Skip to content

S
seminar-breakout
Commit 0ef50616 authored by Yuxin Wu's avatar Yuxin Wu
Browse files
Options

fix fasterrcnn linting

parent 23446308
Show whitespace changes
Inline Side-by-side
Showing with 87 additions and 78 deletions
examples/FasterRCNN/basemodel.py
View file @ 0ef50616
......@@ -27,18 +27,20 @@ def image_preprocess(image, bgr=True):
image = (image - image_mean) / image_std
return image
def get_bn(zero_init=False):
if zero_init:
return lambda x, name: BatchNorm('bn', x, gamma_init=tf.zeros_initializer())
else:
return lambda x, name: BatchNorm('bn', x)
def resnet_shortcut(l, n_out, stride, nl=tf.identity):
data_format = get_arg_scope()['Conv2D']['data_format']
n_in = l.get_shape().as_list()[1 if data_format == 'NCHW' else 3]
if n_in != n_out: # change dimension when channel is not the same
if stride == 2 and 'group3' not in tf.get_variable_scope().name:
l = l[:,:,:-1,:-1]
l = l[:, :, :-1, :-1]
return Conv2D('convshortcut', l, n_out, 1,
stride=stride, padding='VALID', nl=nl)
else:
......@@ -52,7 +54,7 @@ def resnet_bottleneck(l, ch_out, stride):
l, shortcut = l, l
l = Conv2D('conv1', l, ch_out, 1, nl=BNReLU)
if stride == 2 and 'group3' not in tf.get_variable_scope().name:
l = tf.pad(l, [[0,0],[0,0],[0,1],[0,1]])
l = tf.pad(l, [[0, 0], [0, 0], [0, 1], [0, 1]])
l = Conv2D('conv2', l, ch_out, 3, stride=2, nl=BNReLU, padding='VALID')
else:
l = Conv2D('conv2', l, ch_out, 3, stride=stride, nl=BNReLU)
......@@ -70,14 +72,15 @@ def resnet_group(l, name, block_func, features, count, stride):
l = tf.nn.relu(l)
return l
def pretrained_resnet_conv4(image, num_blocks):
assert len(num_blocks) == 3
with argscope([Conv2D, MaxPooling, BatchNorm], data_format='NCHW'), \
argscope(Conv2D, nl=tf.identity, use_bias=False), \
argscope(BatchNorm, use_local_stat=False):
l = tf.pad(image, [[0,0],[0,0],[2,3],[2,3]])
l = tf.pad(image, [[0, 0], [0, 0], [2, 3], [2, 3]])
l = Conv2D('conv0', l, 64, 7, stride=2, nl=BNReLU, padding='VALID')
l = tf.pad(l, [[0,0],[0,0],[0,1],[0,1]])
l = tf.pad(l, [[0, 0], [0, 0], [0, 1], [0, 1]])
l = MaxPooling('pool0', l, shape=3, stride=2, padding='VALID')
l = resnet_group(l, 'group0', resnet_bottleneck, 64, num_blocks[0], 1)
# TODO replace var by const to enable folding
......
