def get_crop_size(smalldataset_id, crop_metric):
imagenetval_id = settings.convert_id_small_to_imagenetval(smalldataset_id)
image_tag = settings.get_ind_name(imagenetval_id)
image_filename = PATH_TO_DATA + settings.folder_name('img') + image_tag + '.JPEG'
im = Image.open(image_filename)
width, height = im.size
# print('IMAGE SHAPE:', width, height)
crop_type = 'proportional' if crop_metric <= 1. else 'constant'
crop_size = c_m_p.get_crop_size(height, crop_metric, crop_type) if height <= width else c_m_p.get_crop_size(width, crop_metric, crop_type)
return crop_size
def minimal_image_distribution(num_imgs, crop_metric, model_name, image_scale, strictness):
resize_dim = 150
minimal_image_aggregation = np.zeros((resize_dim, resize_dim))
# img_ids = random.sample(range(100), num_imgs) # for testing on my machine: only a subset of the maps. TODO remove for full job
img_ids = range(3)
for smalldataset_id in img_ids:
# get bbx dimensions
imagenetval_id = settings.convert_id_small_to_imagenetval(smalldataset_id)
image_tag = settings.get_ind_name(imagenetval_id)
with open(BBX_FILE, 'r') as bbx_file:
all_bbxs = json.load(bbx_file)
crop_dims = [bbx[0] for bbx in all_bbxs[image_tag]] # get all x1, y1, x2, y2 crops
minimal_map_f = PATH_TO_DATA + settings.map_filename(settings.TOP5_MAPTYPE, crop_metric, model_name, image_scale, smalldataset_id)
minimal_map_f = minimal_map_f + '_' + ('l' if strictness == 'loose' else '') + 'map'
minimal_map = np.load(minimal_map_f + '.npy')
image_filename = PATH_TO_DATA + settings.folder_name('img') + image_tag + '.JPEG'
try: # for testing on my machine: if the image is not on my machine, move on. TODO remove for full job
im = Image.open(image_filename)
except OSError:
continue
width, height = im.size
crop_type = 'proportional' if crop_metric <= 1. else 'constant'
crop_size = c_m_p.get_crop_size(height, crop_metric, crop_type) if height <= width else c_m_p.get_crop_size(width, crop_metric, crop_type)
for x1, y1, x2, y2 in crop_dims:
minmap_sub = minimal_map[y1:y2 - crop_size + 1, x1:x2 - crop_size + 1]
minmap_sub = imresize(minmap_sub, (resize_dim, resize_dim))
minimal_image_aggregation += minmap_sub
vis = (minimal_image_aggregation - np.min(minimal_image_aggregation))
vis /= np.max(vis)
vis *= 255.
Image.fromarray(vis).show()
return minimal_image_aggregation
def backprop_crops(image_id, crop_metric, model_name, image_scale):
'''
note: even with the graph inefficiency, it makes more sense for this function to be one image at a time because the indices are different for each one. This would only change if I figured out a programmatic way to find the right indices.
'''
# Get indices for this map
with open(PATH_TO_DATA + settings.BACKPROP_INDICES_FILENAME,
'r') as indfile:
all_inds = json.load(indfile)
indices = all_inds[str(
(crop_metric, model_name, image_scale, image_id))]
with tf.Session() as sess:
# Set up CNN
model = settings.MODELS[model_name]
batch_size = len(indices)
imgs = tf.placeholder(tf.float32,
[batch_size, model.im_size, model.im_size, 3])
network = model(imgs, sess)
# Create backprop objects
true_labels = json.load(
open('caffe_ilsvrc12/' + settings.DATASET + '-labels.json'))
true_label = true_labels[settings.get_ind_name(image_id)]
y = tf.constant([true_label for __ in range(batch_size)])
err = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=y, logits=network.logits)
dy_dx = tf.gradients(err, network.imgs)
# Scale image, get all crops and run backprop
image = Image.open(PATH_TO_DATA + settings.folder_name('img') +
settings.get_ind_name(image_id) + '.JPEG')
width, height = image.size
image = imresize(image,
(int(width * image_scale), int(height * image_scale)))
crop_type = 'proportional' if crop_metric <= 1. else 'constant'
crop_size = get_crop_size(
height, crop_metric,
crop_type) if height <= width else get_crop_size(
width, crop_metric, crop_type)
all_crops = []
for x1, y1 in indices:
crop = image[y1:y1 + crop_size, x1:x1 + crop_size, :]
# crop = image.crop((x1, y1, x1 + crop_size, y1 + crop_size))
all_crops.append(imresize(crop, (model.im_size, model.im_size)))
# all_crops.append(imresize(image, (model.im_size, model.im_size))) # TODO remove after making sure this code is working
backprops = sess.run(
dy_dx,
feed_dict={network.imgs: model.preprocess(np.array(all_crops))})[0]
# Make backprop results visualizable
backprops = backprops - np.min(backprops, axis=(1, 2), keepdims=True)
backprops = backprops / np.max(backprops, axis=(1, 2), keepdims=True)
backprops = backprops * 255.
backprops = backprops.astype(np.uint8)
# savez results
folder = PATH_TO_DATA + settings.map_folder_name(
settings.BACKPROP_MAPTYPE, crop_metric, model_name, image_scale)
filename = PATH_TO_DATA + settings.map_filename(
settings.BACKPROP_MAPTYPE, crop_metric, model_name, image_scale,
image_id)
if not os.path.exists(folder):
os.makedirs(folder)
np.savez(filename,
**{str(indices[i]): backprops[i]
for i in range(len(indices))})
def get_max_diff_adjacent_crops(
start_id,
end_id,
crop_metric,
model_name,
image_scale,
compare_corr=True,
use_top5=True
): # eventually add step size, right now defaults to 1 in code
# compare_correctness: bool indicating whether the final crops should necessarily have different classification correctness
# use_top5: bool indicating whether using top5 correctness or top1 correctness
# returns the two crops. If compare_correctness==True and there are no correctly classified crops, returns None.
crop_type = 'proportional' if crop_metric <= 1. else 'constant'
folders = [
PATH_TO_DATA + settings.maxdiff_folder_name(
'size', crop_metric, model_name, image_scale,
'diff' if compare_corr else 'any', conf)
for conf in ['high', 'low']
]
for folder in folders:
if not os.path.exists(folder):
os.makedirs(folder)
sess = tf.Session()
model = settings.MODELS[model_name]
imgs = tf.placeholder(tf.float32, [None, model.im_size, model.im_size, 3])
network = model(imgs, sess)
image_ids = range(start_id, end_id + 1)
hc_correct = []
hc_incorrect = []
lc_correct = []
lc_incorrect = []
h_activations = {}
l_activations = {}
for image_id in image_ids:
f = open('small-dataset-to-imagenet.txt')
lines = f.readlines()
image_tag = lines[image_id].split(" ", 1)[0]
print(image_tag)
#image_tag = settings.get_ind_name(image_id)
image_filename = PATH_TO_DATA + settings.folder_name(
'img') + image_tag #+ ('.png' if image_id == 50001 else '.JPEG')
image = Image.open(image_filename)
if image.mode != 'RGB':
image = image.convert('RGB')
image = np.asarray(image)
image_label = image_tag[:-5] + '_' + str(
crop_metric) + '_' + model_name + '_'
correctness_type = 'top5' if use_top5 else 'top1'
correctness_filename = PATH_TO_DATA + (settings.map_folder_name(
settings.TOP5_MAPTYPE,
crop_metric) if use_top5 else settings.map_folder_name(
settings.TOP1_MAPTYPE, crop_metric)) + image_label + (
settings.TOP5_MAPTYPE
if use_top5 else settings.TOP1_MAPTYPE) + '.npy'
cor_map = np.load(correctness_filename)
corr_extension = '_diffcorrectness' if compare_corr else '_anycorrectness'
if compare_corr:
if not cor_map.any():
print('%s has no correctly classified crops.' % image_tag)
continue
elif cor_map.all():
print('%s has only correctly classified crops.' % image_tag)
continue
con_map_filename = PATH_TO_DATA + settings.map_folder_name(
settings.CONFIDENCE_MAPTYPE,
crop_metric) + image_label + settings.CONFIDENCE_MAPTYPE + '.npy'
con_map = np.load(con_map_filename)
down_diff = np.diff(
con_map, axis=0) # apparently assumes step_size=1 (adjacency)
up_diff = -1. * down_diff
right_diff = np.diff(con_map)
left_diff = -1. * right_diff
diffs = {
'up': up_diff,
'down': down_diff,
'left': left_diff,
'right': right_diff
}
# TESTER: CLASSIFYING CROPS TO SEE IF THERE'S A DIFFERENCE BETWEEN WHAT THE TESTER REPORTS AND WHAT THIS REPORTS
true_labels = json.load(
open('caffe_ilsvrc12/' + settings.DATASET + '-labels.json'))
while True:
maxes = {
direction: np.unravel_index(np.argmax(diffs[direction]),
diffs[direction].shape)
for direction in diffs
} # map each directional diff to its argmax (index of its maximum confidence diff)
max_dir = max(
[direction for direction in maxes],
key=lambda direction: diffs[direction][tuple(maxes[direction])]
) # get the direction of the diff whose max confidence is the highest out of the four max confidences
# depending on the max-confidence direction, get the argmax of that direction. The more confident crop will be offset by 1 in a way that depends on the direction.
if max_dir == 'up':
up_max = maxes['up']
gcell, cell = (
tuple(up_max), (up_max[0] + 1, up_max[1])
) # up (and left) are like this because when up and left diffs are made, the negation also changes the direction in which your step goes. it goes down -> up; right -> left.
elif max_dir == 'down':
down_max = maxes['down']
cell, gcell = (tuple(down_max), (down_max[0] + 1, down_max[1]))
elif max_dir == 'left':
left_max = maxes['left']
gcell, cell = (tuple(left_max), (left_max[0], left_max[1] + 1))
else:
right_max = maxes['right']
cell, gcell = (tuple(right_max), (right_max[0],
right_max[1] + 1))
diff_correctness = cor_map[cell] != cor_map[gcell]
if diff_correctness or not compare_correctness:
y, x = cell
gy, gx = gcell
height, width, channels = image.shape
crop_size = get_crop_size(
height, proportion) if height <= width else get_crop_size(
width, proportion)
dim_used = 'height' if height <= width else 'width'
cropped = image[cell[0]:cell[0] + crop_size,
cell[1]:cell[1] + crop_size]
gcropped = image[gcell[0]:gcell[0] + crop_size,
gcell[1]:gcell[1] + crop_size]
true_value = true_labels[image_tag[:-5]]
hc = imresize(gcropped, (network.im_size, network.im_size))
lc = imresize(cropped, (network.im_size, network.im_size))
hresult = sess.run(network.pull_layers,
feed_dict={network.imgs: [hc]})
hcprob = hresult[0][0]
h_activations[image_id] = hresult[1:]
result = sess.run(network.pull_layers,
feed_dict={network.imgs: [lc]})
lcprob = lresult[0][0]
l_activations[image_id] = lresult[1:]
hcpreds = (np.argsort(hcprob)[::-1])[0:5]
lcpreds = (np.argsort(lcprob)[::-1])[0:5]
if true_value in hcpreds:
hc_correct.append(image_id)
else:
hc_incorrect.append(image_id)
if true_value in lcpreds:
lc_correct.append(image_id)
else:
lc_incorrect.append(image_id)
maxdiff_folder = settings.maxdiff_folder_name(crop_metric)
np.save(
PATH_TO_DATA + maxdiff_folder + image_label +
'maxdiff_lowconf' + corr_extension, cropped)
np.save(
PATH_TO_DATA + maxdiff_folder + image_label +
'maxdiff_highconf' + corr_extension, gcropped)
break
else:
if max_dir in ['up', 'left']:
diffs[max_dir][
gcell] = -2. # for the diff where that was the argmax, mark the cell containing it to something lower than any real entry (-1. <= real entry <= 1.) This is gcell for up, left and cell for down, right because the lower-indexed cell is always the one that contained the confidence originally
else:
diffs[max_dir][cell] = -2.
