentially along the axis=1 dimension.
\nAt the bottom, the three acceleration features (acc_x, acc_y, acc_z) are concatenated sequentially along the axis=1 dimension as well.
Subsequently, since the input to the image classification model requires three channels, the identical spectrogram image is concatenated three times along the channel axis (axis=-1).
\n
Thank you for sharing. How much does the spectrum improve your lb score?
", "rawMarkdown": "Thank you for sharing. How much does the spectrum improve your lb score?", "votes": 1, "replies": [ { "id": 3267460, "postDate": "2025-08-11T09:00:08.833Z", "content": "i think Mel-Spectrum maybe not a efficient method for this question
", "rawMarkdown": "i think Mel-Spectrum maybe not a efficient method for this question", "votes": 1, "replies": [ { "id": 3267486, "postDate": "2025-08-11T10:05:50.137Z", "content": "Yes, it may not achieve better performance than a 1D CNN, but I am curious about the results when using a 2D CNN.
", "rawMarkdown": "Yes, it may not achieve better performance than a 1D CNN, but I am curious about the results when using a 2D CNN.\n\n" }, { "id": 3267848, "postDate": "2025-08-12T02:27:56.370Z", "content": "Just curious: why is the Mel-Spectrum not suitable for this question?
", "rawMarkdown": "Just curious: why is the Mel-Spectrum not suitable for this question?", "votes": 1 }, { "id": 3268929, "postDate": "2025-08-13T15:57:09.140Z", "content": "i' not sure, i've tried Wavelet Transformer as features, but it doesn't work, I think this approach may have lost some information of the interaction between features.
", "rawMarkdown": "i' not sure, i've tried Wavelet Transformer as features, but it doesn't work, I think this approach may have lost some information of the interaction between features.", "votes": 1 }, { "id": 3269116, "postDate": "2025-08-14T00:59:42.587Z", "content": "Be careful when using the Wavelet Transformer.
\nI recommend using a different method for each of IMU, TMP, and TOF.
\nAlso, do not apply it to gestures — only use it during transitions and pauses.
The Mel-Spectrum is specifically designed for audio processing, where the frequency content is mapped to the Mel scale to reflect human auditory perception.
\nHowever, IMU signals are not audio — they are low-frequency motion/acceleration signals that don’t follow the same perceptual frequency mapping.
\nUsing a Mel-Spectrum on IMU data can distort the relevant signal characteristics, lose important low-frequency motion patterns, and introduce unnecessary bias toward frequency bands that have no perceptual meaning in this context.
then whats the point of using it?
", "rawMarkdown": "then whats the point of using it?", "votes": 1 }, { "id": 3269859, "postDate": "2025-08-15T07:26:39.873Z", "content": "@seoyunje Sounds reasonable. Thank your for sharing: )
", "rawMarkdown": "@seoyunje Sounds reasonable. Thank your for sharing: ) ", "votes": 1 }, { "id": 3269915, "postDate": "2025-08-15T10:13:18.837Z", "content": "it is a different representation in which vision models could outperform 1d cnns, or at least be used as a powerful feature extractor, but i couldn't benefit from that by myself
", "rawMarkdown": "it is a different representation in which vision models could outperform 1d cnns, or at least be used as a powerful feature extractor, but i couldn't benefit from that by myself", "votes": 1 }, { "id": 3270208, "postDate": "2025-08-16T01:59:05.097Z", "content": "There is no spectrum experiment included in the discussions or notebooks of CMI 2025. Therefore, if anyone is curious about extracting the spectrum from IMU signals, I would like to provide a guide on how to do so
", "rawMarkdown": "There is no spectrum experiment included in the discussions or notebooks of CMI 2025. Therefore, if anyone is curious about extracting the spectrum from IMU signals, I would like to provide a guide on how to do so" }, { "id": 3270210, "postDate": "2025-08-16T02:02:28.017Z", "content": "It seems to be harder to outperform 1D CNNs, especially in CMI 2025. Models such as 1D CNNs, RNNs, and Transformers will likely be powerful tools
", "rawMarkdown": "It seems to be harder to outperform 1D CNNs, especially in CMI 2025. Models such as 1D CNNs, RNNs, and Transformers will likely be powerful tools " } ] }, { "id": 3267485, "postDate": "2025-08-11T10:04:30.640Z", "content": "I am currently developing an EfficientNet pipeline based on mel-spectrograms. While optimal performance has not yet been achieved, this work is part of an additional experimental phase
", "rawMarkdown": "I am currently developing an EfficientNet pipeline based on mel-spectrograms. While optimal performance has not yet been achieved, this work is part of an additional experimental phase ", "votes": 1 } ] }, { "id": 3269903, "postDate": "2025-08-15T09:44:15.610Z", "content": "Thankyou for sharing…
", "rawMarkdown": "Thankyou for sharing...", "votes": 1 } ], "comments": [ { "id": 3267398, "author_name": "ironrro", "author_url": "", "post_date": "2025-08-11T05:52:56.597000", "content": "Thank you for sharing. How much does the spectrum improve your lb score?
