def get_trimap_data(trimap_path):
filename1 = trimap_path.split("/")[-1]
filename2 = filename1.split("_")[0] + "_" + filename1.split("_")[1]
outpath = "/mnt/CLAWS1/stamilev/volumes/groundtruth/dense_gt_data/" + filename2 + ".h5"
print outpath
trimap = read_h5(trimap_path)
#dgt= dgt[25:725, 0:700, 0:700]
unique = np.unique(trimap)
unique_elements = []
unique_tuples = []
print "no of unique elements", len(unique)
for i in unique:
mask = trimap == i
noe = np.sum(mask)
print i, noe
unique_elements.append(noe)
unique_tuples.append((i, noe))
assert len(unique)==len(unique_elements)
l = len(unique_elements)
nodes_data = np.zeros((l,2), dtype=np.uint)
for n in range(l):
nodes_data[n,0] = unique_tuples[n][0]
nodes_data[n,1] = unique_tuples[n][1]
#outpath = "/home/stamylew/volumes/groundtruth/dense_gt_data/entire_usable_data.h5"
save_h5(nodes_data, outpath, "labels and occurrence", None)
def get_quality_values(predict, gt, dense_gt, minMemb, minSeg, sigMin, sigWeights, sigSmooth,edgeLengths,nodeFeatures,nodeSizes,
nodeLabels, nodeNumStop, beta, metric, wardness,out):
"""
"""
adjusted_predict = adjust_predict(predict)
# save_h5(adjusted_predict, "/home/stamylew/delme/prob.h5", "data")
relevant_predict, relevant_gt = exclude_ignore_label(adjusted_predict, gt)
acc, pre, rec = accuracy_precision_recall(relevant_predict, relevant_gt)
auc_score = calculate_roc_auc_score(relevant_predict, relevant_gt)
no_of_true_pos, no_of_false_pos = true_and_false_pos(relevant_predict, relevant_gt)
no_of_true_neg, no_of_false_neg = true_and_false_neg(relevant_predict, relevant_gt)
print "#nodes in dense gt", len(np.unique(dense_gt))
pmin= 0.5
segmentation, super_pixels, wsDt_data, agglCl_data = get_segmentation(adjusted_predict,pmin, minMemb, minSeg, sigMin,
sigWeights, sigSmooth, True,False, edgeLengths,
nodeFeatures, nodeSizes, nodeLabels, nodeNumStop,
beta, metric, wardness, out)
save_h5(segmentation, "/home/stamylew/delme/segmap.h5", "data", None)
ri_data = skl.randIndex(segmentation.flatten().astype(np.uint32), dense_gt.flatten().astype(np.uint32), True)
voi_data = skl.variationOfInformation(segmentation.flatten().astype(np.uint32), dense_gt.flatten().astype(np.uint32), True)
sp_ri = skl.randIndex(super_pixels.flatten().astype(np.uint32), dense_gt.flatten().astype(np.uint32), True)
sp_voi = skl.variationOfInformation(super_pixels.flatten().astype(np.uint32), dense_gt.flatten().astype(np.uint32), True)
print
print "super pixels values", (sp_ri, sp_voi)
print
print "segmentation values", (ri_data, voi_data)
return acc, pre, rec, auc_score, ri_data, voi_data, no_of_true_pos, no_of_false_pos, no_of_true_neg, no_of_false_neg, \
wsDt_data, agglCl_data
def make_seg_map(prob_map_path, dense_gt_path):
prob_map = read_h5(prob_map_path)
dense_gt = read_h5(dense_gt_path)
nodes_in_dgt = len(np.unique(dense_gt))
print "nodes in dgt", nodes_in_dgt
seed_map = wsDtSegmentation(prob_map, pmin=0.5, minMembraneSize=10, minSegmentSize=0, sigmaMinima=6, sigmaWeights=1, cleanCloseSeeds=True,
returnSeedsOnly=True)
seg_map, sup_map, wsDt_data, agglCl_data = get_segmentation(prob_map, pmin=0.5, minMemb=10, minSeg=10, sigMin=6, sigWeights=1, sigSmooth=0.1,
