diff --git a/src/finn/qnn-data/cybsec-mlp/state_dict.pth b/src/finn/qnn-data/cybsec-mlp/state_dict.pth
new file mode 100644
index 0000000000000000000000000000000000000000..53c002e3fa6f2ae3e7c8f0abb71fa446d80a8f09
Binary files /dev/null and b/src/finn/qnn-data/cybsec-mlp/state_dict.pth differ
diff --git a/src/finn/qnn-data/cybsec-mlp/validate-unsw-nb15.py b/src/finn/qnn-data/cybsec-mlp/validate-unsw-nb15.py
new file mode 100644
index 0000000000000000000000000000000000000000..2fabc716a66a3cc24697e49aa26ec3bbbb231b43
--- /dev/null
+++ b/src/finn/qnn-data/cybsec-mlp/validate-unsw-nb15.py
@@ -0,0 +1,109 @@
+# Copyright (c) 2020 Xilinx, Inc.
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+# list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+# this list of conditions and the following disclaimer in the documentation
+# and/or other materials provided with the distribution.
+#
+# * Neither the name of Xilinx nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+import argparse
+from driver import io_shape_dict
+from driver_base import FINNExampleOverlay
+import numpy as np
+
+
+def make_unsw_nb15_test_batches(bsize, dataset_root, limit_batches):
+ unsw_nb15_data = np.load(dataset_root + "/unsw_nb15_binarized.npz")["test"][:82000]
+ test_imgs = unsw_nb15_data[:, :-1]
+ test_labels = unsw_nb15_data[:, -1]
+ n_batches = int(test_imgs.shape[0] / bsize)
+ if limit_batches == -1:
+ limit_batches = n_batches
+ test_imgs = test_imgs.reshape(n_batches, bsize, -1)[:limit_batches]
+ test_labels = test_labels.reshape(n_batches, bsize)[:limit_batches]
+ return (test_imgs, test_labels)
+
+
+if __name__ == "__main__":
+ parser = argparse.ArgumentParser(
+ description="Validate top-1 accuracy for FINN-generated accelerator"
+ )
+ parser.add_argument("--batchsize", help="samples per batch", type=int, default=1000)
+ parser.add_argument(
+ "--platform", help="Target platform: zynq-iodma alveo", default="zynq-iodma"
+ )
+ parser.add_argument(
+ "--bitfile",
+ help='name of bitfile (i.e. "resizer.bit")',
+ default="../bitfile/finn-accel.bit",
+ )
+ parser.add_argument(
+ "--dataset_root", help="dataset root dir for download/reuse", default="."
+ )
+ parser.add_argument(
+ "--limit_batches", help="number of batches, -1 for max", type=int, default=-1
+ )
+ # parse arguments
+ args = parser.parse_args()
+ bsize = args.batchsize
+ bitfile = args.bitfile
+ platform = args.platform
+ dataset_root = args.dataset_root
+ limit_batches = args.limit_batches
+
+ print("Loading dataset...")
+ (test_imgs, test_labels) = make_unsw_nb15_test_batches(
+ bsize, dataset_root, limit_batches
+ )
+
+ ok = 0
+ nok = 0
+ n_batches = test_imgs.shape[0]
+ total = n_batches * bsize
+
+ print("Initializing driver, flashing bitfile...")
+
+ driver = FINNExampleOverlay(
+ bitfile_name=bitfile,
+ platform=platform,
+ io_shape_dict=io_shape_dict,
+ batch_size=bsize,
+ )
+
+ n_batches = int(total / bsize)
+
+ print("Starting...")
