def check_expected_removal(leader, particles, retained):
particles.sort(key = FrameParticle.average)
classifier = Salmon(top_k = 2)
classifier._leader = leader
classifier._particles = particles
classifier.remove_identical_particles()
assert classifier._leader is leader
assert classifier._particles == retained, "expected: %s, got %s" % (map(str, retained), map(str, classifier._particles))
def test_frame_halving():
'''
Many halvings and particle challenges happen here.
'''
rng = np.random.RandomState(0)
salmon = Salmon(buffer_size=10, controller=PIDController(Kp=0.02, Td=0.4),
n_samples=5, top_k=7, closeness_bonus=0, rng=rng)
for _ in range(1000):
salmon.camera_input(random_im(rng))
salmon.learn(random_nonwriteable_correction(rng))
def test_best_frame():
classifier = Salmon(top_k = 2)
classifier.candidates_to_clear = []
frames = [create_frame(fid, d, r) for fid, d, r in [(5, 33.0, 0.3), (7, 28.0, 0.7), (6, 31.2, 0.5), (7, 34.9, 0.35), (8, 27.8, 0.6), (11, 29.0, 0.8)]]
# The expected best frame is the one with image_dist 28.0 and response 0.7.
# The frame with image_dist 27.8 is actually better, but because the response is not among the
# top two responses, it should not be picked.
expected_best_frame = frames[1]
best_frame = classifier.get_best_frame(frames)
assert best_frame is expected_best_frame
assert classifier.candidates_to_clear == frames
def test_nonwriteable_correction():
'''
This used to fail because PID corrections were applied
in-place to teacher signal
'''
rng = np.random.RandomState(0)
salmon = Salmon(buffer_size=500, controller=PIDController(Kp=0.02, Td=0.4),
n_samples=40, top_k=7, closeness_bonus=0, rng=rng)
for _ in range(5):
salmon.camera_input(random_im(rng))
salmon.learn(random_nonwriteable_correction(rng))
def test_smoke():
# Note that this is not a joke about smoked salmon
img1 = np.random.randn(40, 40, 3)
img2 = np.roll(img1, shift = 3, axis = 1)
# Add noise to the second image
img2 += np.random.randn(40, 40, 3) * 0.1
motor_cmd = [np.random.randn(), np.random.randn()]
salmon = Salmon()
salmon.camera_input(img1)
salmon.learn(motor_cmd)
result = salmon.camera_input(img2)
# No PID controller involved here, so we expect exactly the same motor command out
# as we provided as a teaching signal
assert np.array_equal(motor_cmd, result)
# We rolled the test image by 3 pixels compared with the training image,
# so we are expecting the phase shift in the horizontal direction
# to be very close to 3.0
best_frame = salmon._leader.frame
phase_x = best_frame.phase_x
print "phase_x", phase_x
assert 2.9 < phase_x < 3.1
# Now add a PID controller
Kp = 0.1
salmon.controller = PIDController(Kp = Kp, Td = 0.0)
result = salmon.camera_input(img2)
# The first array in the list should be the first motor cmd entry given,
# with an addition of the correction from the PID controller
assert motor_cmd[0] + phase_x * Kp == result[0]
# The second array in the list should be untouched
assert np.array_equal(motor_cmd[1], result[1])