Academia.eduAcademia.edu

Figure 14 – uploaded by Md Mohaimenuzzaman

See full PDFdownloadDownload figure
ACDNet, AclNet, Micro-ACDNet and Micro-AclNet performance on different datasets.

Table 14 ACDNet, AclNet, Micro-ACDNet and Micro-AclNet performance on different datasets.

Related Figures (16)

SOTA models for ESC-10, ESC-50, US8K and AE datasets sorted by year of publication. Abbreviations: ATTN (Attention), CO (Cochlea gram), CRP (Cross Recurrence Plot), CT (Chromagram), DGT (The Discrete Gabor Transform), ENS (Ensemble Model), FBE (FilterBan Energies), GT (GammatTone), LP (Log-Power Spectrogram), Mel (Mel Spectrogram), MFCC (Mel-Frequency Cepstral Coeficients), P (Phase-Encoded), Raw (Raw audio wave), Spec (Spectrogram), TEO (Teagerfis Energy Operator).
SOTA models for ESC-10, ESC-50, US8K and AE datasets sorted by year of publication. Abbreviations: ATTN (Attention), CO (Cochlea gram), CRP (Cross Recurrence Plot), CT (Chromagram), DGT (The Discrete Gabor Transform), ENS (Ensemble Model), FBE (FilterBan Energies), GT (GammatTone), LP (Log-Power Spectrogram), Mel (Mel Spectrogram), MFCC (Mel-Frequency Cepstral Coeficients), P (Phase-Encoded), Raw (Raw audio wave), Spec (Spectrogram), TEO (Teagerfis Energy Operator).
ACDNet architecture. Output shape represents (channel, frequency, time), i_len is the input length, n_cls is the number of output classes, sr is the sampling rate in Hz.
ACDNet architecture. Output shape represents (channel, frequency, time), i_len is the input length, n_cls is the number of output classes, sr is the sampling rate in Hz.
Fig. 1. Visual representation of ACDNet model architecture.
Fig. 1. Visual representation of ACDNet model architecture.
CV Accuracy and estimated 95% CI of clas- sification accuracy of ACDNet on ESC-10, ESC-50, US8K and AE datasets. the he_normal [69] initialization method. At the end of the training and validation, the best model is used as the final model. The loss function optimized is shown in Eq. 5. Here, x is the input mini- batch, fg (x) is the approximation and y is the true distribution of labels for the input data. Furthermore, n is the mini-batch size, m is the number of classes, 7 is the learning rate, and 0 < 0 — qo. Table 3
CV Accuracy and estimated 95% CI of clas- sification accuracy of ACDNet on ESC-10, ESC-50, US8K and AE datasets. the he_normal [69] initialization method. At the end of the training and validation, the best model is used as the final model. The loss function optimized is shown in Eq. 5. Here, x is the input mini- batch, fg (x) is the approximation and y is the true distribution of labels for the input data. Furthermore, n is the mini-batch size, m is the number of classes, 7 is the learning rate, and 0 < 0 — qo. Table 3
Parameters, size, and computation requirements for current SOTA models on the ESC-50 dataset. Here, x denotes the size and FLOPs of ACDNet, for the Size and FLOPs column, respectively.
Parameters, size, and computation requirements for current SOTA models on the ESC-50 dataset. Here, x denotes the size and FLOPs of ACDNet, for the Size and FLOPs column, respectively.
ACDNet in the SOTA leaderboard for raw audio classification on ESC-10&50, US8K and AE datasets. Accuracy values with asterisk (*) are reproduced by us.
ACDNet in the SOTA leaderboard for raw audio classification on ESC-10&50, US8K and AE datasets. Accuracy values with asterisk (*) are reproduced by us.
ACDNet in overall SOTA leaderboard of ESC-10&50, US8K and AE datasets. Column ’Acc’ presents model accuracy. Accu- racy values with asterisk (*) are reproduced by us.
ACDNet in overall SOTA leaderboard of ESC-10&50, US8K and AE datasets. Column ’Acc’ presents model accuracy. Accu- racy values with asterisk (*) are reproduced by us.
Fig. 2. (a) Comparison of 80% pruned models from Table 8. (b) Comparison of 85% pruned models from Table 9. For visualization, the variables are linearly transformed to range [50 - max(variable)].
Fig. 2. (a) Comparison of 80% pruned models from Table 8. (b) Comparison of 85% pruned models from Table 9. For visualization, the variables are linearly transformed to range [50 - max(variable)].
Performance of Micro-ACDNet when trained using Knowledge Distillation.
Performance of Micro-ACDNet when trained using Knowledge Distillation.
Models found after 80% channel pruning using magnitude-based ranking, Taylor criteria-based ranking and our hybrid pruning approach.
Models found after 80% channel pruning using magnitude-based ranking, Taylor criteria-based ranking and our hybrid pruning approach.
Models found after 85% channel pruning using magnitude-based ranking, Taylor criteria-based ranking and our hybrid pruning approach.
Models found after 85% channel pruning using magnitude-based ranking, Taylor criteria-based ranking and our hybrid pruning approach.
ACDNet vs AclNet performance after 80% compression.
ACDNet vs AclNet performance after 80% compression.
Micro-ACDNet architecture for input length 30,225 (approximately 1.51s audio @ 20kHz).
Micro-ACDNet architecture for input length 30,225 (approximately 1.51s audio @ 20kHz).
Parameters, size and computation requirements for ACDNet and Micro-ACDNet for approximately 1.51s audio @ 20kHz.
Parameters, size and computation requirements for ACDNet and Micro-ACDNet for approximately 1.51s audio @ 20kHz.
Prediction accuracy on ESC-50 after quantization.
Prediction accuracy on ESC-50 after quantization.
Related topics:
AcousticsAudio Classification
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved