Add Pytorch training code

Author: isarandi

demo.ipynb

@@ -0,0 +1,68 @@
+# run: conda env create --file environment.yml
+name: py10
+channels:
+  - defaults
+dependencies:
+  - python=3.10
+  - pip
+  - cachetools
+  - av<9
+  - cython
+  - ezc3d
+  - ffmpeg
+  - imageio
+  - matplotlib
+  - mkl
+  - trimesh
+  - numba
+  - numpy<2.0
+  - libiconv
+  - pandas
+  - Pillow
+  - scikit-image
+  - scikit-learn
+  - scikit-sparse
+  - tqdm
+  - conda-forge::mayavi>=4.8
+  - conda-forge::PySide6==6.7.1
+  - conda-forge::ffmpeg-python
+  - pip:
+    - distutils
+    - addict
+    - tensorflow==2.15
+    - tensorflow-hub
+    - chumpy
+    - embreex
+    - einops
+    - imageio-ffmpeg
+    - importlib_resources
+    - jpeg4py
+    - more_itertools
+    - opencv-python
+    - pyrender
+    - tetgen
+    - pymeshfix
+    - -e $CODE_DIR/fleras
+    - -e $CODE_DIR/affine_combining_autoencoder
+    - -e $CODE_DIR/cameralib
+    - -e $CODE_DIR/boxlib
+    - -e $CODE_DIR/poseviz
+    - -e $CODE_DIR/smplfitter
+    - -e $CODE_DIR/simplepyutils
+    - -e $CODE_DIR/rlemasklib
+    - -e $CODE_DIR/barecat3
+    - -e $CODE_DIR/tensorflow-inputs
+    - -e $CODE_DIR/posepile
+    - -e $CODE_DIR/blendipose
+    - -e $CODE_DIR/blendify
+    - -e $CODE_DIR/nlf-pipeline
+#    - git+https://github.com/isarandi/fleras.git
+#    - git+https://github.com/isarandi/cameralib.git
+#    - git+https://github.com/isarandi/boxlib.git
+#    - git+https://github.com/isarandi/poseviz.git
+#    - git+https://github.com/isarandi/smplfitter.git
+#    - git+https://github.com/isarandi/simplepyutils.git
+#    - git+https://github.com/isarandi/tf-parallel-map.git
+#    - git+https://github.com/isarandi/BareCat.git
+#    - git+https://github.com/isarandi/rlemasklib.git
+#    - git+https://github.com/isarandi/tensorflow-inputs.git

