我试图在conv5层或使用CVR转换的resnet50模块上微调monodepth模型（https://github.com/mrharicot/monodepth）。在更新我的渐变时，我收到错误说明
ValueError: No gradients provided for any variable: ['<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f00022f60>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f0002cfd0>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f00037b38>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f00037eb8>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f00037438>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f000377b8>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f00037f28>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f00037e48>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f00037f60>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f00037f98>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f00037fd0>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f0003d0b8>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f0003d048>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f0003d080>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f0003d2e8>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f0003d358>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f0003d278>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f0003d400>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f0003d710>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f0003d588>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f0003db00>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f0003db70>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f0003dba8>', '<tensorflow.python.training.optimizer._RefVariableProcessor object at 0x7f8f0003d6a0>']
我正在尝试在最大深度点应用CVR变换。然后用可用的地面实况计算预测深度的损失（l2）。在应用渐变时，我得到上面给出的错误。任何人都可以告诉我哪里出错了？
代码：
width_to_focal = dict()
width_to_focal[1242] = 721.5377
width_to_focal[1241] = 718.856
width_to_focal[1224] = 707.0493
width_to_focal[1238] = 718.3351
width_to_focal[1226] = 707.0912
def post_process_disparity(disp):
_, h, w = disp.shape
l_disp = disp[0,:,:]
r_disp = np.fliplr(disp[1,:,:])
m_disp = 0.5 * (l_disp + r_disp)
l, _ = np.meshgrid(np.linspace(0, 1, w), np.linspace(0, 1, h))
l_mask = 1.0 - np.clip(20 * (l - 0.05), 0, 1)
r_mask = np.fliplr(l_mask)
return r_mask * l_disp + l_mask * r_disp + (1.0 - l_mask - r_mask) * m_disp
def count_text_lines(file_path):
f = open(file_path, 'r')
lines = f.readlines()
f.close()
return len(lines)
def point_encoder_simple(depth_batch):
index = np.zeros((2))
index[:] = unravel_index(depth_batch[:,:].argmin(), depth_batch[:,:].shape)
index[:] = index[::-1]
return index
def train(params):
"""Training loop."""
# with tf.Graph().as_default(), tf.device('/cpu:0'):
with tf.Graph().as_default():
global_step = tf.Variable(0, trainable=False)
# INPUT IMAGE PATH
path = '/home/krishna/datasets/'
# GT IMAGE PATH
f1 = tuple(open('/home/krishna/kitti/file_lists/kitti/kitti_gt.txt','r'))
gt_image_path = []
for line in f1:
gt_image_path.append(path + line[:-1])
# OPTIMIZER
num_training_samples = count_text_lines(args.filenames_file)
num_total_steps = params.num_epochs * num_training_samples
start_learning_rate = args.learning_rate
boundaries = [np.int32((3/5) * num_total_steps), np.int32((4/5) * num_total_steps)]
values = [args.learning_rate, args.learning_rate / 2, args.learning_rate / 4]
learning_rate = tf.train.piecewise_constant(global_step, boundaries, values)
opt_step = tf.train.AdamOptimizer(learning_rate)
print("total number of samples: {}".format(num_training_samples))
print("total number of steps: {}".format(num_total_steps))
dataloader_test = MonodepthDataloader(args.data_path, args.filenames_file, params, args.dataset, 'test')
left_test = dataloader_test.left_image_batch
right_test = dataloader_test.right_image_batch
model_test = MonodepthModel(params, 'test', left_test, right_test)
# SESSION
os.environ["CUDA_VISIBLE_DEVICES"]="0"
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
# SAVER
summary_writer = tf.summary.FileWriter(args.log_directory + '/' + args.model_name, sess.graph)
train_saver = tf.train.Saver()
# COUNT PARAMS
total_num_parameters = 0
for variable in tf.trainable_variables(scope='model/encoder'):
total_num_parameters += np.array(variable.get_shape().as_list()).prod()