+
+ for i in range(n_batches):
+ inp = np.pad(test_imgs[i].astype(np.float32), [(0, 0), (0, 7)], mode="constant")
+ exp = test_labels[i].astype(np.float32)
+ inp = 2 * inp - 1
+ exp = 2 * exp - 1
+ out = driver.execute(inp)
+ matches = np.count_nonzero(out.flatten() == exp.flatten())
+ nok += bsize - matches
+ ok += matches
+ print("batch %d / %d : total OK %d NOK %d" % (i + 1, n_batches, ok, nok))
+
+ acc = 100.0 * ok / (total)
+ print("Final accuracy: %f" % acc)
diff --git a/src/finn/qnn-data/templates/driver/driver_base.py b/src/finn/qnn-data/templates/driver/driver_base.py
index ef16a537ce18c52ea42ce9178a7178e8f8b667dd..9ec03ea5dd726b49b157a92addef05f85f02b644 100644
--- a/src/finn/qnn-data/templates/driver/driver_base.py
+++ b/src/finn/qnn-data/templates/driver/driver_base.py
@@ -37,6 +37,8 @@ from finn.util.data_packing import (
packed_bytearray_to_finnpy,
)
+from finn.util.basic import gen_finn_dt_tensor
+
# Driver base class for FINN-generated dataflow accelerators.
# The particulars of the generated accelerator are specified via the
# io_shape_dict (generated by the MakePYNQDriver transformation).
@@ -344,7 +346,7 @@ class FINNExampleOverlay(Overlay):
res["fclk[mhz]"] = self.clock_dict["clock0"]["frequency"]
res["batch_size"] = self.batch_size
# also benchmark driver-related overheads
- input_npy = np.zeros(self.ishape_normal, dtype=self.idt.to_numpy_dt())
+ input_npy = gen_finn_dt_tensor(self.idt, self.ishape_normal)
start = time.time()
ibuf_folded = self.fold_input(input_npy)
end = time.time()
diff --git a/tests/end2end/test_end2end_cybsec_mlp.py b/tests/end2end/test_end2end_cybsec_mlp.py
new file mode 100644
index 0000000000000000000000000000000000000000..c9259afabfa6fcae0020e378c79ce391c218408f
--- /dev/null
+++ b/tests/end2end/test_end2end_cybsec_mlp.py
@@ -0,0 +1,218 @@
+import torch
+from brevitas.nn import QuantLinear, QuantReLU
+import torch.nn as nn
+import numpy as np
+from brevitas.core.quant import QuantType
+from brevitas.nn import QuantIdentity
+import brevitas.onnx as bo
+from finn.core.modelwrapper import ModelWrapper
+from finn.core.datatype import DataType
+import finn.builder.build_dataflow as build
+import finn.builder.build_dataflow_config as build_cfg
+import os
+import shutil
+from finn.util.test import get_build_env, load_test_checkpoint_or_skip
+import pytest
+from finn.util.basic import make_build_dir
+import pkg_resources as pk
+import json
+import wget
+import subprocess
+
+target_clk_ns = 10
+build_kind = "zynq"
+build_dir = os.environ["FINN_BUILD_DIR"]
+
+
+def get_checkpoint_name(step):
+ if step == "build":
+ # checkpoint for build step is an entire dir
+ return build_dir + "/end2end_cybsecmlp_build"
+ else:
+ # other checkpoints are onnx files
+ return build_dir + "/end2end_cybsecmlp_%s.onnx" % (step)
+
+
+class CybSecMLPForExport(nn.Module):
+ def __init__(self, my_pretrained_model):
+ super(CybSecMLPForExport, self).__init__()
+ self.pretrained = my_pretrained_model
+ self.qnt_output = QuantIdentity(
+ quant_type=QuantType.BINARY, bit_width=1, min_val=-1.0, max_val=1.0
+ )
+
+ def forward(self, x):
+ # assume x contains bipolar {-1,1} elems
