vt-vl-lab / df-net Goto Github PK

View Code? Open in Web Editor NEW

210.0 210.0 30.0 26.67 MB

[ECCV 2018] DF-Net: Unsupervised Joint Learning of Depth and Flow using Cross-Task Consistency

Home Page: http://yuliang.vision/DF-Net/

License: MIT License

Python 82.81% C++ 17.06% Shell 0.13%

depth optical-flow self-supervised-learning

df-net's People

Contributors

Stargazers

Watchers

df-net's Issues

Questions about pre-trained loss and supplementary material

Hello!
Thanks for your great work!
I have a question about pre-train loss. When pre-training flow network unsupervisedly, flow network and depth-pose network are trained respectively using the same loss functions to jointly training, right? And I wonder that do all unsupervised networks need pre-training?
By the way, I try to find the supplemetary material but I failed. Could you tell me where to find it?
Thanks in advance!

Testing on Make3D/ Training on cityscape

Hello @Yuliang-Zou
I have three questions:
1- To test the depth on the kitti dataset, the images were scaled to (height= 160, width = 576).
I am interested to know whether you also rescaled the make3D images using the same values!

2- Should the mean values in the preprocess_image function in the Baselearner, be changed during the testing/eval of the make3D.
These are the mean values:
mean = [104.920005, 110.1753, 114.785955]

3- should these mean values be changed when training on kitti+cityscape dataset. Or is cityscape already included?

Question about Flow_SCALE during testing

Hi Mr zou,firstly thanks for you work on this field.It inspired me a lot. the code:
pred0 = cv2.resize(pred_flow[:,:,0], (W, H), interpolation=cv2.INTER_LINEAR) * (1.0W/old_W)
pred1 = cv2.resize(pred_flow[:,:,1], (W, H), interpolation=cv2.INTER_LINEAR) * (1.0H/old_H)
pred = np.stack((pred0, pred1), axis=-1) * FLOW_SCALE
What does the Flow_SCALE mean?

About the depth consistency loss

Hi, @Yuliang-Zou , Great work!

I seems that the depth consistency loss in your paper doesn't hold, the corresponding pixels' depth can be different between adjacent frames as the camera is also moving, right? Thanks!

Result visualization

Hi, thanks for your works. Can you share the core to convert the predicted result in gray scale to colorful result like this?

Best regards.

InvalidArgumentError: No OpKernel was registered to support Op 'Correlation'

Hi,

I am trying to run training with the code while this error bounced back. The compilation process was good (no error or anything and there's the file correlation_op.so). I'm not sure where things can go wrong?

Caused by op u'flow_prediction/flownet_c_3/Correlation_1', defined at:
  File "train_df.py", line 53, in <module>
    tf.app.run()
  File "/home/hale/anaconda2/envs/tf/lib/python2.7/site-packages/tensorflow/python/platform/app.py", line 48, in run
    _sys.exit(main(_sys.argv[:1] + flags_passthrough))
  File "train_df.py", line 50, in main
    learner.train(FLAGS)
  File "/home/hale/TrimBot/projects/DF-Net/core/DFLearner.py", line 467, in train
    self.build_train_graph()
  File "/home/hale/TrimBot/projects/DF-Net/core/DFLearner.py", line 81, in build_train_graph
    losses, grads = self.single_tower_operation(optim, tgt_image_dp_splits[i], src_image_stack_dp_splits[i], tgt_image_flow_splits[i], src_image_stack_flow_splits[i], tgt_image_splits[i], src_image_stack_splits[i], intrinsics_splits[i], model_idx=i)
  File "/home/hale/TrimBot/projects/DF-Net/core/DFLearner.py", line 148, in single_tower_operation
    tgt2src, src2tgt = flownet(tgt_image_flow, src_image_stack_flow[:,:,:,3*i:3*(i+1)], flownet_spec='C', backward_flow=True, reuse=reuse_variables)
  File "/home/hale/TrimBot/projects/DF-Net/core/UnFlow/src/e2eflow/core/flownet.py", line 86, in flownet
    scoped_block(reuse)
  File "/home/hale/TrimBot/projects/DF-Net/core/UnFlow/src/e2eflow/core/flownet.py", line 49, in scoped_block
    channel_mult=channel_mult)
  File "/home/hale/TrimBot/projects/DF-Net/core/UnFlow/src/e2eflow/core/flownet.py", line 231, in flownet_c
    pad=20, kernel_size=1, max_displacement=20, stride_1=1, stride_2=2)
  File "/home/hale/TrimBot/projects/DF-Net/core/UnFlow/src/e2eflow/ops.py", line 64, in correlation
    return _correlation_module.correlation(first, second, **kwargs)[0]
  File "<string>", line 49, in correlation
  File "/home/hale/anaconda2/envs/tf/lib/python2.7/site-packages/tensorflow/python/framework/op_def_library.py", line 767, in apply_op
    op_def=op_def)
  File "/home/hale/anaconda2/envs/tf/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 2506, in create_op
    original_op=self._default_original_op, op_def=op_def)
  File "/home/hale/anaconda2/envs/tf/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 1269, in __init__
    self._traceback = _extract_stack()

InvalidArgumentError (see above for traceback): No OpKernel was registered to support Op 'Correlation' with these attrs.  Registered devices: [CPU,GPU], Registered kernels:
  <no registered kernels>

	 [[Node: flow_prediction/flownet_c_3/Correlation_1 = Correlation[kernel_size=1, max_displacement=20, pad=20, stride_1=1, stride_2=2, _device="/device:GPU:0"](flow_prediction/flownet_c_features_3/conv3_1/leaky_relu/Maximum, flow_prediction/flownet_c_features_3/conv3/leaky_relu/Maximum)]]

version `CXXABI_1.3.9' not found

Hi,
Python=2.7
Ubuntu 14.04

I was following the instructions in read me file and got this errors

coesip@coesip-HP-Z230-SFF-Workstation:$ cd Desktop/DF-Net-master/
coesip@coesip-HP-Z230-SFF-Workstation:/Desktop/DF-Net-master$ python train_df.py --dataset_dir=/home/coesip/Desktop/DF-Net-master/dataset --checkpoint_dir=/home/coesip/Desktop/DF-Net-master
Traceback (most recent call last):
File "train_df.py", line 6, in
from core import DFLearner
File "/home/coesip/Desktop/DF-Net-master/core/init.py", line 2, in
from .DFLearner import DFLearner
File "/home/coesip/Desktop/DF-Net-master/core/DFLearner.py", line 9, in
from .BaseLearner import BaseLearner
File "/home/coesip/Desktop/DF-Net-master/core/BaseLearner.py", line 10, in
from .nets import *
File "/home/coesip/Desktop/DF-Net-master/core/nets.py", line 7, in
from .utils import flow_inverse_warp
File "/home/coesip/Desktop/DF-Net-master/core/utils.py", line 2, in
import matplotlib.pyplot as plt
File "/home/coesip/anaconda2/lib/python2.7/site-packages/matplotlib/pyplot.py", line 31, in
import matplotlib.colorbar
File "/home/coesip/anaconda2/lib/python2.7/site-packages/matplotlib/colorbar.py", line 32, in
import matplotlib.artist as martist
File "/home/coesip/anaconda2/lib/python2.7/site-packages/matplotlib/artist.py", line 16, in
from .path import Path
File "/home/coesip/anaconda2/lib/python2.7/site-packages/matplotlib/path.py", line 21, in
from . import _path, rcParams
ImportError: /usr/lib/x86_64-linux-gnu/libstdc++.so.6: version `CXXABI_1.3.9' not found (required by /home/coesip/anaconda2/lib/python2.7/site-packages/matplotlib/_path.so)
How to solve this?

Saving Flow and Pose models separately during training

@Yuliang-Zou Currently the training saves the complete session (depth, flow and pose) in the checkpoint. Is there any option to save the learned flow model and pose model separately?

