Release note:
1. Support multi-card parallel computing,
2. Update 'Network_load' and 'Network_save' modules to support access to more complex networks
3. Rectify the debugging when the 'Monitor' records time that is not synchronized with the simulation time in a long time step simulation
4. Add many new neuron models and algorithms to support brain simulation applications
5. Fix an issue where connections are limited within 'Assembly'
6. The learning algorithm of STDP class supports updating after simulation by means of optimizer.step
7. 'Delay' supports reverse gradient transmission
8. Add the 'forward_build' mode. In 'forward_build' mode, the network is built in forward mode to avoid delay. In the original build mode, all connections were built first to resolve loop dependencies, so each connection had a one-step delay.
9. Support to customize which model parameters can be trained
10. Use absolute paths in module import
版本更新:
1. 支持多卡并行计算，
2. 更新Network_load与Network_save模块，添加对更多复杂结构网络的存取支持
3. 修复Monitor在长时间步模拟情况下记录时间与仿真时间不同步的debug
4. 添加许多新的神经元模型及算法，支撑脑仿真应用
5. 修复连接局限于Assembly内部的问题
6. STDP类学习算法支持通过optimizer.step的方式，在仿真后进行更新
7. Delay支持反传梯度
8. 新增forward_build方式，在forward_build模式下，网络构建按照前向模式，避免延迟。原build模式下，为了解决环路依赖问题，优先构建所有connection，因此每一个connection都存在一步延迟问题。
9. 支持 自定义选择哪些模型参数可训练
10. 将模块的引用路径全部统一为绝对路径
beta
Release Note
1. To provide a more concise code for build the network, we changed some parameter names for initialize NeuronGroup and Connection:
For NeuronGroup initialization: neuron_number -> num, neuron_model -> model
For Connection initialization: pre_assembly -> pre, post_assembly -> post
2. We have added interfaces in the frontend that could directly get backend values of certain network components using get_values function: such as V = neuron1.get_values(‘V’).
3. Added Conv related operations such as max_pooling, batchNorm2d, Flatten as Synapse modules, which can be added to conv connections. We have also added a Pool_connection to solely conduct pooling operation.
4. We have added Meta_STDP algorithms that can concurrently train the network with gradient Backprop algorithms and STDP learning rules.
5. PoissonEncoders now generate Poisson spikes on the fly rather than at the beginning of the run, to save memory.
6. We used a new Op class to contain operations in the backend, and added more attributes to backend Ops, such as owner, device and requires_grad.
7. Some bugs are fixed.