def visualizeBestTubeRank(params):
class_labels_map = params.class_labels_map
rel_path = params.rel_path
out_file_html = params.out_file_html
out_dir = params.out_dir
score_info_file = params.score_info_file
[class_labels,class_idx_map]=zip(*class_labels_map);
class_labels=list(class_labels);
class_idx_map=list(class_idx_map);
[score_files,score_files_info]=pickle.load(open(score_info_file,'rb'));
class_idx=np.unique(score_files_info[:,0]);
tubes=np.unique(score_files_info[:,3])
best_tubes_overall=score_files_info[:,3];
out_file=os.path.join(out_dir,'overall.png');
visualize.hist(best_tubes_overall,out_file,bins=tubes,normed=True,xlabel='Tube_idx',ylabel='Frequency',title='Best Tube Over All',cumulative=False)
img_paths=[];
captions=[];
for class_idx_curr in class_idx:
label=class_labels[class_idx_map.index(class_idx_curr)]
out_file=os.path.join(out_dir,label+'.png');
img_paths.append([out_file.replace(rel_path[0],rel_path[1])]);
captions.append([label]);
rel_tubes=score_files_info[score_files_info[:,0]==class_idx_curr,3];
print class_idx_curr,rel_tubes.shape,min(rel_tubes),max(rel_tubes);
visualize.hist(rel_tubes,out_file,bins=tubes,normed=True,xlabel='Tube Idx',ylabel='Frequency',title='Best Tube '+label,cumulative=False)
# visualize.save
visualize.writeHTML(out_file_html,img_paths,captions,400,400);
def script_visualizeHashBinDensity(params):
hash_tables = params.hash_tables
in_files = params.in_files
out_files = params.out_files
out_file_html = params.out_file_html
rel_path = params.rel_path
bins = params.bins
height_width = params.height_width
min_maxs = []
for file_idx, in_file in enumerate(in_files):
densities = pickle.load(open(in_file, 'rb'))
densities = densities.values()
min_maxs.append((min(densities), max(densities)))
visualize.hist(densities,
out_files[file_idx],
bins=bins,
normed=True,
xlabel='Bin Density',
ylabel='Frequency',
title="Hash Bins' Density")
img_files_html = [[out_file.replace(rel_path[0], rel_path[1])]
for out_file in out_files]
captions_html = []
for idx_hash_table, hash_table in enumerate(hash_tables):
caption_curr = str(hash_table) + ' ' + str(min_maxs[idx_hash_table])
captions_html.append([caption_curr])
visualize.writeHTML(out_file_html, img_files_html, captions_html,
height_width[0], height_width[1])
def visualizeScores(path_to_data,out_file):
data=np.load(path_to_data);
scores=data[:,-2];
print np.unique(data[:,-1]);
print np.min(scores);
print np.max(scores);
visualize.hist(scores,out_file,bins=100,normed=True,xlabel='Value',ylabel='Frequency',title='Predicted Scores')
def script_visualizeHashBinDensity(params):
hash_tables = params.hash_tables
in_files = params.in_files
out_files = params.out_files
out_file_html = params.out_file_html
rel_path = params.rel_path
bins = params.bins
height_width = params.height_width
min_maxs=[];
for file_idx,in_file in enumerate(in_files):
densities=pickle.load(open(in_file,'rb'));
densities=densities.values();
min_maxs.append((min(densities),max(densities)))
visualize.hist(densities,out_files[file_idx],bins=bins,normed=True,xlabel='Bin Density',
ylabel='Frequency',title="Hash Bins' Density")
img_files_html=[[out_file.replace(rel_path[0],rel_path[1])] for out_file in out_files];
captions_html=[]
for idx_hash_table,hash_table in enumerate(hash_tables):
caption_curr=str(hash_table)+' '+str(min_maxs[idx_hash_table]);
captions_html.append([caption_curr]);
visualize.writeHTML(out_file_html,img_files_html,captions_html,height_width[0],height_width[1]);
def main():
dir_clusters='/disk2/temp/youtube_clusters_check_nothresh';
out_dir=os.path.join(dir_clusters,'viz');
util.mkdir(out_dir);
script_seeMultipleClusters(dir_clusters,out_dir)
return
script_seeMultipleClusters();
return
dir_clusters='/disk3/maheen_data/debug_networks/clusters_youtube_multiple';