examples/FasterRCNN/coco.py
View file @ 0ef50616
......@@ -18,6 +18,7 @@ __all__ = ['COCODetection', 'COCOMeta']
COCO_NUM_CATEGORY = 80
class _COCOMeta(object):
INSTANCE_TO_BASEDIR = {
'train2014': 'train2014',
......@@ -46,8 +47,10 @@ class _COCOMeta(object):
self.class_id_to_category_id = {
v: k for k, v in self.category_id_to_class_id.items()}
COCOMeta = _COCOMeta()
class COCODetection(object):
def __init__(self, basedir, name):
assert name in COCOMeta.INSTANCE_TO_BASEDIR.keys(), name
......@@ -172,4 +175,4 @@ if __name__ == '__main__':
c = COCODetection('train')
gt_boxes = c.load()
print("#Images:", len(gt_boxes))
c.print_class_histogram(bb)
c.print_class_histogram(gt_boxes)
examples/FasterRCNN/common.py
View file @ 0ef50616
......@@ -10,6 +10,7 @@ from tensorpack.utils import logger
import config
class DataFromListOfDict(RNGDataFlow):
def __init__(self, lst, keys, shuffle=False):
self._lst = lst
......@@ -66,10 +67,11 @@ def box_to_point8(boxes):
Returns:
(nx4)x2
"""
b = boxes[:,[0,1,2,3,0,3,2,1]]
b = boxes[:, [0, 1, 2, 3, 0, 3, 2, 1]]
b = b.reshape((-1, 2))
return b
def point8_to_box(points):
"""
Args:
......@@ -78,8 +80,8 @@ def point8_to_box(points):
nx4 boxes (x1y1x2y2)
"""
p = points.reshape((-1, 4, 2))
minxy = p.min(axis=1) #nx2
maxxy = p.max(axis=1) #nx2
minxy = p.min(axis=1) # nx2
maxxy = p.max(axis=1) # nx2
return np.concatenate((minxy, maxxy), axis=1)
......@@ -90,9 +92,9 @@ def clip_boxes(boxes, shape):
shape: h, w
"""
h, w = shape
boxes[:,[0,1]] = np.maximum(boxes[:,[0,1]], 0)
boxes[:,2] = np.minimum(boxes[:,2], w)
boxes[:,3] = np.minimum(boxes[:,3], h)
boxes[:, [0, 1]] = np.maximum(boxes[:, [0, 1]], 0)
boxes[:, 2] = np.minimum(boxes[:, 2], w)
boxes[:, 3] = np.minimum(boxes[:, 3], h)
return boxes
......
examples/FasterRCNN/config.py
View file @ 0ef50616
......@@ -43,8 +43,8 @@ FASTRCNN_FG_THRESH = 0.5
FASTRCNN_FG_RATIO = (0.1, 0.25)
# testing -----------------------
TEST_PRE_NMS_TOPK= 6000
TEST_POST_NMS_TOPK= 1000
TEST_PRE_NMS_TOPK = 6000
TEST_POST_NMS_TOPK = 1000
FASTRCNN_NMS_THRESH = 0.5
RESULT_SCORE_THRESH = 0.05
RESULTS_PER_IM = 100
examples/FasterRCNN/data.py
View file @ 0ef50616
......@@ -23,9 +23,11 @@ from common import (
box_to_point8, point8_to_box)
import config
class MalformedData(BaseException):
pass
@memoized
def get_all_anchors():
"""
......@@ -61,7 +63,7 @@ def get_all_anchors():
# FSxFSxAx4
assert np.all(field_of_anchors == field_of_anchors.astype('int32'))
field_of_anchors = field_of_anchors.astype('float32')
field_of_anchors[:,:,:,[2,3]] += 1
field_of_anchors[:, :, :, [2, 3]] += 1
return field_of_anchors
......@@ -131,10 +133,11 @@ def get_anchor_labels(anchors, gt_boxes, crowd_boxes):
# Set anchor boxes: the best gt_box for each fg anchor
anchor_boxes = np.zeros((NA, 4), dtype='float32')
fg_boxes = gt_boxes[ious_argmax_per_anchor[fg_inds],:]
fg_boxes = gt_boxes[ious_argmax_per_anchor[fg_inds], :]
anchor_boxes[fg_inds, :] = fg_boxes
return anchor_labels, anchor_boxes
def get_rpn_anchor_input(im, boxes, klass, is_crowd):
"""