+ # shift from {-1,1} -> {0,1} since that is the
+ # input range for the trained network
+ x = (x + torch.tensor([1.0])) / 2.0
+ out_original = self.pretrained(x)
+ out_final = self.qnt_output(out_original) # output as {-1,1}
+ return out_final
+
+
+def test_end2end_cybsec_mlp_export():
+ assets_dir = pk.resource_filename("finn.qnn-data", "cybsec-mlp/")
+ # load up trained net in Brevitas
+ input_size = 593
+ hidden1 = 64
+ hidden2 = 64
+ hidden3 = 64
+ weight_bit_width = 2
+ act_bit_width = 2
+ num_classes = 1
+ model = nn.Sequential(
+ QuantLinear(input_size, hidden1, bias=True, weight_bit_width=weight_bit_width),
+ nn.BatchNorm1d(hidden1),
+ nn.Dropout(0.5),
+ QuantReLU(bit_width=act_bit_width),
+ QuantLinear(hidden1, hidden2, bias=True, weight_bit_width=weight_bit_width),
+ nn.BatchNorm1d(hidden2),
+ nn.Dropout(0.5),
+ QuantReLU(bit_width=act_bit_width),
+ QuantLinear(hidden2, hidden3, bias=True, weight_bit_width=weight_bit_width),
+ nn.BatchNorm1d(hidden3),
+ nn.Dropout(0.5),
+ QuantReLU(bit_width=act_bit_width),
+ QuantLinear(hidden3, num_classes, bias=True, weight_bit_width=weight_bit_width),
+ )
+ trained_state_dict = torch.load(assets_dir + "/state_dict.pth")[
+ "models_state_dict"
+ ][0]
+ model.load_state_dict(trained_state_dict, strict=False)
+ W_orig = model[0].weight.data.detach().numpy()
+ # pad the second (593-sized) dimensions with 7 zeroes at the end
+ W_new = np.pad(W_orig, [(0, 0), (0, 7)])
+ model[0].weight.data = torch.from_numpy(W_new)
+ model_for_export = CybSecMLPForExport(model)
+ export_onnx_path = get_checkpoint_name("export")
+ input_shape = (1, 600)
+ bo.export_finn_onnx(model_for_export, input_shape, export_onnx_path)
+ assert os.path.isfile(export_onnx_path)
+ # fix input datatype
+ finn_model = ModelWrapper(export_onnx_path)
+ finnonnx_in_tensor_name = finn_model.graph.input[0].name
+ finn_model.set_tensor_datatype(finnonnx_in_tensor_name, DataType.BIPOLAR)
+ finn_model.save(export_onnx_path)
+ assert tuple(finn_model.get_tensor_shape(finnonnx_in_tensor_name)) == (1, 600)
+ assert len(finn_model.graph.node) == 30
+ assert finn_model.graph.node[0].op_type == "Add"
+ assert finn_model.graph.node[1].op_type == "Div"
+ assert finn_model.graph.node[2].op_type == "MatMul"
+ assert finn_model.graph.node[-1].op_type == "MultiThreshold"
+
+
+@pytest.mark.slow
+@pytest.mark.vivado
+def test_end2end_cybsec_mlp_build():
+ model_file = get_checkpoint_name("export")
+ load_test_checkpoint_or_skip(model_file)
+ build_env = get_build_env(build_kind, target_clk_ns)
+ output_dir = make_build_dir("test_end2end_cybsec_mlp_build")
+
+ cfg = build.DataflowBuildConfig(
+ output_dir=output_dir,
+ target_fps=1000000,
+ synth_clk_period_ns=target_clk_ns,
+ board=build_env["board"],
+ shell_flow_type=build_cfg.ShellFlowType.VIVADO_ZYNQ,
+ generate_outputs=[
+ build_cfg.DataflowOutputType.ESTIMATE_REPORTS,
+ build_cfg.DataflowOutputType.BITFILE,
+ build_cfg.DataflowOutputType.PYNQ_DRIVER,
+ build_cfg.DataflowOutputType.DEPLOYMENT_PACKAGE,
+ ],
+ )
+ build.build_dataflow_cfg(model_file, cfg)
+ # check the generated files
+ assert os.path.isfile(output_dir + "/time_per_step.json")
+ assert os.path.isfile(output_dir + "/final_hw_config.json")
+ assert os.path.isfile(output_dir + "/driver/driver.py")
+ est_cycles_report = output_dir + "/report/estimate_layer_cycles.json"
+ assert os.path.isfile(est_cycles_report)
+ est_res_report = output_dir + "/report/estimate_layer_resources.json"
+ assert os.path.isfile(est_res_report)
+ assert os.path.isfile(output_dir + "/report/estimate_network_performance.json")
+ assert os.path.isfile(output_dir + "/bitfile/finn-accel.bit")
+ assert os.path.isfile(output_dir + "/bitfile/finn-accel.hwh")
+ assert os.path.isfile(output_dir + "/report/post_synth_resources.xml")
+ assert os.path.isfile(output_dir + "/report/post_route_timing.rpt")
+ # examine the report contents
+ with open(est_cycles_report, "r") as f:
+ est_cycles_dict = json.load(f)
+ assert est_cycles_dict["StreamingFCLayer_Batch_0"] == 80
+ assert est_cycles_dict["StreamingFCLayer_Batch_1"] == 64
+ with open(est_res_report, "r") as f:
+ est_res_dict = json.load(f)
+ assert est_res_dict["total"]["LUT"] == 11360.0
+ assert est_res_dict["total"]["BRAM_18K"] == 36.0
+ shutil.copytree(output_dir + "/deploy", get_checkpoint_name("build"))
+
+
+def test_end2end_cybsec_mlp_run_on_hw():
+ build_env = get_build_env(build_kind, target_clk_ns)
+ assets_dir = pk.resource_filename("finn.qnn-data", "cybsec-mlp/")
+ deploy_dir = get_checkpoint_name("build")
+ if not os.path.isdir(deploy_dir):
+ pytest.skip(deploy_dir + " not found from previous test step, skipping")
+ driver_dir = deploy_dir + "/driver"
+ assert os.path.isdir(driver_dir)
+ # put all assets into driver dir
+ shutil.copy(assets_dir + "/validate-unsw-nb15.py", driver_dir)
+ # put a copy of binarized dataset into driver dir
+ dataset_url = (
+ "https://zenodo.org/record/4519767/files/unsw_nb15_binarized.npz?download=1"
+ )
+ dataset_local = driver_dir + "/unsw_nb15_binarized.npz"
+ if not os.path.isfile(dataset_local):
+ wget.download(dataset_url, out=dataset_local)
+ assert os.path.isfile(dataset_local)
+ # create a shell script for running validation: 10 batches x 10 imgs
+ with open(driver_dir + "/validate.sh", "w") as f:
+ f.write(
+ """#!/bin/bash
+cd %s/driver
+echo %s | sudo -S python3.6 validate-unsw-nb15.py --batchsize=10 --limit_batches=10
+ """
+ % (
+ build_env["target_dir"] + "/end2end_cybsecmlp_build",
+ build_env["password"],
+ )
+ )
+ # set up rsync command
+ remote_target = "%s@%s:%s" % (
+ build_env["username"],
+ build_env["ip"],
+ build_env["target_dir"],
+ )
+ rsync_res = subprocess.run(
+ [
+ "sshpass",
+ "-p",
+ build_env["password"],
+ "rsync",
+ "-avz",
+ deploy_dir,
+ remote_target,
+ ]
+ )
+ assert rsync_res.returncode == 0
+ remote_verif_cmd = [
+ "sshpass",
+ "-p",
+ build_env["password"],
+ "ssh",
+ "%s@%s" % (build_env["username"], build_env["ip"]),
+ "sh",
+ build_env["target_dir"] + "/end2end_cybsecmlp_build/driver/validate.sh",
+ ]
+ verif_res = subprocess.run(
+ remote_verif_cmd,
+ stdout=subprocess.PIPE,
+ universal_newlines=True,
+ input=build_env["password"],
+ )
+ assert verif_res.returncode == 0
+ log_output = verif_res.stdout.split("\n")
+ assert log_output[-3] == "batch 10 / 10 : total OK 93 NOK 7"
+ assert log_output[-2] == "Final accuracy: 93.000000"