clusters_all=util.getFilesInFolder(dir_clusters,'.npy');
clusters_all=[file_curr for file_curr in clusters_all if 'harder' in file_curr];
clusters_all.append(os.path.join(dir_clusters,'clusters_original.npy'));
min_mags=[];
for file_curr in clusters_all:
clusters=np.load(file_curr);
mags=np.power(np.sum(np.power(clusters,2),axis=1),0.5);
min_mag=np.min(mags);
min_mags.append(min_mag);
print min_mags,np.max(min_mags);
thresh=1;
counts=[];
for file_curr in clusters_all:
clusters=np.load(file_curr);
print file_curr
mags=np.power(np.sum(np.power(clusters,2),axis=1),0.5);
count=np.sum(mags<=thresh);
print count
counts.append(count);
print np.mean(counts);
# return
dir_curr='/disk3/maheen_data/debug_networks/figuringClustering';
mag_file=os.path.join(dir_curr,'mags_all.npy');
mags=np.load(mag_file);
print len(mags),np.sum(mags<=thresh);
mags=mags[mags>thresh];
print len(mags);
out_file=os.path.join(dir_curr,'mag_hist_noZero.png');
visualize.hist(mags,out_file,bins=40,normed=True,xlabel='Value',ylabel='Frequency',title='',cumulative=False);
print out_file.replace('/disk3','vision3.cs.ucdavis.edu:1001');
print np.min(mags),np.max(mags),np.mean(mags),np.std(mags);
def script_visualizeRankDifferenceAsHist(params):
out_file_pre = params.out_file_pre;
out_file_html = params.out_file_html;
rel_path = params.rel_path;
class_ids = params.class_id;
layers = params.layer
if not hasattr(class_ids, '__iter__'):
class_ids = [class_ids];
if not hasattr(class_ids, '__iter__'):
layers = [layers];
img_paths_html=[];
captions=[];
for class_id in class_ids:
img_paths_html_row=[];
captions_row=[];
for layer in layers:
params = params._replace(class_id=class_id)
params = params._replace(layer=layer)
output=getNNRankComparisonInfo(params);
indices_difference=experiments_super.getDifferenceInRank(output['img_paths_nn_train'],
output['img_paths_nn_no_train'],
output['nn_rank_train'],
output['nn_rank_no_train']);
xlabel='Difference in Rank (Train - Untrained)';
ylabel='Frequency'
title=class_id+' '+layer
indices_difference=[diff_curr for diffs_curr in indices_difference for diff_curr in diffs_curr]
if len(indices_difference)==0:
continue;
out_file_im=out_file_pre+'_'+str(params.angle)+'_'+str(params.diff)+'_'+str(params.delta)+'_'+class_id+'_'+layer+'.png';
img_paths_html_row.append(out_file_im.replace(rel_path[0],rel_path[1]));
total=len(indices_difference)
sum_less=sum(np.array(indices_difference)<0)/float(total);
sum_less='%0.2f'%(sum_less,)
sum_more=sum(np.array(indices_difference)>=0)/float(total);
sum_more='%0.2f'%(sum_more,)
captions_row.append('Total '+str(total)+', <0: '+sum_less+', >0: '+sum_more);
visualize.hist(indices_difference,out_file=out_file_im,bins=params.bins,normed=params.normed,xlabel=xlabel,ylabel=ylabel,title=title);
img_paths_html.append(img_paths_html_row);
captions.append(captions_row);
visualize.writeHTML(out_file_html,img_paths_html,captions,params.height_width[0],params.height_width[1]);
def makeTifHists(tif_files,out_file_x,out_file_y,bins=range(1,41)): tif_x=[]; tif_y=[]; for idx_file_curr,file_curr in enumerate(tif_files): print idx_file_curr tif_curr=scipy.misc.imread(file_curr); tif_curr_x=list(np.ravel(tif_curr[:,:,0])); tif_curr_y=list(np.ravel(tif_curr[:,:,1])); tif_x=tif_x+tif_curr_x; tif_y=tif_y+tif_curr_y; len(tif_x); len(tif_y); print scipy.stats.mode(tif_x); print scipy.stats.mode(tif_y); visualize.hist(tif_x,out_file_x,bins=bins) visualize.hist(tif_y,out_file_y,bins=bins)
def visualizeBestTubeRank(params):
class_labels_map = params.class_labels_map
rel_path = params.rel_path
out_file_html = params.out_file_html
out_dir = params.out_dir
score_info_file = params.score_info_file
[class_labels, class_idx_map] = zip(*class_labels_map)
class_labels = list(class_labels)
class_idx_map = list(class_idx_map)
[score_files, score_files_info] = pickle.load(open(score_info_file, 'rb'))
class_idx = np.unique(score_files_info[:, 0])
tubes = np.unique(score_files_info[:, 3])
best_tubes_overall = score_files_info[:, 3]
out_file = os.path.join(out_dir, 'overall.png')
visualize.hist(best_tubes_overall,
out_file,
bins=tubes,
normed=True,
xlabel='Tube_idx',
ylabel='Frequency',
title='Best Tube Over All',
cumulative=False)
img_paths = []
captions = []
for class_idx_curr in class_idx:
label = class_labels[class_idx_map.index(class_idx_curr)]
out_file = os.path.join(out_dir, label + '.png')
img_paths.append([out_file.replace(rel_path[0], rel_path[1])])
captions.append([label])
rel_tubes = score_files_info[score_files_info[:, 0] == class_idx_curr,
3]
print class_idx_curr, rel_tubes.shape, min(rel_tubes), max(rel_tubes)
visualize.hist(rel_tubes,
out_file,
bins=tubes,
normed=True,
xlabel='Tube Idx',
ylabel='Frequency',
title='Best Tube ' + label,
cumulative=False)
# visualize.save
visualize.writeHTML(out_file_html, img_paths, captions, 400, 400)
def script_toyNNExperiment(params):
path_to_db = params.path_to_db;
class_id_pascal = params.class_id_pascal;
video_id = params.video_id;
shot_id = params.shot_id;
tube_id = params.tube_id;
numberofVideos = params.numberofVideos;
numberOfFrames = params.numberOfFrames;
out_file_html = params.out_file_html;
rel_path = params.rel_path;
out_file_hist = params.out_file_hist;
gpuFlag = params.gpuFlag;
dtype = params.dtype;
pascal_ids = params.pascal_ids;
video_info = params.video_info;
out_file_pickle = params.out_file_pickle
info_for_extraction=getInfoForFeatureExtractionForVideo(path_to_db,video_info,numberOfFrames);
video_info={class_id_pascal:[video_id]}
info_for_extraction_query=getInfoForExtractionForTube(path_to_db,class_id_pascal,video_id,shot_id,tube_id)
features_train,labels_train,img_paths_train=setUpFeaturesMat(info_for_extraction,dtype=dtype);
features_test,labels_test,img_paths_test=setUpFeaturesMat(info_for_extraction_query,dtype=dtype);
# features_test,labels_test,img_paths_test=setUpFeaturesMat(info_for_extraction,dtype=dtype);
indices,distances=nearest_neighbor.getNearestNeighbors(features_test,features_train,gpuFlag=gpuFlag);
img_paths_html=[];
captions_html=[];
record_wrong=[]
for r in range(indices.shape[0]):
img_paths_row=[img_paths_test[r].replace(rel_path[0],rel_path[1])];
captions_row=[labels_test[r]];
for c in range(indices.shape[1]):
rank=indices[r,c];
img_paths_row.append(img_paths_train[rank].replace(rel_path[0],rel_path[1]))
captions_row.append(labels_train[rank]);
if labels_train[rank]!=labels_test[r]:
record_wrong.append(c);
img_paths_html.append(img_paths_row);
captions_html.append(captions_row);
visualize.writeHTML(out_file_html,img_paths_html,captions_html);
visualize.hist(record_wrong,out_file_hist,bins=20,normed=True,xlabel='Rank of Incorrect Class',ylabel='Frequency',title='')
pickle.dump([features_test,features_train,labels_test,labels_train,img_paths_test,img_paths_train,indices,distances],open(out_file_pickle,'wb'));
def script_saveBestImage():
dir_overlaps = '/disk3/maheen_data/headC_160_noFlow_bbox/mat_overlaps_no_neg_1000';
out_dir='/disk3/maheen_data/debugging_score_and_scale';
util.mkdir(out_dir);
num_to_pick=10;
mat_overlaps = util.getFilesInFolder(dir_overlaps,'.npz');
args=[];
for idx_mat_overlap_file,mat_overlap_file in enumerate(mat_overlaps):
args.append((mat_overlap_file,num_to_pick,idx_mat_overlap_file));
p = multiprocessing.Pool(multiprocessing.cpu_count());
pred_scores_all = p.map(loadAndPickN,args);
pred_scores_all = np.vstack(pred_scores_all);
best_image_all = pred_scores_all[:,1];
out_hist = os.path.join(out_dir,'best_image_hist.png');
