SqueezeDet object detector reimplemented in Theano/Lasagne

papers/squeezedet/LICENSE

@@ -0,0 +1,39 @@
+Parts of the code in Squeezedet_Theano are derived from the
+following version of SqueezeDet by Bichen Wu:
+
+     Author    :  Bichen Wu
+     Date      :  22 November 2017
+     Source    :  https://github.com/BichenWuUCB/squeezeDet/
+     Commit    :  e7c0860 (Wed Nov 22 15:47:02 2017 -0800)
+
+Copyright for portions of the derived code in Squeezedet_Theano are
+held by Bichen Wu, 2016-2107. All other code in Squeezedet_Theano is
+copyright Corvidim, 2017. This code is published in accordance with
+and retains the same license as SqueezeDet by Bichen Wu.
+
+BSD 2-Clause License
+
+Copyright (c) 2016-2017, Bichen Wu
+Copyright (c) 2017, Corvidim
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+* Redistributions of source code must retain the above copyright notice, this
+  list of conditions and the following disclaimer.
+
+* Redistributions in binary form must reproduce the above copyright notice,
+  this list of conditions and the following disclaimer in the documentation
+  and/or other materials provided with the distribution.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

papers/squeezedet/README.md

@@ -0,0 +1,82 @@
+# SqueezeDet_Theano
+
+This is a reimplementation of Bichen Wu's SqueezeDet [1],[2] for
+Theano [3] and Lasagne [4] by
+[@corvidim](https://github.com/corvidim). The version contained in
+this repo corresponds to a snapshot of
+[squeezedet_theano](https://github.com/corvidim/squeezedet_theano)
+0.1.0.
+
+SqueezeDet is particularly well-suited to low-power,
+low-memory-consumption applications. In particular, the trained
+weights take up less than 8Mb!
+
+This release includes:
+
+* Network weights trained on the KITTI dataset (`data/squeezedet_kitti.pkl.gz`, <8MB).
+
+* The detector ("prediction graph" and "interpretation graph"
+of the original TensorFlow implementation); training is not yet
+supported.
+
+
+References:
+
+- [1] https://arxiv.org/abs/1612.01051
+- [2] https://github.com/BichenWuUCB/squeezeDet
+- [3] http://deeplearning.net/software/theano/
+- [4] http://lasagne.readthedocs.io/en/latest/
+
+## Setup
+
+Requirements:
+
+- Python 2.7+ (Python 3 experimental support; tested under Python 3.6.3)
+- Theano 0.9+
+- Lasagne 0.2+
+- Numpy, PIL, Matplotlib
+
+Sample docker files for Ubuntu installation:
+
+- `./setup/sqzdet_theano_dockerfile.txt` (Python 2, Theano 0.10)
+- `./setup/sqzdet_theano_0.9_dockerfile.txt` (Python 2, Theano 0.9)
+- `./setup/sqzdet_theano_py3k_dockerfile.txt` (Python 3, Theano 0.10)
+
+
+To run detection on a default image (`data/sample.png` from the original SqueezeDet demo):
+
+```
+THEANO_FLAGS='floatX=float32' python src/sqz_det_thn.py --network_weights=data/squeezedet_kitti.pkl.gz
+```
+
+Use the `--img_in` flag to specify an image file, for example:
+
+```
+THEANO_FLAGS='floatX=float32' python src/sqz_det_thn.py --network_weights=data/squeezedet_kitti.pkl.gz --img_in=data/Boston_01.png
+```
+
+or specify a comma-separated list of images (all of which must be the same size), for example:
+
+```
+THEANO_FLAGS='floatX=float32' python src/sqz_det_thn.py --network_weights=data/squeezedet_kitti.pkl.gz --img_in=data/sample.png,data/Boston_00.png
+```
+
+Output visualization(s) will be saved in PDF format to the current
+directory (e.g., `./out_sample.pdf`) and displayed (if running
+interactively) as shown below.  To specify a different directory for
+output use the `--out_dir` flag.
+
+## Output visualizations of sample input images
+---
+![output from sample.png](data/output/out_sample.png)
+
+---
+
+![output from Boston_00.png](data/output/out_Boston_00.png)
+
+---
+
+![output from Boston_01.png](data/output/out_Boston_01.png)
+
+---
+

