def readXMsg(sock):
s = sock.recv(4)
if s is None or s == '':
return None
sz = struct.unpack('<i', s)[0]
if sz != 0x20aa30bb:
raise Exception('Bad magic')
s = sock.recv(4)
if s is None or s == '':
raise Exception('Cannot read message sz')
sz = struct.unpack('<i', s)[0]
recvsz = 0
msgstr = ""
while recvsz < sz:
r = sock.recv(sz - recvsz)
msgstr = msgstr + r
recvsz = recvsz + len(r)
xmsg = xdrive_pb2.XMsg()
xmsg.ParseFromString(msgstr)
return xmsg
def img_classify(msg):
global g_inputs
global g_inputbuf
global g_fpgaOutput
global g_weightsBlob
global g_fcWeight
global g_fcBias
# message is a rowset, one col, a list of file names.
rs = msg.rowset
if len(rs.columns) == 0 or rs.columns[0].nrow == 0:
print("Img classify request size is 0.\n")
return None
print("Img classify request size is {0}.\n".format(rs.columns[0].nrow))
# Lock the fpga device. config is protected by this lock as well.
fpga_lock.acquire()
ret = None
for i in range(rs.columns[0].nrow):
fname = rs.columns[0].sdata[i]
print("Running classification for images: {0}\n".format(fname))
print("Prepare inputs ...\n")
# g_batchSize = 1, for now.
print "g_inputs", g_inputs
g_inputs[0] = xdnn_io.loadImageBlobFromFile(str(fname), g_mean,
g_img_h, g_img_w)
print("Quantize inputs ...\n")
quantizeInputs = xdnn.quantizeInputs(g_firstFpgaLayerName, g_inputs,
None, None, g_fpgaCfgFile,
g_scaleB)
print("Prepare inputs for fpga inputs ...\n")
fpgaInputs = xdnn.prepareInputsForFpga(quantizeInputs, g_fpgaCfgFile,
g_scaleB, -1,
g_firstFpgaLayerName)
print("Run FPGA commands ...\n")
xdnn.execute(g_netFile, g_weightsBlob, fpgaInputs, g_fpgaOutput,
g_batchSize, g_fpgaCfgFile, g_scaleB, g_PE)
print("Compute FC ...\n")
fcOutput = xdnn.computeFC(g_fcWeight, g_fcBias, g_fpgaOutput,
g_batchSize, g_outputSize, g_fpgaOutputSize,
g_useBlas)
print("Softmax ...\n")
softmaxOut = xdnn.computeSoftmax(fcOutput, g_batchSize)
ret = get_classification(softmaxOut, fname)
fpga_lock.release()
# Now construct return msg
if ret == None:
print("Return None: ???\n")
return None
retmsg = xdrive_pb2.XMsg()
rs = retmsg.rowset
# return 4 columns, (filename, ordinal, score, class)
col1 = rs.columns.add()
col2 = rs.columns.add()
col3 = rs.columns.add()
col4 = rs.columns.add()
col1.nrow = len(ret)
col2.nrow = len(ret)
col3.nrow = len(ret)
col4.nrow = len(ret)
for i in range(len(ret)):
(a, b, c, d) = ret[i]
# print("Return {0}, {1}, {2}, {3}.\n".format(a, b, c, d))
col1.nullmap.append(False)
col1.sdata.append(a)
col2.nullmap.append(False)
col2.i32data.append(b)
col3.nullmap.append(False)
col3.f64data.append(c)
col4.nullmap.append(False)
col4.sdata.append(d)
return retmsg
#
# googlenet test
#
import sys
import xdrive_pb2, server
if __name__=='__main__':
# Test: an echo server.
sock = server.cli_connect("/tmp/ml.sock")
imgs = ["apple.jpeg", "banana.jpeg", "beer.jpeg", "coffee.jpeg", "egg.jpeg", "salad.jpeg"]
for ii in range(6):
xmsg = xdrive_pb2.XMsg()
col = xmsg.rowset.columns.add()
col.nrow = 1
col.nullmap.append(False)
col.sdata.append(sys.argv[1] + "/test/" + imgs[ii])
server.writeXMsg(sock, xmsg)
ret = server.readXMsg(sock)
col1 = ret.rowset.columns[0]
col2 = ret.rowset.columns[1]
col3 = ret.rowset.columns[2]
col4 = ret.rowset.columns[3]
nrow = ret.rowset.columns[0].nrow
for i in range(nrow):
print("Ret {0}: ({1}, {2}, {3}, {4}).\n".format(
i,
col1.sdata[i],
col2.i32data[i],
def img_classify(msg):
global g_args
global g_ctxt
# message is a rowset, one col, a list of file names.
rs = msg.rowset
if len(rs.columns) == 0 or rs.columns[0].nrow == 0:
print("Img classify request size is 0.\n")
return None
print("Img classify request size is {0}.\n".format(rs.columns[0].nrow))
