def master_nonc(model_dir, viz=False):
model_def = model_dir + 'validate.prototxt'
model_weights = caffe_root + 'models/finetune_flickr_tag/finetune_flickr_tag_1_iter_680000.caffemodel'
# model_weights = caffe_root + 'models/finetune_flickr_tag_2/finetune_flickr_tag_0_iter_357194.caffemodel'
# model_weights = caffe_root + 'models/finetune_flickr_tag/finetune_flickr_tag_3_iter_80000.caffemodel'
# model_weights = caffe_root + 'models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel'
net = caffe.Net(
model_def, # defines the structure of the model
model_weights, # contains the trained weights
caffe.TEST) # use test mode (e.g., don't perform dropout)
# net.step(1)
# print 'Baseline accuracy:{0:.4f}'.format(check_baseline_accuracy(solver.test_nets[0], n_samples/batch_size))
if viz:
transformer = tools.SimpleTransformer(
) # This is simply to add back the bias, re-shuffle the color channels to RGB, and so on...
image_index = 50 # First image in the batch.
plt.figure()
plt.imshow(
transformer.deprocess(
copy(net.blobs['data'].data[image_index, ...])))
gtlist = net.blobs['label'].data[image_index, ...].astype(np.int)
plt.title('GT: {}'.format(classes[np.where(gtlist[0, 0, :])]))
plt.axis('off')
plt.show()
for image_index in range(10, 15):
plt.figure()
plt.imshow(
transformer.deprocess(net.blobs['data'].data[image_index,
...]))
gtlist = net.blobs['label'].data[image_index,
...].astype(np.int)[0, 0, :]
estlist = net.blobs['fc8_flickr'].data[image_index,
...].argsort()[::-1][:5]
plt.title('GT: {} \n EST: {}'.format(classes[np.where(gtlist)],
classes[estlist]))
plt.axis('off')
plt.show()
# predict_tags(net)
fp = open('results_nonc.list', 'w')
for k in (1, 3, 5, 10):
res = check_accuracy_nonc(net, k)
fp.write('k: %d\n' % (k))
fp.write('accuracy precision recall\n')
np.savetxt(fp, res, fmt='%.6f')
fp.close()
def master(model_dir, viz=False):
model_def = model_dir + 'test.prototxt'
# model_weights = caffe_root + 'models/finetune_flickr_tag/finetune_flickr_tag_1_2_iter_40000.caffemodel'
model_weights = caffe_root + 'models/vgg_flickr_tag/vgg_flickr_tag_1_iter_340000.caffemodel'
# model_weights = caffe_root + 'models/finetune_flickr_tag_2/finetune_flickr_tag_0_iter_525999.caffemodel'
# model_weights = caffe_root + 'models/finetune_flickr_tag_2/finetune_flickr_tag_1_iter_240000.caffemodel'
# model_weights = caffe_root + 'models/finetune_flickr_tag/finetune_flickr_tag_3_iter_80000.caffemodel'
# model_weights = caffe_root + 'models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel'
net = caffe.Net(
str(model_def), # defines the structure of the model
str(model_weights), # contains the trained weights
caffe.TEST) # use test mode (e.g., don't perform dropout)
# net.step(1)
# print 'Baseline accuracy:{0:.4f}'.format(check_baseline_accuracy(solver.test_nets[0], n_samples/batch_size))
if viz:
transformer = tools.SimpleTransformer(
) # This is simply to add back the bias, re-shuffle the color channels to RGB, and so on...
image_index = 50 # First image in the batch.
plt.figure()
plt.imshow(
transformer.deprocess(
copy(net.blobs['data'].data[image_index, ...])))
gtlist = net.blobs['label'].data[image_index, ...].astype(np.int)
plt.title('GT: {}'.format(classes[np.where(gtlist[0, 0, :])]))
plt.axis('off')
plt.show()
for image_index in range(10, 15):
plt.figure()
plt.imshow(
transformer.deprocess(net.blobs['data'].data[image_index,
...]))
gtlist = net.blobs['label'].data[image_index,
...].astype(np.int)[0, 0, :]
estlist = net.blobs['fc8_flickr'].data[image_index,
...].argsort()[::-1][:5]
plt.title('GT: {} \n EST: {}'.format(classes[np.where(gtlist)],
classes[estlist]))
plt.axis('off')
plt.show()
predict_tags(net)
## This net uses a python datalayer: 'PascalMultilabelDataLayerSync',
## which is defined in './pycaffe/layers/pascal_multilabel_datalayers.py'.
## Take a look at the code. It's quite straight-forward, and gives you full control over data and labels.
## Now we can load the caffe solver as usual.
solver = caffe.SGDSolver(osp.join(workdir, 'solver.prototxt'))
solver.net.copy_from(
caffe_root +
'models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel')
solver.test_nets[0].share_with(solver.net)
solver.step(1)
## Let's check the data we have loaded.
# This is simply to add back the bias, re-shuffle the color channels to RGB, and so on...
# (this block is just showing, or checking, the ground truth of first image)
transformer = tools.SimpleTransformer()
image_index = 0 # First image in the batch.
plt.figure()
plt.imshow(
transformer.deprocess(copy(solver.net.blobs['data'].data[image_index,
...])))
gtlist = solver.net.blobs['label'].data[image_index, ...].astype(np.int)
print(solver.net.blobs['data'].data[image_index, ...])
print(solver.net.blobs['label'].data[image_index, ...])
print(gtlist)
print(np.where(gtlist))
plt.title('GT: {}'.format(classes[np.where(gtlist)]))
plt.axis('off')
data_layer_params = dict(batch_size=128,
im_shape=[227, 227],
split='val',
pascal_root=pascal_root)
f.write(
caffenet_multilabel(data_layer_params,
'PascalMultilabelDataLayerSync'))
solver = caffe.SGDSolver(osp.join(workdir, 'solver.prototxt'))
solver.net.copy_from(
caffe_root +
'models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel')
solver.test_nets[0].share_with(solver.net)
solver.step(1)
transformer = tools.SimpleTransformer(
) # This is simply to add back the bias, re-shuffle the color channels to RGB, and so on...
image_index = 0 # First image in the batch.
plt.figure()
plt.imshow(
transformer.deprocess(copy(solver.net.blobs['data'].data[image_index,
...])))
gtlist = solver.net.blobs['label'].data[image_index, ...].astype(np.int)
plt.title('GT: {}'.format(classes[np.where(gtlist)]))
plt.axis('off')
def hamming_distance(gt, est):
return sum([1 for (g, e) in zip(gt, est) if g == e]) / float(len(gt))
def check_accuracy(net, num_batches, batch_size=128):