print("number of trainable parameters: {}".format(total_num_parameters))
a = tf.trainable_variables(scope='model/encoder')
variable_list = a[106:]
# INIT
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
coordinator = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coordinator)
# LOAD CHECKPOINT IF SET
if args.checkpoint_path != '':
train_saver.restore(sess, args.checkpoint_path.split(".")[0])
if args.retrain:
sess.run(global_step.assign(0))
print('now testing {} files'.format(num_training_samples))
disparities_pp = np.zeros((num_training_samples, params.height, params.width), dtype=np.float32)
pred_after_tf_disparities_pp = np.zeros((num_training_samples, params.height, params.width), dtype=np.float32)
# LOSS and GRAD INTIALIZER
tower_grads = []
tower_losses = []
reuse_variables = True
with tf.variable_scope(tf.get_variable_scope()):
with tf.device('/gpu:0'):
for step in range(num_training_samples):
# if (step+1)%10==0:
print ('Iteration : ', step+1)
# reading gt image
gt_disp = cv2.imread(gt_image_path[step],-1)
original_height,original_width= gt_disp.shape
gt_disp = gt_disp.astype(np.float32) / 256
# running the model initially
disp = sess.run(model_test.disp_left_est[0])
disparities_pp[step] = post_process_disparity(disp.squeeze())
# resized output
disparities_pp_resize = cv2.resize(disparities_pp[step].squeeze(), (original_width, original_height))
# index of max depth
index_simple = point_encoder_simple(disparities_pp_resize)
# print (index_simple)
# INITIALIZING CVR PARAMETERS
input_ph = tf.placeholder(tf.float32, shape=[None, original_height, original_width, 3], name='input_image')
input_ph_inv = tf.placeholder(tf.float32, shape=[None, original_height, original_width, 1], name='input_image')
fovea_tf = tf.placeholder(tf.float32, shape=[None, 2], name='fovea_tf') ## foveaX, foveaY
alpha_tf = tf.placeholder(tf.float32, shape=[None, 2], name='alpha_tf') ## alphaX, alphaY
transformed_cvr_batch = cvr_forward(input_ph, fovea_tf, alpha_tf, False, original_width, original_height)
inverse_batch = cvr_inverse(input_ph_inv, fovea_tf, alpha_tf, False, original_width, original_height)
# initiatizing fovea and scaling factor
alpha_np = np.zeros((1,2))
fovea_np = np.zeros((1,2))
alpha_np[0,:] = [0.003,0.003]
fovea_np[0,:] = index_simple[::]
# input rgb for cvr transform
image = tf.image.resize_bilinear(left_test, (original_height, original_width))
input_img = tf.cast(image,np.float32).eval(session=sess)
# cvr transform
after_transform = sess.run(transformed_cvr_batch,feed_dict={input_ph:input_img,alpha_tf:alpha_np,fovea_tf:fovea_np})
# input to model (cvr images)
input_image_ = cv2.resize(after_transform[0,:,:,:].squeeze(), (512, 256))
input_image_ = input_image_.astype(np.float32) / 255
input_images = np.stack((input_image_, np.fliplr(input_image_)), 0)
# running model for the CVR transformed image
pred_after_tf = sess.run(model_test.disp_left_est[0], feed_dict={left_test: input_images})
pred_after_tf_disparities_pp[step] = post_process_disparity(pred_after_tf.squeeze())
# input for inverse cvr
pred_after_tf_disparities_pp_expanded = np.expand_dims(pred_after_tf_disparities_pp,axis=3)
pred_after_tf_disparities_pp_resize = tf.image.resize_bilinear(pred_after_tf_disparities_pp_expanded, (original_height, original_width))
output_image = tf.cast(pred_after_tf_disparities_pp_resize,np.float32).eval(session=sess)
# Inverse CVR transform of prediction
after_inv_transform = sess.run(inverse_batch,feed_dict={input_ph_inv:output_image,alpha_tf:alpha_np,fovea_tf:fovea_np})
ow = tf.cast(after_inv_transform[step,:,:,0], tf.float32)
pred_disp = tf.scalar_mul(original_width,ow)
mask = gt_disp > 0
gt = tf.add(gt_disp, tf.subtract(1.0,mask))
gt_depth = tf.multiply(width_to_focal[original_width], tf.divide(0.54,gt))
pred_depth = tf.multiply(width_to_focal[original_width], tf.divide(0.54,pred_disp))
pred_depth_masked = tf.multiply(pred_depth,mask)
gt_depth = tf.multiply(gt_depth,mask)
loss = tf.nn.l2_loss(tf.subtract(gt_depth,pred_depth_masked))
# print (loss)
tower_losses.append(loss)
reuse_variables = True
grads = opt_step.compute_gradients(loss, var_list = variable_list)
tf.summary.scalar('loss', loss, ['model_0'])
# print (grads)
apply_gradient_op = opt_step.apply_gradients(grads, global_step=global_step)
tower_grads.append(grads)