151 |
152 |
153 |
154 |
155 |
156 | 
3 |
2 |
3 |
34 | 
35 |
| 160 | d | ||
| 164 | |
165 | dkeras | 166 | |
| 169 | |
170 | dkeras.dKeras | 171 | |
173 | Distributed deep learning can be essential for production systems where you 161 | need fast inference but don’t want expensive hardware accelerators or when 162 | researchers need to train large models made up of distributable parts.
163 |This becomes a challenge for developers because they’ll need expertise in not 164 | only deep learning but also distributed systems. A production team might also 165 | need a machine learning optimization engineer to use neural network 166 | optimizers in terms of precision changes, layer fusing, or other techniques.
167 |Distributed inference is a simple way to get better inference FPS. The graph 168 | below shows how non-optimized, out-of-box models from default frameworks can 169 | be quickly sped up through data parallelism:
170 |The fastest way to get help with any issues with dKeras is through the official GitHub 172 | repository:
173 | 174 |The official email for the dKeras project is dkerasproject@gmail.com.
175 |dKeras can be installed either through PyPI or from source:
164 |dKeras is currently not available on PyPI, install from source instead
167 |
>>> pip install dkeras
168 | >>> git clone https://github.com/dkeras-project/dkeras.git
174 | >>> cd dkeras
175 | >>> pip install -e .
176 | dKeras maintains the same API as Keras, the only differences are the addition 166 | of functions to maintain distributed systems and arguements.
167 |from tensorflow.keras.applications import ResNet50
168 | from dkeras import dKeras
169 | import numpy as np
170 |
171 | data = np.random.uniform(-1, 1, (100, 224, 224, 3))
172 |
173 | model = dKeras(ResNet50, wait_for_workers=True, n_workers=4)
174 | preds = model.predict(data)
175 |
|
187 |
|
195 |
|
199 |
|
207 |
|
215 |
|
223 |