def experiment_nnFullImage(params):
if params.out_file_pickle is None:
out_file_pickle,out_file_text=setupFilesWithBigObjectForFeatureExtraction(params.db_path,params.class_id,params.threshold,params.out_file_pre,params.path_to_annotation,params.path_to_images)
params=params._replace(out_file_pickle=out_file_pickle)
params=params._replace(out_file_text=out_file_text)
out_file_pickle=params.out_file_pickle;
out_file_text=params.out_file_text;
print out_file_pickle
[img_paths_originals,img_paths_to_keep,class_id_idx_tuples]=pickle.load(open(out_file_pickle,'rb'));
assert len(img_paths_originals)==len(class_id_idx_tuples)
file_names_mat,object_indices=recreateOriginalPaths(params.path_to_annotation,img_paths_to_keep,returnObject_idx=True);
print 'getting azimuths'
azimuths=[getObjectStruct(file_name,object_idx).viewpoint.azimuth_coarse for file_name,object_idx in zip(file_names_mat,object_indices)]
if params.out_file_layers is None:
print 'running layers part'
out_file_layers=caffe_wrapper.saveFeaturesOfLayers(out_file_text,params.path_to_classify,params.gpu_no,params.layers,ext='jpg',out_file=params.out_file_pre,meanFile=params.caffe_mean,deployFile=params.caffe_deploy,modelFile=params.caffe_model)
params=params._replace(out_file_layers=out_file_layers)
out_file_layers=params.out_file_layers;
print 'writing to db'
for layer in params.layers:
vals=np.load(out_file_layers);
indices,distances=nearest_neighbor.doCosineDistanceNN(vals[layer],numberOfN=None);
mani=Pascal3D_Manipulator(params.db_path_out);
mani.openSession();
for idx in range(len(img_paths_originals)):
mani.insert(idx,img_paths_originals[idx],layer,out_file_layers,class_id_idx_tuples[idx][0],class_id_idx_tuples[idx][1],params.caffe_model, azimuth=azimuths[idx],neighbor_index=indices[idx],neighbor_distance=distances[idx],trainedClass=params.trainFlag)
mani.closeSession();
return params
def experiment_nnPatches(params):
out_file_text=params.out_file_text;
class_ids=params.class_id;
img_paths=util.readLinesFromFile(out_file_text)
class_id_idx_tuples=[];
for img_path in img_paths:
class_id=img_path[:img_path.rindex('_')]
class_id=class_id[class_id.rindex('_')+1:];
class_idx=class_ids.index(class_id);
class_idx=params.class_idx[class_idx];
class_id_idx_tuples.append((class_id,class_idx));
file_names_mat,object_indices=recreateOriginalPaths(params.path_to_annotation,img_paths,returnObject_idx=True);
print 'getting azimuths'
azimuths=[getObjectStruct(file_name,object_idx).viewpoint.azimuth_coarse for file_name,object_idx in zip(file_names_mat,object_indices)]
if params.out_file_layers is None:
print 'running layers part'
out_file_layers=caffe_wrapper.saveFeaturesOfLayers(out_file_text,params.path_to_classify,params.gpu_no,params.layers,ext='jpg',out_file=params.out_file_pre,meanFile=params.caffe_mean,deployFile=params.caffe_deploy,modelFile=params.caffe_model,images_dim=params.images_dim)
params=params._replace(out_file_layers=out_file_layers)
out_file_layers=params.out_file_layers;
print 'writing to db'
for layer in params.layers:
vals=np.load(out_file_layers);
indices,distances=nearest_neighbor.doCosineDistanceNN(vals[layer],numberOfN=None);
mani=Pascal3D_Manipulator(params.db_path_out);
mani.openSession();
for idx in range(len(img_paths)):
mani.insert(idx,img_paths[idx],layer,out_file_layers,class_id_idx_tuples[idx][0],class_id_idx_tuples[idx][1],params.caffe_model, azimuth=azimuths[idx],neighbor_index=indices[idx],neighbor_distance=distances[idx],trainedClass=params.trainFlag,commitFlag=False)
mani.closeSession();
return params
def experiment_nnFullImageMixWithImagenet(params):
if params.out_file_text is None:
all_files_info=preprocessImagesFromImagenet(params.imagenet_ids_to_test,params.synset_words,params.val_gt_file,path_to_images_imagenet=params.path_to_images_imagenet)
all_files_info=getImageInfoForMixTestFromPascal3d(params.db_path_in,params.class_id_pascal,all_files_info=all_files_info);
out_file_pickle=params.out_file_pre+'.p';
out_file_text=params.out_file_pre+'.txt';
pickle.dump(all_files_info,open(out_file_pickle,'wb'));
with open(out_file_text,'wb') as f:
for dict_curr in all_files_info:
f.write(dict_curr['img_path']+'\n');
params=params._replace(out_file_pickle=out_file_pickle)
params=params._replace(out_file_text=out_file_text)
out_file_text=params.out_file_text;
out_file_pickle=params.out_file_pickle;
if params.out_file_layers is None:
print 'running layers part'
out_file_layers=caffe_wrapper.saveFeaturesOfLayers(out_file_text,params.path_to_classify,params.gpu_no,params.layers,ext=params.ext,out_file=params.out_file_pre,meanFile=params.caffe_mean,deployFile=params.caffe_deploy,modelFile=params.caffe_model)
params=params._replace(out_file_layers=out_file_layers)
out_file_layers=params.out_file_layers;
all_files_info=pickle.load(open(out_file_pickle,'rb'));
print 'writing to db'
for layer in params.layers:
vals=np.load(out_file_layers);
indices,distances=nearest_neighbor.doCosineDistanceNN(vals[layer],numberOfN=None);
mani=Pascal3D_Manipulator(params.db_path_out);
mani.openSession();
# for idx in range(len(img_paths_originals)):
for idx,dict_curr in enumerate(all_files_info):
mani.insert(idx,dict_curr['img_path'],layer,out_file_layers,dict_curr['class_id'],dict_curr['class_idx'],params.caffe_model, azimuth=dict_curr['azimuth'],neighbor_index=indices[idx],neighbor_distance=distances[idx],trainedClass=params.trainFlag)
mani.closeSession();
return params;
def experiment_nnFullImage(params):
if params.out_file_pickle is None:
out_file_pickle, out_file_text = setupFilesWithBigObjectForFeatureExtraction(
params.db_path, params.class_id, params.threshold,
params.out_file_pre, params.path_to_annotation,
params.path_to_images)
params = params._replace(out_file_pickle=out_file_pickle)
params = params._replace(out_file_text=out_file_text)
out_file_pickle = params.out_file_pickle
out_file_text = params.out_file_text
print out_file_pickle
[img_paths_originals, img_paths_to_keep,
class_id_idx_tuples] = pickle.load(open(out_file_pickle, 'rb'))
assert len(img_paths_originals) == len(class_id_idx_tuples)
file_names_mat, object_indices = recreateOriginalPaths(
params.path_to_annotation, img_paths_to_keep, returnObject_idx=True)
print 'getting azimuths'
azimuths = [
getObjectStruct(file_name, object_idx).viewpoint.azimuth_coarse
for file_name, object_idx in zip(file_names_mat, object_indices)
]
if params.out_file_layers is None:
print 'running layers part'
out_file_layers = caffe_wrapper.saveFeaturesOfLayers(
out_file_text,
params.path_to_classify,
params.gpu_no,
params.layers,
ext='jpg',
out_file=params.out_file_pre,
meanFile=params.caffe_mean,
deployFile=params.caffe_deploy,
modelFile=params.caffe_model)
params = params._replace(out_file_layers=out_file_layers)
out_file_layers = params.out_file_layers
print 'writing to db'
for layer in params.layers:
vals = np.load(out_file_layers)
indices, distances = nearest_neighbor.doCosineDistanceNN(
vals[layer], numberOfN=None)
mani = Pascal3D_Manipulator(params.db_path_out)
mani.openSession()
for idx in range(len(img_paths_originals)):
mani.insert(idx,
img_paths_originals[idx],
layer,
out_file_layers,
class_id_idx_tuples[idx][0],
class_id_idx_tuples[idx][1],
params.caffe_model,
azimuth=azimuths[idx],
neighbor_index=indices[idx],
neighbor_distance=distances[idx],
trainedClass=params.trainFlag)
mani.closeSession()
return params
def experiment_nnPatches(params):
out_file_text = params.out_file_text
class_ids = params.class_id
img_paths = util.readLinesFromFile(out_file_text)
class_id_idx_tuples = []
for img_path in img_paths:
class_id = img_path[:img_path.rindex('_')]
class_id = class_id[class_id.rindex('_') + 1:]
class_idx = class_ids.index(class_id)
class_idx = params.class_idx[class_idx]
class_id_idx_tuples.append((class_id, class_idx))
file_names_mat, object_indices = recreateOriginalPaths(
params.path_to_annotation, img_paths, returnObject_idx=True)
print 'getting azimuths'
azimuths = [
getObjectStruct(file_name, object_idx).viewpoint.azimuth_coarse
for file_name, object_idx in zip(file_names_mat, object_indices)
]
if params.out_file_layers is None:
print 'running layers part'
out_file_layers = caffe_wrapper.saveFeaturesOfLayers(
out_file_text,
params.path_to_classify,
params.gpu_no,
params.layers,
ext='jpg',
out_file=params.out_file_pre,
meanFile=params.caffe_mean,
deployFile=params.caffe_deploy,
modelFile=params.caffe_model,
images_dim=params.images_dim)
params = params._replace(out_file_layers=out_file_layers)
out_file_layers = params.out_file_layers
print 'writing to db'
for layer in params.layers:
vals = np.load(out_file_layers)
indices, distances = nearest_neighbor.doCosineDistanceNN(
vals[layer], numberOfN=None)
mani = Pascal3D_Manipulator(params.db_path_out)
mani.openSession()
for idx in range(len(img_paths)):
mani.insert(idx,
img_paths[idx],
layer,
out_file_layers,
class_id_idx_tuples[idx][0],
class_id_idx_tuples[idx][1],
params.caffe_model,
azimuth=azimuths[idx],
neighbor_index=indices[idx],
neighbor_distance=distances[idx],
trainedClass=params.trainFlag,
commitFlag=False)
mani.closeSession()
return params
def experiment_nnFullImageMixWithImagenet(params):
if params.out_file_text is None:
all_files_info = preprocessImagesFromImagenet(
params.imagenet_ids_to_test,
params.synset_words,
params.val_gt_file,
path_to_images_imagenet=params.path_to_images_imagenet)
all_files_info = getImageInfoForMixTestFromPascal3d(
params.db_path_in,
params.class_id_pascal,
all_files_info=all_files_info)
out_file_pickle = params.out_file_pre + '.p'
out_file_text = params.out_file_pre + '.txt'
pickle.dump(all_files_info, open(out_file_pickle, 'wb'))
with open(out_file_text, 'wb') as f:
for dict_curr in all_files_info:
f.write(dict_curr['img_path'] + '\n')
params = params._replace(out_file_pickle=out_file_pickle)
params = params._replace(out_file_text=out_file_text)
out_file_text = params.out_file_text
out_file_pickle = params.out_file_pickle
if params.out_file_layers is None:
print 'running layers part'
out_file_layers = caffe_wrapper.saveFeaturesOfLayers(
out_file_text,
params.path_to_classify,
params.gpu_no,
params.layers,
ext=params.ext,
out_file=params.out_file_pre,
meanFile=params.caffe_mean,
deployFile=params.caffe_deploy,
modelFile=params.caffe_model)
params = params._replace(out_file_layers=out_file_layers)
out_file_layers = params.out_file_layers
all_files_info = pickle.load(open(out_file_pickle, 'rb'))
print 'writing to db'
for layer in params.layers:
vals = np.load(out_file_layers)
indices, distances = nearest_neighbor.doCosineDistanceNN(
vals[layer], numberOfN=None)
mani = Pascal3D_Manipulator(params.db_path_out)
mani.openSession()
# for idx in range(len(img_paths_originals)):
for idx, dict_curr in enumerate(all_files_info):
mani.insert(idx,
dict_curr['img_path'],
layer,
out_file_layers,
dict_curr['class_id'],
dict_curr['class_idx'],
params.caffe_model,
azimuth=dict_curr['azimuth'],
neighbor_index=indices[idx],
neighbor_distance=distances[idx],
trainedClass=params.trainFlag)
mani.closeSession()
return params
def script_temp():
path_to_val = "/disk2/imagenet/val"
ext = "JPEG"
out_dir = "/disk2/novemberExperiments/nn_imagenet_try"
if not os.path.exists(out_dir):
os.mkdir(out_dir)
in_file_pre = "list_of_ims_for_nn"
in_file_pre = os.path.join(out_dir, in_file_pre)
path_to_classify = ".."
trainFlag = False
# caffe_model='/home/maheenrashid/Downloads/caffe/caffe-rc2/models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel';
caffe_model = "/disk2/octoberExperiments/nn_performance_without_pascal/snapshot_iter_450000.caffemodel"
caffe_deploy = "/disk2/octoberExperiments/nn_performance_without_pascal/deploy.prototxt"
caffe_mean = "/disk2/octoberExperiments/nn_performance_without_pascal/mean.npy"
gpu_no = 0
layers = ["pool5", "fc6", "fc7"]
out_file = "nn_non_trained"
out_file = os.path.join(out_dir, out_file)
db_path_out = "sqlite://///disk2/novemberExperiments/nn_imagenet/nn_imagenet.db"
synset_words = "../../data/ilsvrc12/synset_words.txt"
val_gt_file = "../../data/ilsvrc12/val.txt"
idx_chosen = pickle.load(open("/disk2/novemberExperiments/nn_imagenet/equal_mix_ids.p", "rb"))
im_files_gt_classes = imagenet.selectTestSetByID(val_gt_file, idx_chosen, path_to_val=path_to_val)
im_files = list(zip(*im_files_gt_classes)[0])
gt_classes = list(zip(*im_files_gt_classes)[1])
print len(im_files)
print len(gt_classes)
print len(set(gt_classes))
per_file = len(im_files)
# in_files,_=writeInputImageFiles(im_files,in_file_pre,per_file);
in_files = [in_file_pre + "_" + str(0) + ".txt"]
print in_files
out_files = []
for idx, in_file_curr in enumerate(in_files):
out_file_curr = out_file + "_" + str(idx)
out_files.append(
caffe_wrapper.saveFeaturesOfLayers(
in_file_curr,
path_to_classify,
gpu_no,
layers,
ext=ext,
out_file=out_file_curr,
meanFile=caffe_mean,
deployFile=caffe_deploy,
modelFile=caffe_model,
)
)
print in_files
print out_files
file_list_all = []
for in_file_curr in in_files:
with open(in_file_curr, "rb") as f:
file_list = f.readlines()
file_list_all.extend([file_curr.strip("\n") for file_curr in file_list])
print len(file_list_all)
imagenet_idx_mapped, imagenet_ids_mapped, imagenet_labels_mapped = imagenet.getMappingInfo(
file_list_all, synset_words, val_gt_file
)
print "about to combine"
t = time.time()
val_combined = combineDeepFeaturesFromFiles(out_files, layers)
print time.time() - t
for layer_curr in layers:
print "about to nn for ", layer_curr
t = time.time()
indices, distances = nearest_neighbor.doCosineDistanceNN(val_combined[layer_curr], numberOfN=None)
print time.time() - t
# break;
# return
print indices.shape
print distances.shape
print "writing to db"
mani = Imagenet_Manipulator(db_path_out)
mani.openSession()
for idx in range(len(file_list_all)):
if idx % 100 == 0:
print layer_curr, idx, len(file_list_all)
idx_out_file = idx / per_file
out_file_layers = out_file + "_" + str(idx_out_file) + ".npz"
mani.insert(
idx,
file_list_all[idx],
layer_curr,
out_file_layers,
trainFlag,
imagenet_idx_mapped[idx],
imagenet_ids_mapped[idx],
caffe_model,
class_label_imagenet=imagenet_labels_mapped[idx],
neighbor_index=indices[idx],
neighbor_distance=distances[idx],
)
mani.closeSession()