Dataset not found

Could you please specify which part of the KITTI to be downloaded and how the dataset should be organized? I am having this error

tensorflow.python.framework.errors_impl.NotFoundError: misc/raw/2011_09_30_drive_0028_sync_03/0000002522_cam.txt
	 [[Node: data_loading/ReaderReadV2_1 = ReaderReadV2[_device="/job:localhost/replica:0/task:0/cpu:0"](data_loading/TextLineReaderV2, data_loading/input_producer_1)]]

I've tried downloading the raw+sync KITTI, using the link at this line:
Download the raw dataset download script (1 MB) (thanks to Omid Hosseini for sharing!)

and put them under raw folder, and passed in --dataset_dir=raw

The structure appears to be : like raw/2011_09_30/2011_09_30_drive_0028_sync/ and inside it are these directories:
image_00 , image_01 , image_02 , image_03 , oxts , velodyne_points

There's no 2011_09_30_drive_0028_sync_03 or 0000002522_cam.txt to be found, maybe I need to download from different link or is there any pre-processing / re-organizing required?

How to cacluate the EPE for the optical flow 2012/2015 test sets?

Hello @Yuliang-Zou ,
In the DF-Net paper, it was mentioned that the EPE you achieve on the KITTI 2012 test set is 4.4.
I am not sure how you can obtain this error without the flow_occ ground truth label which is not available for the test set. I'm definitely missing something here but I don't know what.

Also by following the code in test_flow_2012.py, I found that the EPE for the test is ignored here:

# Save for eval
            if 'test' in FLAGS.dataset_dir:
                _, scaled_pred = compute_flow_error(np.array(raw_im0)[:,:,:2], pred_flow_val[0,:,:,:], np.array(raw_im0)[:,:,0])
                png_name = os.path.join(FLAGS.output_dir, new_files[t][0].split('/')[-1])
                write_flow_png(png_name, scaled_pred)

so the compute_flow_error function should return the EPE and the scaled_pred but the EPE is ignored here. On the contrary, for the train set, in the same file, its is used to calculate the EPE in errs[t]:

# Only for training set
            if 'train' in FLAGS.dataset_dir:
                # no occlusion
                #gt_flow, mask = get_flow(new_files[t][0].replace('colored_0', 'flow_noc'))
                # all
                gt_flow, mask = get_flow(new_files[t][0].replace('colored_0', 'flow_occ'))
                errs[t], scaled_pred = compute_flow_error(gt_flow, pred_flow_val[0,:,:,:], mask)

Any clarification would be of great help!

Training with multiple batches on the same GPU

I am trying to train the network on Nvidia 1080Ti graphics card (memory of 11GB). I am unable to run multiple batches on the same GPU. I get the following error.

2018-12-05 13:14:36.687218: I tensorflow/core/common_runtime/bfc_allocator.cc:700] Sum Total of in-use chunks: 9.24GiB
2018-12-05 13:14:36.687230: I tensorflow/core/common_runtime/bfc_allocator.cc:702] Stats:
Limit: 10775137485
InUse: 9919629824
MaxInUse: 10754012160
NumAllocs: 57124
MaxAllocSize: 2548712448

2018-12-05 13:14:36.687794: W tensorflow/core/common_runtime/bfc_allocator.cc:277] ****************************************************************************************************
2018-12-05 13:14:36.687829: W tensorflow/core/framework/op_kernel.cc:1158] Resource exhausted: OOM when allocating tensor with shape[2,32,160,576]

Right now, if I have to train with batch size of 2 I need 2 GPUs. For other networks like monodepth, the requirement is that the batch size should be divisible by the number of GPUs. But for DF-Net batch size has to be same as the number of GPUs used. Have you tried running with a batch size two/four times the number of GPUs?