papers/squeezedet/setup/sqzdet_theano_0.9_dockerfile.txt

@@ -0,0 +1,37 @@
+#Ubuntu CUDA Py2 docker file for sqzdet_theano under Theano 0.9
+# nvidia-docker build -t sqzdet_theano_0.9 . -- file sqzdet_theano_0.9.txt
+
+FROM nvidia/cuda:8.0-cudnn5-devel-ubuntu16.04
+
+RUN apt-get update && apt-get install -y \
+  apt-utils \
+  git \
+  wget
+
+RUN wget --no-check-certificate https://repo.continuum.io/miniconda/Miniconda2-latest-Linux-x86_64.sh -O miniconda2.sh
+
+RUN chmod +x miniconda2.sh && bash ./miniconda2.sh -b -p /opt/miniconda2
+
+ENV PATH /opt/miniconda2/bin:$PATH
+
+RUN conda update conda
+
+# as of 20171226: matplotlib=2.1.1 numpy=1.13.3 pillow=4.3.0
+RUN conda install \
+    matplotlib \
+    numpy \
+    pillow \
+    mkl-service
+
+#NB: we require Lasagne>=0.2, still "bleeding edge" (0.2.dev1 as of 20171225)
+RUN pip install \
+    numpy \
+    easydict \
+    future \
+    Theano==0.9 \
+    --upgrade https://github.com/Lasagne/Lasagne/archive/master.zip
+
+# to avoid RuntimeError('To use MKL 2018 with Theano...')
+ENV MKL_THREADING_LAYER GNU
+
+# nvidia-docker run --rm -it --name st_0  -v /home/ubuntu:/home/ubuntu sqzdet_theano_0.9

papers/squeezedet/setup/sqzdet_theano_dockerfile.txt

@@ -0,0 +1,36 @@
+#Ubuntu CUDA Python2 docker file for sqzdet_theano under Theano 1.0
+# nvidia-docker build -t sqzdet_theano . --file sqzdet_theano.txt
+
+FROM nvidia/cuda:8.0-cudnn6-devel-ubuntu16.04
+
+RUN apt-get update && apt-get install -y \
+  apt-utils \
+  git \
+  wget
+
+RUN wget --no-check-certificate https://repo.continuum.io/miniconda/Miniconda2-latest-Linux-x86_64.sh -O miniconda2.sh
+
+RUN chmod +x miniconda2.sh && bash ./miniconda2.sh -b -p /opt/miniconda2
+
+ENV PATH /opt/miniconda2/bin:$PATH
+
+RUN conda update conda
+
+# as of 20171226: matplotlib=2.1.1 numpy=1.13.3 pillow=4.3.0
+RUN conda install \
+    matplotlib \
+    numpy \
+    pillow \
+    mkl-service
+
+# as of 20171226, bleeding-edge installs below: theano=1.0.1+unknown, lasagne=0.2.dev1
+RUN pip install \
+    easydict \
+    future \
+    --upgrade https://github.com/Theano/Theano/archive/master.zip \
+    --upgrade https://github.com/Lasagne/Lasagne/archive/master.zip
+
+#to avoid RuntimeError('To use MKL 2018 with Theano...')
+ENV MKL_THREADING_LAYER GNU
+
+# nvidia-docker run --rm -it --name st_1  -v /home/ubuntu:/home/ubuntu sqzdet_theano

papers/squeezedet/setup/sqzdet_theano_py3k_dockerfile.txt

@@ -0,0 +1,36 @@
+#Ubuntu CUDA Python3 docker file for sqzdet_theano under Theano 1.0
+# nvidia-docker build -t sqzdet_theano_py3k . --file sqzdet_theano_py3k.txt
+
+FROM nvidia/cuda:8.0-cudnn6-devel-ubuntu16.04
+
+RUN apt-get update && apt-get install -y \
+  apt-utils \
+  git \
+  wget
+
+RUN wget --no-check-certificate https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh -O miniconda3.sh
+
+RUN chmod +x miniconda3.sh && bash ./miniconda3.sh -b -p /opt/miniconda3
+
+ENV PATH /opt/miniconda3/bin:$PATH
+
+RUN conda update conda
+
+# as of 20171226: matplotlib=2.1.1 numpy=1.13.3 pillow=4.3.0
+RUN conda install \
+    matplotlib \
+    numpy \
+    pillow \
+    mkl-service
+
+# as of 20171226, bleeding-edge installs below: theano=1.0.1+unknown, lasagne=0.2.dev1
+RUN pip install \
+    easydict \
+    future \
+    --upgrade https://github.com/Theano/Theano/archive/master.zip \
+    --upgrade https://github.com/Lasagne/Lasagne/archive/master.zip
+
+#to avoid RuntimeError('To use MKL 2018 with Theano...')
+ENV MKL_THREADING_LAYER GNU
+
+# nvidia-docker run --rm -it --name st_3k  -v /home/ubuntu:/home/ubuntu sqzdet_theano_py3k