cleanCloseSeeds=True, returnSeedsOnly=False, edgeLengths=None, nodeFeatures=None,
nodeSizes=None, nodeLabels=None, nodeNumStop=nodes_in_dgt, beta=0, metric='l1',
wardness=0.2, out=None)
print "seeds", len(np.unique(seed_map))
print "seg", len(np.unique(seg_map))
#make folder for seeds
seed_folder = prob_map_path.split("prob")[0] + "seeds/"
if not os.path.exists(seed_folder):
os.mkdir(seed_folder)
seed_map_outpath = seed_folder + prob_map_path.split("/")[-1]
#make folder for super pixels
sup_folder = prob_map_path.split("prob")[0] + "sup_maps/"
if not os.path.exists(sup_folder):
os.mkdir(sup_folder)
sup_map_outpath = sup_folder + prob_map_path.split("/")[-1]
#make folder for segmentations
seg_folder = prob_map_path.split("prob")[0] + "seg_maps/"
if not os.path.exists(seg_folder):
os.mkdir(seg_folder)
seg_map_outpath = seg_folder + prob_map_path.split("/")[-1]
#make folder for segmentation data
data_folder = prob_map_path.split("prob")[0] + "seg_data/"
if not os.path.exists(data_folder):
os.mkdir(data_folder)
data_outpath = data_folder + prob_map_path.split("/")[-1]
#get segmentation data
ri, voi = rand_index_variation_of_information(seg_map, dense_gt)
print "ri, voi", ri , voi
ri_data = np.array((ri))
voi_data = np.array((voi))
ri_voi_data = np.array((ri, voi))
#save seeds, superpixels and segmentations
# save_h5(ri_voi_data, data_outpath, key)
# save_h5(ri_data, data_outpath, "rand index", None)
# save_h5(voi_data, data_outpath, "variation of information", None)
# save_h5(seed_map, seed_map_outpath, "data", None)
save_h5(sup_map,sup_map_outpath, "data", None)
save_h5(seg_map,seg_map_outpath, "data", None)
def save_quality_values(predict_path, gt_path, dense_gt_path, outpath, slices, minMemb=10, minSeg=10, sigMin=6, sigWeights=1,
sigSmooth=0.1,edgeLengths=None, nodeFeatures=None, nodeSizes=None, nodeLabels=None, nodeNumStop=None,
beta=0, metric='l1', wardness=0.2, out=None):
"""
"""
#read predict and groundtruth data
predict_data = read_h5(predict_path, "exported_data")
gt_data = read_h5(gt_path)
dense_gt_data = read_h5(dense_gt_path)
nodes_in_dgt = np.unique(dense_gt_data)
nodeNumStop = nodes_in_dgt
print "nodeNumStop", nodeNumStop
#all_quality_data = get_quality_values(predict_data, gt_data, dense_gt_data)
#get quality values
acc, prec, rec, auc_score, ri, voi, tp, fp, tn, fn, wsDt_data, agglCl_data = get_quality_values(predict_data, gt_data,
dense_gt_data, minMemb, minSeg,
sigMin, sigWeights, sigSmooth,
edgeLengths, nodeFeatures,
nodeSizes, nodeLabels,
nodeNumStop, beta, metric,
wardness, out)
measurements = {"accuracy":acc, "precision":prec, "recall":rec, "auc score":auc_score, "rand index":ri,
"variation of information":voi, "true positives":tp, "false positives":fp, "true negatives":tn,
"false negatives":fn}
# acc, prec, rec, auc_score, tp, fp, tn, fn, wsDt_data, agglCl_data = get_quality_values(predict_data, gt_data, dense_gt_data)
# measurements = {"accuracy":acc, "precision":prec, "recall":rec, "auc score":auc_score,
# "true positives":tp, "false positives":fp, "true negatives":tn,
# "false negatives":fn}
if not os.path.exists(outpath):
print "Output h5 file did not exist."