Args:
......@@ -157,21 +160,21 @@ def get_rpn_anchor_input(im, boxes, klass, is_crowd):
def filter_box_inside(im, boxes):
h, w = im.shape[:2]
indices = np.where(
(boxes[:,0] >= 0) &
(boxes[:,1] >= 0) &
(boxes[:,2] <= w) &
(boxes[:,3] <= h))[0]
(boxes[:, 0] >= 0) &
(boxes[:, 1] >= 0) &
(boxes[:, 2] <= w) &
(boxes[:, 3] <= h))[0]
return indices
crowd_boxes = boxes[is_crowd == 1]
non_crowd_boxes = boxes[is_crowd == 0]
# fHxfWxAx4
featuremap_anchors = ALL_ANCHORS[:featureH,:featureW,:,:]
featuremap_anchors = ALL_ANCHORS[:featureH, :featureW, :, :]
featuremap_anchors_flatten = featuremap_anchors.reshape((-1, 4))
# only use anchors inside the image
inside_ind = filter_box_inside(im, featuremap_anchors_flatten)
inside_anchors = featuremap_anchors_flatten[inside_ind,:]
inside_anchors = featuremap_anchors_flatten[inside_ind, :]
anchor_labels, anchor_boxes = get_anchor_labels(inside_anchors, non_crowd_boxes, crowd_boxes)
......@@ -203,6 +206,7 @@ def read_and_augment_images(ds):
augs = [CustomResize(config.SHORT_EDGE_SIZE, config.MAX_SIZE),
imgaug.Flip(horiz=True)]
ds = AugmentImageComponents(ds, augs, index=(0,), coords_index=(1,))
def unmapf(points):
boxes = point8_to_box(points)
return boxes
......@@ -241,10 +245,12 @@ def get_train_dataflow():
ds = MapData(ds, add_anchor_to_dp)
return ds
def get_eval_dataflow():
imgs = COCODetection.load_many(config.BASEDIR, config.VAL_DATASET, add_gt=False)
# no filter for training
ds = DataFromListOfDict(imgs, ['file_name', 'id'])
def f(fname):
im = cv2.imread(fname, cv2.IMREAD_COLOR)
assert im is not None, fname
......@@ -252,8 +258,8 @@ def get_eval_dataflow():
ds = MapDataComponent(ds, f, 0)
return ds
if __name__ == '__main__':
#logger.setLevel(logging.DEBUG)
from tensorpack.dataflow import PrintData
ds = get_train_dataflow('/datasets01/COCO/060817')
ds = PrintData(ds, 100)
......@@ -261,6 +267,3 @@ if __name__ == '__main__':
ds.reset_state()
for k in ds.get_data():
pass
#import IPython as IP; IP.embed()
examples/FasterRCNN/eval.py
View file @ 0ef50616
......@@ -25,6 +25,7 @@ DetectionResult = namedtuple(
'DetectionResult',
['class_id', 'boxes', 'scores'])
@memoized
def get_tf_nms():
"""
......@@ -59,8 +60,8 @@ def nms_fastrcnn_results(boxes, probs):
if ids.size == 0:
continue
probs_k = probs[ids, klass].flatten()
boxes_k = boxes[ids,:]
selected_ids = nms_func(boxes_k[:,[1,0,3,2]], probs_k)
boxes_k = boxes[ids, :]
selected_ids = nms_func(boxes_k[:, [1, 0, 3, 2]], probs_k)
selected_boxes = boxes_k[selected_ids, :].copy()
ret.append(DetectionResult(klass, selected_boxes, probs_k[selected_ids]))
......@@ -73,7 +74,7 @@ def nms_fastrcnn_results(boxes, probs):
keep_ids = np.where(scores >= score_thresh)[0]
if len(keep_ids):
newret.append(DetectionResult(
klass, boxes[keep_ids,:], scores[keep_ids]))
klass, boxes[keep_ids, :], scores[keep_ids]))
ret = newret
return ret
......@@ -115,8 +116,8 @@ def eval_on_dataflow(df, detect_func):
results = detect_func(img)
for classid, boxes, scores in results:
cat_id = COCOMeta.class_id_to_category_id[classid]
