def write_textfile(p, r, a, tp, tn, fp, fn, threshold, model_base, dir=None):
if dir is None:
protoname = solverproto.replace('.prototxt', '')
dir = 'multilabel_results-' + protoname + '_' + model_base.replace(
'.caffemodel', '')
Utils.ensure_dir(dir)
fname = os.path.join(dir, model_base + 'results.txt')
with open(fname, 'a') as f:
f.write(model_base + ' threshold = ' + str(threshold) + '\n')
f.write('solver:' + solverproto + '\n')
f.write('model:' + caffemodel + '\n')
f.write('categories: ' + str(constants.web_tool_categories) + '\n')
f.write('precision\n')
f.write(str(p) + '\n')
f.write('recall\n')
f.write(str(r) + '\n')
f.write('accuracy\n')
f.write(str(a) + '\n')
f.write('true positives\n')
f.write(str(tp) + '\n')
f.write('true negatives\n')
f.write(str(tn) + '\n')
f.write('false positives\n')
f.write(str(fp) + '\n')
f.write('false negatives\n')
f.write(str(fn) + '\n')
f.close()
def generate_bbfiles_from_dir_of_jsons(json_dir,
imagefiles_dir_of_dirs,
bb_dir,
darknet=True,
category_number=None,
clobber=False):
# only_dirs = [d for d in json_dir if os.path.isdir(os.path.join(json_dir,d))]
json_files = [
f for f in os.listdir(json_dir)
if os.path.isfile(os.path.join(json_dir, f)) and f[-5:] == '.json'
and not 'retrieval' in f
]
class_number = 0
Utils.ensure_dir(bb_dir)
for a_file in json_files:
prefix = a_file[0:-5]
imagefiles_dir = prefix
print('json {} images {} class {}'.format(a_file, imagefiles_dir,
class_number))
full_json_file = os.path.join(json_dir, a_file)
full_images_dir = os.path.join(imagefiles_dir_of_dirs, prefix)
generate_bbfiles_from_json(full_json_file,
full_images_dir,
bb_dir,
darknet=True,
class_number=class_number,
clobber=False)
class_number += 1
def after_nn_processdir(indir,
outdir,
labels=constants.ultimate_21,
filter='.bmp'):
os.listdir(indir)
Utils.ensure_dir(outdir)
masks = [f for f in os.listdir(indir) if filter in f]
print('found {} files in {}'.format(len(masks), indir))
label_dict = {labels[i]: i for i in range(len(labels))}
print label_dict
for f in masks:
print f
fullname = os.path.join(indir, f)
mask = cv2.imread(fullname) #have to worry abt 3chan masks?
if len(mask.shape) == 3:
mask = mask[:, :, 0]
after_mask = pipeline.after_nn_conclusions(mask, label_dict)
after_nn_conclusions_name = os.path.join(outdir,
f[:-4] + '_nnconclusions.bmp')
cv2.imwrite(after_nn_conclusions_name, after_mask)
imutils.show_mask_with_labels(after_nn_conclusions_name,
labels,
save_images=True)
def download_image(prod, feature_name, category_id, max_images):
downloaded_images = 0
directory = os.path.join(category_id, feature_name)
try:
downloaded_images = len([name for name in os.listdir(directory) if os.path.isfile(name)])
except:
pass
if downloaded_images < max_images:
xlarge_url = prod['image']['sizes']['XLarge']['url']
img_arr = Utils.get_cv2_img_array(xlarge_url)
if img_arr is None:
logging.warning("Could not download image at url: {0}".format(xlarge_url))
return
relevance = background_removal.image_is_relevant(img_arr)
if relevance.is_relevant:
logging.info("Image is relevant...")
filename = "{0}_{1}.jpg".format(feature_name, prod["id"])
filepath = os.path.join(directory, filename)
Utils.ensure_dir(directory)
logging.info("Attempting to save to {0}...".format(filepath))
success = cv2.imwrite(filepath, img_arr)
if not success:
logging.info("!!!!!COULD NOT SAVE IMAGE!!!!!")
return 0
# downloaded_images += 1
logging.info("Saved... Downloaded approx. {0} images in this category/feature combination"
.format(downloaded_images))
return 1
else:
# TODO: Count number of irrelevant images (for statistics)
return 0
def dl_photos(photosfile, images_savedir, start_from=0):
listings = get_product_photos(photosfile)
max_items = 5000000
Utils.ensure_dir(images_savedir)
num_cores = multiprocessing.cpu_count()
Parallel(n_jobs=num_cores)(delayed(get_file)(listings[i], i +
1, images_savedir)
for i in range(start_from, len(listings)))
def flickr_get_dates(tag,mintime=0,savedir=None,n_pages=9):
time.sleep(21)
time_inc = 3600*24*1 #1 day
print('getting dates, min is '+str(mintime))
if savedir is None:
savedir = '/home/jeremy/image_dbs/flickr/'+tag+'/'
Utils.ensure_dir(savedir)
compressed_tag = tag.replace(' ','+')
outfile = compressed_tag+'out.txt'
print('outfile:'+outfile)
maxtime = mintime+time_inc
pages=0
oldpages = -1
while(pages<n_pages and maxtime<time.time()):
time.sleep(3)
initial_query = '&tags='+tag+'&min_upload_date='+str(mintime)+'&max_upload_date='+str(maxtime)+'&per_page=500'
initial_query = '&text='+compressed_tag+'&min_upload_date='+str(mintime)+'&max_upload_date='+str(maxtime)+'&per_page=500'
print('trying dates '+str(mintime)+' and '+str(maxtime))
cmd = 'curl -X GET "https://api.flickr.com/services/rest/?method=flickr.photos.search&api_key=d8548143cce923734f4093b4b063bc4f&format=json'+initial_query+'" > ' + outfile
print cmd
res = subprocess.call(cmd,shell=True)
with open(outfile,'r') as f:
content=f.read()
#the info is returned inside a function call like flickerApi({bla:bla...}) so I need to strip the beginning and end stuff
stripped = content[14:-1]
f.close()
try:
d = json.loads(stripped)
except:
print('get_dates json problem')
print('problem josn:'+str(stripped))
maxtime = maxtime + time_inc
continue
# pprint(d)
if not d:
print('no file found')
if not 'photos' in d:
print('no photos field in result, continuing')
phot = d['photos']
if not 'photo' in phot:
print('no photo field in result, continuing')
if 'page' in phot:
print('page '+str(phot['page']))
page = phot['page']
if 'pages' in phot:
print('of total pages '+str(phot['pages']))
pages = phot['pages']
if pages==oldpages:
time_inc = time_inc*1.4
print('got '+str(pages)+' pages')
maxtime = int(maxtime + time_inc)
oldpages=pages
maxtime = maxtime - time_inc
print('returning maxtime:'+str(maxtime)+' pages:'+str(pages))
return maxtime
def get_pd_results_and_save(url=None, filename=None):
if url is not None:
print('getting pd results for ' + url)
image = Utils.get_cv2_img_array(url)
elif filename is not None:
print('getting pd results for ' + filename)
image = cv2.imread(filename)
if image is None:
print('image came back none')
return None
try:
seg_res = pd_falcon_client.pd(image)
except:
print('exception in pd_falcon_client.pd ' + str(sys.exc_info()[0]))
return
# print('segres:'+str(seg_res))
if filename is not None:
imgfilename = os.path.basename(
filename
) #use the actual on-disk filename if thats what we started with, otherwise use the name generated by pd
#this will et saved to /home/jeremy/pd_output/filename
else:
try:
imgfilename = seg_res['filename'] + '.jpg'
except:
print('some error on imgfile name')
imgfilename = str(int(time.time())) + '.jpg'
print('filename ' + imgfilename)
if not 'mask' in seg_res:
print('pd did not return mask')
print seg_res
return
mask = seg_res['mask']
label_dict = seg_res['label_dict']
print('labels:' + str(label_dict))
conversion_utils.count_values(mask)
pose = seg_res['pose']
mask_np = np.array(mask, dtype=np.uint8)
print('masksize ' + str(mask_np.shape))
pose_np = np.array(pose, dtype=np.uint8)
print('posesize ' + str(pose_np.shape))
# print('returned url '+seg_res['url'])
converted_mask = convert_and_save_results(mask_np, label_dict, pose_np,
imgfilename, image, url)
dir = constants.pd_output_savedir
Utils.ensure_dir(dir)
full_name = os.path.join(dir, imgfilename)
# full_name = filename
bmp_name = full_name.strip('.jpg') + (
'.bmp') #this bmp should get generated by convert_and_save_results
# imutils.show_mask_with_labels(bmp_name,constants.fashionista_categories_augmented_zero_based,original_image = full_name,save_images=False)
return mask, label_dict, pose_np, converted_mask
def convert_labels_dir(
indir,
outdir,
jpgdir=None,
converter=constants.
fashionista_aug_zerobased_to_pixlevel_categories_v2,
suffix_in='.png',
suffix_out='_pixlevelv2.bmp',
for_webtool=False,
inlabels=constants.fashionista_categories_augmented_zero_based,
outlabels=constants.pixlevel_categories_v2,
save_legends=True):
'''
convert e..g from paperdoll to ultimate21 or pixlevel_categories_v2 .
Optionally only convert R channel for use with webtool. Don't forget to convert back to all chans after done w webtool
:param dir:
:param converter:
:param input_suffix:
:param for_webtool:
:return:
'''
Utils.ensure_dir(outdir)
files = [
os.path.join(indir, f) for f in os.listdir(indir) if suffix_in in f
]
print('STARTING CONVERT - converting ' + str(len(files)) + ' files in ' +
indir)
for f in files:
print('')
newname = os.path.join(outdir, os.path.basename(f))
newname = newname.replace(suffix_in, suffix_out)
print('converting {} to {} '.format(f, newname))
converted_arr = convert_labels(f,
converter=converter,
for_webtool=for_webtool,
inlabels=inlabels,
outlabels=outlabels)
cv2.imwrite(newname, converted_arr)
#raw_input('ret to cont')
if save_legends:
if jpgdir is None:
jpgdir = indir
orig_imagename = os.path.basename(f).replace(suffix_in, '.jpg')
orig_imagename = os.path.join(jpgdir, orig_imagename)
print('saving legend using {} '.format(orig_imagename))
imutils.show_mask_with_labels(converted_arr,
outlabels,
original_image=orig_imagename,
save_images=True)
def close_html(model_base, dir=None):
if dir is None:
protoname = solverproto.replace('.prototxt', '')
dir = 'multilabel_results-' + protoname + '_' + model_base.replace(
'.caffemodel', '')
Utils.ensure_dir(dir)
htmlname = os.path.join(dir, model_base + 'results.html')
with open(htmlname, 'a') as g:
g.write('</table><br>')
plotfilename = 'multilabel_results' + model_base + '.png'
g.write('<a href=\"' + plotfilename + '\">plot<img src = \"' +
plotfilename + '\" style=\"width:300px\"></a>')
g.write('</html>')
def write_textfile(caffemodel,
solverproto,
threshold,
model_base,
dir=None,
classes=None):
if dir is None:
protoname = solverproto.replace('.prototxt', '')
dir = 'multilabel_results-' + protoname + '_' + model_base.replace(
'.caffemodel', '')
Utils.ensure_dir(dir)
fname = os.path.join(dir, model_base + 'results.txt')
with open(fname, 'a') as f:
f.write(model_base + ' threshold = ' + str(threshold) + '\n')
f.write('solver:' + solverproto + '\n')
f.write('model:' + caffemodel + '\n')
f.write('categories: ' + str(classes) + '\n')
f.close()
def replace_labels_dir(dir_of_masks,
index_conversion,
overwrite=False,
outdir=None,
filter=None):
'''
convert masks in dir from tamara_berg_improved to ultimate_21 labels
#21 cats for direct replacement of VOC systems
#lose the necklace,
#combine tights and leggings
:param dir_of_masks:
:return:
'''
if filter:
masks = [
os.path.join(dir_of_masks, f) for f in os.listdir(dir_of_masks)
if filter in f
]
else:
masks = [
os.path.join(dir_of_masks, f) for f in os.listdir(dir_of_masks)
]
for f in masks:
mask_arr = replace_labels(f, index_conversion)
if mask_arr is None:
logging.warning('got no file')
continue
if len(mask_arr.shape) == 3:
logging.warning('multichannel mask, taking chan 0')
mask_arr = mask_arr[:, :, 0]
if overwrite:
outname = f
else:
indir = os.path.dirname(f)
parentdir = os.path.abspath(os.path.join(indir, os.pardir))
if outdir == None:
outdir = os.path.join(parentdir, '_relabelled')
Utils.ensure_dir(outdir)
inname = os.path.basename(f)
outname = os.path.join(outdir, f)
print('in {} out {}'.format(f, outname))
cv2.imwrite(mask_arr)
def move_the_rest(mask_filename, k):
if not os.path.exists(mask_filename):
print(mask_filename + ' does not exist')
return
indir = os.path.dirname(mask_filename)
parentdir = os.path.abspath(os.path.join(indir, os.pardir))
curdir = os.path.split(indir)[1] #the parent of current dir
print('in {} parent {} cur {}'.format(indir, parentdir, curdir))
if k == ord('d'):
newdir = curdir + '_removed'
dest_dir = os.path.join(parentdir, newdir)
Utils.ensure_dir(dest_dir)
print('REMOVING moving {} to {}'.format(mask_filename, dest_dir))
shutil.move(mask_filename, dest_dir)
elif k == ord('c'):
newdir = curdir + '_needwork'
dest_dir = os.path.join(parentdir, newdir)
Utils.ensure_dir(dest_dir)
print('CLOSE so moving {} to {}'.format(mask_filename, dest_dir))
shutil.move(mask_filename, dest_dir)
elif k == ord('k'):
newdir = curdir + '_kept'
dest_dir = os.path.join(parentdir, newdir)
Utils.ensure_dir(dest_dir)
print('KEEPING moving {} to {}'.format(mask_filename, dest_dir))
shutil.move(mask_filename, dest_dir)
else:
print('doing nothing')
def open_html(modelname,
dir=None,
solverproto='',
caffemodel='',
classlabels=constants.web_tool_categories,
name=None):
model_base = os.path.basename(modelname)
if dir is not None:
Utils.ensure_dir(dir)
htmlname = os.path.join(dir, model_base + 'results.html')
else:
htmlname = os.path.join(model_base, 'results.html')
if name is not None:
htmlname = name
with open(htmlname, 'a') as g:
g.write('<!DOCTYPE html>')
g.write('<html>')
g.write('<head>')
g.write('<title>')
dt = datetime.datetime.today()
g.write(model_base + ' ' + dt.isoformat())
g.write('</title>')
g.write('solver:' + solverproto + '\n' + '<br>')
g.write('model:' + caffemodel + '\n' + '<br>')
g.write('</head>')
# g.write('categories: '+str(constants.web_tool_categories)+'<br>'+'\n')
g.write('<table style=\"width:100%\">\n')
g.write('<tr>\n')
g.write('<th>')
g.write('metric')
g.write('</th>\n')
g.write('<th>')
g.write('fw avg.')