for measurement in measurements.iterkeys():
key = "quality/" + measurement
data = np.zeros((1,))
data[0] = measurements[str(measurement)]
save_h5(data, outpath, key, None)
else:
for measurement in measurements.iterkeys():
key = "quality/" + measurement
data = read_h5(outpath, key)
new_data = np.vstack((data, measurements[measurement]))
save_h5(new_data, outpath, key, None)
for slice in slices:
im_outpath = outpath.split(".")[-2]
predict_im_outpath = im_outpath + "_slice_" + str(slice) + ".png"
gt_im_outpath = im_outpath + "_slice_" + str(slice) + "_gt.png"
plt.imsave(predict_im_outpath, adjust_predict(predict_data)[slice])
plt.imsave(gt_im_outpath, gt_data[slice])
save_h5(wsDt_data, outpath, "segmentation/wsDt data", None)
save_h5([agglCl_data], outpath, "segmentation/agglCl data", None)
fpr, tpr, threshold = draw_roc_curve(predict_data, gt_data)
plt.figure()
plt.plot(fpr, tpr)
plt.savefig(im_outpath + "roc_curve_test.png")
def test_wsDt_agglCl_configs(predict_path, dense_gt_path, pmin=0.5, minMemb=10, minSeg=10, sigMin=2, sigWeights=2, sigSmooth=0.1):
"""
:param predict_path:
:param dense_gt_path:
:param pmin:
:param minMemb:
:param minSeg:
:param sigMin:
:param sigWeights:
:param sigSmooth:
:return:
"""
#get rand index and variation of information
predict_data = read_h5(predict_path)
adjusted_predict_data = adjust_predict(predict_data)
dense_gt_data = read_h5(dense_gt_path)
seg, sup = get_segmentation(adjusted_predict_data,pmin,minMemb,minSeg,sigMin,sigWeights,sigSmooth)
ri, voi = rand_index_variation_of_information(seg, dense_gt_data)
print "ri, voi", ri, voi
config = predict_path.split("/")[-2]
filename = config + "_" + predict_path.split("/")[-1]
outpath_folder = "/home/stamylew/test_folder/q_data/wsDt_agglCl_tests/"
outpath = outpath_folder + filename
key = "sigMin_" + str(sigMin) + "_sigWeights_" + str(sigWeights) + "_sigSmooth_" + str(sigSmooth)
if not os.path.exists(outpath):
print "Output h5 file did not exist."
data = np.zeros((1,2))
data[0,0] = ri
data[0,1] = voi
save_h5(data, outpath, key, None)
else:
old_data = read_h5(outpath, key)
data = np.zeros((1,2))
data[0,0] = ri
data[0,1] = voi
new_data = np.vstack((old_data, data))
save_h5(new_data, outpath, key, None)
def redo_quality2(block_path, dense_gt_path, minMemb, minSeg, sigMin, sigWeights, sigSmooth, edgeLengths=None,
nodeFeatures=None, nodeSizes=None, nodeLabels=None, nodeNumStop=None, beta=0, metric='l1',
wardness=0.2, out=None):
folder_list = [f for f in os.listdir(block_path) if not os.path.isfile(os.path.join(block_path,f))
and ("figure" not in f) and "redone" not in f]
print "folder_list", folder_list
outpath = block_path + "/redone_quality"
if not os.path.exists(outpath):
os.mkdir(outpath)
dense_gt_data = read_h5(dense_gt_path)
nodes_in_dgt = len(np.unique(dense_gt_data))
for f in folder_list:
folder_path = os.path.join(block_path, f)
prob_file = [h for h in os.listdir(folder_path) if "probs" in h]
prob_file_path = os.path.join(folder_path,prob_file[0])
print
print
print prob_file_path
predict_data = read_h5(prob_file_path)
adjusted_predict_data = adjust_predict(predict_data)
print "#nodes in dense gt", nodes_in_dgt
pmin = 0.5
seg, sup, wsDt_data, agglCl_data = get_segmentation(adjusted_predict_data, pmin, minMemb, minSeg, sigMin, sigWeights,
sigSmooth,True,False, edgeLengths,nodeFeatures, nodeSizes,
nodeLabels, nodeNumStop,beta, metric, wardness, out)
nodes_in_seg = len(np.unique(seg))
nodes_in_sup = len(np.unique(sup))
ri, voi = rand_index_variation_of_information(seg, dense_gt_data)
new_folder_path = outpath + "/" + f
if not os.path.exists(new_folder_path):
os.mkdir(new_folder_path)
key = "pmin_"+ str(pmin) + "_minMemb_" + str(minMemb)+ "_minSeg_"+ str(minSeg) +"_sigMin_" + str(sigMin) + "_sigWeights_" + str(sigWeights) \