", "votes": 1, "replies": [ { "id": 3267460, "author_name": "DaoHe Liu", "author_url": "", "post_date": "2025-08-11T09:00:08.833000", "content": "i think Mel-Spectrum maybe not a efficient method for this question
", "votes": 1, "replies": [ { "id": 3267486, "author_name": "Peter", "author_url": "", "post_date": "2025-08-11T10:05:50.137000", "content": "Yes, it may not achieve better performance than a 1D CNN, but I am curious about the results when using a 2D CNN.
", "votes": 0, "replies": [] }, { "id": 3267848, "author_name": "ironrro", "author_url": "", "post_date": "2025-08-12T02:27:56.370000", "content": "Just curious: why is the Mel-Spectrum not suitable for this question?
", "votes": 1, "replies": [] }, { "id": 3268929, "author_name": "DaoHe Liu", "author_url": "", "post_date": "2025-08-13T15:57:09.140000", "content": "i' not sure, i've tried Wavelet Transformer as features, but it doesn't work, I think this approach may have lost some information of the interaction between features.
", "votes": 1, "replies": [] }, { "id": 3269116, "author_name": "Peter", "author_url": "", "post_date": "2025-08-14T00:59:42.587000", "content": "Be careful when using the Wavelet Transformer.
\nI recommend using a different method for each of IMU, TMP, and TOF.
\nAlso, do not apply it to gestures — only use it during transitions and pauses.
The Mel-Spectrum is specifically designed for audio processing, where the frequency content is mapped to the Mel scale to reflect human auditory perception.
\nHowever, IMU signals are not audio — they are low-frequency motion/acceleration signals that don’t follow the same perceptual frequency mapping.
\nUsing a Mel-Spectrum on IMU data can distort the relevant signal characteristics, lose important low-frequency motion patterns, and introduce unnecessary bias toward frequency bands that have no perceptual meaning in this context.
then whats the point of using it?
", "votes": 1, "replies": [] }, { "id": 3269859, "author_name": "ironrro", "author_url": "", "post_date": "2025-08-15T07:26:39.873000", "content": "@seoyunje Sounds reasonable. Thank your for sharing: )
", "votes": 1, "replies": [] }, { "id": 3269915, "author_name": "Taha_Alshatiri", "author_url": "", "post_date": "2025-08-15T10:13:18.837000", "content": "it is a different representation in which vision models could outperform 1d cnns, or at least be used as a powerful feature extractor, but i couldn't benefit from that by myself
", "votes": 1, "replies": [] }, { "id": 3270208, "author_name": "Peter", "author_url": "", "post_date": "2025-08-16T01:59:05.097000", "content": "There is no spectrum experiment included in the discussions or notebooks of CMI 2025. Therefore, if anyone is curious about extracting the spectrum from IMU signals, I would like to provide a guide on how to do so
", "votes": 0, "replies": [] }, { "id": 3270210, "author_name": "Peter", "author_url": "", "post_date": "2025-08-16T02:02:28.017000", "content": "It seems to be harder to outperform 1D CNNs, especially in CMI 2025. Models such as 1D CNNs, RNNs, and Transformers will likely be powerful tools
", "votes": 0, "replies": [] } ] }, { "id": 3267485, "author_name": "Peter", "author_url": "", "post_date": "2025-08-11T10:04:30.640000", "content": "I am currently developing an EfficientNet pipeline based on mel-spectrograms. While optimal performance has not yet been achieved, this work is part of an additional experimental phase
", "votes": 1, "replies": [] } ] }, { "id": 3269903, "author_name": "Sarah Arshad", "author_url": "", "post_date": "2025-08-15T09:44:15.610000", "content": "Thankyou for sharing…