g.write('</th>\n')
if classlabels:
for i in range(len(classlabels)):
g.write('<th>')
g.write(classlabels[i])
g.write('</th>\n')
g.write('</tr>\n')
def precision_accuracy_recall(caffemodel,
solverproto,
outlayer='label',
n_tests=100):
#TODO dont use solver to get inferences , no need for solver for that
caffe.set_mode_gpu()
caffe.set_device(1)
workdir = './'
snapshot = 'snapshot'
# caffemodel = '/home/jeremy/caffenets/multilabel/vgg_ilsvrc_16_multilabel_2/snapshot/train_iter_240000.caffemodel'
# caffemodel = '/home/jeremy/caffenets/multilabel/vgg_ilsvrc_16_multilabel_2/snapshot/train_iter_340000.caffemodel'
model_base = caffemodel.split('/')[-1]
p_all = []
r_all = []
a_all = []
n_all = []
# for t in [0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.85,0.9,0.92,0.95,0.98]:
thresh = [0.1, 0.5, 0.6, 0.7, 0.8, 0.9, 0.95]
# thresh = [0.1,0.5,0.95]
protoname = solverproto.replace('.prototxt', '')
netname = get_netname(solverproto)
if netname:
dir = 'multilabel_results-' + netname + '_' + model_base.replace(
'.caffemodel', '')
dir = dir.replace('"', '') #remove quotes
dir = dir.replace(' ', '') #remove spaces
dir = dir.replace('\n', '') #remove newline
dir = dir.replace('\r', '') #remove return
else:
dir = 'multilabel_results-' + protoname + '_' + model_base.replace(
'.caffemodel', '')
dir = dir.replace('"', '') #remove quotes
dir = dir.replace(' ', '') #remove spaces
dir = dir.replace('\n', '') #remove newline
dir = dir.replace('\r', '') #remove return
print('dir to save stuff in : ' + str(dir))
Utils.ensure_dir(dir)
# open_html(model_base,dir=dir)
positives = True
for t in thresh:
p, r, a, tp, tn, fp, fn = check_accuracy(solverproto,
caffemodel,
threshold=t,
num_batches=n_tests,
outlayer=outlayer)
p_all.append(p)
r_all.append(r)
a_all.append(a)
n_occurences = [tp[i] + fn[i] for i in range(len(tp))]
n_all.append(n_occurences)
write_textfile(p, r, a, tp, tn, fp, fn, t, model_base, dir=dir)
write_html(p,
r,
a,
n_occurences,
t,
model_base,
positives=positives,
dir=dir,
tp=tp,
tn=tn,
fp=fp,
fn=fn)
positives = False
close_html(model_base, dir=dir)
p_all_np = np.transpose(np.array(p_all))
r_all_np = np.transpose(np.array(r_all))
a_all_np = np.transpose(np.array(a_all))
labels = constants.web_tool_categories
plabels = [label + 'precision' for label in labels]
rlabels = [label + 'recall' for label in labels]
alabels = [label + 'accuracy' for label in labels]
important_indices = [3, 5, 7, 10, 11, 13, 17]
#cardigan dress footwear jeans pants skirt top
#['bag', 'belt', 'blazer','cardigan','coat','dress', 'eyewear', 'footwear', 'hat','jacket',
# 'jeans','pants','shorts', 'skirt','stocking','suit','sweater','top','scarf','womens_swimwear_bikini',
# 'womens_swimwear_nonbikini']
p_important = [p_all_np[i] for i in important_indices]
r_important = [r_all_np[i] for i in important_indices]
a_important = [a_all_np[i] for i in important_indices]
labels_important = [labels[i] for i in important_indices]
for i in range(len(important_indices)):
print(constants.web_tool_categories[i] + ' p:' + str(p_important[i]) +
' r:' + str(r_important[i]) + ' a:' + str(a_important[i]))
thresh_all_np = np.array(thresh)
print('shape:' + str(p_all_np.shape))
print('len:' + str(len(p_important)))
markers = [
'^', '<', 'v', '^', '8', 'o', '.', 'x', '|', '+', 0, '4', 3, 4, 'H',
'3', 'p', 'h', '*', 7, '', 5, ',', '2', 1, 6, 's', 'd', '1', '_', 2,
' ', 'D'
]
markers = [
'.', 'x', '|', '^', '+', '<', 0, 'v', '4', 3, '^', '8', 4, 'o', 'H',
'3', 'p', '*', 'h', 7, '', 5, ',', '2', 1, 6, 's', 'd', '1', '_', 2,
' ', 'D'
]
markers_important = ['^', '<', 'v', '^', '8', 'o', 'H', '3', 'p', '*', 'h']
for i in range(len(p_important)):
plt.subplot(311)
print('plotting {} vs {}'.format(p_all_np[i, :], thresh_all_np))
plt.plot(thresh_all_np,
p_important[i],
label=labels_important[i],
linestyle='None',
marker=markers_important[i])
plt.subplot(312) #
plt.plot(thresh_all_np,
r_important[i],
label=labels_important[i],
linestyle='None',
marker=markers_important[i])
plt.subplot(313)
plt.plot(thresh_all_np,
a_important[i],
label=labels_important[i],
linestyle='None',
marker=markers_important[i])
# plt.plot(thresh_all_np,a_all_np[i],label=labels_important[i],linestyle='None',marker=markers_important[i])
# plt.plot(thresh_all_np,p_all_np[i,:],label=labels[i],marker=markers[i])
# plt.subplot(312) #
# plt.plot(thresh_all_np,r_all_np[i,:],label=labels[i],linestyle='None',marker=markers[i])
# plt.subplot(313)
# plt.plot(thresh_all_np,a_all_np[i,:],label=labels[i],linestyle='None',marker=markers[i])
plt.subplot(311)
plt.title('results ' + model_base)
plt.xlabel('threshold')
plt.ylabel('precision')
plt.grid(True)
plt.ylim((0, 1))
plt.subplot(312) #
plt.xlabel('threshold')
plt.ylabel('recall')
plt.grid(True)
plt.ylim((0, 1))
plt.subplot(313)
plt.xlabel('threshold')
plt.ylabel('accuracy')
plt.ylim((0, 1))
plt.grid(True)
plt.legend(loc='center left', bbox_to_anchor=(1, 0.1))
plt.show() #
figname = os.path.join(dir, model_base + '.png')
print('saving figure:' + str(figname))
plt.savefig(figname, bbox_inches='tight')
#
summary_html(dir)
def combine_pixlevel_and_multilabel_using_graylevel(
pixel_graylevels,
multilabel,
multilabel_threshold=0.7,
median_factor=1.0,
multilabel_to_ultimate21_conversion=constants.
binary_classifier_categories_to_ultimate_21,
multilabel_labels=constants.binary_classifier_categories,
face=None,
output_layer='pixlevel_sigmoid_output',
required_image_size=(224, 224),
do_graylevel_zeroing=True):
print(
'combining multilabel w. neurodoll using nfc watch out, required imsize:'
+ str(required_image_size))
thedir = './images'
Utils.ensure_dir(thedir)
if isinstance(url_or_np_array, basestring):
image = url_to_image(url_or_np_array)
orig_filename = os.path.join(thedir,
url_or_np_array.split('/')[-1]).replace(
'.jpg', '')
elif type(url_or_np_array) == np.ndarray:
hash = hashlib.sha1()
hash.update(str(time.time()))
name_base = 'orig' + hash.hexdigest()[:10]
orig_filename = os.path.join(thedir, name_base)
image = url_or_np_array
if image is None:
logging.debug('got None in grabcut_using_neurodoll_output')
print('writing orig to ' + orig_filename + '.jpg')
cv2.imwrite(orig_filename + '.jpg', image)
multilabel = get_multilabel_output(url_or_np_array)
# multilabel = get_multilabel_output_using_nfc(url_or_np_array)
#take only labels above a threshold on the multilabel result
#possible other way to do this: multiply the neurodoll mask by the multilabel result and threshold that product
if multilabel is None:
logging.debug('None result from multilabel')
return None
thresholded_multilabel = [ml > multilabel_threshold
for ml in multilabel] #
logging.info('orig label:' + str(multilabel) + ' len:' +
str(len(multilabel)))
# print('incoming label:'+str(multilabel))
# logging.info('thresholded label:'+str(thresholded_multilabel))
for i in range(len(thresholded_multilabel)):
if thresholded_multilabel[i]:
logging.info(multilabel_labels[i] + ' is over threshold')
# print('multilabel to u21 conversion:'+str(multilabel_to_ultimate21_conversion))
# print('multilabel labels:'+str(multilabel_labels))
#todo - this may be wrong later if we start taking both nd and multilabel into acct. Maybe ml thinks theres nothing there but nd thinks there is...
if np.equal(thresholded_multilabel,
0).all(): #all labels 0 - nothing found
logging.debug('no items found')
return #
pixlevel_categorical_output = graylevel_nd_output.argmax(
axis=2) #the returned mask is HxWxC so take max along C
pixlevel_categorical_output = threshold_pixlevel(
pixlevel_categorical_output) #threshold out the small areas
print('before graylevel zeroing:')
count_values(pixlevel_categorical_output, labels=constants.ultimate_21)
if do_graylevel_zeroing:
graylevel_nd_output = zero_graylevels_not_in_ml(graylevel_nd_output,
multilabel,
threshold=0.7)
pixlevel_categorical_output = graylevel_nd_output.argmax(
axis=2) #the returned mask is HxWxC so take max along C
pixlevel_categorical_output = threshold_pixlevel(
pixlevel_categorical_output) #threshold out the small areas
print('after graylevel zeroing:')
count_values(pixlevel_categorical_output, labels=constants.ultimate_21)
foreground = np.array((pixlevel_categorical_output > 0) * 1)
background = np.array((pixlevel_categorical_output == 0) * 1)
# item_masks = nfc.pd(image, get_all_graylevels=True)
logging.debug('shape of pixlevel categorical output:' +
str(pixlevel_categorical_output.shape))
logging.debug('n_fg {} n_bg {} tot {} w*h {}'.format(
np.sum(foreground), np.sum(background),
np.sum(foreground) + np.sum(background),
pixlevel_categorical_output.shape[0] *
pixlevel_categorical_output.shape[1]))
first_time_thru = True #hack to dtermine image size coming back from neurodoll
# final_mask = np.zeros([224,224])
final_mask = np.zeros(pixlevel_categorical_output.shape[:])
print('final_mask shape ' + str(final_mask.shape))
if face:
y_split = face[1] + 3 * face[3]
else:
# BETTER TO SEND A FACE
y_split = np.round(0.4 * final_mask.shape[0])
print('y split {} face {}'.format(y_split, face))
#the grabcut results dont seem too hot so i am moving to a 'nadav style' from-nd-and-ml-to-results system
#namely : for top , decide if its a top or dress or jacket
# for bottom, decide if dress/pants/skirt
#decide on one bottom
# for i in range(len(thresholded_multilabel)):
# if multilabel_labels[i] in ['dress', 'jeans','shorts','pants','skirt','suit','overalls'] #missing from list is various swimwear which arent getting returned from nd now anyway
#############################################################################################
#Make some conclusions nadav style.
#Currently the decisions are based only on ml results without taking into acct the nd results.
#In future possibly inorporate nd as well, first do a head-to-head test of nd vs ml
#############################################################################################
#1. take winning upper cover, donate losers to winner
#2. take winning upper under, donate losers to winner
#3. take winning lower cover, donate losers to winner.
#4. take winning lower under, donate losers to winner
#5. decide on whole body item (dress, suit, overall) vs. non-whole body (two part e.g. skirt+top) items.