boxes[:,2] -= boxes[:,0]
boxes[:,3] -= boxes[:,1]
boxes[:, 2] -= boxes[:, 0]
boxes[:, 3] -= boxes[:, 1]
for box, score in zip(boxes, scores):
all_results.append({
'image_id': img_id,
......@@ -142,9 +143,3 @@ def print_evaluation_scores(json_file):
cocoEval.evaluate()
cocoEval.accumulate()
cocoEval.summarize()
if __name__ == '__main__':
ds = get_eval_dataflow('/home/yuxinwu/data/COCO/')
print("Size: ", ds.size())
TestDataSpeed(ds, 1000).start()
examples/FasterRCNN/model.py
View file @ 0ef50616
......@@ -13,6 +13,7 @@ from tensorpack.models import Conv2D, FullyConnected
from utils.box_ops import pairwise_iou
import config
def rpn_head(featuremap):
with tf.variable_scope('rpn'), \
argscope(Conv2D, data_format='NCHW',
......@@ -116,6 +117,7 @@ def decode_bbox_target(box_predictions, anchors):
out = tf.squeeze(tf.concat([x1y1, x2y2], axis=2), axis=1, name='output')
return out
@under_name_scope()
def encode_bbox_target(boxes, anchors):
"""
......@@ -179,11 +181,10 @@ def generate_rpn_proposals(boxes, scores, img_shape):
topk_boxes_x1y1, topk_boxes_x2y2 = tf.split(topk_boxes_x1y1x2y2, 2, axis=1)
# nx1x2 each
wbhb = tf.squeeze(topk_boxes_x2y2 - topk_boxes_x1y1, axis=1)
valid = tf.reduce_all(wbhb > config.RPN_MIN_SIZE, axis=1) #n,
valid = tf.reduce_all(wbhb > config.RPN_MIN_SIZE, axis=1) # n,
topk_valid_boxes_x1y1x2y2 = tf.boolean_mask(topk_boxes_x1y1x2y2, valid)
topk_valid_scores = tf.boolean_mask(topk_scores, valid)
topk_valid_boxes_y1x1y2x2 = tf.reshape(
tf.reverse(topk_valid_boxes_x1y1x2y2, axis=[2]),
(-1, 4), name='nms_input_boxes')
......@@ -228,7 +229,7 @@ def sample_fast_rcnn_targets(boxes, gt_boxes, gt_labels):
# find best gt box for each roi box
best_iou_ind = tf.argmax(iou, axis=1) # n, each in 1~m
best_iou = tf.reduce_max(iou, axis=1) # n,
best_gt_boxes = tf.gather(gt_boxes, best_iou_ind) #nx4
best_gt_boxes = tf.gather(gt_boxes, best_iou_ind) # nx4
best_gt_labels = tf.gather(gt_labels, best_iou_ind) # n, each in 1~C
fg_mask = best_iou >= config.FASTRCNN_FG_THRESH
......@@ -255,8 +256,7 @@ def sample_fast_rcnn_targets(boxes, gt_boxes, gt_labels):
# don't have to add gt for training, but add it anyway
fg_inds = tf.reshape(tf.where(fg_mask), [-1])
fg_inds = tf.concat([fg_inds,
tf.cast(
fg_inds = tf.concat([fg_inds, tf.cast(
tf.range(tf.size(gt_labels)) + tf.shape(boxes)[0],
tf.int64)], 0)
num_fg = tf.size(fg_inds)
......@@ -265,7 +265,7 @@ def sample_fast_rcnn_targets(boxes, gt_boxes, gt_labels):
num_fg, name='num_fg')
fg_inds = tf.slice(tf.random_shuffle(fg_inds), [0], [num_fg])
bg_inds = tf.where(tf.logical_not(fg_mask))[:,0]
bg_inds = tf.where(tf.logical_not(fg_mask))[:, 0]
num_bg = tf.size(bg_inds)
num_bg = tf.minimum(config.FASTRCNN_BATCH_PER_IM - num_fg, num_bg)
num_bg = tf.minimum(
......@@ -387,6 +387,7 @@ def fastrcnn_predict_boxes(labels, box_logits):
fg_box_logits = tf.gather_nd(box_logits, tf.stop_gradient(ind_2d))
return fg_ind, fg_box_logits
@under_name_scope()
def fastrcnn_losses(labels, boxes, label_logits, box_logits):
"""
......
examples/FasterRCNN/train.py
View file @ 0ef50616
......@@ -88,7 +88,7 @@ class Model(ModelDesc):
proposal_boxes, gt_boxes, gt_labels)
boxes_on_featuremap = rcnn_sampled_boxes * (1.0 / config.ANCHOR_STRIDE)
roi_resized = roi_align(featuremap, boxes_on_featuremap, 14)
feature_fastrcnn = resnet_conv5(roi_resized) #nxc
feature_fastrcnn = resnet_conv5(roi_resized) # nxc
fastrcnn_label_logits, fastrcnn_box_logits = fastrcnn_head(feature_fastrcnn, config.NUM_CLASS)
fastrcnn_label_loss, fastrcnn_box_loss = fastrcnn_losses(
......@@ -107,7 +107,7 @@ class Model(ModelDesc):
add_moving_summary(k)
else:
roi_resized = roi_align(featuremap, proposal_boxes * (1.0 / config.ANCHOR_STRIDE), 14)
feature_fastrcnn = resnet_conv5(roi_resized) #nxc
feature_fastrcnn = resnet_conv5(roi_resized) # nxc
label_logits, fastrcnn_box_logits = fastrcnn_head(feature_fastrcnn, config.NUM_CLASS)
label_probs = tf.nn.softmax(label_logits, name='fastrcnn_all_probs') # NP,
labels = tf.argmax(label_logits, axis=1)
......@@ -119,7 +119,6 @@ class Model(ModelDesc):
decoded_boxes = decode_bbox_target(fg_box_logits, fg_boxes) # Nfx4, floatbox
decoded_boxes = tf.identity(decoded_boxes, name='fastrcnn_fg_boxes')
def _get_optimizer(self):
lr = symbf.get_scalar_var('learning_rate', 0.003, summary=True)
opt = tf.train.MomentumOptimizer(lr, 0.9)
......@@ -261,7 +260,6 @@ if __name__ == '__main__':
predict(args.load, args.predict)
sys.exit()
logger.set_logger_dir(args.logdir, 'd')
stepnum = 300
warmup_epoch = max(math.ceil(500.0 / stepnum), 5)
......@@ -271,17 +269,19 @@ if __name__ == '__main__':
callbacks=[
PeriodicTrigger(ModelSaver(), every_k_epochs=5),
# linear warmup
ScheduledHyperParamSetter('learning_rate',
ScheduledHyperParamSetter(
'learning_rate',
[(0, 0.003), (warmup_epoch, 0.01)], interp='linear'),
# step decay
ScheduledHyperParamSetter('learning_rate',
[(warmup_epoch, 0.01), ((120000//stepnum) + warmup_epoch, 1e-3), (180000//stepnum, 1e-4)]),
ScheduledHyperParamSetter(
'learning_rate',
[(warmup_epoch, 0.01), (120000 // stepnum, 1e-3), (180000 // stepnum, 1e-4)]),
HumanHyperParamSetter('learning_rate'),
EvalCallback(),
GPUUtilizationTracker(),
],
steps_per_epoch=stepnum,
max_epoch=205000//stepnum,
max_epoch=205000 // stepnum,
session_init=get_model_loader(args.load),
nr_tower=nr_gpu
)
......
examples/FasterRCNN/viz.py
View file @ 0ef50616
......@@ -10,6 +10,7 @@ from tensorpack.utils import viz
from coco import COCOMeta
from utils.box_ops import get_iou_callable
def draw_annotation(img, boxes, klass, is_crowd=None):
labels = []
assert len(boxes) == len(klass)
......@@ -36,12 +37,12 @@ def draw_proposal_recall(img, proposals, proposal_scores, gt_boxes):
gt_boxes: NG
"""
bbox_iou_float = get_iou_callable()
box_ious = bbox_iou_float(gt_boxes, proposals) #ng x np
box_ious = bbox_iou_float(gt_boxes, proposals) # ng x np
box_ious_argsort = np.argsort(-box_ious, axis=1)
good_proposals_ind = box_ious_argsort[:,:3] # for each gt, find 3 best proposals
good_proposals_ind = box_ious_argsort[:, :3] # for each gt, find 3 best proposals
good_proposals_ind = np.unique(good_proposals_ind.ravel())
proposals = proposals[good_proposals_ind,:]
proposals = proposals[good_proposals_ind, :]
tags = list(map(str, proposal_scores[good_proposals_ind]))
img = viz.draw_boxes(img, proposals, tags)
return img, good_proposals_ind
......
examples/tox.ini
View file @ 0ef50616
[flake8]
max-line-length = 120
ignore = F403,F401,F405,F841,E401
exclude = private
exclude = private,
FasterRCNN/utils
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment