Introducing CURRENNT: The Munich Open-Source CUDA RecurREnt Neural Network Toolkit
Authors: Felix Weninger
JMLR 2015 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Benchmarks are given on a noisy speech recognition task from the 2013 2nd CHi ME Speech Separation and Recognition Challenge, where LSTM-RNNs have been shown to deliver best performance. In the result, double digit speedups in bidirectional LSTM training are achieved with respect to a reference single-threaded CPU implementation. We conclude with a benchmark on a word recognition task with convolutive non-stationary noise from the 2013 2nd CHi ME Challenge s track 1 (Vincent et al., 2013)... Results (Figure 1) show that the error rate after 50 epochs is not heavily influenced by the batch size for parallel processing, while speedups of up to 22.2 can be achieved. |
| Researcher Affiliation | Academia | Felix Weninger EMAIL Johannes Bergmann EMAIL Bj orn Schuller EMAIL Machine Learning & Signal Processing, Technische Universit at M unchen, 80290 Munich, Germany . B. Schuller is also with the Department of Computing, Imperial College London, UK. |
| Pseudocode | No | The paper describes the implementation of Deep LSTM-RNNs using mathematical equations and textual descriptions (e.g., 'Deep LSTM-RNNs with N layers are implemented as follows. An input sequence xt is mapped to the output sequence yt, t = 1, . . . , T through the iteration (forward pass):'), but it does not contain a clearly labeled pseudocode block or algorithm. |
| Open Source Code | Yes | CURRENNT is available under the GNU General Public License from http://sourceforge.net/p/currennt. |
| Open Datasets | Yes | We conclude with a benchmark on a word recognition task with convolutive non-stationary noise from the 2013 2nd CHi ME Challenge s track 1 (Vincent et al., 2013) |
| Dataset Splits | No | The paper mentions using a 'Validation set' in Table 1 with 'Validation set error (10 ep.)' and 'Validation set error (50 ep.)', implying a validation set was used. However, it does not provide specific details on how the dataset (CHiME 2013) was split into training, validation, and test sets, such as percentages or sample counts. |
| Hardware Specification | Yes | The GPU is an NVIDIA GTX 560 with 2 GB of RAM. [...] running in a single CPU thread on an Intel Core2Quad PC with 4 GB of RAM. |
| Software Dependencies | No | The paper mentions NVIDIA's CUDA, C++ as the implementation language, JSON for network architecture, and Net CDF for data format, but does not provide specific version numbers for these or any other software dependencies crucial for replication. |
| Experiment Setup | No | The paper states, 'The corresponding Net CDF, network configuration, and training parameter files are distributed with CURRENNT' and refers to 'the documentation for more details' regarding the gradient descent training algorithm. Specific hyperparameter values (e.g., learning rate, optimizer settings, noise parameters) are not explicitly detailed in the main text of the paper, apart from testing different values for P (parallel sequences) as shown in Table 1. |