major refactoring

Author: chrischoy

.gitignore

@@ -1,5 +1,12 @@
+# Python cache
 *.pyc
-*.mexa64
 
+# OS X temporary metadata
+._*
+*.DS_Store
+
+# Wiki page
 wiki/
+
+# Results
 output/

.pylintrc

@@ -0,0 +1,23 @@
+[MESSAGES CONTROL]
+disable=missing-docstring,invalid-name,wildcard-import,unused-wildcard-import,bad-builtin,no-self-use,locally-disabled
+
+[TYPECHECK]
+ignored-classes=numpy,theano
+
+[DESIGN]
+# Maximum number of arguments for function / method
+max-args=20
+# Maximum number of locals for function / method body
+max-locals=30
+# Maximum number of return / yield for function / method body
+max-returns=10
+# Maximum number of branch for function / method body
+max-branches=12
+# Maximum number of statements in function / method body
+max-statements=200
+# Maximum number of attributes for a class (see R0902).
+max-attributes=100
+# Maximum number of public methods for a class (see R0904).
+max-public-methods=200
+# Maximum number of boolean expressions in a if statement
+max-bool-expr=10

README.md

@@ -5,7 +5,7 @@
 1. Download `ShapeNetCore` dataset
 2. Set the `MODEL_ROOT_PATH` in `lib/config.py` to the extracted `ShapeNetCore`
 3. Download model lists from the website. Some models do not have any faces.
-4. Generate dataset `json` file by running `python tools/generate_category_list.py`
+4. Generate dataset `json` file by running `python tools/gen_category_list.py`
 5. Voxelize all the models by running `python tools/voxelize_models.py`
 6. Render all the models by runnning `python tools/render_models.py`. To run this step, you have to setup `blender`.
 7. Set `cfg.DIR.MODEL_PATH`, `cfg.DIR.RENDERING_PATH` and `cfg.DIR.VOXEL_PATH` in `lib/config.py` accoringly

_init_paths.py

@@ -0,0 +1,14 @@
+""" Add required paths to the PYTHONPATH """
+
+import os.path as osp
+import sys
+
+
+def add_path(path):
+    if path not in sys.path:
+        sys.path.insert(0, path)
+
+this_dir = osp.dirname(__file__)
+
+# Add lib to PYTHONPATH
+add_path(osp.join(this_dir, 'lib'))

experiments/scripts/mv_deep_res_gru_net_3x3x3.sh

@@ -18,23 +18,24 @@ echo Logging output to "$LOG"
 
 export THEANO_FLAGS="floatX=float32,device=gpu,assert_no_cpu_op='raise'"
 
-python3 ./tools/train.py \
-       --cfg ./experiments/cfgs/shapenet_1000.yaml \
-       --cfg ./experiments/cfgs/random_crop.yaml \
-       --cfg ./experiments/cfgs/no_random_background.yaml \
-       --cfg ./experiments/cfgs/max_5_views.yaml \
-       --cfg ./experiments/cfgs/local_shapenet.yaml \
-       --cfg ./experiments/cfgs/batch_size_24.yaml \
-       --out $OUT_PATH \
-       --model $NET_NAME \
-       ${*:1}
+python main.py \
+      --batch-size 24 \
+      --cfg ./experiments/cfgs/shapenet_1000.yaml \
+      --cfg ./experiments/cfgs/random_crop.yaml \
+      --cfg ./experiments/cfgs/no_random_background.yaml \
+      --cfg ./experiments/cfgs/max_5_views.yaml \
+      --cfg ./experiments/cfgs/local_shapenet.yaml \
+      --out $OUT_PATH \
+      --model $NET_NAME \
+      ${*:1}
 
-python3 ./tools/test.py \
-       --cfg ./experiments/cfgs/shapenet_1000.yaml \
-       --cfg ./experiments/cfgs/no_random_background.yaml \
-       --cfg ./experiments/cfgs/max_5_views.yaml \
-       --cfg ./experiments/cfgs/local_shapenet.yaml \
-       --cfg ./experiments/cfgs/batch_size_24.yaml \
-       --weights $OUT_PATH/weights.npy \
-       --model $NET_NAME \
-       ${*:1}
+python main.py \
+      --test \
+      --batch-size 24 \
+      --cfg ./experiments/cfgs/shapenet_1000.yaml \
+      --cfg ./experiments/cfgs/no_random_background.yaml \
+      --cfg ./experiments/cfgs/max_5_views.yaml \
+      --cfg ./experiments/cfgs/local_shapenet.yaml \
+      --weights $OUT_PATH/weights.npy \
+      --model $NET_NAME \
+      ${*:1}

lib/config.py

@@ -24,7 +24,7 @@
 __C.CONST.BATCH_SIZE = 36
 __C.CONST.RECNET = 'rec_net'
 __C.CONST.WEIGHTS = '/scratch/chrischoy/mv_deep_res_gru_net_3x3x3/max_5_views_no_rnd_bg/weights.npy.39999.npy'  # when set, load the weights from the file
-__C.CONST.IMAGE_MEAN = np.load('mean.npz')['mean']
+__C.CONST.IMAGE_MEAN = np.load('./models/data/shapenet_mean.npz')['mean']
 
 #
 # Directories
@@ -54,6 +54,8 @@
 __C.TRAIN.WORKER_CAPACITY = 1000  # if use OSG, load only limited number of models at a time
 __C.TRAIN.NUM_RENDERING = 24
 __C.TRAIN.NUM_VALIDATION_ITERATIONS = 24
+__C.TRAIN.VALIDATION_FREQ = 2000
+__C.TRAIN.NAN_CHECK_FREQ = 2000
 __C.TRAIN.RANDOM_NUM_VIEWS = True  # feed in random # views if n_views > 1
 
 __C.QUEUE_SIZE = 15  # maximum number of minibatches that can be put in a data queue
@@ -62,9 +64,7 @@
 __C.TRAIN.RANDOM_CROP = True
 __C.TRAIN.PAD_X = 10
 __C.TRAIN.PAD_Y = 10
-__C.TRAIN.HUE_CHANGE = True
 __C.TRAIN.FLIP = True
-__C.TRAIN.HUE_RANGE = 0.1
 
 # For no random bg images, add random colors
 __C.TRAIN.NO_BG_COLOR_RANGE = [[225, 255], [225, 255], [225, 255]]
@@ -83,7 +83,7 @@
 # after the training almost converged since weight can increase indefinitely
 # (for cross entropy loss). Too high regularization will also hinder training.
 __C.TRAIN.WEIGHT_DECAY = 0.00005
-
+__C.TRAIN.LOSS_LIMIT = 2  # stop training if the loss exceeds the limit
 __C.TRAIN.SAVE_FREQ = 10000   # weights will be overwritten every save_freq
 __C.TRAIN.PRINT_FREQ = 40
 
@@ -93,7 +93,7 @@
 # Scale: scale 1/255
 # Center: subtract mean from the image
 ###############################################################################
-__C.TRAIN.PREPROCESSING_TYPE = 'center' # ['scale', 'center']
+__C.TRAIN.PREPROCESSING_TYPE = 'scale' # ['scale', 'center']
 
 #
 # Testing options

lib/data_augmentation.py

@@ -1,93 +1,26 @@
-#!/usr/bin/env python3
-
 import numpy as np
 from lib.config import cfg
 from PIL import Image
 
 
-def rgb_to_hsv(rgb):
-    # Translated from source of colorsys.rgb_to_hsv
-    # r,g,b should be a numpy arrays with values between 0 and 255
-    # rgb_to_hsv returns an array of floats between 0.0 and 1.0.
-    rgb = rgb.astype(np.float32)
-    hsv = np.zeros_like(rgb)
-    # in case an RGBA array was passed, just copy the A channel
-    hsv[..., 3:] = rgb[..., 3:]
-    r, g, b = rgb[..., 0], rgb[..., 1], rgb[..., 2]
-    maxc = np.max(rgb[..., :3], axis=-1)
-    minc = np.min(rgb[..., :3], axis=-1)
-    hsv[..., 2] = maxc
-    mask = maxc != minc
-    hsv[mask, 1] = (maxc - minc)[mask] / maxc[mask]
-    rc = np.zeros_like(r)
-    gc = np.zeros_like(g)
-    bc = np.zeros_like(b)
-    rc[mask] = (maxc - r)[mask] / (maxc - minc)[mask]
-    gc[mask] = (maxc - g)[mask] / (maxc - minc)[mask]
-    bc[mask] = (maxc - b)[mask] / (maxc - minc)[mask]
-    hsv[..., 0] = np.select([r == maxc, g == maxc],
-                            [bc - gc, 2.0 + rc - bc],
-                            default=4.0 + gc - rc)
-    hsv[..., 0] = (hsv[..., 0] / 6.0) % 1.0
-    return hsv
-
-
-def hsv_to_rgb(hsv):
-    # Translated from source of colorsys.hsv_to_rgb
-    # h,s should be a numpy arrays with values between 0.0 and 1.0
-    # v should be a numpy array with values between 0.0 and 255.0
-    # hsv_to_rgb returns an array of uints between 0 and 255.
-    rgb = np.empty_like(hsv)
-    rgb[..., 3:] = hsv[..., 3:]
-    h, s, v = hsv[..., 0], hsv[..., 1], hsv[..., 2]
-    i = (h * 6.0).astype('uint8')
-    f = (h * 6.0) - i
-    p = v * (1.0 - s)
-    q = v * (1.0 - s * f)
-    t = v * (1.0 - s * (1.0 - f))
-    i = i % 6
-    conditions = [s == 0.0, i == 1, i == 2, i == 3, i == 4, i == 5]
-    rgb[..., 0] = np.select(conditions, [v, q, p, p, t, v], default=v)
-    rgb[..., 1] = np.select(conditions, [v, v, v, q, p, p], default=t)
-    rgb[..., 2] = np.select(conditions, [v, p, t, v, v, q], default=p)
-    return rgb.astype('uint8')
-
-
-def shift_hue(arr, hout):
-    hsv = rgb_to_hsv(arr)
-    hsv[0, ...] += hout  # change hue
-    hsv[0, ...] = np.max(np.min(hsv[0, ...], 1), 0)
-    rgb = hsv_to_rgb(hsv)
-    return rgb
-
-
 def image_transform(img, crop_x, crop_y, crop_loc=None, color_tint=None):
     """
     Takes numpy.array img
     """
 