print('INTERNAL TEST')
print("High confidence crop correctly classified:", hc_correct)
print('High confidence crop incorrectly classified:', hc_incorrect)
print('Low confidence crop correctly classified:', lc_correct)
print('Low confidence crop incorrectly classified:', lc_incorrect)
sess.close()
return h_activations, l_activations
def get_maxdiff_size_crops(start_id,
end_id,
crop_metric,
model_name,
image_scale,
compare_corr=True):
# For each image id in range(start_id, end_id + 1), finds the crop of size crop_metric and the crop ~2 pixels smaller that are maximally different in confidence. If compare_corr is True, it necessarily finds crops where the smaller one is classified incorrectly and the larger one is classified correctly. This uses the top5 maps for the two scales.
crop_type = 'proportional' if crop_metric <= 1. else 'constant'
folders = [
PATH_TO_DATA + settings.maxdiff_folder_name(
'size', crop_metric, model_name, image_scale,
'diff' if compare_corr else 'any', conf)
for conf in ['high', 'low']
]
for folder in folders:
if not os.path.exists(folder):
os.makedirs(folder)
map_folder = PATH_TO_DATA + settings.maxdiff_folder_name(
'size', crop_metric, model_name, image_scale, 'map')
if not os.path.exists(map_folder):
os.makedirs(map_folder)
# with tf.Session() as sess: # TODO use for testing
# model = settings.MODELS[model_name]
# imgs = tf.placeholder(tf.float32, [None, model.im_size, model.im_size, 3])
# network = model(imgs, sess)
true_labels = json.load(
open('caffe_ilsvrc12/' + settings.DATASET + '-labels.json'))
for image_id in range(start_id, end_id + 1):
image_tag = settings.get_ind_name(image_id)
image_filename = PATH_TO_DATA + settings.folder_name(
'img') + image_tag + '.JPEG'
true_class = true_labels[image_tag]
im = Image.open(image_filename)
if im.mode != 'RGB':
im = im.convert('RGB')
width, height = im.size
im = im.resize((int(width * image_scale), int(height * image_scale)))
width, height = im.size
im = np.asarray(im)
# Get the small crop_metric, crop sizes for the large and small crop_metrics
size_dim = height if height <= width else width
large_size = get_crop_size(
size_dim, crop_metric,
crop_type) if height <= width else get_crop_size(
width, crop_metric, crop_type)
small_metric = 0.194 if crop_metric == 0.2 else 0.394 # TODO change to be a calculation and command
small_size = get_crop_size(size_dim, small_metric, crop_type)
metrics = [crop_metric, small_metric]
# Get the correctness maps (top5, may become a choice between top5 and top1 in the future), and if the call requires diff correctness, check that that's possible
corr_fns = [
PATH_TO_DATA +
settings.map_filename(settings.TOP5_MAPTYPE, metric, model_name,
image_scale, image_id) + '.npy'
for metric in metrics
] # TODO change to allow choice of top1 or top5
lcor, scor = cor_maps = [np.load(corr_fn) for corr_fn in corr_fns]
if compare_corr:
for cor_map in cor_maps:
if not cor_map.any():
print('%s has no correctly classified crops.' % image_tag)
continue
elif cor_map.all():
print('%s has only correctly classified crops.' %
image_tag)
continue
# Get confidence maps
con_fns = [
PATH_TO_DATA +
settings.map_filename(settings.CONFIDENCE_MAPTYPE, metric,
model_name, image_scale, image_id) + '.npy'
for metric in metrics
]
lcon, scon = [np.load(con_fn) for con_fn in con_fns]
# Calculate difference matrices
lrows, lcols = lcon.shape
offset = large_size - small_size # get the metric that bottom and right are off by
tl_sub = scon[:lrows, :
lcols] # for top left, get the top left small crops that correspond to big crops - same shape. It's all of them because the large crops are all adjacent, even if the difference in pixels is >1.
