def RunDPU(kernel, img, count):
"""
DPU run function
kernel: dpu kernel
img: image to be run
count : test rounds count
"""
"""Create DPU Tasks from DPU Kernel"""
task = n2cube.dpuCreateTask(kernel, 0)
while count < 1000:
"""Load image to DPU"""
dputils.dpuSetInputImage2(task, KERNEL_CONV_INPUT, img)
"""Get input Tesor"""
tensor = n2cube.dpuGetInputTensor(task, KERNEL_CONV_INPUT)
"""Model run on DPU"""
n2cube.dpuRunTask(task)
"""Get the output tensor size from FC output"""
size = n2cube.dpuGetOutputTensorSize(task, KERNEL_FC_OUTPUT)
"""Get the output tensor channel from FC output"""
channel = n2cube.dpuGetOutputTensorChannel(task, KERNEL_FC_OUTPUT)
softmax = np.zeros(size, dtype=float32)
"""Get FC result"""
conf = n2cube.dpuGetOutputTensorAddress(task, KERNEL_FC_OUTPUT)
"""Get output scale of FC"""
outputScale = n2cube.dpuGetOutputTensorScale(task, KERNEL_FC_OUTPUT)
"""Run softmax"""
softmax = n2cube.dpuRunSoftmax(conf, channel, size // channel,
outputScale)
l.acquire()
count = count + threadnum
l.release()
"""Destroy DPU Tasks & free resources"""
n2cube.dpuDestroyTask(task)
def predict_label(imfile):
task = n2cube.dpuCreateTask(kernel, 0)
# Set client to get file from S3
s3_client.download_file(BUCKET, imfile, image_folder + imfile)
img_obj = os.path.join(image_folder, imfile)
#To get it from local path
#img_file = os.path.join(image_folder, imfile)
img = cv2.imread(img_obj)
img = cv2.resize(img, (IMG_DIMS, IMG_DIMS))
img = img.astype(np.float32)
img = (img/255.0)
"""Get input Tensor"""
tensor = n2cube.dpuGetInputTensor(task, KERNEL_CONV_INPUT)
input_len = n2cube.dpuGetInputTensorSize(task, KERNEL_CONV_INPUT)
"""Set input Tesor"""
n2cube.dpuSetInputTensorInHWCFP32(task, KERNEL_CONV_INPUT, img, input_len)
"""Model run on DPU"""
n2cube.dpuRunTask(task)
"""Get the output tensor size from FC output"""
size = n2cube.dpuGetOutputTensorSize(task, KERNEL_FC_OUTPUT)
"""Get the output tensor channel from FC output"""
channel = n2cube.dpuGetOutputTensorChannel(task, KERNEL_FC_OUTPUT)
softmax = np.zeros(size,dtype=np.float32)
"""Get FC result"""
conf = n2cube.dpuGetOutputTensorAddress(task, KERNEL_FC_OUTPUT)
"""Get output scale of FC"""
outputScale = n2cube.dpuGetOutputTensorScale(task, KERNEL_FC_OUTPUT)
"""Run softmax"""
softmax = n2cube.dpuRunSoftmax(conf, channel, size // channel, outputScale)
#print("softmax =", softmax)
n2cube.dpuDestroyTask(task)
return slabels[np.argmax(softmax)].strip('\n')
def main():
print("STARTING UNETv2 on DPU...")
if USE_DPU:
# Attach to DPU driver
n2cube.dpuOpen()
# Load DPU Kernel and create a task
kernel = n2cube.dpuLoadKernel(KERNEL_CONV)
task = n2cube.dpuCreateTask(kernel, 0)
# load and preprocess images and load segmentation labels
assert os.path.isdir(IMG_TEST_DIR)
#print(IMG_TEST_DIR)
x_test, y_test, img_file, seg_file = dpu_get_data(IMG_TEST_DIR,
SEG_TEST_DIR,
cfg.NUM_CLASSES,
cfg.WIDTH, cfg.HEIGHT)
y_pred = []
# process all images
for i in range(len(x_test)):
# opened image as BGR, convert it to RGB
#B,G,R = cv2.split(x_test[i])
#imageRun = cv2.merge((R,G,B))
imageRun = x_test[i]
imageRun = imageRun.reshape(
(imageRun.shape[0] * imageRun.shape[1] * imageRun.shape[2]))
input_len = len(imageRun)
if USE_DPU:
# load pre-processed image as DPU input
n2cube.dpuSetInputTensorInHWCFP32(task, CONV_INPUT_NODE, imageRun,
input_len)
dpu_in = n2cube.dpuGetInputTensor(task, CONV_INPUT_NODE)
ti_scale = n2cube.dpuGetTensorScale(dpu_in)
ti_h = n2cube.dpuGetTensorHeight(dpu_in)
ti_w = n2cube.dpuGetTensorWidth(dpu_in)
ti_sz = n2cube.dpuGetTensorSize(dpu_in)
ti_ch = n2cube.dpuGetTensorChannel(dpu_in)
if (i == 0):
print(
"Input tensor=%3d ch=%3d H=%3d W=%3d Size=%6d scale=%4d" %
(i, ti_ch, ti_h, ti_w, ti_sz, ti_scale))
# run DPU task
n2cube.dpuRunTask(task)
# get output tensor address
dpu_out = n2cube.dpuGetOutputTensorAddress(task, CONV_OUTPUT_NODE)
# get number of channels in output tensor
to_ch = n2cube.dpuGetOutputTensorChannel(task, CONV_OUTPUT_NODE)
# get size in bytes of output tensor
to_sz = n2cube.dpuGetOutputTensorSize(task, CONV_OUTPUT_NODE)
# get width output tensor
to_w = n2cube.dpuGetOutputTensorWidth(task, CONV_OUTPUT_NODE)
# get height output tensor
to_h = n2cube.dpuGetOutputTensorHeight(task, CONV_OUTPUT_NODE)
# get output tensor scale
to_scale = n2cube.dpuGetOutputTensorScale(task, CONV_OUTPUT_NODE)
softmax = np.zeros(to_sz, dtype=np.float32)
if (i == 0):
print("Output tensor=%3d ch=%3d H=%3d W=%3d Size=%6d" %
(i, to_ch, to_h, to_w, to_sz))
print("Output tensor scaling factor", to_scale)
softmax = n2cube.dpuRunSoftmax(dpu_out, to_ch, to_sz // to_ch,
to_scale)
prediction = softmax.reshape((to_h, to_w, to_ch))
y_pred.append(prediction)
if (i == 0):
print("prediction shape: ", prediction.shape)
# Calculate intersection over union for each segmentation class
y_pred = np.asarray(y_pred)
y_test = np.asarray(y_test)
print("y_pred shape: ", y_pred.shape)
print("y_test shape: ", y_test.shape)
y_predi = np.argmax(y_pred, axis=3)
y_testi = np.argmax(y_test, axis=3)
print("shape of y_testi and y_predi ", y_testi.shape, y_predi.shape)
dpu_IoU(y_testi, y_predi)
# print results
print("Processed", len(x_test), "images")
print("FINISHED")
if USE_DPU:
# Destroy DPU Kernel & detach
n2cube.dpuDestroyKernel(kernel)
n2cube.dpuClose()