     # Slight translation
     if cfg.TRAIN.RANDOM_CROP and not crop_loc:
-        crop_loc = [0]*2
-        crop_loc[0] = np.random.randint(0, crop_y)  # corner position row
-        crop_loc[1] = np.random.randint(0, crop_x)  # corner position column
+        crop_loc = [np.random.randint(0, crop_y),
+                    np.random.randint(0, crop_x)]
 
     if crop_loc:
         cr, cc = crop_loc
-        height, width, channel = img.shape
+        height, width, _ = img.shape
         img_h = height - crop_y
         img_w = width - crop_x
-        img = img[cr:cr+img_h, cc:cc+img_w, :]
+        img = img[cr:cr+img_h, cc:cc+img_w]
         # depth = depth[cr:cr+img_h, cc:cc+img_w]
 
-    if cfg.TRAIN.HUE_CHANGE and not color_tint:
-        # color tint
-        color_tint = (np.random.rand() - 0.5) * cfg.TRAIN.HUE_RANGE
-
-    if color_tint:
-        # Hue change
-        img = shift_hue(img, color_tint)
-
     if cfg.TRAIN.FLIP and np.random.rand() > 0.5:
         img = img[:, ::-1, ...]
 
@@ -124,9 +57,7 @@ def add_random_background(im, background_img_fns):
 
 
 def add_random_color_background(im, color_range):
-    r = np.random.randint(color_range[0][0], color_range[0][1] + 1)
-    g = np.random.randint(color_range[1][0], color_range[1][1] + 1)
-    b = np.random.randint(color_range[2][0], color_range[2][1] + 1)
+    r, g, b = [np.random.randint(color_range[i][0], color_range[i][1] + 1) for i in range(3)]
 
     if isinstance(im, Image.Image):
         im = np.array(im)
@@ -137,15 +68,14 @@ def add_random_color_background(im, color_range):
     alpha = (np.expand_dims(im[:, :, 3], axis=2) == 0).astype(np.float)
     im = im[:, :, :3]
     bg_color = np.array([[[r, g, b]]])
-    return alpha * bg_color + im
+    return alpha * bg_color + (1 - alpha) * im
 
 
 def test(fn):
     import matplotlib.pyplot as plt
     cfg.TRAIN.RANDOM_CROP = True
-    cfg.TRAIN.HUE_CHANGE = True
     im = Image.open(fn)
-    im = np.asarray(im)[:, :, 0:3]
-    imt = image_transform(im, 50, 50)
+    im = np.asarray(im)[:, :, :3]
+    imt = image_transform(im, 10, 10)
     plt.imshow(imt)
     plt.show()

lib/data_io.py

@@ -68,11 +68,12 @@ def model_names(model_path, model_file):
         models = model_names(model_path, model_file)
         num_models = len(models)
 
-        portioned_models = models[int(num_models * dataset_portion[0]):
-                                  int(num_models * dataset_portion[1])]
+        portioned_models = models[
+            int(num_models * dataset_portion[0]):
+            int(num_models * dataset_portion[1])
+        ]
 
-        category_name_pair.extend([(cat['id'], model_id)
-                                   for model_id in portioned_models])
+        category_name_pair.extend([(cat['id'], model_id) for model_id in portioned_models])
 
     print('lib/data_io.py: model paths from %s' % (cfg.DATASET))