train ERROR

when I use train_df.py to train, It outputs:
/bin/sh: 1: nvcc: not found
Traceback (most recent call last):
File "train_df.py", line 6, in
from core import DFLearner
File "/home/boyob/DepthPredict/DF-Net/core/init.py", line 2, in
from .DFLearner import DFLearner
File "/home/boyob/DepthPredict/DF-Net/core/DFLearner.py", line 15, in
from .UnFlow import flownet
File "/home/boyob/DepthPredict/DF-Net/core/UnFlow/init.py", line 1, in
from .src import flownet
File "/home/boyob/DepthPredict/DF-Net/core/UnFlow/src/init.py", line 1, in
from .e2eflow import flownet
File "/home/boyob/DepthPredict/DF-Net/core/UnFlow/src/e2eflow/init.py", line 1, in
from .core import flownet
File "/home/boyob/DepthPredict/DF-Net/core/UnFlow/src/e2eflow/core/init.py", line 1, in
from .flownet import flownet
File "/home/boyob/DepthPredict/DF-Net/core/UnFlow/src/e2eflow/core/flownet.py", line 5, in
from ..ops import correlation
File "/home/boyob/DepthPredict/DF-Net/core/UnFlow/src/e2eflow/ops.py", line 55, in
compile(n)
File "/home/boyob/DepthPredict/DF-Net/core/UnFlow/src/e2eflow/ops.py", line 35, in compile
subprocess.check_output(nvcc_cmd, shell=True)
File "/usr/lib/python2.7/subprocess.py", line 574, in check_output
raise CalledProcessError(retcode, cmd, output=output)
subprocess.CalledProcessError: Command 'nvcc -std=c++11 -c -gencode=arch=compute_30,code=sm_30 -o backward_warp_op.cu.o backward_warp_op.cu.cc -I /usr/local/lib/python2.7/dist-packages/tensorflow/include -D GOOGLE_CUDA=1 -x cu -Xcompiler -fPIC' returned non-zero exit status 127
Could you help me to solve this? Any suggestion would be a great appreciate!

Do you have the generated ground truth depth map of Kitti?

I notice you use the ground truth depth map of Kitti for display. Could you please share them or share the method to produce? The depth map I generated is not correct.

error with test_flownet_2012.py/test_flownet_2015.py

I was able to get the same results for the depth net by running the following commands mentioned in the repo:
python test_kitti_depth.py --dataset_dir=/path/to/your/data --output_dir=/path/to/save/your/prediction --ckpt_file=/DF-Net-master/pretrained/dfnet - --split="test"
and
python kitti_eval/eval_depth.py --pred_file=/path/to/your/prediction --split="test"

However when I run the test Flownet command on kitti 2012 OR on kitti 2015:
python test_flownet_2012.py --dataset_dir=/path/to/your/data --ckpt_file=/DF-Net-master/pretrained/unflow_pre

I get the following error:

  File "test_flownet_2015.py", line 131, in <module>
    tf.app.run()
  File "/system/user/taha/pycharmprojects/DF-Net-master/env/lib/python3.6/site-packages/tensorflow/python/platform/app.py", line 48, in run
    _sys.exit(main(_sys.argv[:1] + flags_passthrough))
  File "test_flownet_2015.py", line 90, in main
    saver.restore(sess, FLAGS.ckpt_file)
  File "/system/user/taha/pycharmprojects/DF-Net-master/env/lib/python3.6/site-packages/tensorflow/python/training/saver.py", line 1548, in restore
    {self.saver_def.filename_tensor_name: save_path})
  File "/system/user/taha/pycharmprojects/DF-Net-master/env/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 789, in run
    run_metadata_ptr)
  File "/system/user/taha/pycharmprojects/DF-Net-master/env/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 997, in _run
    feed_dict_string, options, run_metadata)
  File "/system/user/taha/pycharmprojects/DF-Net-master/env/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1132, in _do_run
    target_list, options, run_metadata)
  File "/system/user/taha/pycharmprojects/DF-Net-master/env/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1152, in _do_call
    raise type(e)(node_def, op, message)
tensorflow.python.framework.errors_impl.InvalidArgumentError: No OpKernel was registered to support Op 'Correlation' with these attrs.  Registered devices: [CPU], Registered kernels:
  device='GPU'

	 [[Node: flownet_c/Correlation = Correlation[kernel_size=1, max_displacement=20, pad=20, stride_1=1, stride_2=2](flownet_c_features/conv3/leaky_relu/Maximum, flownet_c_features/conv3_1/leaky_relu/Maximum)]]

Caused by op 'flownet_c/Correlation', defined at:
  File "test_flownet_2015.py", line 131, in <module>
    tf.app.run()
  File "/system/user/taha/pycharmprojects/DF-Net-master/env/lib/python3.6/site-packages/tensorflow/python/platform/app.py", line 48, in run
    _sys.exit(main(_sys.argv[:1] + flags_passthrough))
  File "test_flownet_2015.py", line 78, in main
    pred_flows = flownet(im1_pl, im2_pl, flownet_spec='C')
  File "/system/user/taha/pycharmprojects/DF-Net-master/core/UnFlow/src/e2eflow/core/flownet.py", line 86, in flownet
    scoped_block(reuse)
  File "/system/user/taha/pycharmprojects/DF-Net-master/core/UnFlow/src/e2eflow/core/flownet.py", line 43, in scoped_block
    channel_mult=channel_mult)
  File "/system/user/taha/pycharmprojects/DF-Net-master/core/UnFlow/src/e2eflow/core/flownet.py", line 231, in flownet_c
    pad=20, kernel_size=1, max_displacement=20, stride_1=1, stride_2=2)
  File "/system/user/taha/pycharmprojects/DF-Net-master/core/UnFlow/src/e2eflow/ops.py", line 64, in correlation
    return _correlation_module.correlation(first, second, **kwargs)[0]
  File "<string>", line 49, in correlation
  File "/system/user/taha/pycharmprojects/DF-Net-master/env/lib/python3.6/site-packages/tensorflow/python/framework/op_def_library.py", line 767, in apply_op
    op_def=op_def)
  File "/system/user/taha/pycharmprojects/DF-Net-master/env/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 2506, in create_op
    original_op=self._default_original_op, op_def=op_def)
  File "/system/user/taha/pycharmprojects/DF-Net-master/env/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 1269, in __init__
    self._traceback = _extract_stack()

InvalidArgumentError (see above for traceback): No OpKernel was registered to support Op 'Correlation' with these attrs.  Registered devices: [CPU], Registered kernels:
  device='GPU'

	 [[Node: flownet_c/Correlation = Correlation[kernel_size=1, max_displacement=20, pad=20, stride_1=1, stride_2=2](flownet_c_features/conv3/leaky_relu/Maximum, flownet_c_features/conv3_1/leaky_relu/Maximum)]]

Any ideas on how to fix it?
@Yuliang-Zou

Pretraining the network

Hi @Yuliang-Zou Thanks for sharing the code for the reimplementation. I wanted to replicate the results as mentioned in the paper. Is it possible for you to share the code changes required for 2 frames?
Can you also please share the results after pretraining the depth and flow network with 2 frame setup?

flownet produces non-deterministic output

Hello @roboticsbala @Yuliang-Zou @lhoangan

using this command:
python test_flownet_2012.py --dataset_dir=/publicdata/kitti/optical_flow_2012/training/ --ckpt_file=pretrained/unflowc_pre

I still have the same problem that I get non consistent values for the EPE error with flow_2012. I traced the reason and found out that the flownet returns different values for the same input in test_flownet_2012.py

Also, as suggested by @roboticsbala in #4 , I already deleted "-gencode=arch=compute_30,code=sm_30" from the nvcc_cmd in ops.py.

Now the lines 31 to 38 in the ops.py are:

cuda_lib64_path_arg = "-L /system/apps/biosoft/cuda-8.0/lib64"
        nvcc_cmd = "nvcc -std=c++11 -c -o {} -I {} -D GOOGLE_CUDA=1 -x cu -Xcompiler -fPIC -I $/system/apps/biosoft/cuda-8.0/include"
        nvcc_cmd = nvcc_cmd.format(" ".join([fn_cu_o, fn_cu_cc]),
                                   tf_inc)
        subprocess.check_output(nvcc_cmd, shell=True)

        gcc_cmd = "{} -std=c++11 -shared -o {} -I {} -fPIC -lcudart -D GOOGLE_CUDA=1 {}"
        gcc_cmd = gcc_cmd.format('g++',
                                " ".join([fn_so, fn_cu_o, fn_cc]),
                                 tf_inc,
                                 cuda_lib64_path_arg)

I am using python 3.6.1
Cuda 8.0
cudnn 5.1
tensorflow 1.2.0
gcc version 4.8.5 20150623 (Red Hat 4.8.5-36) (GCC)

Do you have any idea how to fix it?

Originally posted by @YasserTahaSW in #4 (comment)

Kitti 2015 optical flow results?

Hi Yuliang, thank you for sharing the code, that's very helpful! Can you share your results of kitti flow 2015 in the kitti evaluation format? I notice you share the Visual Results on KITTI15 but not in the format for kitti evaluation, it will be really appreciated if you can also share the results in the format of kitti evaluation. Thank you in advance!

The shuffle operation in the data loading process.

Hi, thanks for your inspiring work! I am new to tensorflow and optical flow estimation, and I am trying to learn your code recently. I don't quite understand here in your code: the order of the data sequence is shuffled. In this way, can we ensure that a batch of images read from the sequence is consecutive?
Because tensorflow has few documents, I don't know if my understanding is correct. Looking forward to your response.

Test pose code

Hi,

Great work! Can you please share a script to do pose evaluation (similar to test_kitti_depth.py)? I tried loading the weights and creating the model with image sizes as [128, 416], [256, 832], [384, 1248], but all of them resulted in size mismatch errors.

Thanks,
Sai

saving depth images

Hi @Yuliang-Zou ,
I have a problem wit saving the depth images correctly.
To save the depth images, I added the following code in the test_kitti_depth.py:

            for b in range(FLAGS.batch_size):
                idx = t + b
                if idx >= len(test_files):
                    break
                pred_all.append(pred['depth'][b,:,:,0])

                #added the following two lines

                png_name = os.path.join(FLAGS.output_dir, str(idx)+'.jpg')
                pyplot.imsave(png_name, (pred['depth'][b,:,:,0]),vmin = 1e-3, vmax = 80)

The problem i get is that the saved images are not the same as the one posted on your project page. They have the same overall structure but the overall intensities are different.
For example, image 57 in the test split should look like:

However, my output is:

The car and the trees are there but the intensities are completely different...

So, I wonder if you can point out whats wrong or if you can share the code where you saved the images.

Thanks a lot!

compilation error

I have installed all the configurations as specified on the page i.e.Python 3.6, TensorFlow 1.2.0, g++ 4.9, CUDA 8.0, In Ubuntu using anaconda virtual environment. But I am running into a compilation error:

backward_warp_op.cu.cc:5:54: fatal error: tensorflow/core/framework/register_types.h: No such file or directory
compilation terminated.
Traceback (most recent call last):
File "/visinf/home/sgoyal/Desktop/DF-Net-master/core/UnFlow/src/e2eflow/ops.py", line 53, in
op_lib = tf.load_op_library(lib_path)
File "/visinf/home/sgoyal/anaconda3/envs/sak/lib/python3.6/site-packages/tensorflow/python/framework/load_library.py", line 64, in load_op_library
None, None, error_msg, error_code)
tensorflow.python.framework.errors_impl.NotFoundError: ./backward_warp_op.so: cannot open shared object file: No such file or directory

Can you please help me out with this?

@jbhuang0604 @sanketloke @Yuliang-Zou @LemonATsu @gaochen315

Receive loss nan when training

I've started training the network when I received all nan value for loss from the first iteration.

python train_df.py --dataset_dir=misc/dataset/kitti_5frame_1152_320 --checkpoint_dir=checkpoint

I'm not sure if it's the effect of using 1 GPU and setting batch size to 1. Do you have any insight to this?