#6. if wholebody beats two-part - donate all non-whole-body pixels to whole body (except upper-cover (jacket/blazer etc) and lower under-stockings)
#? if no upper cover and no upper under and no whole-body: take max of all those and donate losers to winner
#upper_cover: jacket, coat, blazer etc
#upper under: shirt, top, blouse etc
#lower cover: skirt, pants, shorts
#lower under: tights, leggings
whole_body_indexlist = [
multilabel_labels.index(s) for s in ['dress', 'suit', 'overalls']
] #swimsuits could be added here
upper_cover_indexlist = [
multilabel_labels.index(s)
for s in ['cardigan', 'coat', 'jacket', 'sweater', 'sweatshirt']
]
upper_under_indexlist = [multilabel_labels.index(s) for s in ['top']]
lower_cover_indexlist = [
multilabel_labels.index(s)
for s in ['jeans', 'pants', 'shorts', 'skirt']
]
lower_under_indexlist = [multilabel_labels.index(s) for s in ['stocking']]
final_mask = np.copy(pixlevel_categorical_output)
logging.info('size of final mask ' + str(final_mask.shape))
print('wholebody indices:' + str(whole_body_indexlist))
for i in whole_body_indexlist:
print multilabel_labels[i]
whole_body_ml_values = np.array(
[multilabel[i] for i in whole_body_indexlist])
print('wholebody ml_values:' + str(whole_body_ml_values))
whole_body_winner = whole_body_ml_values.argmax()
whole_body_winner_value = whole_body_ml_values[whole_body_winner]
whole_body_winner_index = whole_body_indexlist[whole_body_winner]
print('winning index:' + str(whole_body_winner) + ' mlindex:' +
str(whole_body_winner_index) + ' value:' +
str(whole_body_winner_value))
print('uppercover indices:' + str(upper_cover_indexlist))
for i in upper_cover_indexlist:
print multilabel_labels[i]
upper_cover_ml_values = np.array(
[multilabel[i] for i in upper_cover_indexlist])
print('upper_cover ml_values:' + str(upper_cover_ml_values))
upper_cover_winner = upper_cover_ml_values.argmax()
upper_cover_winner_value = upper_cover_ml_values[upper_cover_winner]
upper_cover_winner_index = upper_cover_indexlist[upper_cover_winner]
print('winning upper_cover:' + str(upper_cover_winner) + ' mlindex:' +
str(upper_cover_winner_index) + ' value:' +
str(upper_cover_winner_value))
print('upperunder indices:' + str(upper_under_indexlist))
for i in upper_under_indexlist:
print multilabel_labels[i]
upper_under_ml_values = np.array(
[multilabel[i] for i in upper_under_indexlist])
print('upper_under ml_values:' + str(upper_under_ml_values))
upper_under_winner = upper_under_ml_values.argmax()
upper_under_winner_value = upper_under_ml_values[upper_under_winner]
upper_under_winner_index = upper_under_indexlist[upper_under_winner]
print('winning upper_under:' + str(upper_under_winner) + ' mlindex:' +
str(upper_under_winner_index) + ' value:' +
str(upper_under_winner_value))
print('lowercover indices:' + str(lower_cover_indexlist))
for i in lower_cover_indexlist:
print multilabel_labels[i]
lower_cover_ml_values = np.array(
[multilabel[i] for i in lower_cover_indexlist])
print('lower_cover ml_values:' + str(lower_cover_ml_values))
lower_cover_winner = lower_cover_ml_values.argmax()
lower_cover_winner_value = lower_cover_ml_values[lower_cover_winner]
lower_cover_winner_index = lower_cover_indexlist[lower_cover_winner]
print('winning lower_cover:' + str(lower_cover_winner) + ' mlindex:' +
str(lower_cover_winner_index) + ' value:' +
str(lower_cover_winner_value))
print('lowerunder indices:' + str(lower_under_indexlist))
for i in lower_under_indexlist:
print multilabel_labels[i]
lower_under_ml_values = np.array(
[multilabel[i] for i in lower_under_indexlist])
print('lower_under ml_values:' + str(lower_under_ml_values))
lower_under_winner = lower_under_ml_values.argmax()
lower_under_winner_value = lower_under_ml_values[lower_under_winner]
lower_under_winner_index = lower_under_indexlist[lower_under_winner]
print('winning lower_under:' + str(lower_under_winner) + ' mlindex:' +
str(lower_under_winner_index) + ' value:' +
str(lower_under_winner_value))
#for use later, decide on a winner between upper cover and upper under
if upper_under_winner_value > upper_cover_winner_value:
upper_winner_value = upper_under_winner_value
upper_winner_index = upper_under_winner_index
else:
upper_winner_value = upper_cover_winner_value
upper_winner_index = upper_cover_winner_index
#for use later, decide on a winner between lower cover and lower under
if lower_under_winner_value > lower_cover_winner_value:
lower_winner_value = lower_under_winner_value
lower_winner_index = lower_under_winner_index
else:
lower_winner_value = lower_cover_winner_value
lower_winner_index = lower_cover_winner_index
upper_winner_nd_index = multilabel_to_ultimate21_conversion[
upper_winner_index]
lower_winner_nd_index = multilabel_to_ultimate21_conversion[
lower_winner_index]
print('upper winner {} nd {} val {} lower winner {} nd {} val {}'.format(
upper_winner_index, upper_winner_nd_index, upper_winner_value,
lower_winner_index, lower_winner_nd_index, lower_winner_value))
#1. take max upper cover , donate losers to winner
#this actually might not be always right, e.g. jacket+ sweater
#todo - #1 - 4 can be put into a function since they are nearly identical
neurodoll_upper_cover_index = multilabel_to_ultimate21_conversion[
upper_cover_winner_index]
if neurodoll_upper_cover_index is None:
logging.warning('nd upper cover index {} has no conversion '.format(
upper_cover_winner_index))
else:
n = len(final_mask[final_mask == neurodoll_upper_cover_index])
logging.debug('donating to upper cover winner, initial n :' + str(n) +
' for ndindex ' + str(neurodoll_upper_cover_index) +
' ml index ' + str(upper_cover_winner_index) +
', checking mls ' + str(upper_cover_indexlist))
for i in upper_cover_indexlist: #whole_body donated to upper_under
nd_index = multilabel_to_ultimate21_conversion[i]
if nd_index is None:
logging.warning('ml index {} has no conversion '.format(i))
continue
x = final_mask[final_mask == nd_index]
final_mask[final_mask == nd_index] = neurodoll_upper_cover_index
n = len(final_mask[final_mask == neurodoll_upper_cover_index])
logging.info(
'upper cover ndindex {} {} donated to upper cover winner nd {} , now {} pixels, lenx {} '
.format(nd_index, constants.ultimate_21[nd_index],
neurodoll_upper_cover_index, n, len(x)))
#2. take max upper under, donate losers to winner
neurodoll_upper_under_index = multilabel_to_ultimate21_conversion[
upper_under_winner_index]
if neurodoll_upper_under_index is None:
logging.warning('nd upper cover index {} has no conversion '.format(
upper_under_winner_index))
else:
n = len(final_mask[final_mask == neurodoll_upper_under_index])
logging.debug('donating to upper under winner, initial n :' + str(n) +
' for ndindex ' + str(neurodoll_upper_under_index) +
' ml index ' + str(upper_under_winner_index) +
', checking mls ' + str(upper_under_indexlist))
for i in upper_under_indexlist: #upper under losers donated to upper under winner
nd_index = multilabel_to_ultimate21_conversion[i]
print('nd index {} ml index {}'.format(nd_index, i))
if nd_index is None:
logging.warning('ml index {} has no conversion '.format(i))
continue
final_mask[final_mask == nd_index] = neurodoll_upper_under_index
n = len(final_mask[final_mask == neurodoll_upper_under_index])
logging.info(
'upper under ndindex {} {} donated to upper under winner nd {}, now {} pixels'
.format(nd_index, constants.ultimate_21[nd_index],
neurodoll_upper_under_index, n))
#3. take max lower cover, donate losers to winner.
neurodoll_lower_cover_index = multilabel_to_ultimate21_conversion[
lower_cover_winner_index]
if neurodoll_lower_cover_index is None:
logging.warning('nd lower cover index {} has no conversion '.format(
lower_cover_winner_index))
else:
n = len(final_mask[final_mask == neurodoll_lower_cover_index])
logging.debug('donating to lower cover winner, initial n :' + str(n) +
' for ndindex ' + str(neurodoll_lower_cover_index) +
' ml index ' + str(lower_cover_winner_index) +
', checking mls ' + str(lower_cover_indexlist))
for i in lower_cover_indexlist: #lower cover losers donated to lower cover winner
nd_index = multilabel_to_ultimate21_conversion[i]
if nd_index is None:
logging.warning('ml index {} has no conversion '.format(i))
continue
final_mask[final_mask == nd_index] = neurodoll_lower_cover_index
n = len(final_mask[final_mask == neurodoll_lower_cover_index])
logging.info(
'lower cover ndindex {} {} donated to lower cover winner nd {}, now {} pixels'
.format(nd_index, constants.ultimate_21[nd_index],
neurodoll_lower_cover_index, n))
#4. take max lower under, donate losers to winner.
neurodoll_lower_under_index = multilabel_to_ultimate21_conversion[
lower_under_winner_index]
if neurodoll_lower_under_index is None:
logging.warning('nd lower under index {} has no conversion '.format(
lower_under_winner_index))
else:
n = len(final_mask[final_mask == neurodoll_lower_under_index])
logging.debug('donating to lower under winner, initial n :' + str(n) +
' for ndindex ' + str(neurodoll_lower_under_index) +
' ml index ' + str(lower_under_winner_index) +
', checking mls ' + str(lower_under_indexlist))
for i in lower_under_indexlist: #lower under losers donated to lower under winner
nd_index = multilabel_to_ultimate21_conversion[i]
if nd_index is None:
logging.warning('ml index {} has no conversion '.format(i))
continue
final_mask[final_mask == nd_index] = neurodoll_lower_under_index
n = len(final_mask[final_mask == neurodoll_lower_under_index])
logging.info(
'lower under ndindex {} {} donated to lower under winner nd {}, now {} pixels'
.format(nd_index, constants.ultimate_21[nd_index],
neurodoll_lower_under_index, n))
logging.debug('after step 4, pixelcounts look like:')
count_values(final_mask, labels=constants.ultimate_21)
#########################
# 5. WHOLEBODY VS TWO-PART
# decide on whole body item (dress, suit, overall) vs. non-whole body items.
# case 1 wholebody>upper_under>lower_cover
# case 2 upper_under>wholebody>lower_cover
# case 3 lower_cover>wholebody>upper-under
# case 4 lower_cover,upper_under > wholebody
#consider reducing this to nadav's method:
# whole_sum = np.sum([item.values()[0] for item in mask_sizes['whole_body']])
# partly_sum = np.sum([item.values()[0] for item in mask_sizes['upper_under']]) +\
# np.sum([item.values()[0] for item in mask_sizes['lower_cover']])
# if whole_sum > partly_sum:
# donate partly to whole
# its a little different tho since in multilabel you cant compare directly two items to one , e.g. if part1 = 0.6, part2 = 0.6, and whole=0.99, you
# should prob go with whole even tho part1+part2>whole
#########################
neurodoll_wholebody_index = multilabel_to_ultimate21_conversion[
whole_body_winner_index]
if neurodoll_wholebody_index is None:
logging.warning(
'nd wholebody index {} ml index {} has no conversion '.format(
neurodoll_wholebody_index, whole_body_winner_index))
#first case - wholebody > upper_under > lowercover
#donate all non-whole-body pixels to whole body (except upper-cover (jacket/blazer etc) and lower under-stockings)
elif (whole_body_winner_value > upper_under_winner_value) and (
whole_body_winner_value > lower_cover_winner_value
) and whole_body_winner_value > multilabel_threshold:
logging.info(
'case 1. one part {} wins over upper cover {} and lower cover {}'.
