前言
上午写了anchor构造部分，上了一节课，下午继续肝，现在就是SSD的网络构造部分了，如果想知道SSD的模型结构可以看我写的原理里面有结构图，这里就不复制过来了。额算了还是整过来吧：
ssd_model_fn（）
def ssd_model_fn(features, labels, mode, params):
"""model_fn for SSD to be used with our Estimator."""
shape = labels['shape']
loc_targets = labels['loc_targets']
cls_targets = labels['cls_targets']
match_scores = labels['match_scores']
global global_anchor_info
decode_fn = global_anchor_info['decode_fn']
num_anchors_per_layer = global_anchor_info['num_anchors_per_layer']
all_num_anchors_depth = global_anchor_info['all_num_anchors_depth']
# bboxes_pred = decode_fn(loc_targets[0])
# bboxes_pred = [tf.reshape(preds, [-1, 4]) for preds in bboxes_pred]
# bboxes_pred = tf.concat(bboxes_pred, axis=0)
# save_image_op = tf.py_func(save_image_with_bbox,
# [ssd_preprocessing.unwhiten_image(features[0]),
# tf.clip_by_value(cls_targets[0], 0, tf.int64.max),
# match_scores[0],
# bboxes_pred],
# tf.int64, stateful=True)
# with tf.control_dependencies([save_image_op]):
#print(all_num_anchors_depth)
with tf.variable_scope(params['model_scope'], default_name=None, values=[features], reuse=tf.AUTO_REUSE):
backbone = ssd_net.VGG16Backbone(params['data_format'])
feature_layers = backbone.forward(features, training=(mode == tf.estimator.ModeKeys.TRAIN))
#print(feature_layers)
location_pred, cls_pred = ssd_net.multibox_head(feature_layers, params['num_classes'], all_num_anchors_depth, data_format=params['data_format'])
if params['data_format'] == 'channels_first':
cls_pred = [tf.transpose(pred, [0, 2, 3, 1]) for pred in cls_pred]
location_pred = [tf.transpose(pred, [0, 2, 3, 1]) for pred in location_pred]
cls_pred = [tf.reshape(pred, [tf.shape(features)[0], -1, params['num_classes']]) for pred in cls_pred]
location_pred = [tf.reshape(pred, [tf.shape(features)[0], -1, 4]) for pred in location_pred]
cls_pred = tf.concat(cls_pred, axis=1)
location_pred = tf.concat(location_pred, axis=1)
cls_pred = tf.reshape(cls_pred, [-1, params['num_classes']])
location_pred = tf.reshape(location_pred, [-1, 4])
with tf.device('/cpu:0'):
with tf.control_dependencies([cls_pred, location_pred]):
with tf.name_scope('post_forward'):
#bboxes_pred = decode_fn(location_pred)
bboxes_pred = tf.map_fn(lambda _preds : decode_fn(_preds),
tf.reshape(location_pred, [tf.shape(features)[0], -1, 4]),
dtype=[tf.float32] * len(num_anchors_per_layer), back_prop=False)
#cls_targets = tf.Print(cls_targets, [tf.shape(bboxes_pred[0]),tf.shape(bboxes_pred[1]),tf.shape(bboxes_pred[2]),tf.shape(bboxes_pred[3])])
bboxes_pred = [tf.reshape(preds, [-1, 4]) for preds in bboxes_pred]
bboxes_pred = tf.concat(bboxes_pred, axis=0)
flaten_cls_targets = tf.reshape(cls_targets, [-1])
flaten_match_scores = tf.reshape(match_scores, [-1])
flaten_loc_targets = tf.reshape(loc_targets, [-1, 4])
# each positive examples has one label
positive_mask = flaten_cls_targets > 0
n_positives = tf.count_nonzero(positive_mask)
batch_n_positives = tf.count_nonzero(cls_targets, -1)
batch_negtive_mask = tf.equal(cls_targets, 0)#tf.logical_and(tf.equal(cls_targets, 0), match_scores > 0.)