environment_comfortable_py10.yml

@@ -0,0 +1,118 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# This was tested on Ubuntu 18.04 and 22.04
+sudo apt update
+sudo apt install build-essential --yes wget curl gfortran git ncurses-dev libncursesw5-dev unzip tar libxcb-xinerama0
+
+################
+# This part is ONLY needed if you'll work with the original Human3.6M files.
+# For this, we need to install the [CDF library](https://cdf.gsfc.nasa.gov/) since the annotations are in CDF files.
+# We will use [SpacePy](https://spacepy.github.io/) as a wrapper, which in turn depends on this CDF library.
+CDF_VERSION=39_0
+wget "https://spdf.gsfc.nasa.gov/pub/software/cdf/dist/cdf${CDF_VERSION}/linux/cdf${CDF_VERSION}-dist-cdf.tar.gz"
+tar xf "cdf${CDF_VERSION}-dist-cdf.tar.gz"
+rm "cdf${CDF_VERSION}-dist-cdf.tar.gz"
+pushd "cdf${CDF_VERSION}-dist"
+make OS=linux ENV=gnu CURSES=no FORTRAN=no UCOPTIONS=-O2 SHARED=yes -j"$(nproc)" all
+# If you have sudo rights, simply run `sudo make install`. If you have no `sudo` rights, set the env var
+# `CDF_LIB` to `cdf${CDF_VERSION}-dist/src/lib`
+# add the export line to ~/.bashrc for permanent effect, or use GNU Stow.
+# mv src ~/.local/stow/cdf-${CDF_VERSION}
+# stow cdf-${CDF_VERSION}
+# The following will work temporarily:
+export CDF_LIB=$PWD/src/lib
+popd
+####################
+
+# Micromamba is the simplest way to install the dependencies
+# If you don't have it yet, install it as follows:
+export MAMBA_ROOT_PREFIX=$HOME/micromamba
+mkdir -p $MAMBA_ROOT_PREFIX
+curl -Ls https://micro.mamba.pm/api/micromamba/linux-64/latest | tar -xvj -C $MAMBA_ROOT_PREFIX bin/micromamba
+# Execute and add to ~/.bashrc the following two lines
+export MAMBA_ROOT_PREFIX=$HOME/micromamba
+eval "$($MAMBA_ROOT_PREFIX/bin/micromamba shell hook -s posix)"
+#
+
+# Create a new environment and install the dependencies
+envsubst < environment_comfortable_py10.yml > env_subst.yml
+micromamba env create --name=nlf --file=env_subst.yml -y
+micromamba activate nlf
+pip install --no-build-isolation git+https://github.com/spacepy/spacepy
+
+
+micromamba install -y -c conda-forge \
+    cachetools \
+    cython \
+    ezc3d \
+    ffmpeg \
+    imageio \
+    matplotlib \
+    mkl \
+    trimesh \
+    numba \
+    "numpy<2.0" \
+    libiconv \
+    pandas \
+    pillow \
+    scikit-image \
+    scikit-learn \
+    scikit-sparse \
+    tqdm \
+    conda-forge::mayavi>=4.8 \
+    conda-forge::PySide6==6.7.1 \
+    conda-forge::ffmpeg-python
+
+pip install \
+    setuptools \
+    addict \
+    tensorflow==2.15 \
+    tensorflow-hub \
+    torch \
+    torchvision \
+    torchdata \
+    chumpy \
+    embreex \
+    einops \
+    imageio-ffmpeg \
+    importlib_resources \
+    jpeg4py \
+    more_itertools \
+    opencv-python \
+    pyrender \
+    tetgen \
+    pymeshfix
+
+
+# Optional:
+# Install libjpeg-turbo for faster JPEG decoding.
+# wget https://sourceforge.net/projects/libjpeg-turbo/files/2.0.5/libjpeg-turbo-2.0.5.tar.gz
+# Then compile it.
+# Or use the repo:
+# git clone https://github.com/libjpeg-turbo/libjpeg-turbo.git
+# cd libjpeg-turbo
+# PACKAGE_NAME=libjpeg-turbo
+# TARGET=$HOME/.local
+# sudo apt install nasm
+# cmake -DCMAKE_INSTALL_PREFIX="$TARGET"  -DCMAKE_POSITION_INDEPENDENT_CODE=ON -G"Unix Makefiles" .
+# TEMP_DESTDIR=$(mktemp --directory --tmpdir="$STOW_DIR")
+# make -j "$(nproc)" install DESTDIR="$TEMP_DESTDIR"
+# mv -T "$TEMP_DESTDIR/$TARGET" "$STOW_DIR/$PACKAGE_NAME"
+# rm -rf "$TEMP_DESTDIR"
+# stow "$PACKAGE_NAME" --target="$TARGET"
+
+ - posepile image barecat
+ - anno barecats (4)
+ - code projects
+ - micromamba, env install
+ - cuda, cudnn copy
+ - set up project initializer bashrc command that sets envvars, activates env, cd to project
+ - wacv23_models
+ - stuff from $DATA_ROOT/cache
+ - projects/localizerfields
+
+
+ #sudo apt-get install libavformat-dev libavdevice-dev
+#pip install av --no-binary av
+# https://stackoverflow.com/questions/72604912/cant-show-image-with-opencv-when-importing-av

install_dependencies.sh

@@ -1,12 +1,12 @@
+import cv2
 import numpy as np
 from simplepyutils import FLAGS
 