format(whole_body_winner_value, upper_cover_winner_value,
lower_cover_winner_value))
n = len(final_mask[final_mask == neurodoll_wholebody_index])
logging.info('n in final mask from wholebody alone:' + str(n))
for i in upper_cover_indexlist:
#jackets etc can occur with dress/overall so dont donate these
pass
for i in upper_under_indexlist: #donate upper_under to whole_body
#todo fix the case of suit (which can have upper_under)
#ideally, do this for dress - suit and overalls can have upper_under
nd_index = multilabel_to_ultimate21_conversion[i]
if nd_index is None:
logging.debug(
'upper cover nd index for {} has no conversion '.format(i))
continue
#add upper cover item to wholebody mask
final_mask[final_mask == nd_index] = neurodoll_wholebody_index
logging.info('adding upperunder nd index {} '.format(nd_index))
n = final_mask[final_mask == neurodoll_wholebody_index]
logging.info('n in final mask from wholebody:' + str(n))
for i in lower_cover_indexlist: #donate lower_cover to whole_body
nd_index = multilabel_to_ultimate21_conversion[i]
final_mask[final_mask == nd_index] = neurodoll_wholebody_index
logging.info('adding lowercover nd index {} '.format(nd_index))
n = final_mask[final_mask == neurodoll_wholebody_index]
logging.info('n in final mask from wholebody alone:' + str(n))
for i in lower_under_indexlist:
#not doing this for stockings which is currently the only lower under
pass
logging.debug('after case one pixel values look like')
count_values(final_mask, labels=constants.ultimate_21)
# second case - upper_under > wholebody > lowercover
# here its not clear who to sack - the wholebody or the upper_under
# so i arbitrarily decided to sack th whole_body in favor of the upper_under since upper_under is higher
# EXCEPT if the wholebody is overalls , in which case keep overalls, upper_under and upper_cover, donate lower_cover/under to overalls
# otherwise if wholebody is e.g. dress then add dress to upper_under and lower_cover
elif (whole_body_winner_value < upper_under_winner_value) and (
whole_body_winner_value > lower_cover_winner_value) and (
whole_body_winner_value > multilabel_threshold):
logging.info(
'case 2. one part {} < upper under {} but > lower cover {}'.format(
whole_body_winner_value, upper_under_winner_value,
lower_cover_winner_value))
#if overalls, donate loewr_cover and lower_under to overalls
if whole_body_winner_index == multilabel_labels.index('overalls'):
neurodoll_whole_body_index = multilabel_to_ultimate21_conversion[
whole_body_winner_index]
n = len(final_mask[final_mask == neurodoll_wholebody_index])
logging.info('n in final mask from wholebody (overall) alone:' +
str(n))
for i in upper_cover_indexlist:
pass #upper cover ok with overalls
for i in upper_under_indexlist:
pass #upper under ok with overalls
for i in lower_cover_indexlist: #lower cover donated to overalls
nd_index = multilabel_to_ultimate21_conversion[i]
if nd_index is None:
logging.warning('ml index {} has no conversion '.format(i))
continue
final_mask[final_mask == nd_index] = neurodoll_wholebody_index
logging.info(
'uppercover nd index {} donated to overalls'.format(
nd_index))
n = len(final_mask[final_mask == neurodoll_wholebody_index])
logging.info('n in final mask from wholebody alone:' + str(n))
for i in lower_under_indexlist: #lower under donated to overalls - this can conceivably go wrong e.g. with short overalls and stockings
nd_index = multilabel_to_ultimate21_conversion[i]
if nd_index is None:
logging.warning('ml index {} has no conversion '.format(i))
continue
final_mask[final_mask == nd_index] = neurodoll_wholebody_index
logging.info(
'uppercover nd index {} donated to overalls'.format(
nd_index))
n = len(final_mask[final_mask == neurodoll_wholebody_index])
logging.info('n in final mask from wholebody alone:' + str(n))
#not overalls, so donate whole_body to upper_under - maybe not to lower_under . Not clear what to do actually.
else: #not overalls
if upper_winner_nd_index is None:
logging.warning(
'nd wholebody index {} ml index {} has no conversion '.
format(upper_winner_nd_index, upper_winner_index))
else:
n1 = len(final_mask[final_mask == upper_winner_nd_index])
n2 = len(final_mask[final_mask == lower_winner_nd_index])
logging.info(
'n in final mask from wholebody before donation to upper winner {} px,nd {} (lower {} px,nd {}'
.format(n1, upper_winner_nd_index, n2,
lower_winner_nd_index))
#todo - actually only wholebody pixels in the upper half of the image should be donated
for i in whole_body_indexlist: #whole_body donated to upper_under
nd_index = multilabel_to_ultimate21_conversion[i]
if nd_index is None:
logging.warning(
'ml index {} has no conversion (4upper)'.format(i))
continue #donate upper pixels to upper_winner
logging.debug(
'3. donating nd {} in top of wholebody to upper_under and bottom to lower_under'
.format(nd_index))
logging.debug(
'first adding from upper part of nd {} to nd {}, ysplit {}'
.format(nd_index, upper_winner_nd_index, y_split))
for y in range(0, final_mask.shape[0]):
if y <= y_split:
for x in range(0, final_mask.shape[1]):
if final_mask[y][x] == nd_index:
final_mask[y][x] = upper_winner_nd_index
n1 = len(final_mask[final_mask == upper_winner_nd_index])
n2 = len(final_mask[final_mask == lower_winner_nd_index])
logging.info(
'n in final mask from wholebody after donation to upper winner {} px,nd {} (lower {} px,nd {}'
.format(n1, upper_winner_nd_index, n2,
lower_winner_nd_index))
# donate whole-body pixels to lower winner
if lower_winner_nd_index is None:
logging.warning(
'nd wholebody index {} ml index {} has no conversion '.
format(upper_winner_nd_index, upper_winner_index))
else:
n1 = len(final_mask[final_mask == upper_winner_nd_index])
n2 = len(final_mask[final_mask == lower_winner_nd_index])
logging.info(
'n in final mask from wholebody after donation to upper winner {} px,nd {} , lower winner {} px, nd {} '
.format(n1, upper_winner_nd_index, n2,
lower_winner_nd_index))
#todo - actually only wholebody pixels in the upper half of the image should be donated
for i in whole_body_indexlist: #whole_body donated to upper_under
nd_index = multilabel_to_ultimate21_conversion[i]
if nd_index is None:
logging.warning(
'ml index {} has no conversion (4lower)'.format(i))
continue
#donate upper pixels to upper_winner
logging.debug(
'3. donating nd {} in botto of wholebody to upper_under and bottom to lower_under'
.format(nd_index))
logging.debug(
'second, adding from lower part of nd {} to nd {}, ysplit {}'
.format(nd_index, lower_winner_nd_index, y_split))
for y in range(0, final_mask.shape[0]):
if y > y_split:
for x in range(0, final_mask.shape[1]):
if final_mask[y][x] == nd_index:
final_mask[y][x] = lower_winner_nd_index
#donate upper pixels to lower_winner
n1 = len(final_mask[final_mask == upper_winner_nd_index])
n2 = len(final_mask[final_mask == lower_winner_nd_index])
logging.info(
'n in final mask from wholebody donation to upper {} and lower {}:'
.format(n1, n2))
logging.debug('after case two pixel values look like')
count_values(final_mask, labels=constants.ultimate_21)
# third case - lowercover > wholebody > upper_under
# here its not clear who to sack - the lowercover or the wholebody
# so i arbitrarily decided to sack the whole_body in favor of the lowercover since lowercover is higher
# donate lower part of wholebody to lowerwinner and upper part to upper winner
# this can be combined with second case I guess as there is nothing different - whole body gets added to lower/upper winners
elif (whole_body_winner_value < lower_cover_winner_value) and (
whole_body_winner_value > upper_under_winner_value
) and whole_body_winner_value > multilabel_threshold:
logging.info(
'case 3. one part {} > upper under {} and < lower cover {}'.format(
whole_body_winner_value, upper_under_winner_value,
lower_cover_winner_value))
if upper_winner_nd_index is None:
logging.warning(
'nd wholebody index {} ml index {} has no conversion '.format(
upper_winner_nd_index, upper_winner_index))
else:
n1 = len(final_mask[final_mask == upper_winner_nd_index])
n2 = len(final_mask[final_mask == lower_winner_nd_index])
logging.info(
'n in final mask from wholebody before donation to upper winner {} px,nd {} (lower {} px,nd {}'
.format(n1, upper_winner_nd_index, n2, lower_winner_nd_index))
#todo - actually only wholebody pixels in the upper half of the image should be donated
for i in whole_body_indexlist: #whole_body donated to upper_under
nd_index = multilabel_to_ultimate21_conversion[i]
if nd_index is None:
logging.warning(
'ml index {} has no conversion (4upper)'.format(i))
continue #donate upper pixels to upper_winner
logging.debug(
'3. donating nd {} in top of wholebody to upper_under and bottom to lower_under'
.format(nd_index))
logging.debug(
'first adding from upper part of nd {} to nd {}, ysplit {}'
.format(nd_index, upper_winner_nd_index, y_split))
for y in range(0, final_mask.shape[0]):
if y <= y_split:
for x in range(0, final_mask.shape[1]):
if final_mask[y][x] == nd_index:
final_mask[y][x] = upper_winner_nd_index
n1 = len(final_mask[final_mask == upper_winner_nd_index])
n2 = len(final_mask[final_mask == lower_winner_nd_index])
logging.info(
'n in final mask from wholebody after donation to upper winner {} px,nd {} (lower {} px,nd {}'
.format(n1, upper_winner_nd_index, n2, lower_winner_nd_index))
# donate whole-body pixels to lower winner
if lower_winner_nd_index is None:
logging.warning(
'nd wholebody index {} ml index {} has no conversion '.format(
upper_winner_nd_index, upper_winner_index))
else:
n1 = len(final_mask[final_mask == upper_winner_nd_index])
n2 = len(final_mask[final_mask == lower_winner_nd_index])
logging.info(
'n in final mask from wholebody after donation to upper winner {} px,nd {} , lower winner {} px, nd {} '
.format(n1, upper_winner_nd_index, n2, lower_winner_nd_index))
#todo - actually only wholebody pixels in the upper half of the image should be donated
for i in whole_body_indexlist: #whole_body donated to upper_under
nd_index = multilabel_to_ultimate21_conversion[i]
if nd_index is None:
logging.warning(
'ml index {} has no conversion (4lower)'.format(i))
continue
#donate upper pixels to upper_winner
logging.debug(
'3. donating nd {} in botto of wholebody to upper_under and bottom to lower_under'
.format(nd_index))
logging.debug(
'second, adding from lower part of nd {} to nd {}, ysplit {}'
.format(nd_index, lower_winner_nd_index, y_split))
for y in range(0, final_mask.shape[0]):
if y > y_split:
for x in range(0, final_mask.shape[1]):
if final_mask[y][x] == nd_index:
final_mask[y][x] = lower_winner_nd_index
#donate upper pixels to lower_winner
n1 = len(final_mask[final_mask == upper_winner_nd_index])
n2 = len(final_mask[final_mask == lower_winner_nd_index])
logging.info(
'n in final mask from wholebody donation to upper {} and lower {}:'
.format(n1, n2))
# fourth case - lowercover , upper_under > wholebody
# sack wholebody in favor of upper and lower
# donate top of wholebody to greater of upper cover/upper under (yes this is arbitrary and possibly wrong)
# donate bottom pixels of wholebody to greater of lower cover/lower under (again somewhat arbitrary)
# this also could get combined with #2,3 I suppose
# neurodoll_upper_cover_index = multilabel_to_ultimate21_conversion[upper_cover_winner_index] #
logging.debug('after case three pixel values look like')
count_values(final_mask, labels=constants.ultimate_21)
elif (whole_body_winner_value < lower_cover_winner_value) and (
whole_body_winner_value < upper_under_winner_value):
logging.info(
'case 4.one part {} < upper under {} and < lower cover {}'.format(
whole_body_winner_value, upper_under_winner_value,
lower_cover_winner_value))
neurodoll_lower_cover_index = multilabel_to_ultimate21_conversion[
lower_cover_winner_index]
# donate whole-body pixels to upper winner
if upper_winner_nd_index is None:
logging.warning(
'nd wholebody index {} ml index {} has no conversion '.format(
upper_winner_nd_index, upper_winner_index))
else:
n1 = len(final_mask[final_mask == upper_winner_nd_index])
n2 = len(final_mask[final_mask == lower_winner_nd_index])
logging.info(
'n in final mask from wholebody before donation to upper winner {} px,nd {} (lower {} px,nd {}'
.format(n1, upper_winner_nd_index, n2, lower_winner_nd_index))
#todo - actually only wholebody pixels in the upper half of the image should be donated
for i in whole_body_indexlist: #whole_body donated to upper_under
nd_index = multilabel_to_ultimate21_conversion[i]
if nd_index is None:
logging.warning(
'ml index {} has no conversion (4upper)'.format(i))
continue #donate upper pixels to upper_winner
logging.debug(
'4. donating nd {} in top of wholebody to upper_under and bottom to lower_under'
.format(nd_index))
logging.debug(
'first adding from upper part of nd {} to nd {}, ysplit {}'
.format(nd_index, upper_winner_nd_index, y_split))
for y in range(0, final_mask.shape[0]):
if y <= y_split:
for x in range(0, final_mask.shape[1]):
if final_mask[y][x] == nd_index:
final_mask[y][x] = upper_winner_nd_index
n1 = len(final_mask[final_mask == upper_winner_nd_index])
n2 = len(final_mask[final_mask == lower_winner_nd_index])
logging.info(
'n in final mask from wholebody after donation to upper winner {} px,nd {} (lower {} px,nd {}'
.format(n1, upper_winner_nd_index, n2, lower_winner_nd_index))
# donate whole-body pixels to lower winner
if lower_winner_nd_index is None:
logging.warning(
'nd wholebody index {} ml index {} has no conversion '.format(
upper_winner_nd_index, upper_winner_index))
else:
n1 = len(final_mask[final_mask == upper_winner_nd_index])
n2 = len(final_mask[final_mask == lower_winner_nd_index])
logging.info(
'n in final mask from wholebody after donation to upper winner {} px,nd {} , lower winner {} px, nd {} '
.format(n1, upper_winner_nd_index, n2, lower_winner_nd_index))
#todo - actually only wholebody pixels in the upper half of the image should be donated
for i in whole_body_indexlist: #whole_body donated to upper_under
nd_index = multilabel_to_ultimate21_conversion[i]
if nd_index is None:
logging.warning(
'ml index {} has no conversion (4lower)'.format(i))
continue
#donate upper pixels to upper_winner
logging.debug(
'4. donating nd {} in botto of wholebody to upper_under and bottom to lower_under'
.format(nd_index))
logging.debug(
'second, adding from lower part of nd {} to nd {}, ysplit {}'
.format(nd_index, lower_winner_nd_index, y_split))
for y in range(0, final_mask.shape[0]):
if y > y_split:
for x in range(0, final_mask.shape[1]):
if final_mask[y][x] == nd_index:
final_mask[y][x] = lower_winner_nd_index
#donate upper pixels to lower_winner
n1 = len(final_mask[final_mask == upper_winner_nd_index])
n2 = len(final_mask[final_mask == lower_winner_nd_index])
logging.info(
'n in final mask from wholebody donation to upper {} and lower {}:'
.format(n1, n2))
logging.debug('after case four pixel values look like')
count_values(final_mask, labels=constants.ultimate_21)
foreground = np.array(
(pixlevel_categorical_output > 0) * 1) #*1 turns T/F into 1/0
final_mask = final_mask * foreground # only keep stuff that was part of original fg - this is already true
# unless we start adding pixvalues that didn't 'win'
#7. if no lower cover and no whole-body was decided upon above: take max of lowercover items , donate losers to winner
#8. take at most one lower under, donate losers to winner
if (0):
for i in range(len(thresholded_multilabel)):
if thresholded_multilabel[i]:
neurodoll_index = multilabel_to_ultimate21_conversion[i]
if neurodoll_index is None:
print('no mapping from index {} (label {}) to neurodoll'.
format(i, multilabel_labels[i]))
continue
nd_pixels = len(pixlevel_categorical_output[
pixlevel_categorical_output == neurodoll_index])
print(
'index {} webtoollabel {} newindex {} neurodoll_label {} was above threshold {} (ml value {}) nd_pixels {}'
.format(i, multilabel_labels[i], neurodoll_index,
constants.ultimate_21[neurodoll_index],
multilabel_threshold, multilabel[i], nd_pixels))
gray_layer = graylevel_nd_output[:, :, neurodoll_index]
print('gray layer size:' + str(gray_layer.shape))
# item_mask = grabcut_using_neurodoll_output(url_or_np_array,neurodoll_index,median_factor=median_factor)
if nd_pixels > 0: #possibly put a threshold here, too few pixels and forget about it
item_mask = grabcut_using_neurodoll_graylevel(
url_or_np_array,
gray_layer,
median_factor=median_factor)
#the grabcut results dont seem too hot so i am moving to a 'nadav style' from-nd-and-ml-to-results system
#namely : for top , decide if its a top or dress or jacket
# for bottom, decide if dress/pants/skirt
pass
else:
print('no pixels in mask, skipping')
if item_mask is None:
continue
item_mask = np.multiply(item_mask, neurodoll_index)
if first_time_thru:
final_mask = np.zeros_like(item_mask)
first_time_thru = False
unique_to_new_mask = np.logical_and(
item_mask != 0, final_mask == 0
) #dealing with same pixel claimed by two masks. if two masks include same pixel take first, don't add the pixel vals together
unique_to_new_mask = np.multiply(unique_to_new_mask,
neurodoll_index)
final_mask = final_mask + unique_to_new_mask
# cv2.imshow('mask '+str(i),item_mask)
# cv2.waitKey(0)
timestamp = int(10 * time.time())
#write file (for debugging)
name = orig_filename + '_combinedoutput.png'
print('combined png name:' + name + ' orig filename ' + orig_filename)
cv2.imwrite(name, final_mask)
nice_output = imutils.show_mask_with_labels(name,
constants.ultimate_21,
save_images=True,
original_image=orig_filename +
'.jpg',
visual_output=False)
# nice_output = imutils.show_mask_with_labels(name,constants.ultimate_21,save_images=True)
#save graymask, this should be identical to nd except no threshold on low amt of pixels
graymask_filename = orig_filename + '_origmask.png'
print('original mask file:' + graymask_filename)
cv2.imwrite(graymask_filename, pixlevel_categorical_output)
nice_output = imutils.show_mask_with_labels(graymask_filename,
constants.ultimate_21,
save_images=True,
original_image=orig_filename +
'.jpg',
visual_output=False)
count_values(final_mask, labels=constants.ultimate_21)
return final_mask
subprocess.call(copycmd, shell=True)
#generate report filename
if type == 'pixlevel':
outname = os.path.join(
outdir, outdir[2:] + '_netoutput.txt'
) #TODO fix the shell script to not look for this, then it wont be needed
if type == 'multilabel':
outname = os.path.join(outdir, outdir[2:] + '_mlresults.html')
if type == 'single_label':
outdir = outdir + '_' + cat
outname = os.path.join(outdir, outdir[2:] + '_' + cat + '_slresults.txt')
loss_outputname = os.path.join(outdir, outdir[2:] + '_loss.txt')
print('outname:{}\n lossname {}\n outdir {}\n'.format(outname, loss_outputname,
outdir))
Utils.ensure_dir(outdir)
time.sleep(0.1)
Utils.ensure_file(loss_outputname)
#copycmd = 'cp -r '+outdir + ' ' + host_dirname
scpcmd = 'rsync -avz ' + outdir + ' [email protected]:/var/www/results/' + type + '/'
i = 0
losses = []
iters = []
loss_avg = [0] * n_iter
accuracy_avg = [0] * n_iter
tot_iters = 0
#instead of taking steps its also possible to do
#solver.solve()
def get_hydra_output(url_or_image_arr,
out_dir='./',
orig_size=(256, 256),
crop_size=(224, 224),
mean=(104.0, 116.7, 122.7),
gpu=1,
save_data=True,
save_path='/data/jeremy/caffenets/hydra/production/saves',
detection_thresholds=constants.hydra_tg_thresholds,
url=None):
'''
start net, get a bunch of results. DONE: resize to e.g. 250x250 (whatever was done in training) and crop to dims
:param url_or_image_arr_list:#
:param prototxt:
:param caffemodel:
:param out_dir:
:param dims:
:param output_layers:
:param mean:
:return:
'''
detection_thresholds = None
start_time = time.time()
caffe.set_mode_gpu()
caffe.set_device(gpu)
# print('params:'+str(hydra_net.params))
out_layers = hydra_net.outputs
out_layers = put_in_numeric_not_alphabetic_order(out_layers)
# print('out layers: '+str(out_layers))
j = 0
output_names = constants.hydra_tg_heads
# load image, resize, crop, subtract mean, and make dims C x H x W for Caffe
im = Utils.get_cv2_img_array(url_or_image_arr)
if im is None:
logging.warning('could not get image ' + str(url_or_image_arr))
return None
if isinstance(url_or_image_arr, basestring):
print('get_hydra_output working on:' + url_or_image_arr)
print('img size:' + str(im.shape))
im = imutils.resize_keep_aspect(im, output_size=orig_size)
im = imutils.center_crop(im, crop_size)
in_ = np.array(im, dtype=np.float32)
if len(in_.shape) != 3:
print('got 1-chan image, skipping')
return
elif in_.shape[2] != 3:
print('got n-chan image, skipping - shape:' + str(in_.shape))
return
in_ -= mean
in_ = in_.transpose((2, 0, 1)) #W,H,C -> C,W,H
hydra_net.blobs['data'].reshape(1, *in_.shape)
hydra_net.blobs['data'].data[...] = in_
# run net and take argmax for prediction
hydra_net.forward()
out = {}
i = 0
for output_layer in out_layers:
one_out = hydra_net.blobs[output_layer].data[
0] #not sure why the data is nested [1xN] matrix and not a flat [N] vector
second_neuron = copy.copy(
one_out[1]
) #the copy is required - if you dont do it then out gets over-written with each new one_out
second_neuron = round(float(second_neuron), 3)
# print('type:'+str(type(second_neuron)))
name = output_names[i]
if detection_thresholds is None:
out[name] = second_neuron
print('{} for category {} {}'.format(second_neuron, i, name))
elif second_neuron > detection_thresholds[i]:
out[name] = second_neuron
print('{} is past threshold {} for category {} {}'.format(
second_neuron, detection_thresholds[i], i, name))
logging.debug('output for {} {} is {}'.format(output_layer, name,
second_neuron))
# print('final till now:'+str(all_outs)+' '+str(all_outs2))
i = i + 1
logging.debug('all output:' + str(out))
logging.debug('elapsed time:' + str(time.time() - start_time))
if save_data:
if isinstance(url_or_image_arr, basestring):
filename = url_or_image_arr.replace('https://', '').replace(
'http://', '').replace('/', '_')
url = url_or_image_arr
else:
n_chars = 6
filename = ''.join(
random.choice(string.ascii_uppercase + string.digits)
for _ in range(n_chars)) + '.jpg'
if url is None:
url = 'not_from_url'
Utils.ensure_dir(save_path)
imgname = os.path.join(save_path, filename)
if imgname[:-4] != '.jpg':
imgname = imgname + '.jpg'
cv2.imwrite(imgname, im)
# out['imgname']=filename
out['url'] = url
textfile = os.path.join(save_path, 'output.txt')
with open(textfile, 'a') as fp:
json.dump(out, fp, indent=4)
fp.close()
print('wrote image to {} and output text to {}'.format(
imgname, textfile))
return out
def single_label_acc(caffemodel,
testproto,
net=None,
label_layer='label',
estimate_layer='loss',
n_tests=100,
gpu=0,
classlabels=constants.web_tool_categories_v2,
save_dir=None):
#TODO dont use solver to get inferences , no need for solver for that
#DONE
print('checking accuracy of net {} using proto {}'.format(
caffemodel, testproto))
n_classes = len(classlabels)
print('nclasses {} labels {}'.format(n_classes, classlabels))
if net is None:
caffe.set_mode_gpu()
caffe.set_device(gpu)
# net = caffe.Net(testproto,caffemodel, caffe.TEST) #apparently this is how test is chosen instead of train if you use a proto that contains both test and train
if caffemodel == '':
caffemodel = None #hack to keep things working, ideally change refs to caffemodel s.t. None is ok
net = caffe.Net(
testproto, caffe.TEST
) #apparently this is how test is chosen instead of train if you use a proto that contains both test and train
else:
net = caffe.Net(
testproto, caffe.TEST, weights=caffemodel
) #apparently this is how test is chosen instead of train if you use a proto that contains both test and train
#Net('train_val.prototxt', 1, weights='')
if caffemodel is not '' and caffemodel is not None:
model_base = caffemodel.split('/')[-1]
else:
model_base = 'scratch'
protoname = testproto.replace('.prototxt', '')
netname = multilabel_accuracy.get_netname(testproto)
if netname:
name = 'single_label_' + netname + '_' + model_base.replace(
'.caffemodel', '')
else:
name = 'single_label_' + protoname + '_' + model_base.replace(
'.caffemodel', '')
name = name.replace('"', '') #remove quotes
name = name.replace(' ', '') #remove spaces
name = name.replace('\n', '') #remove newline
name = name.replace('\r', '') #remove return
htmlname = name + '.html'
if save_dir is not None:
Utils.ensure_dir(save_dir)
htmlname = os.path.join(save_dir, htmlname)
print('htmlname : ' + str(htmlname))
# Utils.ensure_dir(dir)
confmat = check_accuracy(net,
n_classes,
n_tests=n_tests,
label_layer=label_layer,
estimate_layer=estimate_layer)
open_html(htmlname,
testproto,
caffemodel,
netname=netname,
classlabels=classlabels) #
write_confmat_to_html(htmlname, confmat, classlabels=classlabels)
acc_list = []
for i in range(n_classes):
p, r, a = precision_recall_accuracy(confmat, i)
acc_list.append(a)
write_pra_to_html(htmlname, p, r, a, i, classlabels[i])
close_html(htmlname)
# accs = np.array(acc_list)
# mean_acc = np.mean(accs)
acc = overall_acc_from_confmat(confmat)
return acc
def local_yolo(img_arr,
url='',
classes=constants.hls_yolo_categories,
save_results=True):
# item = {'object':label,'bbox':[xmin,ymin,xmax,ymax],'confidence':'>'+str(thresh)}
print('started local pyyolo detect, thresh=' + str(thresh) + ' hier ' +
str(hier_thresh))
save_path = '/data/jeremy/pyyolo/results/'
if img_arr is None:
print('got None img array!!')
return
img_original = copy.copy(img_arr)
if len(img_arr.shape) == 2: #got 1-chan(gray) image
print('got gray img')
img_arr_bgr = np.zeros([img_arr.shape[0], img_arr.shape[1], 3])
img_arr_bgr[:, :] = img_arr #copy bw image into all channels
print('sizes: {} {}'.format(img_arr_bgr, img_arr))
img_arr = img_arr_bgr
print('img arr size {}'.format(img_arr.shape))
max_image_size = 1200 #1280 fails
resized = False
if img_arr.shape[0] > max_image_size or img_arr.shape[
1] > max_image_size: #maybe this is causing prob at http://a.abcnews.com/images/US/150815_yt_phillypd_33x16_1600.jpg
maxside = max(img_arr.shape)
reduction_factor = maxside / 1000.0 #force maxside to 1000
original_size = img_arr.shape
resized = True
img_arr = cv2.resize(img_arr,
(int(img_arr.shape[1] / reduction_factor),
int(img_arr.shape[0] / reduction_factor)))
print('reduced size to {}'.format(img_arr.shape))
#generate randonm filename
hash = hashlib.sha1()
hash.update(str(time.time()))
if save_results:
timestr = time.strftime("%Y%m%d.%H%M%S")
img_filename = timestr + hash.hexdigest()[:5] + '.jpg'
Utils.ensure_dir(save_path)
img_path = os.path.join(save_path, img_filename)
print('detect_yolo_pyyolo saving file ' + str(img_path))
try:
cv2.imwrite(img_path, img_arr)
except:
print('some trouble saving image,' + str(sys.exc_info()[0]))
relevant_items = []
###############
#action is here
###############
print('getting results from get_pyyolo_results')
yolo_results = get_local_pyyolo_results(img_arr)
print('got results from get_pyyolo_results')
if resized:
img_arr = cv2.resize(img_arr, (original_size[1], original_size[0]))
print('de-resizing image')
for item in yolo_results:
print(item)
if resized:
print('de-resizing bbs')
item['bbox'][0] = int(item['bbox'][0] * reduction_factor)
item['bbox'][1] = int(item['bbox'][1] * reduction_factor)
item['bbox'][2] = int(item['bbox'][2] * reduction_factor)
item['bbox'][3] = int(item['bbox'][3] * reduction_factor)
print('fixed size back to original {}'.format(item))
xmin = item['bbox'][0]
ymin = item['bbox'][1]
xmax = item['bbox'][2]
ymax = item['bbox'][3]
assert xmin < xmax, 'xmin not < xmax!!!'
assert ymin < ymax, 'xmin not < xmax!!!'
relevant_items.append(item)
imutils.bb_with_text(img_arr,
[xmin, ymin, (xmax - xmin),
(ymax - ymin)], item['object'])
if save_results:
marked_imgname = img_path.replace('.jpg', '_bb_yolos.jpg')
json_name = img_path.replace('.jpg', '.json')
print('pyyolo bb image being writtten to ' + str(marked_imgname))
print('pyyolo bb data being writtten to ' + str(json_name))
try:
with open(json_name, 'w') as fp:
json.dump(yolo_results, fp, indent=4)
r = cv2.imwrite(marked_imgname, img_arr)
#skip saving bb image to conserve space
print('imgwrite result ' + str(r))
except:
print('some trouble saving bb image or data:' +
str(sys.exc_info()[0]))
output_like_api = {'annotations': relevant_items}
print('detect yolo returning:' + str(output_like_api))
return output_like_api
def write_html(p,
r,
a,
n,
threshold,
modelname,
positives=False,
dir=None,
name=None,
classlabels=None):
model_base = os.path.basename(modelname)
if dir is not None:
Utils.ensure_dir(dir)
htmlname = os.path.join(dir, model_base + 'results.html')
else:
htmlname = os.path.join(model_base, 'results.html')
if name is not None:
htmlname = name
open_html(model_base, dir=dir, classlabels=classlabels)
with open(htmlname, 'a') as g:
fwavp = 0
fwavr = 0
fwava = 0
n_p = 0
n_r = 0
n_a = 0
fwavn = 0
n_sum = 0
#calculate frequency-weighted averages
for i in range(len(p)):
if not np.isnan(p[i]):
fwavp = fwavp + p[i] * n[i]
n_p = n_p + n[i]
if not np.isnan(r[i]):
fwavr = fwavr + r[i] * n[i]
n_r = n_r + n[i]
if not np.isnan(a[i]):
fwava = fwava + a[i] * n[i]
n_a = n_a + n[i]
n_sum = n_sum + n[i]
print(
'n sum {} fwavp {} fwavr {} fwava {} before division np {} nr {} na {} '
.format(n_sum, fwavp, fwavr, fwava, n_p, n_r, n_a))
fwavp = fwavp / float(n_p)
fwavr = fwavr / float(n_r)
fwava = fwava / float(n_a)
fwavn = n_sum / float(len(p))
print('frequency weighted averages p {} r {} acc {} n {}'.format(
fwavp, fwavr, fwava, fwavn))
g.write('frequency weighted averages p {} r {} acc {} n {}'.format(
round(fwavp, 3), round(fwavr, 3), round(fwava, 3), round(fwavn,
3)))
#write line with n_positives
if (positives):
g.write('<tr>\n')
g.write('<td>')
g.write('n_positives')
g.write('</td>\n')
g.write('<td>')
g.write(str(round(fwavn, 3)))
g.write('</td>\n')
for i in range(len(p)):
g.write('<td>')
g.write(str(int(n[i])))
g.write('</td>\n')
g.write('</tr>\n<br>\n')
#write line with threshold
# g.write('<table style=\"width:100%\">\n')
g.write('<b>')
g.write('<tr>\n')
g.write('<td>')
g.write('threshold\n')
g.write('</td>')
g.write('<td>')
g.write('')
g.write('</td>\n')
for i in range(len(p)):
g.write('<td>')
g.write(str(round(threshold, 3)))
g.write('</td>\n')
g.write('</tr>\n')
g.write('</b>')
#write row with precision
g.write('<tr>\n')
g.write('<td>')
g.write('precision')
g.write('</td>\n')
g.write('<td>')
g.write(str(round(fwavp, 3)))
g.write('</td>\n')
for i in range(len(p)):
g.write('<td>')
g.write(str(round(p[i], 3)))
g.write('</td>\n')
g.write('</tr>\n')
#write row with recall
g.write('<tr>\n')
g.write('<td>')
g.write('recall')
g.write('</td>\n')
g.write('<td>')
g.write(str(round(fwavr, 3)))
g.write('</td>\n')
for i in range(len(p)):
g.write('<td>')
g.write(str(round(r[i], 3)))
g.write('</td>\n')
g.write('</tr>\n')
#write row with accuracy
g.write('<tr>\n')
g.write('<td>')
g.write('accuracy')
g.write('</td>\n')
g.write('<td>')
g.write(str(round(fwava, 3)))
g.write('</td>\n')
for i in range(len(p)):
g.write('<td>')
g.write(str(round(a[i], 3)))
g.write('</td>\n')
g.write('</tr>\n<br>\n')
g.write('<tr><td><br/></td></tr>') #blank row
def selectsiya(dir):
alone_dir = os.path.join(dir,'alone')
Utils.ensure_dir(alone_dir)
delete_dir = os.path.join(dir,'delete_these')
Utils.ensure_dir(delete_dir)
good_images_dir = os.path.join(dir,'good_images')
Utils.ensure_dir(good_images_dir)
files = [f for f in os.listdir(dir) if 'jpg' in f]
# print('files:'+str(files))
n = 0
i = 0
start_time=time.time()
while i < len(files):
f = files[i]
print('i='+str(i))
count_curdir = len([g for g in os.listdir(dir) if os.path.isfile(os.path.join(dir, g))])
count_alonedir = len([g for g in os.listdir(alone_dir) if os.path.isfile(os.path.join(alone_dir, g))])
count_deletedir = len([g for g in os.listdir(delete_dir) if os.path.isfile(os.path.join(delete_dir, g))])
count_goodimagesdir = len([g for g in os.listdir(good_images_dir) if os.path.isfile(os.path.join(good_images_dir, g))])
print(str(n)+' done of '+ str(count_curdir)+' files, '+str(count_goodimagesdir)+' good images, '+str(count_alonedir)+' alone, '+str(count_deletedir)+' deleted, tpi='+str((time.time()-start_time)/(i+1)))
fullfile = os.path.join(dir,f)
print('file:'+str(fullfile))
try:
img_arr = cv2.imread(fullfile)
except:
print('something bad happened trying to imread')
continue
try:
h,w = img_arr.shape[0:2]
img_arr = cv2.resize(img_arr,dsize=(w*2,h*2))
cv2.imshow('candidate',img_arr)
except:
print('something bad happened trying to imshow')
destname = os.path.join(delete_dir, f)
print('source:'+fullfile+' dest:'+destname)
os.rename(fullfile,destname)
print('(d)elete (a)lone (b)ack (space)nothing (q)uit ')
c = cv2.waitKey(0)
if c == ord('b'):
print('go back')
prev_dir = os.path.join(dir, files[i-1])
os.rename(destname, prev_dir)
print('moved image to main dir')
i=i-1
continue
elif c == ord('d'):
print('delete')
destname = os.path.join(delete_dir, f)
print('source:'+fullfile+' dest:'+destname)
os.rename(fullfile,destname)
elif c == ord('a'):
print('alone')
destname = os.path.join(alone_dir, f)
print('source:'+fullfile+' dest:'+destname)
os.rename(fullfile,destname)
elif c == ord(' '):
print('good image')
destname = os.path.join(good_images_dir, f)
print('source:' + fullfile + ' dest:' + destname)
os.rename(fullfile, destname)
elif c == ord('q'):
print('quit')
cv2.destroyAllWindows()
return
else:
print('nothing')
n = n + 1
i = i + 1
def get_pixlevel_confmat_using_falcon(images_and_labels_file,
labels=constants.ultimate_21,
save_dir='./nd_output'):
with open(images_and_labels_file, 'r') as fp:
lines = fp.readlines()
n_cl = len(labels)
hist = np.zeros((n_cl, n_cl))
loss = 0
print('n channels: ' + str(n_cl))
start_time = time.time()
for line in lines:
imagefile = line.split()[0]
gtfile = line.split()[1]
img_arr = cv2.imread(imagefile)
if img_arr is None:
logging.warning('could not get image data from ' + imagefile)
continue
gt_data = cv2.imread(gtfile)
if gt_data is None:
logging.warning('could not get gt data from ' + gtfile)
continue
if len(gt_data.shape) == 3:
logging.warning('got 3 chan image for mask, taking chan 0 ' +
gtfile)
gt_data = gt_data[:, :, 0]
dic = nfc.pd(img_arr)
if not dic['success']:
logging.debug('nfc pd not a success')
continue
net_data = dic['mask']
print('sizes of gt {} net output {}'.format(gt_data.shape,
net_data.shape))
hist += jrinfer.fast_hist(gt_data.flatten(), net_data.flatten(), n_cl)
if save_dir:
Utils.ensure_dir(save_dir)
gt_name = os.path.basename(imagefile)[:-4] + '_gt_legend.jpg'
gt_name = os.path.join(save_dir, gt_name)
ndout_name = os.path.basename(imagefile)[:-4] + '_ndout_legend.jpg'
ndout_name = os.path.join(save_dir, ndout_name)
imutils.show_mask_with_labels(gt_data,
labels,
original_image=imagefile,
save_images=True,
visual_output=False,
savename=gt_name)
imutils.show_mask_with_labels(
net_data,
labels,
original_image=imagefile,
save_images=True,
visual_output=False,
savename=ndout_name) # compute the loss as well
results_dict = jrinfer.results_from_hist(hist,
save_file=os.path.join(
save_dir, 'output.html'))
print results_dict
elapsed_time = time.time() - start_time
print('elapsed time: ' + str(elapsed_time) + ' tpi:' +
str(float(elapsed_time) / len(lines)))
return hist
def do_pixlevel_accuracy(caffemodel,
n_tests,
layer,
classes=constants.ultimate_21,
testproto=None,
solverproto=None,
iter=0,
savepics=True,
gpu=0,
output_layer='output'):
#to do accuracy we prob dont need to load solver
caffemodel_base = os.path.basename(caffemodel)
dir = 'pixlevel_results-' + caffemodel_base.replace('.caffemodel', '')
Utils.ensure_dir(dir)
if savepics:
picsdir = os.path.join(dir, caffemodel_base + '_output')
Utils.ensure_dir(picsdir)
else:
picsdir = False
htmlname = os.path.join(dir, dir + '.html')
detailed_outputname = htmlname[:-5] + '.txt'
print('saving net of {} {} to dir {} html {} and file {}'.format(
caffemodel, solverproto, picsdir, htmlname, detailed_outputname))
n_images = range(n_tests)
if gpu is not None:
caffe.set_device(gpu)
else:
caffe.set_mode_cpu()
if (solverproto is not None): #do this the old way with sgdsolver
solver = caffe.SGDSolver(solverproto)
solver.net.copy_from(caffemodel)
print('using net defined by {} and {} '.format(solverproto,
caffemodel))
answer_dict = jrinfer.seg_tests(solver,
n_images,
output_layer=output_layer,
gt_layer='label',
outfilename=detailed_outputname)
elif (testproto is not None): #try using net without sgdsolver
net = caffe.Net(testproto, caffemodel, caffe.TEST)
answer_dict = jrinfer.do_seg_tests(net,
iter,
picsdir,
n_images,
output_layer=output_layer,
gt_layer='label',
outfilename=detailed_outputname)
# in_ = np.array(im, dtype=np.float32)
# net.blobs['data'].reshape(1, *in_.shape)
# net.blobs['data'].data[...] = in_
# # run net and take argmax for prediction
# net.forward()
# out = net.blobs['seg-score'].data[0].argmax(axis=0)
open_html(htmlname, caffemodel, solverproto, classes, answer_dict)
write_html(htmlname, answer_dict)
close_html(htmlname)
def dosolve(weights,
solverproto,
testproto,
type='single_label',
steps_per_iter=1,
n_iter=200,
n_loops=200,
n_tests=1000,
cat=None,
classlabels=None,
baremetal_hostname='brainiK80a',
solverstate=None):
if classlabels is None:
classlabels = ['not_' + cat, cat]
caffe.set_device(int(sys.argv[1]))
caffe.set_mode_gpu()
solver = caffe.get_solver(solverproto)
if weights is not None:
solver.net.copy_from(weights)
if solverstate is not None:
solver.restore(
solverstate) #see https://github.com/BVLC/caffe/issues/3651
#No need to use solver.net.copy_from(). .caffemodel contains the weights. .solverstate contains the momentum vector. Both are needed to restart training. If you restart training without momentum, the loss will spike up and it will take ~50k iterations to recover. At test time you only need .caffemodel.
training_net = solver.net
solver.test_nets[0].share_with(
solver.net) #share train weight updates with testnet
test_net = solver.test_nets[0] # more than one testnet is supported
net_name = multilabel_accuracy.get_netname(testproto)
docker_hostname = socket.gethostname()
datestamp = datetime.datetime.strftime(datetime.datetime.now(),
'time%H.%M_%d-%m-%Y')
prefix = baremetal_hostname + '_' + net_name + '_' + docker_hostname + '_' + datestamp
#get netname, copy train/test to outdir
tt = caffe_utils.get_traintest_from_proto(solverproto)
print('netname {} train/test {}'.format(net_name, tt))
#detailed_jsonfile = detailed_outputname[:-4]+'.json'
if weights:
weights_base = os.path.basename(weights)
else:
weights_base = '_noweights_'
threshold = 0.5
if net_name:
outdir = type + '_' + prefix + '_' + weights_base.replace(
'.caffemodel', '')
else:
outdir = type + '_' + prefix + '_' + testproto + '_' + weights_base.replace(
'.caffemodel', '')
outdir = outdir.replace('"', '') #remove quotes
outdir = outdir.replace(' ', '') #remove spaces
outdir = outdir.replace('\n', '') #remove newline
outdir = outdir.replace('\r', '') #remove return
outdir = './' + outdir
#generate report filename, outdir to save everything (loss, html etc)
if type == 'pixlevel':
outname = os.path.join(
outdir, outdir[2:] + '_netoutput.txt'
) #TODO fix the shell script to not look for this, then it wont be needed
if type == 'multilabel':
outname = os.path.join(outdir, outdir[2:] + '_mlresults.html')
if type == 'single_label':
outdir = outdir + '_' + cat
outname = os.path.join(outdir,
outdir[2:] + '_' + cat + '_slresults.txt')
loss_outputname = os.path.join(outdir, outdir[2:] + '_loss.txt')
print('outname:{}\n lossname {}\n outdir {}\n'.format(
outname, loss_outputname, outdir))
Utils.ensure_dir(outdir)
time.sleep(0.1)
Utils.ensure_file(loss_outputname)
#copy training and test files to outdir
if tt is not None:
if len(tt) == 1: #copy single traintest file to dir of info
copycmd = 'cp ' + tt[0] + ' ' + outdir
subprocess.call(copycmd, shell=True)
else: #copy separate train and test files to dir of info
copycmd = 'cp ' + tt[0] + ' ' + outdir
subprocess.call(copycmd, shell=True)
copycmd = 'cp ' + tt[1] + ' ' + outdir
subprocess.call(copycmd, shell=True)
#cpoy solverproto to results dir
if solverproto is not None:
copycmd = 'cp ' + solverproto + ' ' + outdir
subprocess.call(copycmd, shell=True)
#copy test proto to results dir
if testproto is not None:
copycmd = 'cp ' + testproto + ' ' + outdir
subprocess.call(copycmd, shell=True)
#copy this file too
copycmd = 'cp solve.py ' + outdir
subprocess.call(copycmd, shell=True)
#copycmd = 'cp -r '+outdir + ' ' + host_dirname
scpcmd = 'rsync -avz ' + outdir + ' [email protected]:/var/www/results/' + type + '/'
i = 0
losses = []
iters = []
loss_avg = np.zeros(n_iter)
accuracy_avg = np.zeros(n_iter)
tot_iters = 0
#instead of taking steps its also possible to do
#solver.solve()
if type == 'multilabel':
multilabel_accuracy.open_html(
weights,
dir=outdir,
solverproto=solverproto,
caffemodel=weights,
classlabels=constants.web_tool_categories_v2,
name=outname)
for _ in range(n_loops):
for i in range(n_iter):
solver.step(steps_per_iter)
# loss = solver.net.blobs['score'].data
loss = solver.net.blobs['loss'].data
loss_avg[i] = loss
losses.append(loss)
tot_iters = tot_iters + steps_per_iter * n_iter
if type == 'single_label':
accuracy = solver.net.blobs['accuracy'].data
accuracy_avg[i] = accuracy
print('iter ' + str(i * steps_per_iter) + ' loss:' +
str(loss) + ' acc:' + str(accuracy))
else:
print('iter ' + str(i * steps_per_iter) + ' loss:' + str(loss))
averaged_loss = np.mean(loss_avg)
if type == 'single_label':
averaged_acc = np.mean(accuracy_avg)
s = 'avg loss over last {} steps is {}, acc:{}'.format(
n_iter * steps_per_iter, averaged_loss, averaged_acc)
print(s)
s2 = '{}\t{}\t{}\n'.format(tot_iters, averaged_loss, averaged_acc)
else:
s = 'avg loss over last {} steps is {}'.format(
n_iter * steps_per_iter, averaged_loss)
print(s)
s2 = '{}\t{}\n'.format(tot_iters, averaged_loss)
#for test net:
# solver.test_nets[0].forward() # test net (there can be more than one)
with open(loss_outputname, 'a+') as f:
f.write(str(int(time.time())) + '\t' + s2)
f.close()
# progress_plot.lossplot(loss_outputname) this hits tkinter problem
if type == 'multilabel':
precision, recall, accuracy, tp, tn, fp, fn = multilabel_accuracy.check_acc(
test_net,
num_samples=n_tests,
threshold=0.5,
gt_layer='labels',
estimate_layer='prob')
print('solve.py: p {} r {} a {} tp {} tn {} fp {} fn {}'.format(
precision, recall, accuracy, tp, tn, fp, fn))
n_occurences = [tp[i] + fn[i] for i in range(len(tp))]
multilabel_accuracy.write_html(precision,
recall,
accuracy,
n_occurences,
threshold,
weights,
positives=True,
dir=outdir,
name=outname)
elif type == 'pixlevel':
# number of tests for pixlevel
val = range(0, n_tests) #
jrinfer.seg_tests(solver,
val,
output_layer='mypixlevel_output',
gt_layer='label',
outfilename=outname,
save_dir=outdir)
elif type == 'single_label':
acc = single_label_accuracy.single_label_acc(
weights,
testproto,
net=test_net,
label_layer='label',
estimate_layer='fc4_0',
n_tests=n_tests,
classlabels=classlabels,
save_dir=outdir)
# test_net = solver.test_nets[0] # more than one testnet is supported
# testloss = test_net.blobs['loss'].data
try:
testloss = test_net.blobs['loss'].data
except:
print('n o testloss available')
testloss = 0
with open(loss_outputname, 'a+') as f:
f.write('test\t' + str(int(time.time())) + '\t' +
str(tot_iters) + '\t' + str(testloss) + '\t' +
str(acc) + '\n')
f.close()
#
# subprocess.call(copycmd,shell=True)
subprocess.call(scpcmd, shell=True)
# imgfiles = [f.split('.')[0] for f in os.listdir(path) if os.path.isfile(os.path.join(path, f))]
labelfiles = [f for f in os.listdir(labelpath) if 'png' in f]
imgfiles = [f for f in os.listdir(imgpath) if 'jpg' in f]
print('nfiles:' + str(len(labelfiles)))
# fashionista_ordered_categories = constants.fashionista_categories
#in case it changes in future - as of 2/16 this list goes a little something like this:
#fashionista_categories = ['null','tights','shorts','blazer','t-shirt','bag','shoes','coat','skirt','purse',
# 'boots', 'blouse','jacket','bra','dress','pants','sweater','shirt','jeans','leggings','scarf','hat',
# 'top','cardigan','accessories','vest','sunglasses','belt','socks','glasses','intimate',
# 'stockings','necklace','cape','jumper','sweatshirt','suit','bracelet','heels','wedges','ring',
# 'flats','tie','romper','sandals','earrings','gloves','sneakers','clogs','watch','pumps','wallet',
# 'bodysuit','loafers','hair','skin']
# colorbars(fashionista_ordered_categories)
Utils.ensure_dir(savedir)
for file in imgfiles:
full_img = os.path.join(imgpath, file)
full_lbl = os.path.join(labelpath, file.replace('.jpg', '.png'))
if not os.path.isfile(full_img):
print('img {} is not there '.format(full_img))
continue
if not os.path.isfile(full_lbl):
print('img {} is not there '.format(full_lbl))
continue
save_clothing_parts(full_lbl, full_img, savedir=savedir)
# raw_input('enter')
# show_pd_results(fullpath)
# raw_input('enter for next')
def detect_yolo_shell(self, img_arr, url='',classes=constants.hls_yolo_categories,save_results=True):
#RETURN dict like: ({'object':class_name,'bbox':bbox,'confidence':round(float(score),3)})
# relevant_bboxes.append({'object':class_name,'bbox':bbox,'confidence':round(float(score),3)})
print('started defense_falcon_rcnn.detect_yolo')
hash = hashlib.sha1()
hash.update(str(time.time()))
# img_filename = 'incoming.jpg'
yolo_path = '/data/jeremy/darknet_python/darknet'
# cfg_path = '/data/jeremy/darknet_python/cfg/yolo-voc_544.cfg'
cfg_path = '/data/jeremy/darknet_python/cfg/yolo-voc_608.cfg'
# weights_path = '/data/jeremy/darknet_python/yolo-voc_544_95000.weights'
weights_path = '/data/jeremy/darknet_python/yolo-voc_608_46000.weights'
save_path = './results/'
detections_path = 'detections.txt' #these are getting stored in local dir it seems
img_filename = hash.hexdigest()[:10]+'.jpg'
Utils.ensure_dir(save_path)
img_path = os.path.join(save_path,img_filename)
print('save file '+str(img_path))
cv2.imwrite(img_path,img_arr)
try:
os.remove(detections_path) #this is a temp file to hold current detection - erase then write
except:
print('file {} doesnt exist'.format(detections_path))
cmd = yolo_path+' detect '+cfg_path+' '+weights_path+' '+img_path
subprocess.call(cmd, shell=True) #blocking call
time.sleep(0.1) #wait for file to get written
relevant_bboxes = []
if not os.path.exists(detections_path):
return []
with open(detections_path,'r') as fp:
lines = fp.readlines()
fp.close()
saved_detections = img_path.replace('.jpg','.txt')
with open(saved_detections,'w') as fp2: #copy into another file w unique name so we can delete original
fp2.writelines(lines)
fp2.close()
for line in lines:
label_index,confidence,xmin,ymin,xmax,ymax = line.split()
label_index=int(label_index)
label=classes[label_index]
confidence=float(confidence)
xmin=int(xmin)
xmax=int(xmax)
ymin=int(ymin)
ymax=int(ymax)
item = {'object':label,'bbox':[xmin,ymin,xmax,ymax],'confidence':round(float(confidence),3)}
if label == 'person':
cropped_image = img_arr[ymin:ymax, xmin:xmax]
# print('crop:{} {}'.format(item["bbox"],cropped_image.shape))
# get hydra results
try:
hydra_output = self.get_hydra_output(cropped_image)
if hydra_output:
item['details'] = hydra_output
except:
print "Hydra failed " + traceback.format_exc()
relevant_bboxes.append(item)
if save_results:
imutils.bb_with_text(img_arr,[xmin,ymin,(xmax-xmin),(ymax-ymin)],label)
if save_results:
marked_imgname = img_path.replace('.jpg','_bbs.jpg')
print('bbs writtten to '+str(marked_imgname))
cv2.imwrite(marked_imgname,img_arr)
self.write_log(url,relevant_bboxes)
print relevant_bboxes
return relevant_bboxes
def convert_and_save_results(mask,
label_names,
pose,
filename,
img,
url,
forwebtool=True):
'''
This saves the mask using the labelling fashionista_categories_augmented_zero_based
:param mask:
:param label_names:
:param pose:
:param filename:
:param img:
:param url:
:return:
'''
fashionista_ordered_categories = constants.fashionista_categories_augmented_zero_based #constants.fashionista_categories
h, w = mask.shape[0:2]
new_mask = np.ones(
(h, w, 3), dtype=np.uint8
) * 255 # anything left with 255 wasn't dealt with in the following conversion code
print('new mask size:' + str(new_mask.shape))
success = True #assume innocence until proven guilty
print('attempting convert and save, shapes:' + str(mask.shape) + ' new:' +
str(new_mask.shape))
for label in label_names: # need these in order
if label in fashionista_ordered_categories:
fashionista_index = fashionista_ordered_categories.index(
label) + 0 # number by 0=null, 55=skin , not 1=null,56=skin
pd_index = label_names[label]
pixlevel_v2_index = constants.fashionista_aug_zerobased_to_pixlevel_categories_v2[
fashionista_index]
pixlevel_index = constants.fashionista_aug_zerobased_to_pixlevel_categories_v4_for_web[
fashionista_index]
if pixlevel_index is None:
pixlevel_index = 0 #map unused categories (used in fashionista but not pixlevel v2) to background
# new_mask[mask==pd_index] = fashionista_index
# print('old index '+str(pd_index)+' for '+str(label)+': gets new index:'+str(fashionista_index)+':' + fashionista_ordered_categories[fashionista_index]+ ' and newer index '+str(pixlevel_v2_index)+':'+constants.pixlevel_categories_v2[pixlevel_v2_index])
new_mask[mask == pd_index] = pixlevel_index
else:
print('label ' + str(label) + ' not found in regular cats')
success = False
if 255 in new_mask:
print('didnt fully convert mask')
success = False
if success:
try: #write orig file
# conversion_utils.count_values(new_mask,labels=constants.pixlevel_categories_v2)
conversion_utils.count_values(
new_mask, labels=constants.pixlevel_categories_v4_for_web)
dir = constants.pd_output_savedir
Utils.ensure_dir(dir)
full_name = os.path.join(dir, filename)
# full_name = filename
print('writing output img to ' + str(full_name))
cv2.imwrite(full_name, img)
except:
print('fail in try 1, ' + sys.exc_info()[0])
try: #write rgb mask
# bmp_name = full_name.replace('.jpg','_pixv2.png')
bmp_name = full_name.replace('.jpg', '_pixv4.png')
print('writing mask bmp to ' + str(bmp_name))
cv2.imwrite(bmp_name, new_mask)
imutils.show_mask_with_labels(
new_mask,
labels=constants.pixlevel_categories_v4,
original_image=full_name,
save_images=True)
if socket.gethostname() != 'extremeli-evolution-1':
command_string = 'scp ' + bmp_name + ' [email protected]:/var/www/js-segment-annotator/data/pd_output/pd/'
subprocess.call(command_string, shell=True)
command_string = 'scp ' + full_name + ' [email protected]:/var/www/js-segment-annotator/data/pd_output/pd/'
subprocess.call(command_string, shell=True)
except:
print('fail in try 2, ' + str(sys.exc_info()[0]))
try: #write webtool mask
if forwebtool:
webtool_mask = copy.copy(new_mask)
webtool_mask[:, :,
0] = 0 #zero out the B,G for webtool - leave only R
webtool_mask[:, :,
1] = 0 #zero out the B,G for webtool - leave only R
# bmp_name=full_name.replace('.jpg','_pixv2_webtool.png')
bmp_name = full_name.replace('.jpg', '_pixv4_webtool.png')
print('writing mask bmp to ' + str(bmp_name))
cv2.imwrite(bmp_name, webtool_mask)
command_string = 'scp ' + bmp_name + ' [email protected]:/var/www/js-segment-annotator/data/pd_output/pd'
subprocess.call(command_string, shell=True)
except:
print('fail in try 3, ' + str(sys.exc_info()[0]))
try:
pose_name = full_name.strip('.jpg') + '.pose'
with open(pose_name, "w+") as outfile:
print('succesful open, attempting to write pose')
poselist = pose[0].tolist()
# json.dump([1,2,3], outfile, indent=4)
json.dump(poselist, outfile, indent=4)
if url is not None:
url_name = full_name.strip('.jpg') + '.url'
print('writing url to ' + str(url_name))
with open(url_name, "w+") as outfile2:
print('succesful open, attempting to write:' + str(url))
outfile2.write(url)
except:
print('fail in convert_and_save_results dude, bummer')
print(str(sys.exc_info()[0]))
return
else:
print(
'didnt fully convert mask, or unkown label in convert_and_save_results'
)
success = False
return
def flickr_dl(tag,avoid_these_terms=None,n_pages = 20000,start_page=1,savedir=None):
'''
https://www.flickr.com/services/api/flickr.photos.search.html tags (Optional)
A comma-delimited list of tags. Photos with one or more of the tags listed will be returned. You can exclude results that match a term by prepending it with a - character.