# Lock the fpga device. config is protected by this lock as well.
fpga_lock.acquire()
ret = []
if is_deploymode():
firstInput = g_ctxt['fpgaInput'].itervalues().next()
firstOutput = g_ctxt['fpgaOutput'].itervalues().next()
for i in xrange(0, rs.columns[0].nrow, g_args['batch_sz']):
pl = []
for j in range(g_args['batch_sz']):
fname = str(rs.columns[0].sdata[i + j])
print("Running classification for {0}-th images: {1}\n".format(
i + j, fname))
if is_deploymode():
firstInput[j, ...], _ = xdnn_io.loadImageBlobFromFile(
fname, g_args['img_raw_scale'], g_args['img_mean'],
g_args['img_input_scale'], g_ctxt['inShape'][2],
g_ctxt['inShape'][3])
else:
g_ctxt['batch_array'][j,
...], _ = xdnn_io.loadImageBlobFromFile(
fname, g_args['img_raw_scale'],
g_args['img_mean'],
g_args['img_input_scale'],
g_ctxt['in_shape'][2],
g_ctxt['in_shape'][1])
pl.append(fname)
if is_deploymode():
g_ctxt['fpgaRT'].execute(g_ctxt['fpgaInput'], g_ctxt['fpgaOutput'])
xdnn.computeFC(g_ctxt['fcWeight'], g_ctxt['fcBias'], firstOutput,
g_ctxt['fcOutput'])
else:
g_ctxt['fpgaRT'].execute(g_ctxt['batch_array'],
g_ctxt['fpgaOutput'])
xdnn.computeFC(g_ctxt['fcWeight'], g_ctxt['fcBias'],
g_ctxt['fpgaOutput'], g_args['batch_sz'],
g_args['outsz'], g_args['fpgaoutsz'],
g_ctxt['fcOutput'])
softmaxOut = xdnn.computeSoftmax(g_ctxt['fcOutput'])
ret = ret + get_classification(softmaxOut, pl, g_ctxt['labels'])
fpga_lock.release()
retmsg = xdrive_pb2.XMsg()
rs = retmsg.rowset
# return 4 columns, (filename, ordinal, score, class)
col1 = rs.columns.add()
col2 = rs.columns.add()
col3 = rs.columns.add()
col4 = rs.columns.add()
col1.nrow = len(ret)
col2.nrow = len(ret)
col3.nrow = len(ret)
col4.nrow = len(ret)
for i in range(len(ret)):
# print("Return {0}, {1}, {2}, {3}.\n".format(a, b, c, d))
col1.nullmap.append(False)
col1.sdata.append(ret[i][0])
col2.nullmap.append(False)
col2.i32data.append(ret[i][1])
col3.nullmap.append(False)
col3.f64data.append(ret[i][2])
col4.nullmap.append(False)
col4.sdata.append(ret[i][3])
return retmsg
def obj_detect(msg):
global g_qIn, g_qOut
rs = msg.rowset
if len(rs.columns) == 0 or rs.columns[0].nrow == 0:
print("Obj deection req size is 0.\n")
return None
# Input, will be a video file, start time, for how long.
fname = rs.columns[0].sdata[0]
start = rs.columns[1].f32data[0]
duration = rs.columns[2].f32data[0]
ret = []
# use opencv to get frames
print ("Obj dectect on file {0}: start {1}, length {2}.\n", fname, rs.columns[1].f32data[0], rs.columns[2].f32data[0])
vc = cv2.VideoCapture(fname)
# 5: fps.
fps = vc.get(5)
if start > 1.0:
# set 0: position to milissec.
# set 1: postiion to frame number
vc.set(0, start * 1000)
i = 0
while i <= duration * fps:
i += 1
ok, frame = vc.read()
if not ok:
break
if (i - 1) % g_skip == 0:
# got a frame, do some transformation, then send it to FPGA.
inputs = np.zeros((g_batchSize, g_imgc*g_imgh*g_imgw), dtype = np.float32)
inputs[0] = load_yoloimg(frame)
fpga_lock.acquire()
g_qIn.put(inputs)
outputs = g_qOut.get()
fpga_lock.release()
# running the rest of yolo layer in CPU.
outputs = outputs.reshape(g_anchor_boxes, g_outc, g_outh, g_outw)
# sigmoid
outputs[:,0:2,:,:] = sigmoid(outputs[:,0:2,:,:])
outputs[:,4,:,:] = sigmoid(outputs[:,4,:,:])
for box in range(g_anchor_boxes):
outputs[box,5:,:,:] = softmax(outputs[box,5:,:,:])
bboxes = nms.do_baseline_nms(outputs.flat,
frame.shape[1], frame.shape[0],
g_imgw, g_imgh,
g_outw, g_outh,
g_anchor_boxes, g_classes,
g_scorethresh, g_iouthresh
)
for j in range(len(bboxes)):
cls = coconames(bboxes[j]['classid'])
if cls is None:
continue
llx = bboxes[j]['ll']['x']
lly = bboxes[j]['ll']['y']
urx = bboxes[j]['ur']['x']
ury = bboxes[j]['ur']['y']
# very tall/wide objects, we don't want to covering bbox
if ((urx-llx) > frame.shape[1] * 0.5) or ((lly - ury) > frame.shape[0] * 0.5):
continue
# and avoid objects less than 30x30.
if (urx-llx > 30) and (lly-ury > 30):
objimg = frame[ury:lly, llx:urx]
objimg_str = cv2.imencode('.jpg', objimg)[1].tostring()
objimg_str = base64.b64encode(objimg_str)
ret.append((i, cls, bboxes[j]['prob'],
llx, lly, urx, ury,
objimg_str))
vc.release()
# return resuts
retmsg = xdrive_pb2.XMsg()
rs = retmsg.rowset
col1 = rs.columns.add()
col2 = rs.columns.add()
col3 = rs.columns.add()
col4 = rs.columns.add()
col5 = rs.columns.add()
col6 = rs.columns.add()
col7 = rs.columns.add()
col8 = rs.columns.add()
col1.nrow = len(ret)
col2.nrow = len(ret)
col3.nrow = len(ret)
col4.nrow = len(ret)
col5.nrow = len(ret)
col6.nrow = len(ret)
col7.nrow = len(ret)
col8.nrow = len(ret)
for r in ret:
col1.nullmap.append(False)
col1.i32data.append(r[0])
col2.nullmap.append(False)
col2.sdata.append(r[1])
col3.nullmap.append(False)
col3.f32data.append(r[2])
col4.nullmap.append(False)
col4.f32data.append(r[3])
col5.nullmap.append(False)
col5.f32data.append(r[4])
col6.nullmap.append(False)
col6.f32data.append(r[5])
col7.nullmap.append(False)
col7.f32data.append(r[6])
col8.nullmap.append(False)
col8.sdata.append(r[7])
return retmsg
def img_classify(msg):
global g_cInputBuffer
global g_cFpgaInputBuffer
# message is a rowset, one col, a list of file names.
rs = msg.rowset
if len(rs.columns) == 0 or rs.columns[0].nrow == 0:
print("Img classify request size is 0.\n")
return None
print("Img classify request size is {0}.\n".format(rs.columns[0].nrow))
# Lock the fpga device. config is protected by this lock as well.
fpga_lock.acquire()
ret = None
for i in range(rs.columns[0].nrow):
fname = rs.columns[0].sdata[i]
print("Running classification for images: {0}\n".format(fname))
print("Prepare inputs ...\n")
# g_batchSize = 1, for now.
config["g_inputs"][0] = pyxfdnn_io.loadImageBlobFromFile(
fname, config["img_mean"], g_imgh, g_imgw)
print("Quantize inputs ...\n")
quantizeInputs = pyxfdnn.quantizeInputs(
config["firstfpgalayer"], config["g_inputs"], g_cInputBuffer,
g_cFpgaInputBuffer, config["quantizecfg"], config["scaleB"])
print("Prepare inputs for fpga inputs ...\n")
fpgaInputs = pyxfdnn.prepareInputsForFpga(quantizeInputs,
config["quantizecfg"],
config["scaleB"], -1,
config["firstfpgalayer"])
print("Run FPGA commands ...\n")
pyxfdnn.execute(
config["fpgacommands"], config["weightsBlob"], fpgaInputs,
config["g_fpgaOutput"], g_batchSize, config["quantizecfg"],
config["scaleB"]
#
# This is freaking insane. What is PE?
#
# Xilinx notebook uses PE = 0, which works for a few images then crash.
# Xilinx example batch_classify.py says do not supply this PE paramenter,
# then default is -1. Runs fine for many images.
#
# , config["PE"]
#
)
print("Compute FC ...\n")
fcOut = pyxfdnn.computeFC(config["fcWeight"], config["fcBias"],
config["g_fpgaOutput"], g_batchSize,
config["outsz"], config["fpgaoutsz"],
config["useblas"])
print("Softmax ...\n")
softmaxOut = pyxfdnn.computeSoftmax(fcOut, g_batchSize)
ret = get_classification(softmaxOut, fname, config)
fpga_lock.release()
# Now construct return msg
if ret == None:
print("Return None: ???\n")
return None
retmsg = xdrive_pb2.XMsg()
rs = retmsg.rowset
# return 4 columns, (filename, ordinal, score, class)
col1 = rs.columns.add()
col2 = rs.columns.add()
col3 = rs.columns.add()
col4 = rs.columns.add()
col1.nrow = len(ret)
col2.nrow = len(ret)
col3.nrow = len(ret)
col4.nrow = len(ret)
for i in range(len(ret)):
(a, b, c, d) = ret[i]
# print("Return {0}, {1}, {2}, {3}.\n".format(a, b, c, d))
col1.nullmap.append(False)
col1.sdata.append(a)
col2.nullmap.append(False)
col2.i32data.append(b)
col3.nullmap.append(False)
col3.f64data.append(c)
col4.nullmap.append(False)
col4.sdata.append(d)
return retmsg