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "3267383": "# **1 Step. Interpolate Sequence Length**\n\nOn average, each sequence has a length of 69, with considerable variability extending up to approximately 700.\n\nTo standardize the input length, interpolation is applied to resample all sequences to a fixed length of 1024\n\n\n\n- The reason for varying sequence lengths is that, although measurements are taken at consistent intervals, the duration of each action differs.\n- Consequently, applying interpolation to resample sequences to a length of 2048 can result in a slower effective speed compared to the original motion.\n- ***Additionally, when applying interpolation, it is also valid to extract the last 50 values and perform the resampling on this segment of uniform length.***\n___\n# **2 Step. Calculate with Sampling Ratio**\n\nIn the previous step, the sequence length was arbitrarily fixed at 1024. However, to generate a Mel Spectrogram, an appropriate sampling ratio is required. Since the competition dataset does not provide this sampling ratio, it can be set to one of the powers of two: 32, 64, 128 or 256\n\nBecause the spectrogram contains frequency information up to half of the actual sampling rate:\n\n- If the sampling ratio is too low, high-frequency information will be insufficient.\n- Conversely, if the sampling ratio is too high, the spectrogram will contain excessive empty or redundant regions.\n\nAnalysis of the original signals reveals that the frequency band is relatively low. Therefore, by gradually increasing the sampling ratio and searching for an optimal value that sufficiently captures the frequency information in the original signal, it was confirmed that beyond a sampling ratio of 128, changes in the spectrogram information become negligible.\n\n| Sampling Ratio | Frequency Coverage | Effect on Spectrogram | Notes |\n| --- | --- | --- | --- |\n| 32 | Very low | Significant loss of high-frequency components | May miss important details |\n| 64 | Low | Some high-frequency info retained | Still somewhat limited |\n| 128 | Moderate | Most relevant frequency info captured | Optimal minimum threshold |\n| 256 | High | Minimal change compared to 128 | Slightly more detail |\n| 512 | Very high | Excessive empty space in spectrogram | Diminishing returns |\n___\n\n# **3 Step. Spectrogram Overview**\n\nWe utilize the `librosa` library to generate spectrograms. For each unique `sequence_id`, a single spectrogram is created from a sequence with a fixed length of 2048. The resulting spectrograms have dimensions of 128 (frequency bins) by 192 (time frames).\n\nThe core function used is:\n\n```python\n\nmel_spec = librosa.feature.melspectrogram(\n y=x,\n sr=128,\n hop_length=len(x)//64,\n n_fft=128,\n n_mels=32,\n fmin=0,\n fmax=64,\n win_length=128\n)\n\n```\n\n- **`y`** represents the input time-series signal.\n- **`sr`** denotes the sampling frequency; in this competition, EEG signals are sampled at 200 Hz.\n- **`hop_length`** determines the number of samples between successive frames, effectively controlling the width of the output spectrogram (`width = len(x)/hop_length`).\n- **`n_fft`** defines the window size for the Fourier transform, thus influencing the vertical resolution and overall quality of the spectrogram.\n- **`n_mels`** specifies the number of Mel frequency bands, setting the spectrogram height (`height = n_mels`).\n- **`fmin`** and **`fmax`** set the minimum and maximum frequency bounds captured by the spectrogram.\n- **`win_length`** controls the length of the windowed signal segment used for each Fourier transform, affecting the horizontal resolution and spectral quality.