-import nlf.tf.augmentation.color as coloraug
-from nlf.tf import improc, util
-from nlf.tf.augmentation import voc_loader
-from nlf.tf.augmentation.border import augment_border
-from nlf.tf.util import TRAIN
-import cv2
+import nlf.common.augmentation.color as coloraug
+from nlf.common import improc, util
+from nlf.common.util import TRAIN
+from nlf.common.augmentation import voc_loader
+from nlf.common.augmentation.border import augment_border
 
 
 def augment_appearance(im, learning_phase, occlude_prob, border_value, rng):
@@ -46,8 +46,9 @@ def augment_appearance(im, learning_phase, occlude_prob, border_value, rng):
             if FLAGS.jpeg_aug_prob:
                 im = jpeg_artifact(im, jpeg_rng)
 
-        if FLAGS.augment_border and border_value is not None:
-            im = augment_border(im, border_value, border_rng)
+        if FLAGS.augment_border_prob and border_value is not None:
+            if border_rng.uniform(0.0, 1.0) < FLAGS.augment_border_prob:
+                im = augment_border(im, border_value, border_rng)
 
     return im
 

nlf/pt/scripts/__init__.py nlf/common/__init__.pynlf/pt/scripts/__init__.py renamed to nlf/common/__init__.py

@@ -1,14 +1,16 @@
 import functools
+import os.path as osp
 
+import barecat
 import cameralib
 import cv2
 import numpy as np
 import simplepyutils as spu
 from posepile.paths import DATA_ROOT
 from simplepyutils import FLAGS
-import barecat
-from nlf.tf import improc, util
-import os.path as osp
+
+from nlf.common import improc, util
+
 
 @functools.lru_cache()
 def get_inria_holiday_background_paths():

nlf/tf/augmentation/__init__.py nlf/common/augmentation/__init__.pynlf/tf/augmentation/__init__.py renamed to nlf/common/augmentation/__init__.py

@@ -1,6 +1,5 @@
 import cv2
 import numpy as np
-from simplepyutils import FLAGS
 
 
 def augment_border(im, border_value, rng):
@@ -11,15 +10,15 @@ def augment_border(im, border_value, rng):
     size = 0.08
     angle = np.deg2rad(8)
 
-    def get_angle():
+    def random_angle():
         if rng.uniform(0, 1) < 0.9:
             return rng.uniform(-angle, angle)
         else:
             return rng.uniform(-angle, angle) * 1.5
 
     # top:
     if rng.uniform(0, 1) < p:
-        alpha = get_angle()
+        alpha = random_angle()
         d = rng.uniform(0, h * size)
         y1 = - np.tan(alpha) * h / 2 + d
         y2 = + np.tan(alpha) * h / 2 + d
@@ -30,7 +29,7 @@ def get_angle():
 
     # bottom:
     if rng.uniform(0, 1) < p:
-        alpha = get_angle()
+        alpha = random_angle()
         d = rng.uniform(0, h * size)
         y1 = h - np.tan(alpha) * h / 2 - d
         y2 = h + np.tan(alpha) * h / 2 - d
@@ -41,7 +40,7 @@ def get_angle():
 
     # left:
     if rng.uniform(0, 1) < p:
-        alpha = get_angle()
+        alpha = random_angle()
         d = rng.uniform(0, w * size)
         x1 = - np.tan(alpha) * w / 2 + d
         x2 = + np.tan(alpha) * w / 2 + d
@@ -52,7 +51,7 @@ def get_angle():
 
     # right:
     if rng.uniform(0, 1) < p:
-        alpha = get_angle()
+        alpha = random_angle()
         d = rng.uniform(0, w * size)
         x1 = w - np.tan(alpha) * w / 2 - d
         x2 = w + np.tan(alpha) * w / 2 - d

nlf/tf/augmentation/appearance.py nlf/common/augmentation/appearance.pynlf/tf/augmentation/appearance.py renamed to nlf/common/augmentation/appearance.py

