def all_iotnets():
for spacename in networks.ARCH_SPACES:
print(spacename)
# Check if you can construct the network.
law_config = networks.get_law_config(spacename, "s1")
net_args = networks.sample_from_law(spacename, law_config, 0)
net = networks.build(spacename, 12, net_args)
print("OK")
def demo():
for arch in ['ResNet18', 'ResNet101', 'ResNet152']:
print("REFERENCE model: {}".format(arch))
net = networks.build(arch, 10, net_args=None)
ts = latency_run(net)
spacename = "random_net_mobilenetv2"
for law in ["s1", "s3", "s6"]:
print("OUR Constraint ARCH searchspace, based on {} with sampling law {}".format(spacename, law))
law_config = networks.get_law_config(spacename, law)
net_args = networks.sample_from_law(spacename, law_config, 0)
net = networks.build(spacename, 10, net_args)
ts = latency_run(net)
def run_all():
# Open configuration file
configs = load_json("configs/NetConfigs.json")
results = []
# Run all configurations
for c in configs:
law_config = networks.get_law_config(c["spacename"], c["law"])
net_args = networks.sample_from_law(c["spacename"], law_config,
c["id"])
net = networks.build(c["spacename"], 10, net_args)
ts = latency_run(net, device="cpu", bs=1, rep=10, verbose=True)
# Clean the data
c["t_avg"] = np.mean(1000 * ts[5:])
c["t_std"] = np.std(1000 * ts[5:])
results.append(c)
# Save results (with configurations)
write_json("results.json", results)
def run_model(config, root_models, device="cpu"):
# Defines the model topology
law_config = networks.get_law_config(config["spacename"], config["law"])
net_args = networks.sample_from_law(config["spacename"], law_config,
config["id"])
net = networks.build(config["spacename"], 10, net_args)
net = net.to(device)
# Defines where the model is stored
ckpt_model = os.path.join(root_models,
"{}_ckpt.t7".format(config["runid"]))
# Loading model
mode = "None" if device == "cuda" else "rm"
net, acc = load_net(net, ckpt_model, device, mode=mode, verbose=False)
print("Loaded net acc: {}".format(acc))
# Latency measurement
ts = latency_run(net, device="cpu", bs=1, rep=10, verbose=True)
return np.mean(1000 * ts[5:])