\n\n---\n\n#### Effect of `win_length`\n\nA shorter `win_length` relative to `n_fft` results in less tapering of the actual signal and more zero-padding. This causes subtle differences in spectral leakage patterns. To avoid the complexity of additional zero-padding, it is common practice to set `win_length` equal to `n_fft`.\n\n---\n\n#### Effect of `hop_length`\n\n- A smaller `hop_length` (e.g., 256) leads to greater overlap between frames, producing smoother transitions along the time axis and better capturing rapid transient events (such as chirps). However, this increases the number of frames and computational cost.\n- Conversely, a larger `hop_length` (e.g., 1024) reduces overlap, yielding fewer frames and lower time resolution, which may miss quick temporal changes.\n\n---\n\n#### Frequency Bounds (`fmin` and `fmax`)\n\nSetting `fmin` to 50 Hz excludes low-frequency noise such as mechanical rumble or microphone handling artifacts. Setting `fmax` to 14,000 Hz filters out very high-frequency noise sources (e.g., insect sounds or wind artifacts), thereby focusing the analysis on the most relevant frequency range.\n\n---\n\n#### Number of Mel Bands (`n_mels`)\n\n- A higher number of Mel bands (e.g., 128 or 256) provides finer spectral resolution, yielding more detailed spectrograms at the expense of increased memory usage and computational load.\n- A lower number of Mel bands (e.g., 40 or 64) results in a coarser frequency representation, reducing resource demands but potentially obscuring subtle spectral differences in similar sounds.\n\nImportantly, if the number of Mel bands (`n_mels`) equals or exceeds the maximum frequency parameter (`fmax`), the frequency resolution becomes unrealistic, leading to distortion or loss of meaningful spectral information.\n\n`import librosa\n\ndef spectrum_from_imu(sequence, sr=CFG.sampling_rate):\n\n img_acc = np.zeros((32, 64, len(acc_list)),dtype='float32') \n img_rot = np.zeros((32, 48, len(rot_list)), dtype='float32')\n \n for i, col in enumerate(acc_list):\n x = sequence[col].values\n #x = x[len(x)//2:]\n x = inter_signal(x)\n # RAW SPECTROGRAM\n mel_spec = librosa.feature.melspectrogram(y=x, sr=sr, hop_length=len(x)//64,\n n_fft=64, n_mels=32, fmin=0, fmax=64, win_length=64)\n # LOG TRANSFORM\n width = (mel_spec.shape[1]//8)*8 \n\n # STANDARDIZE TO -1 TO 1\n mel_spec_db = librosa.power_to_db(mel_spec, ref=np.max).astype(np.float32)[:,:width]\n mel_spec_db = (mel_spec_db+40)/40\n img_acc[:,:,i] = mel_spec_db\n\n for i, col in enumerate(rot_list):\n x = sequence[col].values\n #x = x[len(x)//2:]\n x = inter_signal(x)\n\n # RAW SPECTROGRAM\n mel_spec = librosa.feature.melspectrogram(y=x, sr=sr, hop_length=len(x)//48, \n n_fft=64, n_mels=32, fmin=0, fmax=64, win_length=64)\n\n width = (mel_spec.shape[1]//8)*8 \n # LOG TRANSFORM\n mel_spec_db = librosa.power_to_db(mel_spec, ref=np.max).astype(np.float32)[:,:width]\n \n # STANDARDIZE TO -1 TO 1\n mel_spec_db = (mel_spec_db+40)/40\n img_rot[:,:,i] = mel_spec_db\n \n return img_acc, img_rot`\n\n___\n\n# **4 Step. How to make Spectrogram with IMU Device**\n\nThe IMU device records a total of seven features: acc_x, acc_y, acc_z, rot_x, rot_y, rot_z, and rot_w.\n\nOur objective is to generate spectrograms for each of these seven features.\n\n\n\nAt the top, the four rotation features (rot_x, rot_y, rot_z, rot_w) are concatenated sequentially along the axis=1 dimension.\nAt the bottom, the three acceleration features (acc_x, acc_y, acc_z) are concatenated sequentially along the axis=1 dimension as well.\n\nSubsequently, since the input to the image classification model requires three channels, the identical spectrogram image is concatenated three times along the channel axis (axis=-1).\n\n", "3267398": "Thank you for sharing. How much does the spectrum improve your lb score?", "3269903": "Thankyou for sharing..." } }