def _model_infer_to_count(self, lstm_output, mask_list, model_params):
wd = model_params["weight_decay"]
desmap_scale = model_params["desmap_scale"]
count_list = list()
with tf.variable_scope("count_fc"):
for i, output in enumerate(lstm_output):
if i != 0:
tf.get_variable_scope().reuse_variables()
count_bias = mf.fully_connected_layer(output, 1, wd, "fc")
count_bias_decay = tf.mul(tf.nn.l2_loss(count_bias),
wd,
name='bias_decay_%d' % i)
tf.add_to_collection("losses", count_bias_decay)
pred_img = self._filter_mask(self.predict_list[i][-1],
mask_list[i])
image_sum = tf.expand_dims(tf.reduce_sum(pred_img, [1, 2, 3]),
1) / desmap_scale
count = count_bias + image_sum
tf.summary.scalar("check/image_sum/%d" % (i),
tf.reduce_mean(image_sum))
tf.summary.scalar("check/count_bias/%d" % (i),
tf.reduce_mean(count_bias))
count_list.append(count)
return count_list
def model_infer(self, data_ph, model_params):
input_ph = data_ph.get_input()
leaky_param = model_params["leaky_param"]
wd = model_params["weight_decay"]
image_ph_0 = tf.image.resize_images(input_ph, [72, 72])
stage_0 = self._single_hydra_cnn(image_ph_0, model_params, 0)
print(stage_0)
image_ph_1 = iuf.batch_center_crop_frac(image_ph_0, 0.9)
image_ph_1 = tf.image.resize_images(image_ph_1, [72, 72])
stage_1 = self._single_hydra_cnn(image_ph_1, model_params, 1)
print(stage_1)
image_ph_2 = iuf.batch_center_crop_frac(image_ph_0, 0.8)
image_ph_2 = tf.image.resize_images(image_ph_2, [72, 72])
stage_2 = self._single_hydra_cnn(image_ph_2, model_params, 2)
print(stage_2)
concat_feature = tf.concat(1, [stage_0, stage_1, stage_2])
print(concat_feature)
fc6 = mf.fully_connected_layer(concat_feature, 1024, wd, "fc6")
fc7 = mf.fully_connected_layer(fc6, 1024, wd, "fc7")
fc8 = mf.fully_connected_layer(fc7, 1024, wd, "fc8")
batch_size = model_params["batch_size"]
output = tf.expand_dims(tf.reshape(fc8, [batch_size, 32, 32]), 3)
print(output)
l_ph_r = model_params["label_ph_row"]
l_ph_c = model_params["label_ph_col"]
resized_output = tf.image.resize_images(output, [l_ph_r, l_ph_c])
print(resized_output)
with tf.variable_scope("image_sum"):
self._add_image_sum(data_ph.get_input(), data_ph.get_label(),
resized_output, data_ph.get_mask())
self.predict_list = list()
self.predict_list.append(resized_output)
def _model_infer_to_count(self, lstm_output, model_params):
wd = model_params["weight_decay"]
count_list = list()
with tf.variable_scope("count_fc"):
for i, output in enumerate(lstm_output):
if i != 0:
tf.get_variable_scope().reuse_variables()
count = mf.fully_connected_layer(output, 1, wd, "fc")
count_list.append(count)
return count_list