{'alpha_image_loss': 0.85,
 'batch_size': 1,
 'beta1': 0.9,
 'checkpoint_dir': 'checkpoint',
 'ckpt_dp': 'misc/pretrained/cs_5frame_pre',
 'ckpt_flow': 'misc/pretrained/unflowc_pre',
 'ckpt_pose': None,
 'continue_train': True,
 'cross_consistency': 0.5,
 'dataset_dir': 'misc/dataset/kitti_5frame_1152_320',
 'depth_consistency': 0.2,
 'fix_pose': False,
 'flow_consistency': 0.2,
 'flow_smooth_weight': 3.0,
 'img_height': 320,
 'img_width': 1152,
 'learning_rate': 0.0001,
 'max_steps': 100000,
 'num_gpus': 1,
 'save_latest_freq': 5000,
 'scale_normalize': False,
 'seq_length': 5,
 'smooth_weight': 3.0,
 'summary_freq': 100}
WARNING:tensorflow:From /home/hale/TrimBot/projects/DF-Net/core/DFLearner.py:475: all_variables (from tensorflow.python.ops.variables) is deprecated and will be removed after 2017-03-02.
Instructions for updating:
Please use tf.global_variables instead.
2018-09-18 14:42:53.853974: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.
2018-09-18 14:42:53.854021: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.
2018-09-18 14:42:53.854031: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
2018-09-18 14:42:53.854049: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX2 instructions, but these are available on your machine and could speed up CPU computations.
2018-09-18 14:42:53.854057: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use FMA instructions, but these are available on your machine and could speed up CPU computations.
2018-09-18 14:42:54.272654: I tensorflow/core/common_runtime/gpu/gpu_device.cc:940] Found device 0 with properties: 
name: GeForce GTX TITAN X
major: 5 minor: 2 memoryClockRate (GHz) 1.076
pciBusID 0000:03:00.0
Total memory: 11.91GiB
Free memory: 10.38GiB
2018-09-18 14:42:54.272711: I tensorflow/core/common_runtime/gpu/gpu_device.cc:961] DMA: 0 
2018-09-18 14:42:54.272721: I tensorflow/core/common_runtime/gpu/gpu_device.cc:971] 0:   Y 
2018-09-18 14:42:54.272734: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1030] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce GTX TITAN X, pci bus id: 0000:03:00.0)
Trainable variables: 
depth_net_res50/Conv/weights:0
depth_net_res50/Conv/biases:0
...
flownet_c/flow3_up2/biases:0
flownet_c/flow2/weights:0
flownet_c/flow2/biases:0
('parameter_count =', 99228758)
Resume flow_net from previous checkpoint: misc/pretrained/unflowc_pre
Resume depth_net and pose_net from previous checkpoint: misc/pretrained/cs_5frame_pre
2018-09-18 14:44:23.043925: I tensorflow/core/kernels/cuda_solvers.cc:137] Creating CudaSolver handles for stream 0x3f12fac0
2018-09-18 14:44:30.411640: W tensorflow/core/common_runtime/bfc_allocator.cc:217] Allocator (GPU_0_bfc) ran out of memory trying to allocate 3.22GiB. The caller indicates that this is not a failure, but may mean that there could be performance gains if more memory is available.
2018-09-18 14:44:31.111228: I tensorflow/core/common_runtime/gpu/pool_allocator.cc:247] PoolAllocator: After 1378 get requests, put_count=1240 evicted_count=1000 eviction_rate=0.806452 and unsatisfied allocation rate=0.898403
2018-09-18 14:44:31.111302: I tensorflow/core/common_runtime/gpu/pool_allocator.cc:259] Raising pool_size_limit_ from 100 to 110
2018-09-18 14:45:26.740231: I tensorflow/core/common_runtime/gpu/pool_allocator.cc:247] PoolAllocator: After 2095 get requests, put_count=2145 evicted_count=1000 eviction_rate=0.4662 and unsatisfied allocation rate=0.464439
2018-09-18 14:45:26.740317: I tensorflow/core/common_runtime/gpu/pool_allocator.cc:259] Raising pool_size_limit_ from 256 to 281
/home/hale/TrimBot/projects/DF-Net/core/flowlib.py:257: RuntimeWarning: invalid value encountered in greater
  idxUnknow = (abs(u) > UNKNOWN_FLOW_THRESH) | (abs(v) > UNKNOWN_FLOW_THRESH)
Epoch: [ 1] [  100/34841] time: 2.5545/it loss: nan
Epoch: [ 1] [  200/34841] time: 2.1562/it loss: nan
Epoch: [ 1] [  300/34841] time: 2.1207/it loss: nan
Epoch: [ 1] [  400/34841] time: 2.1037/it loss: nan
Epoch: [ 1] [  500/34841] time: 2.1077/it loss: nan
Epoch: [ 1] [  600/34841] time: 2.1089/it loss: nan
Epoch: [ 1] [  700/34841] time: 2.1245/it loss: nan
Epoch: [ 1] [  800/34841] time: 2.1175/it loss: nan
Epoch: [ 1] [  900/34841] time: 2.1038/it loss: nan
Epoch: [ 1] [ 1000/34841] time: 2.1104/it loss: nan
...

models

Could you give me your petrained-depth-models ? I want to use it to draw some figures and I want to ask that how to make kitti depth gt image?

vt-vl-lab / df-net Goto Github PK

df-net's People

Contributors

Stargazers

Watchers

Forkers

df-net's Issues

Recommend Projects

Recommend Topics

Recommend Org