+ "_sigSmooth_" + str(sigSmooth) + "_nodeNumStop_" + str(nodeNumStop)
save_h5(seg, new_folder_path + "/segmentation.h5", key)
save_h5(sup, new_folder_path + "/super_pixels.h5", key)
seg_data = np.zeros((2,3))
seg_data[0,0] = ri
seg_data[0,1] = voi
seg_data[1,0] = nodes_in_dgt
seg_data[1,1] = nodes_in_sup
seg_data[1,2] = nodes_in_seg
save_h5(seg_data, new_folder_path + "/seg_data", key, None)
def test(ilp, files, gt_path, dense_gt_path, labels="", loops=3, weights="", repeats=1, outpath= "",
t_cache = "", p_cache = ""):
""" Run test
:param: ilp : path to ilp project to use for training
:param: files : path to file to do batch prediction on
:param: gt_path : path to trimap groundtruth
:param: dense_gt_path : path to dense groundtruth
:param: labels : amount of labeled pixels to use in training
:param: loops : amount of autocontext loops
:param: weights : weighting of the labels over the autocontext loops
:param: repeats : amount of repeat runs of the test
:param: outpath : outpath for test outputs
:param: t_cache : path to training data cache
:param: p_cache : path to prediction data cache
"""
hostname = socket.gethostname()
print "test information:"
print
print "hostname:", hostname
print
print "ilp_file:", ilp
print
print "files to predict:", files
print
print "groundtruth:", gt_path
print
print "dense groundtruth:", dense_gt_path
print
print "labels:", labels
print
print "loops:", loops
print
print "weights:", weights
print
print "repeats:", repeats
# Collect the test specifications
ilp_split = ilp.split(".")[-2]
training_file = "training file: "+ilp_split.split("/")[-1]
gt_split = gt_path.split(".")[-2]
trimap_file = "trimap file: " + gt_split.split("/")[-1]
# Assign paths
filesplit = files.split(".")[-2]
filename = filesplit.split("/")[-1]
prediction_file = "prediction file: "+ filename
test_folder_path = assign_path(hostname)[5]
if t_cache == "":
t_cache = test_folder_path + "/t_cache"
if p_cache == "":
p_cache = test_folder_path + "/p_cache/" + filename
if not os.path.exists(p_cache):
os.mkdir(p_cache)
if outpath == "":
output = test_folder_path + "/q_data"
else:
output = outpath
print
print "outpath:", outpath
print
print "t_cache:", t_cache
print
print "p_cache:", p_cache
# Create file tags
if labels == "":
label_tag = "all"
else:
# assert labels == check_ilp_labels(ilp)
true_labels = check_ilp_labels(ilp)
label_tag = true_labels
if weights == "":
weight_tag = "none"
else:
weight_tag = str(weights)
# Make folder for quality data
if "manual" in ilp:
filename += "_manual"
if "hand_drawn" in ilp:
filename += "_hand_drawn"
if "clever" in ilp:
filename += "_clever"
if "less_feat" in ilp:
filename += "_less_feat"
#change labels for every test
# if labels != "":
# print
# print "reducing labels to " + str(labels)
# ilp = reduce_labels_in_ilp(ilp, labels)
file_dir = output + "/" + filename
if not os.path.exists(file_dir):
os.mkdir(file_dir)
# Overwrite folder directory
# file_dir = "/mnt/CLAWS1/stamilev/delme"
# Check if file directory exists, if not make such directory
if not os.path.exists(file_dir):
print
print "Output folder did not exist."
os.mkdir(file_dir)
print
print "New one named " + file_dir + " was created."
q_outpath = file_dir + "/n_" + str(loops) + "_l_" + str(label_tag) + "_w_" + weight_tag
prob_folder = q_outpath + "/prob_files"
q_data_outpath = q_outpath + "/n_" + str(loops) + "_l_" + str(label_tag) + "_w_" + weight_tag + ".h5"
# Check if test directory exists, if not make such directory
if not os.path.exists(q_outpath):
print
print "Output h5 file did not exist"
os.mkdir(q_outpath)
os.mkdir(prob_folder)
print
print "New one named " + q_outpath + " was created."
if not os.path.exists(prob_folder):
os.mkdir(prob_folder)
# Run the test
for i in range(repeats):
print
print "round of repeats %d of %d" % (i+1, repeats)
#train on ilp project
ac_train(ilp, labels, loops, weights, t_cache, outpath)
print
print "training completed"
#batch predict files
ac_batch_predict(files, t_cache, p_cache, overwrite = "")
print
print "batch prediction completed"
#save prob files and quality data
prob_file = [x for x in os.listdir(p_cache) if ("probs" in x)]
predict_path = p_cache + "/" + prob_file[0]
predict_data = adjust_predict(read_h5(predict_path))
prob_path = prob_folder + "/prob_"+ str(i+1)+ ".h5"
# if os.path.exists(prob_path):
# # TODO:not working yet
# prob_path = prob_path.split("/")[-2] + "/new_" + prob_path.split("/")[-1]
# print prob_path
save_h5(predict_data, prob_path, "data")
save_quality_values(predict_path, gt_path, dense_gt_path, q_data_outpath, (0,49,99))
#save test specification data
save_h5([training_file, prediction_file, trimap_file], q_data_outpath, "used files", None)
save_h5([labels],q_data_outpath, "autocontext_parameters/#labels")
save_h5([loops], q_data_outpath, "autocontext_parameters/#loops")
save_h5([str(weight_tag)], q_data_outpath, "autocontext_parameters/weights")
print
print "quality data saved"
#save configuration data
# call(["cp", predict_path, q_outpath])
save_h5(["pmin", "minMemb", "minSeg", "sigMin", "sigWeights", "sigSmooth", "cleanCloseSeeds", "returnSeedsOnly"],
q_data_outpath, "segmentation/wsDt parameters", None)
save_h5(["edge_weights", "edgeLengths", "nodeFeatures", "nodeSizes", "nodeLabels", "nodeNumStop", "beta", "metric",
"wardness", "out"], q_data_outpath, "segmentation/agglCl parameters", None)
print
print "quality data saved"
def compress(input, ouput):
file = read_h5(input[0], input[1])
save_h5(file, output[0], output[1], "lzf")
# for l in range(t-1):
# data[i+l,j,k] = 1
# if data[i,j,k] == 1 and data[i+t,j,k] == 0:
# for l in range(t):
# data[i+l,j,k] = 1
# return data
def label_neurons(data, neuron_label):
for i in range(data.shape[0]):
for j in range(data.shape[1]):
for k in range(data.shape[2]):
if data[i,j,k] != 1 and data[i,j,k] != 0:
data[i,j,k] = neuron_label
return data
def filter_membrane(dense_gt_data):
prepared_data = prepare_data(dense_gt_data)
print "prepared data unique", np.unique(prepared_data)
memb = create_membrane(prepared_data)
memb_data = label_neurons(memb, 0)
return memb_data, dense_gt_data
if __name__ == '__main__':
data = read_h5("/home/stamylew/volumes/groundtruth/dense_groundtruth/100p_cube1_dense_gt.h5", "data")
memb = filter_membrane(data)
save_h5(memb, "/home/stamylew/volumes/groundtruth/memb/100p_cube1_memb.h5", "data", "lzf")
print "done"
# -*- coding: utf-8 -*-
"""
Created on Tue Sep 22 14:31:38 2015
@author: stamylew
"""
import numpy as np
from python_functions.handle_h5.handle_h5 import read_h5, save_h5
def threshold(data, threshold_membrane, threshold_neurons):
bin_data = np.zeros(data.shape)
for i in range(data.shape[0]):
for j in range(data.shape[1]):
for k in range(data.shape[2]):
if data[i,j,k] <= threshold_membrane:
bin_data[i,j,k] = 0
if data[i,j,k] > threshold_membrane and data[i,j,k] <= threshold_neurons:
bin_data[i,j,k] = 0.5
if data[i,j,k] > threshold_neurons:
bin_data[i,j,k] = 1
return bin_data
if __name__ == '__main__':
f = read_h5("/home/stamylew/volumes/100x100block_of_validation_sample_Probabilities.h5", "exported_data")
data = threshold(f, 0.4, 0.6)
save_h5(data, "/home/stamylew/volumes/100x100block_of_validation_sample_Probabilities.h5", "bin", None)
print "done"