def __init__(self, input_img, bitfile, model_path1, model_path2, numfpga,
numclus):
"""
In this example MDLA will be capable of taking an input image
and running that image on all clusters
"""
print('{}{}{}...'.format(CP_Y, 'Initializing MDLA', CP_0))
################################################################################
# Initialize 2 Micron DLA
self.dla1 = microndla.MDLA()
self.dla2 = microndla.MDLA()
# Run the network in batch mode (one image on all clusters)
self.dla1.SetFlag('clustersbatchmode', '0')
self.dla2.SetFlag('clustersbatchmode', '0')
self.batch, self.height, self.width, self.channels = input_img.shape
self.dla1.SetFlag('nclusters', str(numclus))
self.dla2.SetFlag('nclusters', str(numclus))
# Compile the NN and generate instructions <save.bin> for MDLA
self.dla1.Compile(model_path1)
self.dla2.Compile(model_path2)
print('{}{}{}!!!'.format(CP_C,
'Successfully generated binaries for MDLA',
CP_0))
# Send the generated instructions to MDLA
# Send the bitfile to the FPGA only during the first run
# Otherwise bitfile is an empty string
print('{}{}{}!!!'.format(CP_G, 'MDLA initialization complete', CP_0))
print('{:-<80}'.format(''))
def __init__(self, input_img, bitfile, model_path1, model_path2, numfpga, numclus):
"""
In this example MDLA will be capable of taking multiple input images
and running that images through 2 models on 1 fpga
"""
print('{}{}{}...'.format(CP_Y, 'Initializing MDLA', CP_0))
# Initialize 1 Micron DLA
self.dla = microndla.MDLA()
self.dla2 = microndla.MDLA()
# Run the network in batch mode (one image on all clusters)
self.batch, self.height, self.width, self.channels = input_img.shape
# Compile the NN and generate instructions <save.bin> for MDLA
self.dla.SetFlag({'nclusters': numclus, 'clustersbatchmode': 1})
self.dla2.SetFlag({'nclusters': numclus, 'clustersbatchmode': 1, 'firstcluster': numclus})
#self.dla.SetFlag('debug', 'bw') # Comment it out to see detailed output from compiler
self.dla.Compile(model_path1)
self.dla2.Compile(model_path2)
print('{}{}{}!!!'.format(CP_C, 'Successfully generated binaries for MDLA', CP_0))
# Send the generated instructions to MDLA
# Send the bitfile to the FPGA only during the first run
# Otherwise bitfile is an empty string
print('{}{}{}!!!'.format(CP_G, 'MDLA initialization complete', CP_0))
print('{:-<80}'.format(''))
def __init__(self, input_img, n_classes, bitfile, model_path):
"""
In this example MDLA will be capable of taking an input image
and running that image on all clusters
"""
print('{}{}{}...'.format(CP_Y, 'Initializing MDLA', CP_0))
################################################################################
# Initialize Micron DLA
self.dla = microndla.MDLA()
if bitfile and bitfile != '':
self.dla.SetFlag('bitfile', bitfile)
print('{}{}{}'.format(CP_C, 'Finished loading bitfile on FPGA',
CP_0))
# Run the network in no-batch mode (one image on all clusters)
self.dla.SetFlag('clustersbatchmode', '1')
# TODO Uncomment this line to see detailed compiler output
#self.dla.SetFlag('debug', 'b')
self.height, self.width, self.channels = input_img.shape
# Compile the NN and generate instructions <save.bin> for MDLA
self.dla.Compile(model_path)
print('{}{}{}'.format(CP_C, 'Successfully generated binaries for MDLA',
CP_0))
# Send the generated instructions to MDLA
# Send the bitfile to the FPGA only during the first run
# Otherwise bitfile is an empty string
print('\n{}{}{}!!!'.format(CP_G, 'MDLA initialization complete', CP_0))
print('{:-<80}'.format(''))
# Allocate space for output if the model
self.n_classes = n_classes # Number of expected output planes/classes
def __init__(self, input_img, bitfile, model_path, numfpga, numclus):
"""
In this example MDLA will be capable of taking an input image
and running that image on all clusters
"""
print('{}{}{}...'.format(CP_Y, 'Initializing MDLA', CP_0))
# Initialize Micron DLA
self.dla = microndla.MDLA()
self.batch, self.height, self.width, self.channels = input_img.shape
# Run the network in batch mode (two images, one on each cluster)
image_per_cluster=self.batch/numclus/numfpga
if image_per_cluster==1:
self.dla.SetFlag('clustersbatchmode', '0')
else:
self.dla.SetFlag('imgs_per_cluster', str(image_per_cluster))
self.dla.SetFlag('nfpgas', str(numfpga))
self.dla.SetFlag('nclusters', str(numclus))
if bitfile and bitfile != '':
self.dla.SetFlag('bitfile', bitfile)
print('{}{}{}'.format(CP_C, 'Finished loading bitfile on FPGA', CP_0))
#self.dla.SetFlag('debug', 'b') # Uncomment it to see detailed output from compiler
# Compile the NN and generate instructions <save.bin> for MDLA
sz = "{:d}x{:d}x{:d}x{:d}".format(self.batch, self.channels, self.height, self.width)
self.dla.Compile(model_path, 'save.bin', sz)
print('{}{}{}!!!'.format(CP_C, 'Successfully generated binaries for MDLA', CP_0))
# Send the generated instructions to MDLA
# Send the bitfile to the FPGA only during the first run
# Otherwise bitfile is an empty string
self.dla.Init('save.bin')
print('{}{}{}!!!'.format(CP_G, 'MDLA initialization complete', CP_0))
print('{:-<80}'.format(''))
def __init__(self,
input_img,
bitfile,
model_path,
numfpga=1,
numclus=1,
nobatch=False):
print('Initializing MDLA')
self.dla = microndla.MDLA() # initialize MDLA
sz = "{:d}x{:d}x{:d}".format(224, 224,
1) # input size from the ONNX model
if nobatch: # Check if you need to run one image on whole fpga or not
self.dla.SetFlag('clustersbatchmode', '1')
self.dla.SetFlag('nclusters', str(numclus))
self.dla.SetFlag('nfpgas', str(numfpga))
if bitfile and bitfile != '':
self.dla.SetFlag('bitfile', bitfile)
#self.dla.SetFlag('debug', 'b') # Comment it out for detailed output from compiler
self.dla.Compile(
model_path, 'save.bin'
) # Compile the NN and generate instructions <save.bin> for MDLA
print('\nSuccesfully generated binaries for MDLA')
self.dla.Init(
'save.bin'
) # Send instruction to FPGA and load bitfile if necessary
print('MDLA initialization complete\n')
def __init__(self,
input_img,
bitfile,
model_path,
numfpga=1,
nobatch=False):
self.dla = microndla.MDLA()
b, h, w, c = input_img.shape
if nobatch:
self.dla.SetFlag('clustersbatchmode', '1')
assert b == 1, "Input batch should be equal to 1 for nobatch mode"
self.dla.SetFlag('nfpgas', str(numfpga))
if bitfile and bitfile != '':
self.dla.SetFlag('bitfile', bitfile)
self.dla.Compile(model_path)
self.cfg = yolov3_cfg
self.grids = []
self.n = []
self.anchors = []
self.strides = []
self.na = 3
self.no = 85
self.create_grids(h, w)
def __init__(self,
model_path,
class_names,
res,
bitfile,
numclus=4,
threshold=0.5,
disp_time=1):
self.thr = threshold
self.times = deque(maxlen=25)
self.disp_time = disp_time
# Load class names from file
with open(class_names, 'r') as f:
self.labels = f.readlines()
for i in range(len(self.labels)):
self.labels[i] = self.labels[i].rstrip()
# Initialize Micron DLA
self.dla = microndla.MDLA()
self.res = res
w, h, c = res
# Run the network in batch mode (one image on all clusters)
self.dla.SetFlag('clustersbatchmode', '1')
# Compile the NN and generate instructions <save.bin> for MDLA
if bitfile and bitfile != '':
self.dla.SetFlag('bitfile', bitfile)
self.dla.SetFlag('nclusters', str(numclus))
sz = '{:1}x{:d}x{:d}x{:d}'.format(1, c, h, w)
self.dla.Compile(model_path, 'save.bin', sz)
# Init fpga with compiled machine code
self.dla.Init('save.bin')
# Model has 10 outputs that each need to be reshaped to the following sizes
self.output_shapes = [
(1, 720, int(h / 8 + .5), int(w / 8 + .5)),
(1, 720, int(h / 16 + .5), int(w / 16 + .5)),
(1, 720, int(h / 32 + .5), int(w / 32 + .5)),
(1, 720, int(h / 64 + .5), int(w / 64 + .5)),
(1, 720, int(h / 128 + .5), int(w / 128 + .5)),
(1, 36, int(h / 8 + .5), int(w / 8 + .5)),
(1, 36, int(h / 16 + .5), int(w / 16 + .5)),
(1, 36, int(h / 32 + .5), int(w / 32 + .5)),
(1, 36, int(h / 64 + .5), int(w / 64 + .5)),
(1, 36, int(h / 128 + .5), int(w / 128 + .5)),
]
def ieprocess(image_file, network_file):
# load image and resize it:
img = Image.open(image_file)
#Resize it to the size expected by the network
img = img.resize((224, 224), resample=Image.BILINEAR)
#Convert to numpy float