@@ -0,0 +1,104 @@
+import cv2
+import numpy as np
+import numba
+
+
+def augment_color(im, rng, out_dtype=None):
+    if out_dtype is None:
+        out_dtype = im.dtype
+
+    if im.dtype == np.uint8:
+        im = cv2.divide(im, (255, 255, 255, 255), dtype=cv2.CV_32F)
+
+    augmentation_functions = [
+        augment_brightness,
+        augment_contrast,
+        augment_hue,
+        augment_saturation,
+    ]
+    rng.shuffle(augmentation_functions)
+
+    colorspace = 'rgb'
+    for fn in augmentation_functions:
+        colorspace = fn(im, colorspace, rng)
+
+    if colorspace != 'rgb':
+        cv2.cvtColor(im, cv2.COLOR_HSV2RGB, dst=im)
+
+    np.clip(im, 0, 1, out=im)
+
+    if out_dtype == np.uint8:
+        return (im * 255).astype(np.uint8)
+    else:
+        return im
+
+
+def augment_brightness(im, in_colorspace, rng):
+    if in_colorspace != 'rgb':
+        cv2.cvtColor(im, cv2.COLOR_HSV2RGB, dst=im)
+
+    im += rng.uniform(-0.125, 0.125)
+    return 'rgb'
+
+
+def augment_contrast(im, in_colorspace, rng):
+    if in_colorspace != 'rgb':
+        cv2.cvtColor(im, cv2.COLOR_HSV2RGB, dst=im)
+    # im -= 0.5
+    # im *= rng.uniform(0.5, 1.5)
+    # im += 0.5
+    _augment_contrast_nb(im, rng.uniform(0.5, 1.5))
+    return 'rgb'
+
+
+@numba.njit(cache=True)
+def _augment_contrast_nb(im, factor):
+    im_flat = im.reshape(-1)
+    factor32 = np.float32(factor)
+    offset = np.float32(-0.5) * factor32 + np.float32(0.5)
+    for i in range(im_flat.shape[0]):
+        im_flat[i] = im_flat[i] * factor32 + offset
+
+
+def augment_hue(im, in_colorspace, rng):
+    if in_colorspace != 'hsv':
+        np.clip(im, 0, 1, out=im)
+        cv2.cvtColor(im, cv2.COLOR_RGB2HSV, dst=im)
+    # hue = im[:, :, 0]
+    # hue += rng.uniform(-72, 72)
+    # hue[hue < 0] += 360
+    # hue[hue > 360] -= 360
+    _augment_hue_nb(im, rng.uniform(-72, 72))
+    return 'hsv'
+
+
+@numba.njit(cache=True)
+def _augment_hue_nb(im, offset):
+    im_flat = im.reshape(-1, 3)
+    for i in range(im_flat.shape[0]):
+        im_flat[i, 0] += offset
+        if im_flat[i, 0] < 0:
+            im_flat[i, 0] += 360
+        elif im_flat[i, 0] > 360:
+            im_flat[i, 0] -= 360
+
+
+def augment_saturation(im, in_colorspace, rng):
+    if in_colorspace != 'hsv':
+        np.clip(im, 0, 1, out=im)
+        cv2.cvtColor(im, cv2.COLOR_RGB2HSV, dst=im)
+
+    # saturation = im[:, :, 1]
+    # saturation *= rng.uniform(0.5, 1.5)
+    # saturation[saturation > 1] = 1
+    _augment_saturation_nb(im, rng.uniform(0.5, 1.5))
+    return 'hsv'
+
+
+@numba.njit(cache=True)
+def _augment_saturation_nb(im, factor):
+    im_flat = im.reshape(-1, 3)
+    for i in range(im_flat.shape[0]):
+        im_flat[i, 1] *= factor
+        if im_flat[i, 1] > 1:
+            im_flat[i, 1] = 1