tr_sub = scon[:lrows, offset:lcols +
offset] # for top right, everything that is 'offset' cols over
bl_sub = scon[offset:lrows + offset, :lcols]
br_sub = scon[offset:lrows + offset, offset:lcols + offset]
ctoffset = int(offset / 2)
ct_sub = scon[ctoffset:lrows + ctoffset, ctoffset:lcols + ctoffset]
diffs = {
'tl': lcon -
tl_sub, # use subtraction because we are looking for increase in conf from increase in size
'tr': lcon - tr_sub,
'bl': lcon - bl_sub,
'br': lcon - br_sub,
'ct': lcon - ct_sub
}
# Make map of the largest size change in confidence across all directions of shrinking
change_map = np.maximum.reduce(list(diffs.values()))
np.save(map_folder + str(image_id), change_map)
# Find maxdiff pair by searching for maximally different pairs until one with different correctness is found (if diffcor. Else, this will terminate after one loop as the first pair found will be the maximally different one and therefore the right one for anycor.)
while True:
maxes = {
corner: np.unravel_index(np.argmax(diffs[corner]),
diffs[corner].shape)
for corner in diffs
} # map each corner diff to its argmax (index of maximum confidence diff)
max_dir = max(
[corner for corner in maxes],
key=lambda corner: diffs[corner][tuple(maxes[corner])]
) # get the corner id of the diff whose max change is the highest out of the four max changes
# getting the indices of the maximal confidence increase. Indices are based on the size of the large crop size map. The first index is the index for the large crop, and the second is for the small crop.
corner_max = maxes[max_dir]
if max_dir == 'tl':
lcell, scell = tuple(corner_max), tuple(corner_max)
elif max_dir == 'tr':
lcell, scell = tuple(corner_max), (corner_max[0],
corner_max[1] + offset)
elif max_dir == 'bl':
lcell, scell = tuple(corner_max), (corner_max[0] + offset,
corner_max[1])
elif max_dir == 'br':
lcell, scell = tuple(corner_max), (corner_max[0] + offset,
corner_max[1] + offset)
else:
lcell, scell = tuple(corner_max), (corner_max[0] + ctoffset,
corner_max[1] + ctoffset)
diff_corr = lcor[lcell] != scor[scell]
if diff_corr or not compare_corr:
sy, sx = scell
ly, lx = lcell
lcropped = im[lcell[0]:lcell[0] + large_size,
lcell[1]:lcell[1] + large_size]
scropped = im[scell[0]:scell[0] + small_size,
scell[1]:scell[1] + small_size]
# lcropped = imresize(lcropped, (network.im_size, network.im_size))
# scropped = imresize(scropped, (network.im_size, network.im_size))
# result = sess.run(network.probs, feed_dict={network.imgs: np.array([lcropped, scropped])}) # run and see if it works later. Without this, the sess isn't actually being used - this is for internal test.
break
else: # if that location wasn't diffcorr, set the diff's entry to -2.
diffs[max_dir][lcell] = -2.
lfolder, sfolder = folders
np.save(lfolder + str(image_id), lcropped)
np.save(sfolder + str(image_id), scropped)