Usage¶

Converting from Standard Nengo¶

NengoFPGA is an extension of Nengo core. Networks and models are described using the traditional Nengo workflow and a single ensemble, including PES learning, can be replaced with an FPGA ensemble using the FpgaPesEnsembleNetwork class. For example, consider the following example of a learned communication channel built with standard Nengo:

import nengo
import numpy as np

def input_func(t):
return [np.sin(t * 2*np.pi), np.cos(t * 2*np.pi)]

with nengo.Network() as model:

# Input stimulus
input_node = nengo.Node(input_func)

# "Pre" ensemble of neurons, and connection from the input
pre = nengo.Ensemble(50, 2)
nengo.Connection(input_node, pre)

# "Post" ensemble of neurons, and connection from "Pre"
post = nengo.Ensemble(50, 2)
conn = nengo.Connection(pre, post)

# Create an ensemble for the error signal
# Error = actual - target = "post" - input
error = nengo.Ensemble(50, 2)
nengo.Connection(post, error)
nengo.Connection(input_node, error, transform=-1)

# Add the learning rule on the pre-post connection
conn.learning_rule_type = nengo.PES(learning_rate=1e-4)

# Connect the error into the learning rule
nengo.Connection(error, conn.learning_rule)


The Nengo code above creates two neural ensembles, pre and post, and forms a PES-learning connection between these two ensembles. The weights of this connection are modulated by an error signal computed by a third neural ensemble (error).

NengoFPGA can be used to replace the pre ensemble with an ensemble that will run on the FPGA. Converting the Nengo model above into a NengoFPGA model proceeds in three steps:

1. Replacing the desired neural ensemble with and FPGA ensemble.

2. Making the appropriate connections to and from the FPGA ensemble.

3. If desired (i.e., if learning is required), making the connections to and from an error-computing neural ensemble.

Constructing the FPGA Ensemble¶

To use the FPGA ensemble, first import the FpgaPesEnsembleNetwork class:

from nengo_fpga.networks import FpgaPesEnsembleNetwork


In the code above, the pre ensemble is to be replaced by the FPGA ensemble. The standard Nengo code for the pre ensemble was:

# "Pre" ensemble of neurons, and connection from the input
pre = nengo.Ensemble(50, 2)


and this is replaced with the FpgaPesEnsembleNetwork class. Since learning is desired in the above model, the learning rule definition on the pre-post connection (conn.learning_rule_type = nengo.PES(learning_rate=1e-4)) has been removed and rolled into the FpgaPesEnsembleNetwork constructor.

# "Pre" ensemble & learning rule
ens_fpga = FpgaPesEnsembleNetwork('de1', n_neurons=50,
dimensions=2,
learning_rate=1e-4)


Notice that the ens_fpga ensemble maintains the same arguments as the original pre ensemble and the learning rule which it encompasses – 50 neurons, 2 dimensions, and a learning rate of 1e-4. The ens_fpga has an additional argument, in this case 'de1', which specifies the desired FPGA device (see NengoFPGA Software Configuration for more details).

Connecting the FPGA Ensemble¶

With the FPGA ensemble created, the connections to and from the original pre ensemble will have to be updated. The original connections are defined as:

# Connection from input to "pre" ensemble
nengo.Connection(input_node, pre)

# Connection from "pre" to "post" ensemble
conn = nengo.Connection(pre, post)


and are replaced with the slightly modified FPGA versions:

# Connection from input to "pre" (FPGA) ensemble
nengo.Connection(input_node, ens_fpga.input)  # Note the added '.input'

# Connection from "pre" (FPGA) to "post" ensemble
nengo.Connection(ens_fpga.output, post)  # Note the added '.output'


The NengoFPGA connections are very similar to the original Nengo connections with the exception that they use the interfaces of the FpgaPesEnsembleNetwork object. The ens_fpga.input and ens_fpga.output replace the input and output of the original pre ensemble.

