def do_inference(onnx_model):
import time
num_tests = 100
num_images = 100
total_tests = 0
total_elapsed = 0
#import ipdb as pdb; pdb.set_trace()
sess = rt.InferenceSession(onnx_model)
input_name = sess.get_inputs()[0].name
pass_cnt = 0
utility.print_banner("Starting Inference Engine")
for test_id in range(0, num_tests):
images, labels = load_images(num_images)
inf_start = time.time()
pred_onx = sess.run(None, {input_name: images.astype(np.float32)})[0]
inf_end = time.time()
elapsed_time = inf_end - inf_start
total_elapsed += elapsed_time
utility.print_log("batch {} took {:2.4f}ms to complete".format(
test_id, elapsed_time * 1000))
for i in range(0, num_images):
pred = pred_onx[i].argmax()
res = "FAIL"
if labels[i] == pred:
res = "PASS"
pass_cnt += 1
total_tests += 1
utility.print_log("actual={} pred={} res={}".format(
labels[i], pred, res),
verbosity="VERB_HIGH")
avg_inf = 1000.0 * float(total_elapsed) / total_tests
utility.print_banner(
"Accuracy = {}% out of {} tests, avg inference={:2.4}ms per image".
format(100 * pass_cnt / (float(num_images) * num_tests),
float(num_images) * num_tests, avg_inf))
def run_inference(exec_net, input_blob, output_blob, images, labels):
# import ipdb as pdb; pdb.set_trace()
util.print_banner("Running Openvino Inference on {0} images".format(
images.shape[0]),
color='green',
verbosity="VERB_LOW")
data_counts = images.shape[0]
hit_counts = 0
for i in range(data_counts):
res = exec_net.infer(inputs={input_blob: images[i]})
pred = res[output_blob].argmax()
if pred == labels[i]:
hit_counts += 1
accuracy = float(hit_counts) / float(data_counts)
print_perf_counts(exec_net)
return accuracy
def export_to_onnx(config):
util.print_banner("exporting {0} to ONNX".format(config['trained_model']),
color='green',
verbosity="VERB_LOW")
# import ipdb as pdb; pdb.set_trace()
# model object
model = models.LENET(config)
# load weights
trained_model = torch.load(config['trained_model'])
model.load_state_dict(trained_model)
# create the imput placeholder for the model
# note: we have to specify the size of a batch of input images
input_placeholder = torch.randn(1, 1, 28, 28)
# export
onnx_model_fname = config['experiment_name'] + ".onnx"
torch.onnx.export(model, input_placeholder, onnx_model_fname)
print('{} exported!'.format(onnx_model_fname))
# print_onnx(onnx_model_fname)
return onnx_model_fname
def optimize_model(expr_name, onnx_file, model_xml, weight_bin):
run_opt = False
if (model_xml == None):
util.print_log("Could not find xml model", id_str="warning")
run_opt = True
if (weight_bin == None):
util.print_log("Could not find binary weights", id_str="warning")
run_opt = True
if run_opt:
util.print_banner(
"Running OpenVino optimizer on {0}".format(onnx_file),
color='green',
verbosity="VERB_LOW")
cmd = "python {0}/deployment_tools/model_optimizer/mo.py --input_model={1} --model_name {2}".format(
openvino_inst_path, onnx_file, expr_name)
util.run_command(cmd, verbosity="VERB_LOW")
model_xml, weight_bin = expr_name + ".xml", expr_name + ".bin"
# load model
# import ipdb as pdb; pdb.set_trace()
return model_xml, weight_bin
def export_to_onnx(args):
# import ipdb as pdb; pdb.set_trace()
util.print_banner("exporting {0} to ONNX".format(args['arch']), color='green', verbosity="VERB_LOW")
# model object
model = models.__dict__[args['arch']](num_classes=10)
#model = torch.nn.DataParallel(model).cuda()
# load weights
checkpoint = torch.load(args['chkpoint'])
model = load_state_dict(model, checkpoint['state_dict'])
#model.load_state_dict(checkpoint['state_dict'])
# create the imput placeholder for the model
# note: we have to specify the size of a batch of input images
input_placeholder = torch.randn(1, 3, 32, 32)
# export
onnx_model_fname = args['arch'] + ".onnx"
#import ipdb as pdb; pdb.set_trace()
torch.onnx.export(model, input_placeholder, onnx_model_fname)
print('{} exported!'.format(onnx_model_fname))
# print_onnx(onnx_model_fname)
return onnx_model_fname
def optimize_model(expr_name, onnx_file, model_xml, weight_bin, opt_ops=""):
# import ipdb as pdb; pdb.set_trace()
run_opt = False
options = ""
if (model_xml == None):
util.print_log("No xml model was provided", id_str="warning")
run_opt = True
if (weight_bin == None):
util.print_log("No binary weight file was provided", id_str="warning")
run_opt = True
if run_opt:
util.print_banner("Running OpenVino optimizer on {0}".format(onnx_file), color='green', verbosity="VERB_LOW")
options += "--input_model={0} ".format(onnx_file)
options += "--model_name {0} ".format(expr_name)
options += (" --"+opt_ops[0]) if len(opt_ops)==1 else "".join(" --"+e for e in opt_ops)
cmd = "python {0}/deployment_tools/model_optimizer/mo_onnx.py {1}".format(openvino_inst_path, options)
util.run_command(cmd, verbosity="VERB_LOW")
model_xml, weight_bin = expr_name+".xml", expr_name+".bin"
# load model
# import ipdb as pdb; pdb.set_trace()
return model_xml, weight_bin
coverage = args['coverage']
debug = args['debug']
waveform = args['waveform']
clean = args['clean']
# import ipdb as pdb; pdb.set_trace()
silence = args['silence']
verbosity = args['verbosity']
if verbosity is None:
verbosity = 'VERB_LOW'
if util.get_platform(verbosity=verbosity) != "linux":
util.print_log("This script works only on a Linux platform",
"ERROR",
verbosity="VERB_LOW")
sys.exit()
if clean:
util.print_banner("Cleaning project", verbosity=verbosity)
util.clean_proj(files_to_clean)
if not os.path.exists(result_dir):
util.print_log("Creating a result directory in {0}".format(result_dir),
"INFO",
verbosity="VERB_LOW")
os.makedirs(result_dir)
if simulator == None:
util.print_log("You need to provide Simulator name",
"ERROR",
verbosity="VERB_LOW")
sys.exit()
if simulator.lower() == "xilinx":
# For Xilinx tools we need to specify top level for creating snapshots which is needed
# by simulator and synthesis tools
if top_level == None:
waveform = args['waveform']
clean = args['clean']
silence = args['silence']
verbosity = args['verbosity']
compileonly = args['compileonly']
elabonly = args['elabonly']
if verbosity is None:
verbosity = 'VERB_LOW'
if util.get_platform(verbosity=verbosity) != "linux":
util.print_log("This script works only on a Linux platform",
"ERROR",
verbosity="VERB_LOW")
sys.exit()
if clean:
util.print_banner("Cleaning project", verbosity=verbosity)
util.clean_proj(files_to_clean)
if not os.path.exists(result_dir):
util.print_log("Creating a result directory in {0}".format(result_dir),
"INFO",
verbosity="VERB_LOW")
os.makedirs(result_dir)
if simulator == None:
util.print_log("You need to provide Simulator name",
"ERROR",
verbosity="VERB_LOW")
sys.exit()
# Load Verilog macros file, if specified
vlogmacros = ""
if vlogmacros_file is not None:
window[w_cnt, :, :] = image[x:x + w_width, y:y + w_height]
w_cnt += 1
return window
if __name__ == '__main__':
args = parse_args()
verbosity = args['verbosity']
verbosity = verbosity if verbosity != None else "VERB_HIGH"
input_img_width = int(args['input_img_width'])
input_img_height = int(args['input_img_height'])
kernel_size = int(args['kernel_size'])
input_img = open(args['input_img'], "r").read()
output_window = open(args['output_window'], "r").read()
util.print_banner("Testing Window Slider Output ", verbosity='VERB_LOW')
util.print_log("input image: {0} ".format(args['input_img']))
util.print_log("simulation output: {0} ".format(args['output_window']))
input_img = input_img.replace("\n", ",").split(",")
input_img = input_img[:-1]
input_img = [int(x) for x in input_img]
input_img = np.asarray(input_img)
input_img = input_img.reshape((input_img_height, input_img_width))
sw_output = gen_window(image=input_img,
w_height=kernel_size,
w_width=kernel_size,
stride=1)
output_window = output_window.replace("1'b", "").replace("'{", "").replace(
"}", "").split("\n")
total_num_windows = ((input_img_height - kernel_size) + 1) * (
ip = dep.split("/")[-1].split("\n")[0]
f.write("../rtl/" + ip + "\n")
f.close()
#=======================================================================
# Main
#=======================================================================
if __name__ == '__main__':
cmd_to_run = ""
__args__ = parse_args()
project_name = __args__['project_name']
path_for_proj = __args__['path_proj']
path_fpga = __args__['path_fpga']
verbosity = __args__['vebosity']
util.print_banner("Creating {0} Project".format(project_name),
verbosity="VERB_LOW")
# import ipdb as pdb; pdb.set_trace()
if path_for_proj == None:
default_path_for_proj = os.getcwd() + "/"
util.print_log("Using current path for creating project: {0}".format(
default_path_for_proj))
path_for_proj = default_path_for_proj
if path_fpga == None:
util.print_log("Using default path to fpga directory {0}".format(
default_path_to_fpga_dir))
path_fpga = default_path_to_fpga_dir
# Check if project has already been created
if os.path.isdir("{0}{1}".format(path_for_proj, project_name)):
util.print_log("Project path {0}{1} already exist!".format(
path_for_proj, project_name),
id_str="ERROR")
util.print_log("saving outputs to {0}".format(output_filename))
if __name__ == '__main__':
args = parse_args()
output_filename = args['output_filename']
operation_type = args['operation_type']
verbosity = args['verbosity']
if output_filename == None:
output_filename = "output.txt"
if operation_type == None:
operation_type = "re_run"
if verbosity == None:
verbosity = "VERB_HIGH"
util.print_banner("Running HW simulator", verbosity="VERB_LOW")
if operation_type.lower() == "re_run":
util.print_log("re-running the model...", verbosity="VERB_LOW")
windows, W1, W2, W3, b1, b2, b3 = hw.load_params(
"image.npy", "w1.npy", "w2.npy", "w3.npy", "b1.npy", "b2.npy",
"b3.npy", w_width, w_height, stride)
elif operation_type.lower() == "normal":
util.print_log("running in normal model...", verbosity="VERB_LOW")
windows, W1, W2, W3, b1, b2, b3 = hw.gen_random_params(
input_file_name, image_h, image_w, w_width, w_height, stride,
l1_input, l2_input, l3_input, output_size)
hw.save_model(W1, W2, W3)
hw.gen_sv_weight_file("weight.svh", W1, W2, W3, b1, b2, b3)
out = run_model(windows, W1, W2, W3, b1, b2, b3, verbosity)
save_output(out, output_filename)