def loadFCWeightsBias(data_dir):
weight = []
fname = "%s/fc" % data_dir
if not os.path.exists(fname):
nearMatch = getNearFileMatchWithPrefix(data_dir, "fc")
if nearMatch:
fname = nearMatch
if os.path.exists(fname):
with open(fname, 'r') as f:
line = f.read()
vals = line.strip().split(' ')
weight = [float(v) for v in vals]
else:
print(("No FC layers found in %s" % data_dir))
return (None, None)
bias = []
fname = "%s/fc_bias" % data_dir
if not os.path.exists(fname):
nearMatch = getNearFileMatchWithPrefix(data_dir, "fc_bias")
if nearMatch:
fname = nearMatch
with open(fname, 'r') as f:
line = f.read()
vals = line.strip().split(' ')
bias = [float(v) for v in vals]
(weightFpga, fpgaHandle) = xdnn.makeFPGAFloatArray(len(weight))
(biasFpga, fpgaHandle) = xdnn.makeFPGAFloatArray(len(bias))
weightFpga[:] = weight
biasFpga[:] = bias
return (weightFpga, biasFpga)
def prepareOutput(output_sz, batch_sz):
"""
Allocate memory for outputs.
:param output_sz: Number of elements in the output volume returned by FPGA for a single inference. This is specific to the network.
:type output_sz: int.
:param batch_sz: Number of images to be processed simultaneously.
:type batch_sz: int.
:returns: numpy.ndarray -- 1D Array large enough to hold output_sz*batch_sz elements.
"""
(fpgaOutput, fpgaHandle) = xdnn.makeFPGAFloatArray(output_sz * batch_sz)
return fpgaOutput
def init_fpga():
global g_inputs
global g_inputbuf
global g_fpgaOutput
global g_weightsBlob
global g_fcWeight
global g_fcBias
print(" --- INIT FPGA --- \n")
print("xclbin: {0}.\n".format(g_xclbin))
print("xdnnLib: {0}.\n".format(g_xdnnLib))
ret = xdnn.createManager(g_xdnnLib)
if ret != True:
raise SystemExit("Error: xdnn createManager failed.")
(g_fcWeight, g_fcBias) = xdnn_io.loadFCWeightsBias(g_xdnnTestDataDir)
ret = xdnn.createHandle(g_xclbin, "kernelSxdnn_0", g_xdnnLib, g_numDevices)
if ret:
raise SystemExit("ERROR: Unable to create handle to FPGA")
else:
print("INFO: Sucessfully create handle to FPGA.")
# magics. See ml-suite/notebooks tutorial. Should we overwrite PE?
args = {
'datadir': g_xdnnTestDataDir,
'quantizecfg': g_fpgaCfgFile,
'scaleA': g_scaleA,
'scaleB': g_scaleB,
'PE': -1,
'netcfg': g_netFile
}
print(" --- load weights --- \n")
g_weightsBlob = xdnn_io.loadWeightsBiasQuant(args)
print(" --- read lable file --- \n")
with open(g_lableFile, 'r') as f:
for line in f:
g_labelarray.append(line.strip())
print(" --- prepare inputs --- \n")
g_inputs = np.zeros((g_batchSize, g_img_c * g_img_h * g_img_w),
dtype=np.float32)
g_inputbuf = np.zeros((g_batchSize, g_img_c, g_img_h, g_img_w),
dtype=np.float32)
print "g_inputs", g_inputs
print(" --- prepare outputs --- \n")
g_fpgaOutput, fpgaHandle = xdnn.makeFPGAFloatArray(g_fpgaOutputSize *
g_batchSize)
def prepareOutput(num):
(fpgaOutput, fpgaHandle) \
= xdnn.makeFPGAFloatArray(g_fpgaOutputSize * num)
return fpgaOutput
def prepareOutput(output_sz, batch_sz):
(fpgaOutput, _) = xdnn.makeFPGAFloatArray(output_sz * batch_sz)
return fpgaOutput