papers/squeezedet/src/config.py

@@ -0,0 +1,142 @@
+# Author: Bichen Wu ([email protected]) 08/25/2016
+
+"""Base Model configurations"""
+
+import os
+import os.path as osp
+import numpy as np
+from easydict import EasyDict as edict
+
+def base_model_config(dataset='PASCAL_VOC'):
+  assert dataset.upper()=='PASCAL_VOC' or dataset.upper()=='KITTI', \
+      'Currently only support PASCAL_VOC or KITTI dataset'
+
+  cfg = edict()
+
+  # Dataset used to train/val/test model. Now support PASCAL_VOC or KITTI
+  cfg.DATASET = dataset.upper()
+
+  if cfg.DATASET == 'PASCAL_VOC':
+    # object categories to classify
+    cfg.CLASS_NAMES = ('aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus',
+                       'car', 'cat', 'chair', 'cow', 'diningtable', 'dog',
+                       'horse', 'motorbike', 'person', 'pottedplant', 'sheep',
+                       'sofa', 'train', 'tvmonitor')
+  elif cfg.DATASET == 'KITTI':
+    cfg.CLASS_NAMES = ('car', 'pedestrian', 'cyclist')
+
+  # number of categories to classify
+  cfg.CLASSES = len(cfg.CLASS_NAMES)    
+
+  # ROI pooling output width
+  cfg.GRID_POOL_WIDTH = 7
+
+  # ROI pooling output height
+  cfg.GRID_POOL_HEIGHT = 7
+
+  # parameter used in leaky ReLU
+  cfg.LEAKY_COEF = 0.1
+
+  # Probability to keep a node in dropout
+  cfg.KEEP_PROB = 0.5
+
+  # image width
+  cfg.IMAGE_WIDTH = 224
+
+  # image height
+  cfg.IMAGE_HEIGHT = 224
+
+  # anchor box, array of [cx, cy, w, h]. To be defined later
+  cfg.ANCHOR_BOX = []
+
+  # number of anchor boxes
+  cfg.ANCHORS = len(cfg.ANCHOR_BOX)
+
+  # number of anchor boxes per grid
+  cfg.ANCHOR_PER_GRID = -1
+
+  # batch size
+  cfg.BATCH_SIZE = 20
+
+  # Only keep boxes with probability higher than this threshold
+  cfg.PROB_THRESH = 0.005
+
+  # Only plot boxes with probability higher than this threshold
+  cfg.PLOT_PROB_THRESH = 0.5
+
+  # Bounding boxes with IOU larger than this are going to be removed
+  cfg.NMS_THRESH = 0.2
+
+  # Pixel mean values (BGR order) as a (1, 1, 3) array. Below is the BGR mean
+  # of VGG16
+  cfg.BGR_MEANS = np.array([[[103.939, 116.779, 123.68]]])
+
+  # loss coefficient for confidence regression
+  cfg.LOSS_COEF_CONF = 1.0
+
+  # loss coefficient for classification regression
+  cfg.LOSS_COEF_CLASS = 1.0
+
+  # loss coefficient for bounding box regression
+  cfg.LOSS_COEF_BBOX = 10.0
+
+  # reduce step size after this many steps
+  cfg.DECAY_STEPS = 10000
+
+  # multiply the learning rate by this factor
+  cfg.LR_DECAY_FACTOR = 0.1
+
+  # learning rate
+  cfg.LEARNING_RATE = 0.005
+
+  # momentum
+  cfg.MOMENTUM = 0.9
+
+  # weight decay
+  cfg.WEIGHT_DECAY = 0.0005
+
+  # wether to load pre-trained model
+  cfg.LOAD_PRETRAINED_MODEL = True
+
+  # path to load the pre-trained model
+  cfg.PRETRAINED_MODEL_PATH = ''
+
+  # print log to console in debug mode
+  cfg.DEBUG_MODE = False
+
+  # a small value used to prevent numerical instability
+  cfg.EPSILON = 1e-16
+
+  # threshold for safe exponential operation
+  cfg.EXP_THRESH=1.0
+
+  # gradients with norm larger than this is going to be clipped.
+  cfg.MAX_GRAD_NORM = 10.0
+
+  # Whether to do data augmentation
+  cfg.DATA_AUGMENTATION = False
+
+  # The range to randomly shift the image widht
+  cfg.DRIFT_X = 0
+
+  # The range to randomly shift the image height
+  cfg.DRIFT_Y = 0
+
+  # Whether to exclude images harder than hard-category. Only useful for KITTI
+  # dataset.
+  cfg.EXCLUDE_HARD_EXAMPLES = True
+
+  # small value used in batch normalization to prevent dividing by 0. The
+  # default value here is the same with caffe's default value.
+  cfg.BATCH_NORM_EPSILON = 1e-5
+
+  # number of threads to fetch data
+  cfg.NUM_THREAD = 4
+
+  # capacity for FIFOQueue
+  cfg.QUEUE_CAPACITY = 100
+
+  # indicate if the model is in training mode
+  cfg.IS_TRAINING = False
+
+  return cfg