:param tag:
:param avoid_these_terms:
:param n_pages:
:param start_page:
:param savedir:
:return:
'''
results_per_page = 500
max_pages_returned = int(float(4900)/results_per_page)
compressed_tag = tag.replace(' ','+')
if savedir is None:
savedir = '/home/jeremy/image_dbs/flickr/'+compressed_tag+'/'
Utils.ensure_dir(savedir)
outfile = compressed_tag+'out.txt'
n_dl = 0
mintime = 1262307661 #jan 2010
n_dates = n_pages/max_pages_returned
for dateloop in range(n_dates):
time.sleep(1)
#
maxtime = flickr_get_dates(tag,mintime,savedir=savedir,n_pages=max_pages_returned)
print('mintime '+str(mintime)+' maxtime:'+str(maxtime))
for i in range(start_page,start_page+n_pages):
query = '&tags='+tag+'&page='+str(i)+'&min_upload_date='+str(mintime)+'&max_upload_date='+str(maxtime)+'&per_page='+str(results_per_page)
query = '&text='+compressed_tag+'&page='+str(i)+'&min_upload_date='+str(mintime)+'&max_upload_date='+str(maxtime)+'&per_page='+str(results_per_page)
print query
#kyle key 6351acc69daa0868c61319df617780c0 secret b7a74cf16401856b
cmd = 'curl -X GET "https://api.flickr.com/services/rest/?method=flickr.photos.search&api_key=d8548143cce923734f4093b4b063bc4f&format=json'+query+'" > ' + outfile
print cmd
res = subprocess.call(cmd,shell=True)
print('res:'+str(type(res))+':'+str(res))
with open(outfile,'r') as f:
content=f.read()
#the info is returned inside a function call like flickerApi({bla:bla...}) so I need to strip the beginning and end stuff
stripped = content[14:-1]
f.close()
try:
d = json.loads(stripped)
except:
print('json problem '+str(stripped))
continue
pprint(d)
if not d:
print('no file found')
continue
if not 'photos' in d:
print('no photos field in result, continuing')
continue
phot = d['photos']
if not 'photo' in phot:
print('no photo field in result, continuing')
continue
if 'page' in phot:
print('page '+str(phot['page']))
page = phot['page']
if 'pages' in phot:
print('of total pages '+str(phot['pages']))
pages = phot['pages']
if page and pages and page>pages:
print('beyond last page')
return
imgs = phot['photo']
l = len(imgs)
# print imgs
print(str(l)+' images found')
skip_this = False
for j in range(l):
# time.sleep(0.05)
nth_img = imgs[j]
if avoid_these_terms:
#go thru the entire dict and check if terms to avoid is in there somewhere
for k,v in nth_img.iteritems():
for item in avoid_these_terms:
# print('item:'+item+' k,v:'+str(k)+':'+str(v)+' type:'+str(type(v)))
if v and (type(v) is str or type(v) is unicode) and item in v.lower():
skip_this = True
print('SKIPPING due to :'+str(k)+':'+str(v))
break
if skip_this:
break
# raw_input('ret to cont')
# print nth_img
if skip_this:
continue
if not 'farm' in nth_img:
print('no farm found, continuing')
continue
farm = nth_img['farm']
if not 'id' in nth_img:
print('no id found, continuing')
continue
id = nth_img['id']
if not 'server' in nth_img:
print('no server found, continuing')
continue
server = nth_img['server']
if not 'secret' in nth_img:
print('no secret found, continuing')
continue #
secret = nth_img['secret']
url = 'https://farm'+str(farm)+'.staticflickr.com/'+str(server)+'/'+str(id)+'_'+str(secret)+'.jpg'
# print url
savename=str(id)+'.'+str(server)+'.'+str(farm)+'.jpg'
savename = compressed_tag + savename
savename = os.path.join(savedir,savename)
Utils.ensure_dir(savedir)
if os.path.exists(savename):
print(savename+' exists!!')
continue
# savename=savename[:-4]+'.b.jpg'
# raw_intput('check the flies')
save_img_at_url(url,savename=savename)
n_dl = n_dl + 1
n_files = len(os.listdir(savedir))
print('n dl:'+str(n_dl)+' n_files:'+str(n_files)+' '+savename)
mintime = maxtime
def getty_dl(searchphrase,avoid_these_terms=None,n_pages = 20000,savedir=None):
if savedir is None:
savedir = '/home/jeremy/image_dbs/getty/'+searchphrase+'/'
Utils.ensure_dir(savedir)
#do a first curl to set the page size
# cmd = 'curl -X GET -H "Api-Key: r6zm5n78dguspxkg2ss4xvje" https://api.gettyimages.com/v3/search/images?page_size=100000 > resout1.txt'
# res = subprocess.call(cmd,shell=True)
#next curl with the right phrase, all subsequent ones with ?page= to get next results from same query
outfile = searchphrase+'out.txt'
for i in range(n_pages):
query = '?phrase='+searchphrase+'&page='+str(i+1)
print query
cmd = 'curl -X GET -H "Api-Key: r6zm5n78dguspxkg2ss4xvje" "https://api.gettyimages.com/v3/search/images'+query+ '" > ' + outfile
print cmd
res = subprocess.call(cmd,shell=True)
with open(outfile,'r') as f:
d = json.load(f)
f.close()
pprint(d)
if not d:
print('no file found')
continue
if not 'images' in d:
print('no images field in result, continuing')
continue
imgs = d['images']
l = len(imgs)
# print imgs
print l
skip_this = False
for j in range(l):
time.sleep(0.05)
nth_img = imgs[j]
if avoid_these_terms:
skip_this = False
#go thru the entire dict and check if terms to avoid is in there somewhere
for k,v in nth_img.iteritems():
for item in avoid_these_terms:
# print('item:'+item+' k,v:'+str(k)+':'+str(v)+' type:'+str(type(v)))
if v and (type(v) is str or type(v) is unicode) and item in v.lower():
skip_this = True
print('SKIPPING due to :'+str(k)+':'+str(v))
break
if skip_this:
break
# raw_input('ret to cont')
# print nth_img
if skip_this:
continue
if not 'display_sizes' in nth_img:
print('no display sizes field found, continuing')
continue
ds = nth_img['display_sizes']
# print ds
first = ds[0]
# print first
uri = first['uri']
# print uri
clean_url = uri.split('?')[0]
# print(clean_url)
savename=clean_url.split('?')[0]
savename=savename.split('/')[-1]
savename = searchphrase + savename
savename = os.path.join(savedir,savename)
Utils.ensure_dir(savedir)
# print(savename)
save_img_at_url(uri,savename=savename)
def detect_yolo_pyyolo(self, img_arr, url='',classes=constants.hls_yolo_categories,save_results=True):
# item = {'object':label,'bbox':[xmin,ymin,xmax,ymax],'confidence':'>'+str(thresh)}
print('started pyyolo detect, thresh='+str(thresh)+' hier '+str(hier_thresh))
save_path = '/data/jeremy/pyyolo/results/'
if img_arr is None:
print('got None img array!!')
None
if len(img_arr.shape) == 2: #got 1-chan(gray) image
print('got gray img')
img_arr_bgr=np.zeros([img_arr.shape[0],img_arr.shape[1],3])
img_arr_bgr=img_arr
print('sizes: {} {}'.format(img_arr_bgr,img_arr))
img_arr = img_arr_bgr
print('img arr size {}'.format(img_arr.shape))
max_image_size=1200 #1280 fails
resized=False
if img_arr.shape[0]>max_image_size or img_arr.shape[1]>max_image_size: #maybe this is causing prob at http://a.abcnews.com/images/US/150815_yt_phillypd_33x16_1600.jpg
maxside=max(img_arr.shape)
reduction_factor = maxside/1000.0 #force maxside to 1000
original_size=img_arr.shape
resized=True
img_arr = cv2.resize(img_arr,(int(img_arr.shape[1]/reduction_factor),int(img_arr.shape[0]/reduction_factor)))
print('reduced size to {}'.format(img_arr.shape))
#generate randonm filename
hash = hashlib.sha1()
hash.update(str(time.time()))
if save_results:
timestr = time.strftime("%Y%m%d.%H%M%S")
img_filename = timestr+hash.hexdigest()[:5]+'.jpg'
Utils.ensure_dir(save_path)
img_path = os.path.join(save_path,img_filename)
print('detect_yolo_pyyolo saving file '+str(img_path))
try:
cv2.imwrite(img_path,img_arr)
except:
print('some trouble saving image,'+str(sys.exc_info()[0]))
relevant_items = []
print('getting results from get_pyyolo_results')
yolo_results = self.get_pyyolo_results(img_arr)
print('got results from get_pyyolo_results')
for item in yolo_results:
print(item)
if resized:
img_arr = cv2.resize(img_arr,(original_size[1],original_size[0]))
item['bbox'][0]=int(item['bbox'][0]*reduction_factor)
item['bbox'][1]=int(item['bbox'][1]*reduction_factor)
item['bbox'][2]=int(item['bbox'][2]*reduction_factor)
item['bbox'][3]=int(item['bbox'][3]*reduction_factor)
print('fixed size back to original {}'.format(item))
xmin=item['bbox'][0]
ymin=item['bbox'][1]
xmax=item['bbox'][2]
ymax=item['bbox'][3]
assert xmin<xmax,'xmin not < xmax!!!'
assert ymin<ymax,'xmin not < xmax!!!'
##### TAKING OUT RELEVANT ITEMS ON ROYS SUGGESTION
use_hydra=False
if use_hydra:
if item['object'] == 'person':
cropped_image = img_arr[ymin:ymax, xmin:xmax]
# print('crop:{} {}'.format(item["bbox"],cropped_image.shape))
# get hydra results
try:
hydra_output = self.get_hydra_output(cropped_image)
if hydra_output:
item['details'] = hydra_output
except:
print "Hydra call from pyyolo defense falcon failed " + traceback.format_exc()
relevant_items.append(item)
if save_results:
imutils.bb_with_text(img_arr,[xmin,ymin,(xmax-xmin),(ymax-ymin)],item['object'])
if save_results:
marked_imgname = img_path.replace('.jpg','_bb_yolos.jpg')
print('pyyolo bbs being writtten to '+str(marked_imgname))
try:
pass #no real need to save bb file as no extra info there
# r=cv2.imwrite(marked_imgname,img_arr)
#skip saving bb image to conserve space
# print('write result '+str(r))
except:
print('some trouble saving bb image,'+str(sys.exc_info()[0]))
txtname=img_path.replace('.jpg','.txt')
self.write_log(url,relevant_items,filename=txtname)
print('detect yolo returning:'+str(relevant_items))
return relevant_items