Connecting the Error Ensemble¶

In the original Nengo model, a neural ensemble was used to compute the error signal that drives the PES learning rule. Using NengoFPGA, this neural ensemble is still needed, and the only change required is to modify the connections from this error ensemble to the FPGA ensemble. The original Nengo model defined the error ensemble and associated connections as:

# Create an ensemble for the error signal
# Error = actual - target = "post" - input
error = nengo.Ensemble(50, 2)
nengo.Connection(post, error)
nengo.Connection(input_node, error, transform=-1)

# Add the learning rule on the pre-post connection
conn.learning_rule_type = nengo.PES(learning_rate=1e-4)

# Connect the error into the learning rule
nengo.Connection(error, conn.learning_rule)


The NengoFPGA equivalent code would be:

# Create an ensemble for the error signal
# Error = actual - target = "post" - input
error = nengo.Ensemble(50, 2)  # Remains unchanged
nengo.Connection(post, error)  # Remains unchanged
nengo.Connection(input_node, error, transform=-1)  # Remains unchanged

# Connect the error into the learning rule
nengo.Connection(error, ens_fpga.error)  # Note the added '.error'


Note that – as mentioned previously – in the NengoFPGA equivalent code, the learning_rule_type definition of the pre-post connection has been removed as this is declared in the FpgaPesEnsembleNetwork object.

Final NengoFPGA Model¶

Altogether the NengoFPGA version of the learned communication channel would look something like this:

import nengo
import numpy as np

from nengo_fpga.networks import FpgaPesEnsembleNetwork

def input_func(t):
return [np.sin(t * 2*np.pi), np.cos(t * 2*np.pi)]

with nengo.Network() as model:

# Input stimulus
input_node = nengo.Node(input_func)

# "Pre" ensemble of neurons, and connection from the input
ens_fpga = FpgaPesEnsembleNetwork('de1', n_neurons=50,
dimensions=2,
learning_rate=1e-4)
nengo.Connection(input_node, ens_fpga.input)  # Note the added '.input'

# "Post" ensemble of neurons, and connection from "Pre"
post = nengo.Ensemble(50, 2)
conn = nengo.Connection(ens_fpga.output, post)  # Note the added '.output'

# Create an ensemble for the error signal
# Error = actual - target = "post" - input
error = nengo.Ensemble(50, 2)
nengo.Connection(post, error)
nengo.Connection(input_node, error, transform=-1)

# Connect the error into the learning rule
nengo.Connection(error, ens_fpga.error)  # Note the added '.error'


Basic Use¶

NengoFPGA is designed to work with Nengo GUI, however you can see also run as a script if you prefer not to use the GUI. In either case, if the FPGA device is not correctly configured, or the NengoFPGA backend is not selected, the FpgaPesEnsembleNetwork will be converted to run as standard Nengo objects and a warning will be printed.

For any questions please visit the Nengo Forum.

Note

Ensure you’ve configured your board and NengoFPGA as outlined in the Getting Started Guide.

Using the GUI¶

To view and run your networks, simply pass nengo_fpga as the backend to Nengo GUI:

nengo <my_file.py> -b nengo_fpga


This should open the GUI in a browser and display the network from my_file.py. You can begin execution by clicking the play button in the bottom left corner. this may take a few moments to establish a connection and initialize the FPGA device.

Scripting¶

If you are not using Nengo GUI, you can use the nengo_fpga.Simulator in Nengo’s scripting environment as well. Consider the following example of running a standard Nengo network:

import nengo

with nengo.Network() as model:

with nengo.Simulator(model) as sim:
sim.run(1)


Simply replace the Simulator with the one from NengoFPGA:

import nengo
import nengo_fpga

with nengo.Network() as model:

# Including an FpgaPesEnsembleNetwork

with nengo_fpga.Simulator(model) as sim:
sim.run(1)


Maximum Model Size¶

When running Nengo models on other hardware there is no set limit to model or network size. The system will continue to allocate resources (like memory) until it runs out which leads to different limits depending on the capabilities of your hardware. On the other hand, the NengoFPGA design is fixed and therefore we must provision resources up front. As a result, we have specific upper bounds which are chosen such that the resource allocation balances performance and flexibility for the given architecture. We store all neuron parameters on-chip giving us bounds based on specific memory requirements:

• The maximum number of neurons, N, used to allocate memory for things like neuron activity and bias.

• The maximum number of representational dimensions (input or output), D, used to allocate memory for things like the input and output vector.

• The maximum product of neurons and dimensions, NxD, used to allocate memory for things like encoder and decoder matrices.

These maximum model size values are summarized in the hardware-specific documentation: