def run_simulation(config):
if len(sys.argv) == 4:
platform_idx = int(sys.argv[2])
device_idx = int(sys.argv[3])
dev = cl.get_platforms()[platform_idx].get_devices()[device_idx]
print("Using device:", dev)
ctx = cl.Context(devices=[dev])
else:
ctx = cl.create_some_context(interactive=False)
queue = cl.CommandQueue(ctx)
mf = cl.mem_flags
ifmt_f = cl.ImageFormat(cl.channel_order.RGBA, cl.channel_type.FLOAT)
gs = config['grid_size']
wgs = 16
clctx = CLContext(ctx, queue, ifmt_f, gs, wgs)
sigs = set_up_signals(config)
colony, feats, blurs2, blurs4, rdctns, colours = build_programs(clctx,
config)
sigs_a = sigs[:, :, :4].reshape(gs*gs*4)
sigs_b = sigs[:, :, 4:].reshape(gs*gs*4)
ifmt_f = cl.ImageFormat(cl.channel_order.RGBA, cl.channel_type.FLOAT)
ibuf_1a = cl.Image(ctx, mf.READ_WRITE, ifmt_f, (gs, gs))
ibuf_1b = cl.Image(ctx, mf.READ_WRITE, ifmt_f, (gs, gs))
ibuf_2a = cl.Image(ctx, mf.READ_WRITE, ifmt_f, (gs, gs))
ibuf_2b = cl.Image(ctx, mf.READ_WRITE, ifmt_f, (gs, gs))
cl.enqueue_copy(queue, ibuf_1a, sigs_a, origin=(0, 0), region=(gs, gs))
cl.enqueue_copy(queue, ibuf_1b, sigs_b, origin=(0, 0), region=(gs, gs))
for i in config.get('dump_images', []):
dump_image(clctx, colours, ibuf_1a, ibuf_1b, i, 0)
n_iters = config['iterations']
feature_v = np.zeros((n_iters, 8, 4), np.float32)
for iteration in range(n_iters):
colony.genome(queue, (gs, gs), (wgs, wgs),
ibuf_1a, ibuf_1b, ibuf_2a, ibuf_2b)
colony.convolve_x(queue, (gs, gs), (wgs, wgs), ibuf_2a, ibuf_1a)
colony.convolve_y(queue, (gs, gs), (wgs, wgs), ibuf_1a, ibuf_2a)
ibuf_1a, ibuf_2a = ibuf_2a, ibuf_1a
ibuf_1b, ibuf_2b = ibuf_2b, ibuf_1b
feature_v[iteration] = features.get_features(
clctx, feats, rdctns, blurs2, blurs4, ibuf_1b)
for i in config.get('dump_images', []):
dump_image(clctx, colours, ibuf_1a, ibuf_1b, i, iteration+1)
if config.get('early_stop') and iteration > 1:
m = np.max(feature_v[iteration])
if m < 0.1:
break
if config.get('dump_final_image'):
for i in genome.get_used_genes(config["genome"]):
dump_image(clctx, colours, ibuf_1a, ibuf_1b,
i, iteration+1, config["genome"])
return iteration, feature_v
def process(config):
# Run the simulation, get the resulting feature vector evolution
niters, fvs = run.run_simulation(config)
# ignore first two FVs due to initial conditions
fvs = fvs[2:]
# now find mean for each feature
fvs = np.mean(fvs, axis=0)
# sum up all the blob counts
blobs = np.sum(fvs[1:6], axis=0)
# reform fvs with iteration count and combined blobs
niters = np.ones_like(blobs) * niters
fvs = np.array((niters, fvs[0], blobs, fvs[6], fvs[7]))
# normalise
gs = config["grid_size"]
iters = config["iterations"]
norms = np.array((iters, (gs**2)/2, (gs**2)/(26*4*5), 0.25, 8.0))
norms = np.kron(norms, np.ones(4)).reshape(5, 4)
fvs /= norms
# weight
scores = np.dot(config["weights"], fvs)
# Take the average weight of the genomes in use
sigs_used = np.array(genome.get_used_genes(config["genome"]))
sigs_used = sigs_used[sigs_used >= 4] - 4
if sigs_used.size == 0:
return 0.0
mask = np.ones_like(scores).astype(np.bool)
mask[:, sigs_used] = 0
score = np.ma.masked_array(scores, mask).mean()
return float(score)