img = np.array(img).astype(np.float32) / 255
#Transpose to plane-major, as required by our API
img = np.ascontiguousarray(img.transpose(2, 0, 1))
# print(img)
print('Image shape:', img.shape)
#Normalize images
stat_mean = list([0.485, 0.456, 0.406])
stat_std = list([0.229, 0.224, 0.225])
for i in range(3):
img[i] = (img[i] - stat_mean[i]) / stat_std[i]
#Create and initialize the Inference Engine object
ie = microndla.MDLA()
#Compile to a file
swnresults = ie.Compile("{:d}x{:d}x{:d}".format(224, 224, 3), network_file,
'save.bin')
#Init fpga
nresults = ie.Init('save.bin', '')
#Create the storage for the result and run one inference
result = np.ndarray(swnresults, dtype=np.float32)
ie.Run(img, result)
#Convert to numpy and print top-5
idxs = (-result).argsort()
rstring = []
with open("categories.txt") as f:
categories = f.read().splitlines()
for i in range(5):
rstring.append(
str(categories[idxs[i]]) + ', ' + str(result[idxs[i]]))
#Free
ie.Free()
return rstring
hidden2 = (torch.randn(1, 1, nh2), torch.randn(1, 1, nh2)) # clean out hidden state
inputs = torch.cat(inputs).view(len(inputs), 1, -1)
modelL = LSTMm()
# Export onnx
torch.onnx.export(modelL, (inputs, hidden, hidden2), "model.onnx")
# Run in pytorch
out = modelL(inputs, hidden, hidden2)
result_pyt = out[0]
result_pyt = result_pyt.permute(1, 0, 2).contiguous()
result_pyt = result_pyt.view(1,-1)
result_pyt = result_pyt.detach().numpy()
#Create and initialize the Inference Engine object
ie = microndla.MDLA()
ie.SetFlag('debug','bw')
#Compile to a file
ie.Compile('model.onnx', 'model.bin')
#Init fpga
ie.Init('model.bin')
np.random.seed(1)
img = inputs.numpy().transpose(1, 0, 2)
hid = [hidden[0].numpy().transpose(1,0,2),
hidden[1].numpy().transpose(1,0,2)]
hid2 = [hidden2[0].numpy().transpose(1,0,2),
hidden2[1].numpy().transpose(1,0,2)]
def forward(self, x):
y = self.op(x)
return y
w = args.w
i = args.i
k = args.k
s = args.s
p = args.p
inVec1 = torch.randn(1, i, w, w, dtype=torch.float32)
modelMax = Maxpool(k, s, p)
torch.onnx.export(modelMax, inVec1, "net_maxpool.onnx")
sf = microndla.MDLA()
if args.verbose:
sf.SetFlag('debug', 'b') #debug options
# Compile to generate binary
sf.Compile('net_maxpool.onnx', 'net_maxpool.bin')
sf.Init("./net_maxpool.bin")
in_1 = np.ascontiguousarray(inVec1)
result = sf.Run(in_1)
outhw = modelMax(inVec1)
result_pyt = outhw.detach().numpy()
if args.verbose:
print("pytorch : {}".format(result_pyt))
print("hw : {}".format(result))
import microndla
import sys
import PIL
from PIL import Image
import numpy as np
from argparse import ArgumentParser
parser = ArgumentParser(description="Micron DLA Load bitfile")
_ = parser.add_argument
_('bitfile', type=str, default='', help='Path to the bitfile')
_('-f',
'--fpga',
type=str,
default='',
help='Select fpga type to use: 511 or 852')
_('-n', '--nfpga', type=str, default='1', help='number of fpgas used')
args = parser.parse_args()
ie = microndla.MDLA() # create MDLA obj
#ie.SetFlag('debug', 'bw') # select fpga type
if args.fpga == "511" or args.fpga == "852":
ie.SetFlag('fpgaid', args.fpga) # select fpga type
ie.SetFlag('nfpgas', args.nfpga) # select fpga type
ie.SetFlag('bitfile', args.bitfile) # load bitfile
ie.Free() # free MDLA obj
print('done')
for i in range(2):
#Convert to numpy float
img[i] = np.array(img[i]).astype(np.float32) / 255
#Transpose to plane-major, as required by our API
img[i] = np.ascontiguousarray(img[i].transpose(2, 0, 1))
#Normalize images
stat_mean = list([0.485, 0.456, 0.406])
stat_std = list([0.229, 0.224, 0.225])
for j in range(3):
img[i][j] = (img[i][j] - stat_mean[j]) / stat_std[j]
#Create and initialize the Inference Engine object
nclus = 2
ie = microndla.MDLA()
ie2 = microndla.MDLA()
ie.SetFlag({'nclusters': nclus, 'clustersbatchmode': 1})
ie2.SetFlag({
'nclusters': nclus,
'firstcluster': nclus,
'clustersbatchmode': 1
})
#Compile to a file
ie.Compile(args.modelpath1)
ie2.Compile(args.modelpath2, MDLA=ie)
#Create the storage for the result and run one inference
ie.PutInput(img[0], None)
ie2.PutInput(img[1], None)
