This is a prednet practice using leido/pytorch-prednet modified code and preprocessed KITTI data.
- image format, from integer 0-255 to float 0-1.0
- train data, sequence_start_mode from unique to all
- ConvLSTMCell, sigmoid to hard_sigmoid
- train learning rate, 0.001/0.0001 switch to use lr_scheduler
- test data, last one to sequence, and to add extrapolated
- to add log output
- to add cpu for function check. Actual train needs GPU.
- to add option load-last and last-epoch to resume train.
- to add preprocess_data.py that make a dataset from mp4 files
A loss curve is shown. This result does not reach to 0.0178 (training) 0.0236(validation), which other reports value using original prednet code.
prednet_kitti_data_colab_practice1.ipynb is a sample notebook of train and test using google colaboratory.
Upper side is actual and lower side is predicted.
Lower side last 3 images are extrapolated.
Regarding to original prednet code license, please see License_prednet.txt.
Regarding to kitti dataset license, please refer to Creative Commons Attribution-NonCommercial-ShareAlike 3.0 License.
Regarding to hickle.py, please see LICENSE_hickle.
Regarding to hard_sigmoid , please see LICENSE_torchMoji.
There is no license description of leido/pytorch-prednet and Kaixhin/lstms.py.
Some code has benn changed and follow following,
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.