visualize.hist(best_image_all,out_hist,bins=7,normed=True,xlabel='Value',ylabel='Frequency',title='Best Image Hist');
print out_hist.replace('/disk3','vision3.cs.ucdavis.edu:1001');
def script_visualizeRankDifferenceAsHist(params):
out_file_pre = params.out_file_pre
out_file_html = params.out_file_html
rel_path = params.rel_path
class_ids = params.class_id
layers = params.layer
if not hasattr(class_ids, '__iter__'):
class_ids = [class_ids]
if not hasattr(class_ids, '__iter__'):
layers = [layers]
img_paths_html = []
captions = []
for class_id in class_ids:
img_paths_html_row = []
captions_row = []
for layer in layers:
params = params._replace(class_id=class_id)
params = params._replace(layer=layer)
output = getNNRankComparisonInfo(params)
indices_difference = experiments_super.getDifferenceInRank(
output['img_paths_nn_train'], output['img_paths_nn_no_train'],
output['nn_rank_train'], output['nn_rank_no_train'])
xlabel = 'Difference in Rank (Train - Untrained)'
ylabel = 'Frequency'
title = class_id + ' ' + layer
indices_difference = [
diff_curr for diffs_curr in indices_difference
for diff_curr in diffs_curr
]
if len(indices_difference) == 0:
continue
out_file_im = out_file_pre + '_' + str(params.angle) + '_' + str(
params.diff) + '_' + str(
params.delta) + '_' + class_id + '_' + layer + '.png'
img_paths_html_row.append(
out_file_im.replace(rel_path[0], rel_path[1]))
total = len(indices_difference)
sum_less = sum(np.array(indices_difference) < 0) / float(total)
sum_less = '%0.2f' % (sum_less, )
sum_more = sum(np.array(indices_difference) >= 0) / float(total)
sum_more = '%0.2f' % (sum_more, )
captions_row.append('Total ' + str(total) + ', <0: ' + sum_less +
', >0: ' + sum_more)
visualize.hist(indices_difference,
out_file=out_file_im,
bins=params.bins,
normed=params.normed,
xlabel=xlabel,
ylabel=ylabel,
title=title)
img_paths_html.append(img_paths_html_row)
captions.append(captions_row)
visualize.writeHTML(out_file_html, img_paths_html, captions,
params.height_width[0], params.height_width[1])
def script_visualizeRankDifferenceAsHist(params):
out_file_pre = params.out_file_pre
out_file_html = params.out_file_html
rel_path = params.rel_path
class_ids = params.class_id
layers = params.layer
if not hasattr(class_ids, "__iter__"):
class_ids = [class_ids]
if not hasattr(layers, "__iter__"):
layers = [layers]
img_paths_html = []
captions = []
for class_id in class_ids:
print class_id
img_paths_html_row = []
captions_row = []
for layer in layers:
print layer
params = params._replace(class_id=class_id)
params = params._replace(layer=layer)
output = getNNRankComparisonInfo(params)
indices_difference = experiments_super.getDifferenceInRank(
output["img_paths_nn_train"],
output["img_paths_nn_no_train"],
output["nn_rank_train"],
output["nn_rank_no_train"],
)
xlabel = "Difference in Rank (Train - Untrained)"
ylabel = "Frequency"
title = class_id + " " + layer
indices_difference = [diff_curr for diffs_curr in indices_difference for diff_curr in diffs_curr]
if len(indices_difference) == 0:
continue
out_file_im = out_file_pre + "_" + class_id + "_" + layer + ".png"
img_paths_html_row.append(out_file_im.replace(rel_path[0], rel_path[1]))
total = len(indices_difference)
sum_less = sum(np.array(indices_difference) < 0) / float(total)
sum_less = "%0.2f" % (sum_less,)
sum_more = sum(np.array(indices_difference) >= 0) / float(total)
sum_more = "%0.2f" % (sum_more,)
captions_row.append("Total " + str(total) + ", <0: " + sum_less + ", >0: " + sum_more)
visualize.hist(
indices_difference,
out_file=out_file_im,
bins=params.bins,
normed=params.normed,
xlabel=xlabel,
ylabel=ylabel,
title=title,
)
img_paths_html.append(img_paths_html_row)
captions.append(captions_row)
visualize.writeHTML(out_file_html, img_paths_html, captions, params.height_width[0], params.height_width[1])
