updating to use CNTKs python wrappers and an improved selective search
This commit is contained in:
Patrick Buehler
Родитель
7edd3276a1
Коммит
bc28ace57c
79
3_runCntk.py
79
3_runCntk.py
|
|
@ -1,68 +1,37 @@
|
|||
import os, sys, importlib
|
||||
import shutil, time
|
||||
import subprocess
|
||||
import PARAMETERS
|
||||
locals().update(importlib.import_module("PARAMETERS").__dict__)
|
||||
from PARAMETERS import *
|
||||
import numpy as np
|
||||
import os, sys, pdb
|
||||
|
||||
|
||||
#TODO: move all cntk imports (and hence also code) inside here and then remove from helpers.py
|
||||
|
||||
####################################
|
||||
# Parameters
|
||||
# MAIN
|
||||
####################################
|
||||
print "classifier = " + classifier
|
||||
cntkCmdStrPattern = "{0}/cntk.exe configFile={1}config.cntk currentDirectory={1}"
|
||||
makeDirectory(modelDir)
|
||||
print ("classifier = " + classifier)
|
||||
print ("cntk_lr_per_image = " + str(cntk_lr_per_image))
|
||||
|
||||
|
||||
|
||||
####################################
|
||||
# Main
|
||||
####################################
|
||||
deleteAllFilesInDirectory(cntkFilesDir + "/tmp", None)
|
||||
shutil.copy(os.path.join(cntkResourcesDir, "config.cntk"), cntkFilesDir)
|
||||
|
||||
#generate cntk command string
|
||||
cmdStr = cntkCmdStrPattern.format(cntkBinariesDir, cntkFilesDir, classifier)
|
||||
cmdStr += " ImageH={} ImageW={}".format(cntk_padHeight, cntk_padWidth)
|
||||
cmdStr += " NumLabels={0} NumTrainROIs={1} NumTestROIs={1}".format(len(classes), cntk_nrRois)
|
||||
cmdStr += " TrainROIDim={} TrainROILabelDim={}".format(4*cntk_nrRois, cntk_nrRois * cntk_featureDimensions[classifier])
|
||||
cmdStr += " TestROIDim={} TestROILabelDim={}".format( 4*cntk_nrRois, cntk_nrRois * cntk_featureDimensions[classifier])
|
||||
# optionally retrain DNN
|
||||
# if the classifier is svm, then simply return the 4096-floats penultimate layer as model
|
||||
# otherwise add new output layer, retrain the DNN, and return this new model.
|
||||
if classifier == 'svm':
|
||||
cmdStr += " [Train=[SGD=[maxEpochs=0]]]" #no need to train the network if just using it as featurizer
|
||||
cmdStr += " [WriteTest=[outputNodeNames=(z.fcOut.h2.y)]]"
|
||||
cmdStr += " [WriteTrain=[outputNodeNames=(z.fcOut.h2.y)]]"
|
||||
boSkipTraining = True
|
||||
else:
|
||||
boSkipTraining = False
|
||||
model = init_train_fast_rcnn(cntk_padHeight, cntk_padWidth, nrClasses, cntk_nrRois, cntk_mb_size, cntk_max_epochs,
|
||||
cntk_lr_per_image, cntk_l2_reg_weight, cntk_momentum_time_constant, cntkFilesDir, boSkipTraining)
|
||||
|
||||
#run cntk
|
||||
tstart = datetime.datetime.now()
|
||||
os.environ['ACML_FMA'] = str(0)
|
||||
print cmdStr
|
||||
pid = subprocess.Popen(cmdStr, cwd = cntkFilesDir) #, creationflags=subprocess.CREATE_NEW_CONSOLE)
|
||||
pid.wait()
|
||||
print ("Time running cntk [s]: " + str((datetime.datetime.now() - tstart).total_seconds()))
|
||||
# write model to disk
|
||||
model_path = os.path.join(modelDir, "frcn_" + classifier + ".model")
|
||||
print("Writing model to %s" % model_path)
|
||||
model.save_model(model_path)
|
||||
|
||||
#delete model files written during cntk training
|
||||
filenames = getFilesInDirectory(cntkFilesDir + "/tmp/", postfix = None)
|
||||
for filename in filenames:
|
||||
if filename.startswith('Fast-RCNN.'):
|
||||
os.remove(cntkFilesDir + "/tmp/" + filename)
|
||||
assert pid.returncode == 0, "ERROR: cntk ended with exit code {}".format(pid.returncode)
|
||||
|
||||
#parse cntk output
|
||||
# compute output of every image and write to disk
|
||||
image_sets = ["test", "train"]
|
||||
for image_set in image_sets:
|
||||
print "Parsing CNTK output for image set: " + image_set
|
||||
cntkImgsListPath = cntkFilesDir + image_set + ".txt"
|
||||
outParsedDir = cntkFilesDir + image_set + "_" + classifier + "_parsed/"
|
||||
if classifier == 'svm':
|
||||
cntkOutputPath = cntkFilesDir + image_set + ".z.fcOut.h2.y"
|
||||
elif classifier == 'nn':
|
||||
cntkOutputPath = cntkFilesDir + image_set + ".z"
|
||||
else:
|
||||
error
|
||||
|
||||
#write cntk output for each image to separate file
|
||||
makeDirectory(outParsedDir)
|
||||
parseCntkOutput(cntkImgsListPath, cntkOutputPath, outParsedDir, cntk_nrRois, cntk_featureDimensions[classifier],
|
||||
saveCompressed = True, skipCheck = True) #, skip5Mod = 0)
|
||||
run_fast_rcnn(model, image_set, cntk_padHeight, cntk_padWidth, nrClasses, cntk_nrRois, cntkFilesDir, outParsedDir)
|
||||
|
||||
#delete cntk output file which can be very large and are no longer needed
|
||||
deleteFile(cntkOutputPath)
|
||||
print "DONE."
|
||||
print("DONE.")
|
||||
|
|
@ -0,0 +1,70 @@
|
|||
import os, sys, importlib
|
||||
import shutil, time
|
||||
import subprocess
|
||||
import PARAMETERS
|
||||
locals().update(importlib.import_module("PARAMETERS").__dict__)
|
||||
|
||||
|
||||
|
||||
####################################
|
||||
# Parameters
|
||||
####################################
|
||||
print("classifier = " + classifier)
|
||||
cntkCmdStrPattern = "{0}/cntk.exe configFile={1}config.cntk currentDirectory={1}"
|
||||
|
||||
|
||||
|
||||
####################################
|
||||
# Main
|
||||
####################################
|
||||
deleteAllFilesInDirectory(cntkFilesDir + "/tmp", None)
|
||||
shutil.copy(os.path.join(cntkResourcesDir, "config.cntk"), cntkFilesDir)
|
||||
|
||||
#generate cntk command string
|
||||
cmdStr = cntkCmdStrPattern.format(cntkBinariesDir, cntkFilesDir, classifier)
|
||||
cmdStr += " ImageH={} ImageW={}".format(cntk_padHeight, cntk_padWidth)
|
||||
cmdStr += " NumLabels={0} NumTrainROIs={1} NumTestROIs={1}".format(len(classes), cntk_nrRois)
|
||||
cmdStr += " TrainROIDim={} TrainROILabelDim={}".format(4*cntk_nrRois, cntk_nrRois * cntk_featureDimensions[classifier])
|
||||
cmdStr += " TestROIDim={} TestROILabelDim={}".format( 4*cntk_nrRois, cntk_nrRois * cntk_featureDimensions[classifier])
|
||||
if classifier == 'svm':
|
||||
cmdStr += " [Train=[SGD=[maxEpochs=0]]]" #no need to train the network if just using it as featurizer
|
||||
cmdStr += " [WriteTest=[outputNodeNames=(z.fcOut.h2.y)]]"
|
||||
cmdStr += " [WriteTrain=[outputNodeNames=(z.fcOut.h2.y)]]"
|
||||
|
||||
#run cntk
|
||||
tstart = datetime.datetime.now()
|
||||
os.environ['ACML_FMA'] = str(0)
|
||||
print(cmdStr)
|
||||
pid = subprocess.Popen(cmdStr, cwd = cntkFilesDir) #, creationflags=subprocess.CREATE_NEW_CONSOLE)
|
||||
pid.wait()
|
||||
print ("Time running cntk [s]: " + str((datetime.datetime.now() - tstart).total_seconds()))
|
||||
|
||||
#delete model files written during cntk training
|
||||
filenames = getFilesInDirectory(cntkFilesDir + "/tmp/", postfix = None)
|
||||
for filename in filenames:
|
||||
if filename.startswith('Fast-RCNN.'):
|
||||
os.remove(cntkFilesDir + "/tmp/" + filename)
|
||||
assert pid.returncode == 0, "ERROR: cntk ended with exit code {}".format(pid.returncode)
|
||||
|
||||
#parse cntk output
|
||||
print("classifier = " + classifier)
|
||||
image_sets = ["test", "train"]
|
||||
for image_set in image_sets:
|
||||
print("Parsing CNTK output for image set: " + image_set)
|
||||
cntkImgsListPath = cntkFilesDir + image_set + ".txt"
|
||||
outParsedDir = cntkFilesDir + image_set + "_" + classifier + "_parsed/"
|
||||
if classifier == 'svm':
|
||||
cntkOutputPath = cntkFilesDir + image_set + ".z.fcOut.h2.y"
|
||||
elif classifier == 'nn':
|
||||
cntkOutputPath = cntkFilesDir + image_set + ".z"
|
||||
else:
|
||||
error
|
||||
|
||||
#write cntk output for each image to separate file
|
||||
makeDirectory(outParsedDir)
|
||||
parseCntkOutput(cntkImgsListPath, cntkOutputPath, outParsedDir, cntk_nrRois, cntk_featureDimensions[classifier],
|
||||
saveCompressed = True, skipCheck = False) #, skip5Mod = 0)
|
||||
|
||||
#delete cntk output file which can be very large and are no longer needed
|
||||
deleteFile(cntkOutputPath)
|
||||
print("DONE.")
|
||||
|
|
@ -0,0 +1,94 @@
|
|||
# -*- coding: utf-8 -*-
|
||||
import sys, os, importlib, random
|
||||
import PARAMETERS
|
||||
import json
|
||||
|
||||
locals().update(importlib.import_module("PARAMETERS").__dict__)
|
||||
|
||||
####################################
|
||||
# Parameters
|
||||
####################################
|
||||
imgPath = r"C:\Users\pabuehle\Desktop\newImgs\WIN_20160803_11_48_26_Pro.jpg"
|
||||
svm_experimentName = 'exp1'
|
||||
classifier = 'nn'
|
||||
|
||||
# no need to change these parameters
|
||||
boAddSelectiveSearchROIs = True
|
||||
boAddGridROIs = True
|
||||
boFilterROIs = True
|
||||
boUseNonMaximaSurpression = True
|
||||
|
||||
|
||||
####################################
|
||||
# Mainpip
|
||||
####################################
|
||||
# load cntk model
|
||||
print("Loading DNN..")
|
||||
tstart = datetime.datetime.now()
|
||||
model_path = os.path.join(modelDir, "frcn_" + classifier + ".model")
|
||||
model = load_model(model_path)
|
||||
print("Time loading DNN [ms]: " + str((datetime.datetime.now() - tstart).total_seconds() * 1000))
|
||||
|
||||
# load trained svm
|
||||
if classifier == "svm":
|
||||
print("Loading svm weights..")
|
||||
tstart = datetime.datetime.now()
|
||||
svmWeights, svmBias, svmFeatScale = loadSvm(trainedSvmDir, svm_experimentName)
|
||||
print("Time loading svm [ms]: " + str((datetime.datetime.now() - tstart).total_seconds() * 1000))
|
||||
else:
|
||||
svmWeights, svmBias, svmFeatScale = (None, None, None)
|
||||
|
||||
# compute ROIs
|
||||
tstart = datetime.datetime.now()
|
||||
imgOrig = imread(imgPath)
|
||||
currRois = computeRois(imgOrig, boAddSelectiveSearchROIs, boAddGridROIs, boFilterROIs, ss_kvals, ss_minSize,
|
||||
ss_max_merging_iterations, ss_nmsThreshold,
|
||||
roi_minDimRel, roi_maxDimRel, roi_maxImgDim, roi_maxAspectRatio, roi_minNrPixelsRel,
|
||||
roi_maxNrPixelsRel, grid_nrScales, grid_aspectRatios, grid_downscaleRatioPerIteration)
|
||||
currRois = currRois[:cntk_nrRois] # only keep first cntk_nrRois rois
|
||||
print("Time roi computation [ms]: " + str((datetime.datetime.now() - tstart).total_seconds() * 1000))
|
||||
|
||||
# prepare DNN inputs
|
||||
tstart = datetime.datetime.now()
|
||||
imgPadded = imresizeAndPad(imgOrig, cntk_padWidth, cntk_padHeight)
|
||||
_, _, roisCntk = getCntkInputs(imgPath, currRois, None, train_posOverlapThres, nrClasses, cntk_nrRois, cntk_padWidth, cntk_padHeight)
|
||||
arguments = {
|
||||
model.arguments[0]: [np.ascontiguousarray(np.array(imgPadded, dtype=np.float32).transpose(2, 0, 1))], # convert to CNTK's HWC format
|
||||
model.arguments[1]: [np.array(roisCntk)]
|
||||
}
|
||||
print("Time cnkt input generation [ms]: " + str((datetime.datetime.now() - tstart).total_seconds() * 1000))
|
||||
|
||||
# run DNN model
|
||||
print("Running model..")
|
||||
tstart = datetime.datetime.now()
|
||||
dnnOutputs = model.eval(arguments)[0][0]
|
||||
dnnOutputs = dnnOutputs[:len(currRois)] # remove the zero-padded rois
|
||||
print("Time running model [ms]: " + str((datetime.datetime.now() - tstart).total_seconds() * 1000))
|
||||
|
||||
# score all ROIs
|
||||
tstart = datetime.datetime.now()
|
||||
labels, scores = scoreRois(classifier, dnnOutputs, svmWeights, svmBias, svmFeatScale, len(classes),
|
||||
decisionThreshold = vis_decisionThresholds[classifier])
|
||||
print("Time making prediction [ms]: " + str((datetime.datetime.now() - tstart).total_seconds() * 1000))
|
||||
|
||||
# perform non-maxima surpression
|
||||
tstart = datetime.datetime.now()
|
||||
nmsKeepIndices = []
|
||||
if boUseNonMaximaSurpression:
|
||||
nmsKeepIndices = applyNonMaximaSuppression(nmsThreshold, labels, scores, currRois)
|
||||
print("Non-maxima surpression kept {:4} of {:4} rois (nmsThreshold={})".format(
|
||||
len(nmsKeepIndices), len(labels), nmsThreshold))
|
||||
print("Time non-maxima surpression [ms]: " + str((datetime.datetime.now() - tstart).total_seconds() * 1000))
|
||||
|
||||
# visualize results
|
||||
imgDebug = visualizeResults(imgPath, labels, scores, currRois, cntk_padWidth, cntk_padHeight,
|
||||
classes, nmsKeepIndices, boDrawNegativeRois=False, boDrawNmsRejectedRois=False)
|
||||
imshow(imgDebug, waitDuration=0, maxDim=800)
|
||||
#imwrite(imgDebug, visualizationDir + "/" + classifier + "_" + str(imgIndex) + os.path.basename(imgPath))
|
||||
|
||||
# create json-encoded string of all detections
|
||||
outDict = [{"label": str(l), "score": str(s), "left": str(r[0]), "top": str(r[1]), "right": str(r[2]), "bottom": str(r[3])} for l,s, r in zip(labels, scores, currRois)]
|
||||
outJsonString = json.dumps(outDict)
|
||||
print("Json-encoded detections: " + outJsonString[:120] + "...")
|
||||
|
||||
print("DONE.")
|
||||
Загрузка…
Ссылка в новой задаче