def measure_top1_accuracy(model_chkpt, dataset, parent_chkpt=None):
if dataset == "cifar10":
trainx, trainy, testx, testy, valx, valy = cifar.load_cifar_data(
"/workspace/finn/dataset", download=True, one_hot=False
)
elif dataset == "mnist":
trainx, trainy, testx, testy, valx, valy = mnist.load_mnist_data(
"/workspace/finn/dataset", download=True, one_hot=False
)
else:
raise Exception("Unrecognized dataset")
# move from dataset_loader layout to ONNX layout: NHWC -> NCHW
testx = testx.transpose(0, 3, 1, 2)
model = ModelWrapper(model_chkpt)
iname = model.graph.input[0].name
oname = model.graph.output[0].name
if parent_chkpt is None:
ishape = model.get_tensor_shape(iname)
else:
parent_model = ModelWrapper(parent_chkpt)
parent_iname = parent_model.graph.input[0].name
ishape = parent_model.get_tensor_shape(parent_iname)
ok = 0
nok = 0
n_batches = testx.shape[0]
for i in range(n_batches):
tdata = testx[i].reshape(ishape).astype(np.float32)
exp = testy[i].item()
if parent_chkpt is not None:
y = execute_parent(parent_chkpt, model_chkpt, tdata)
else:
y = execute_onnx(model, {iname: tdata}, False)[oname]
ret = y.item()
if ret == exp:
ok += 1
else:
nok += 1
if i % 10 == 0:
print("%d : OK %d NOK %d " % (i, ok, nok))
acc_top1 = ok * 100.0 / (ok + nok)
warnings.warn("Final OK %d NOK %d top-1 %f" % (ok, nok, acc_top1))
return acc_top1
def main(args):
trainx, trainy, testx, testy, valx, valy = mnist.load_mnist_data(
"./data", download=False, one_hot=False)
accel = tfc_w1a1_mnist()
print("Expected input shape and datatype: %s %s" %
(str(accel.ishape_normal), str(accel.idt)))
print("Expected output shape and datatype: %s %s" %
(str(accel.oshape_normal), str(accel.odt)))
batch_size = 1000
total = testx.shape[0]
accel.batch_size = batch_size
n_batches = int(total / batch_size)
batch_imgs = testx.reshape(n_batches, batch_size, -1)
batch_labels = testy.reshape(n_batches, batch_size)
obuf_normal = np.empty_like(accel.obuf_packed_device)
print("Ready to run validation, test images tensor has shape %s" %
str(batch_imgs.shape))
print("Accelerator buffer shapes are %s for input, %s for output" %
(str(accel.ishape_packed), str(accel.oshape_packed)))
ok = 0
nok = 0
for i in range(n_batches):
ibuf_normal = batch_imgs[i].reshape(accel.ishape_normal)
exp = batch_labels[i]
obuf_normal = accel.execute(ibuf_normal)
ret = np.bincount(obuf_normal.flatten() == exp.flatten())
nok += ret[0]
ok += ret[1]
print("batch %d / %d : total OK %d NOK %d" % (i, n_batches, ok, nok))
acc = 100.0 * ok / (total)
print("Final accuracy: {}%".format(acc))
return 0
default="resizer.bit")
parser.add_argument("--dataset_root",
help="dataset root dir for download/reuse",
default="/tmp")
# parse arguments
args = parser.parse_args()
bsize = args.batchsize
dataset = args.dataset
bitfile = args.bitfile
platform = args.platform
dataset_root = args.dataset_root
if dataset == "mnist":
from dataset_loading import mnist
trainx, trainy, testx, testy, valx, valy = mnist.load_mnist_data(
dataset_root, download=True, one_hot=False)
elif dataset == "cifar10":
from dataset_loading import cifar
trainx, trainy, testx, testy, valx, valy = cifar.load_cifar_data(
dataset_root, download=True, one_hot=False)
else:
raise Exception("Unrecognized dataset")
test_imgs = testx
test_labels = testy
ok = 0
nok = 0
total = test_imgs.shape[0]