Loop#

Performs a recurrent network computation.

Examples#

Loop
'''
This example creates a Loop consisting of an ElementWise layer that is used as an accumulator.
The accumalter value is named `accumaltor_value`, and a the added value for each iteration is named `accumaltor_added_value`.
The Loop stop condition is a counter initialized to `num_iterations`, which is implemented using the TripLimit layer.
The expected output is `accumaltor_value` + `num_iterations`*`accumaltor_added_value`
'''
num_iterations = 3
trip_limit = network.add_constant(shape=(), weights=trt.Weights(np.array([num_iterations], dtype=np.dtype("i4"))))
accumaltor_value = network.add_input("input1", dtype=trt.float32, shape=(2, 3))
accumaltor_added_value = network.add_input("input2", dtype=trt.float32, shape=(2, 3))
loop = network.add_loop()
# setting the ITripLimit layer to stop after `num_iterations` iterations
loop.add_trip_limit(trip_limit.get_output(0), trt.TripLimit.COUNT)
# initialzing the IRecurrenceLayer with a init value
rec = loop.add_recurrence(accumaltor_value)
# eltwise inputs are 'accumaltor_added_value', and the IRecurrenceLayer output.
eltwise = network.add_elementwise(accumaltor_added_value, rec.get_output(0), op=trt.ElementWiseOperation.SUM)
# wiring the IRecurrenceLayer with the output of eltwise.
# The IRecurrenceLayer output would now be `accumaltor_value` for the first iteration, and the eltwise output for any other iteration
rec.set_input(1, eltwise.get_output(0))
# marking the IRecurrenceLayer output as the Loop output
loop_out = loop.add_loop_output(rec.get_output(0), trt.LoopOutput.LAST_VALUE)
# marking the Loop output as the network output
network.mark_output(loop_out.get_output(0))

inputs[accumaltor_value.name] = np.array(
    [
        [2.7, -4.9, 23.34],
        [8.9, 10.3, -19.8],
    ])
inputs[accumaltor_added_value.name] = np.array(
    [
        [1.1, 2.2, 3.3],
        [-5.7, 1.3, 4.6],
    ])

outputs[loop_out.get_output(0).name] = eltwise.get_input(0).shape
expected[loop_out.get_output(0).name] = inputs[accumaltor_value.name] + inputs[accumaltor_added_value.name] * num_iterations