batch_n_negtives = tf.count_nonzero(batch_negtive_mask, -1)
batch_n_neg_select = tf.cast(params['negative_ratio'] * tf.cast(batch_n_positives, tf.float32), tf.int32)
batch_n_neg_select = tf.minimum(batch_n_neg_select, tf.cast(batch_n_negtives, tf.int32))
# hard negative mining for classification
predictions_for_bg = tf.nn.softmax(tf.reshape(cls_pred, [tf.shape(features)[0], -1, params['num_classes']]))[:, :, 0]
prob_for_negtives = tf.where(batch_negtive_mask,
0. - predictions_for_bg,
# ignore all the positives
0. - tf.ones_like(predictions_for_bg))
topk_prob_for_bg, _ = tf.nn.top_k(prob_for_negtives, k=tf.shape(prob_for_negtives)[1])
score_at_k = tf.gather_nd(topk_prob_for_bg, tf.stack([tf.range(tf.shape(features)[0]), batch_n_neg_select - 1], axis=-1))
selected_neg_mask = prob_for_negtives >= tf.expand_dims(score_at_k, axis=-1)
# include both selected negtive and all positive examples
final_mask = tf.stop_gradient(tf.logical_or(tf.reshape(tf.logical_and(batch_negtive_mask, selected_neg_mask), [-1]), positive_mask))
total_examples = tf.count_nonzero(final_mask)
cls_pred = tf.boolean_mask(cls_pred, final_mask)
location_pred = tf.boolean_mask(location_pred, tf.stop_gradient(positive_mask))
flaten_cls_targets = tf.boolean_mask(tf.clip_by_value(flaten_cls_targets, 0, params['num_classes']), final_mask)
flaten_loc_targets = tf.stop_gradient(tf.boolean_mask(flaten_loc_targets, positive_mask))
predictions = {
'classes': tf.argmax(cls_pred, axis=-1),
'probabilities': tf.reduce_max(tf.nn.softmax(cls_pred, name='softmax_tensor'), axis=-1),
'loc_predict': bboxes_pred }
cls_accuracy = tf.metrics.accuracy(flaten_cls_targets, predictions['classes'])
metrics = {'cls_accuracy': cls_accuracy}
# Create a tensor named train_accuracy for logging purposes.
tf.identity(cls_accuracy[1], name='cls_accuracy')
tf.summary.scalar('cls_accuracy', cls_accuracy[1])
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions)
# Calculate loss, which includes softmax cross entropy and L2 regularization.
#cross_entropy = tf.cond(n_positives > 0, lambda: tf.losses.sparse_softmax_cross_entropy(labels=flaten_cls_targets, logits=cls_pred), lambda: 0.)# * (params['negative_ratio'] + 1.)
#flaten_cls_targets=tf.Print(flaten_cls_targets, [flaten_loc_targets],summarize=50000)
cross_entropy = tf.losses.sparse_softmax_cross_entropy(labels=flaten_cls_targets, logits=cls_pred) * (params['negative_ratio'] + 1.)
# Create a tensor named cross_entropy for logging purposes.
tf.identity(cross_entropy, name='cross_entropy_loss')
tf.summary.scalar('cross_entropy_loss', cross_entropy)
#loc_loss = tf.cond(n_positives > 0, lambda: modified_smooth_l1(location_pred, tf.stop_gradient(flaten_loc_targets), sigma=1.), lambda: tf.zeros_like(location_pred))
loc_loss = modified_smooth_l1(location_pred, flaten_loc_targets, sigma=1.)
#loc_loss = modified_smooth_l1(location_pred, tf.stop_gradient(gtargets))
loc_loss = tf.reduce_mean(tf.reduce_sum(loc_loss, axis=-1), name='location_loss')
tf.summary.scalar('location_loss', loc_loss)
tf.losses.add_loss(loc_loss)
l2_loss_vars = []
for trainable_var in tf.trainable_variables():
if '_bn' not in trainable_var.name:
if 'conv4_3_scale' not in trainable_var.name:
l2_loss_vars.append(tf.nn.l2_loss(trainable_var))
else:
l2_loss_vars.append(tf.nn.l2_loss(trainable_var) * 0.1)
# Add weight decay to the loss. We exclude the batch norm variables because
# doing so leads to a small improvement in accuracy.
total_loss = tf.add(cross_entropy + loc_loss, tf.multiply(params['weight_decay'], tf.add_n(l2_loss_vars), name='l2_loss'), name='total_loss')
if mode == tf.estimator.ModeKeys.TRAIN:
global_step = tf.train.get_or_create_global_step()
lr_values = [params['learning_rate'] * decay for decay in params['lr_decay_factors']]
learning_rate = tf.train.piecewise_constant(tf.cast(global_step, tf.int32),
[int(_) for _ in params['decay_boundaries']],
lr_values)
truncated_learning_rate = tf.maximum(learning_rate, tf.constant(params['end_learning_rate'], dtype=learning_rate.dtype), name='learning_rate')
# Create a tensor named learning_rate for logging purposes.
tf.summary.scalar('learning_rate', truncated_learning_rate)
optimizer = tf.train.MomentumOptimizer(learning_rate=truncated_learning_rate,
momentum=params['momentum'])
optimizer = tf.contrib.estimator.TowerOptimizer(optimizer)
# Batch norm requires update_ops to be added as a train_op dependency.
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize(total_loss, global_step)
else:
train_op = None
return tf.estimator.EstimatorSpec(
mode=mode,
predictions=predictions,
loss=total_loss,
train_op=train_op,
eval_metric_ops=metrics,
scaffold=tf.train.Scaffold(init_fn=get_init_fn()))
现在我们开始逐句解析：
"""model_fn for SSD to be used with our Estimator."""
shape = labels['shape']
loc_targets = labels['loc_targets']
cls_targets = labels['cls_targets']
match_scores = labels['match_scores']
global global_anchor_info
decode_fn = global_anchor_info['decode_fn']
num_anchors_per_layer = global_anchor_info['num_anchors_per_layer']
all_num_anchors_depth = global_anchor_info['all_num_anchors_depth']
这里是和input_pipeline()函数末尾对应的，是全局变量，在input_pipeline()函数末尾传入，在这里取出，这里分析一下代表什么吧，loc_targets预测的边框、cls_targets预测的分类、match_scores得分、decode_fn= ymin, xmin, ymax, xmax、num_anchors_per_layer每个特征图anchor数、all_num_anchors_depth 特征图每个像素点anchor数。
with tf.variable_scope(params['model_scope'], default_name=None, values=[features], reuse=tf.AUTO_REUSE):
backbone = ssd_net.VGG16Backbone(params['data_format'])
feature_layers = backbone.forward(features, training=(mode == tf.estimator.ModeKeys.TRAIN))
#print(feature_layers)
location_pred, cls_pred = ssd_net.multibox_head(feature_layers, params['num_classes'], all_num_anchors_depth, data_format=params['data_format'])
if params['data_format'] == 'channels_first':
cls_pred = [tf.transpose(pred, [0, 2, 3, 1]) for pred in cls_pred]
location_pred = [tf.transpose(pred, [0, 2, 3, 1]) for pred in location_pred]
cls_pred = [tf.reshape(pred, [tf.shape(features)[0], -1, params['num_classes']]) for pred in cls_pred]
location_pred = [tf.reshape(pred, [tf.shape(features)[0], -1, 4]) for pred in location_pred]
cls_pred = tf.concat(cls_pred, axis=1)
location_pred = tf.concat(location_pred, axis=1)
cls_pred = tf.reshape(cls_pred, [-1, params['num_classes']])
location_pred = tf.reshape(location_pred, [-1, 4])
第一句是把数据传入VGG16网络，这个其实可以分析，但是差不多每个基础的文本检测算法前期特征提取都是用的VGG16，所以相当成熟这里不想再说了。
backbone.forward用于获得6个特征图也就是经过了后续的网络，6个特征图shape分别为(1,38,38,512),(1,19,19,1024),(1,10,10,512),(1,5,5,256),(1,3,3,256),(1,1,1,256)。然后获得坐标预测和分类预测，对他格式转换6个特征图融合，得到最后的cls_pred，location_pred。
with tf.device('/cpu:0'):
设置使用的设备。
with tf.control_dependencies([cls_pred, location_pred]):
控制操作的运行顺序，得到这两个参数才会运行。
#bboxes_pred = decode_fn(location_pred)
bboxes_pred = tf.map_fn(lambda _preds : decode_fn(_preds),
tf.reshape(location_pred, [tf.shape(features)[0], -1, 4]),
dtype=[tf.float32] * len(num_anchors_per_layer), back_prop=False)
#cls_targets = tf.Print(cls_targets, [tf.shape(bboxes_pred[0]),tf.shape(bboxes_pred[1]),tf.shape(bboxes_pred[2]),tf.shape(bboxes_pred[3])])
bboxes_pred = [tf.reshape(preds, [-1, 4]) for preds in bboxes_pred]
bboxes_pred = tf.concat(bboxes_pred, axis=0)
flaten_cls_targets = tf.reshape(cls_targets, [-1])
flaten_match_scores = tf.reshape(match_scores, [-1])
flaten_loc_targets = tf.reshape(loc_targets, [-1, 4])
lambda声明函数然后调用，作用是把偏移量转化为坐标，后面的操作就是维度变换所有的都降一个维度。
positive_mask = flaten_cls_targets > 0
n_positives = tf.count_nonzero(positive_mask)
batch_n_positives = tf.count_nonzero(cls_targets, -1)
这里实现的是统计实际正样本个数。
batch_negtive_mask = tf.equal(cls_targets, 0)#tf.logical_and(tf.equal(cls_targets, 0), match_scores > 0.)
batch_n_negtives = tf.count_nonzero(batch_negtive_mask, -1)
batch_n_neg_select = tf.cast(params['negative_ratio'] * tf.cast(batch_n_positives, tf.float32), tf.int32)
batch_n_neg_select = tf.minimum(batch_n_neg_select, tf.cast(batch_n_negtives, tf.int32))
这里是统计负样本个数，然后确定选择个数，至于这里的区别选择需要保证正负样本比例1:3。
predictions_for_bg = tf.nn.softmax(tf.reshape(cls_pred, [tf.shape(features)[0], -1, params['num_classes']]))[:, :, 0]
选择一部分负样本进行训练。
prob_for_negtives = tf.where(batch_negtive_mask,
0. - predictions_for_bg,
# ignore all the positives
0. - tf.ones_like(predictions_for_bg))
将后面的部分用-1代替也就是未训练部分。
final_mask = tf.stop_gradient(tf.logical_or(tf.reshape(tf.logical_and(batch_negtive_mask, selected_neg_mask), [-1]), positive_mask))
total_examples = tf.count_nonzero(final_mask)
这是筛选出最后的正负样本。
cls_pred = tf.boolean_mask(cls_pred, final_mask) # 从cls_pred中取final_mask为true的值，分类预测
location_pred = tf.boolean_mask(location_pred, tf.stop_gradient(positive_mask)) # 从clocation_pred中取positive_mask为true的值，回归预测框
只就是我们熟悉的mask掩膜操作。
cls_accuracy = tf.metrics.accuracy(flaten_cls_targets, predictions['classes'])
计算正确率。
cross_entropy = tf.losses.sparse_softmax_cross_entropy(labels=flaten_cls_targets, logits=cls_pred) * (params['negative_ratio'] + 1.)
计算交叉熵损失并且乘以权重。
loc_loss = modified_smooth_l1(location_pred, flaten_loc_targets, sigma=1.)
#loc_loss = modified_smooth_l1(location_pred, tf.stop_gradient(gtargets))
loc_loss = tf.reduce_mean(tf.reduce_sum(loc_loss, axis=-1), name='location_loss')
计算预测框回归smooth_l1损失。
total_loss = tf.add(cross_entropy + loc_loss, tf.multiply(params['weight_decay'], tf.add_n(l2_loss_vars), name='l2_loss'), name='total_loss')
计算总损失。
if mode == tf.estimator.ModeKeys.TRAIN:
global_step = tf.train.get_or_create_global_step()
lr_values = [params['learning_rate'] * decay for decay in params['lr_decay_factors']]
learning_rate = tf.train.piecewise_constant(tf.cast(global_step, tf.int32),
[int(_) for _ in params['decay_boundaries']],
lr_values)
truncated_learning_rate = tf.maximum(learning_rate, tf.constant(params['end_learning_rate'], dtype=learning_rate.dtype), name='learning_rate')
# Create a tensor named learning_rate for logging purposes.
tf.summary.scalar('learning_rate', truncated_learning_rate)
optimizer = tf.train.MomentumOptimizer(learning_rate=truncated_learning_rate,
momentum=params['momentum'])
optimizer = tf.contrib.estimator.TowerOptimizer(optimizer)
# Batch norm requires update_ops to be added as a train_op dependency.
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize(total_loss, global_step)
else:
train_op = None
这是一些基础参数的设置。需要注意的是只有这块才是真正的开始训练类似于session。
return tf.estimator.EstimatorSpec(
mode=mode,
predictions=predictions,
loss=total_loss,
train_op=train_op,
eval_metric_ops=metrics,
scaffold=tf.train.Scaffold(init_fn=get_init_fn()))
最后的返回值。
最后
这里就是我想写的所有解析了，虽然并不是所有代码但是涉及了大部分重要部分，让我们总结一下，数据预处理+anchor建立+模型+损失，希望能对大家有所帮助，在这里向各位前辈致以诚挚敬意。
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
