Mrcnn

.gitmodules

@@ -4,6 +4,8 @@
 [submodule "lmnet/third_party/coco"]
 	path = lmnet/third_party/coco
 	url = https://github.com/pdollar/coco.git
+	ignore = dirty
 [submodule "lmnet/third_party/tensorflow-on-arm"]
 	path = lmnet/third_party/tensorflow-on-arm
 	url = https://github.com/lhelontra/tensorflow-on-arm.git
+

dlk/argmax.py

@@ -0,0 +1,9 @@
+import sys
+import numpy as np
+array=[]
+for line in sys.stdin:
+    if 'Output' in line or 'test' in line:
+        continue
+    array.extend([float(n) for n in line.strip('[] \r\n').split()])
+print(array)
+print(np.argmax(array))

lmnet/.gitignore

@@ -0,0 +1,5 @@
+dataset/
+images/
+log
+outputs/
+saved/

lmnet/README.md

@@ -213,7 +213,7 @@ Currently, this feature is implemented for quantized classification of cifar10 a
 If you don't want to use this feature, set TRAIN_VALIDATION_SAVING_SIZE to zero.  
 
 The KEEP_CHECKPOINT_MAX is equivalent to 'max_to_keep' of tensorflow train.Saver parameter which indicates the maximum number of recent checkpoint files to keep. As new files are created, older files are deleted. 
-If None or 0, no checkpoints are deleted from the filesystem but only the last one is kept in the checkpoint file. Defaults to 5 (that is, the 5 most recent checkpoint files are kept.)
+If None or 0, no checkpoints but only the last one is kept in the checkpoint file. Defaults to 5 (that is, the 5 most recent checkpoint files are kept.)
 
 To apply this feature to another dataset, the dataset file should define another available subset called train_validation_saving, which is split from the original train dataset in the dataset file. Also a dataset parameter TRAIN_VALIDATION_SAVING_SIZE should be included in the config file.
 
@@ -225,7 +225,7 @@ Exporting a trained model to proto buffer files and meta config yaml.
 In the case with `images` option, create each layer output value npy files in `export/{restore_path}/{image_size}/{image_name}/**.npy` for debug.
 
 * Load config file from saved experiment dir.
-* Export config file to yaml. See also [Config specification](docs/specification/config.md).
+* Export config file to yaml. See also [Config specification](../docs/specification/config.md).
   * `config.yaml` can be used for training and evaluation in python. i.e. [classification.yaml](configs/example/classification.yaml) is exported from [classification.py](configs/example/classification.py)
   * `meta.yaml` include only few parameter for application such as demo. i.e. [classification_meta.yaml](configs/example/classification_meta.yaml) is exported from [classification.py](configs/example/classification.py)
 * Save the model protocol buffer files (tf) for DLK converter.
@@ -328,7 +328,7 @@ Save the predictions npy, json, images results to output dir.
 * json: `{output_dir}/json/{batch number}.json`
 * images: `{output_dir}/images/{some type}/{input image file name}`
 
-The output predictions Tensor(npy) and json format depends on task type. Plsease see [Output Data Specification](docs/specification/output_data.md).
+The output predictions Tensor(npy) and json format depends on task type. Plsease see [Output Data Specification](../docs/specification/output_data.md).
 
 ```
 python3 executor/predict.py -h

lmnet/configs/core/instance_segmentation/lm_resnet_quantize_imagenet.py

@@ -0,0 +1,116 @@
+# -*- coding: utf-8 -*-
+# Copyright 2019 The Blueoil Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# =============================================================================
+from easydict import EasyDict
+
+from lmnet.common import Tasks
+from lmnet.networks.instance_segmentation_old.resnet18_backbone import ResnetQuantize
+#from lmnet.networks.classification.lm_resnet import LmResnetQuantize
+from lmnet.datasets.ilsvrc_2012 import Ilsvrc2012
+from lmnet.data_processor import Sequence
+from lmnet.pre_processor import (
+    Resize,
+    DivideBy255,
+)
+from lmnet.data_augmentor import (
+    Brightness,
+    Color,
+    Contrast,
+    Crop,
+    FlipLeftRight,
+    Hue,
+)
+
+from lmnet.quantizations import (
+    binary_mean_scaling_quantizer,
+    linear_mid_tread_half_quantizer,
+)
+
+IS_DEBUG = False
+
+NETWORK_CLASS = ResnetQuantize
+DATASET_CLASS = Ilsvrc2012
+
+IMAGE_SIZE = [224, 224]
+BATCH_SIZE = 4
+DATA_FORMAT = "NHWC"
+TASK = Tasks.CLASSIFICATION
+CLASSES = DATASET_CLASS.classes
+
+# MAX_STEPS = 2000000
+# SAVE_CHECKPOINT_STEPS = 50000
+# KEEP_CHECKPOINT_MAX = 5
+# TEST_STEPS = 50000
+# SUMMARISE_STEPS = 50000
+# USE_RECOVERY = True
+# pretrain
+IS_PRETRAIN = False
+PRETRAIN_VARS = []
+PRETRAIN_DIR = ""
+PRETRAIN_FILE = ""
+
+
+# for debug
+MAX_STEPS = 100
+# BATCH_SIZE = 31
+SAVE_CHECKPOINT_STEPS = 10
+KEEP_CHECKPOINT_MAX = 5
+TEST_STEPS = 10
+SUMMARISE_STEPS = 2
+# IS_DEBUG = True
+
+PRE_PROCESSOR = Sequence([
+    Resize(size=IMAGE_SIZE),
+    DivideBy255()
+])
+POST_PROCESSOR = None
+
+NETWORK = EasyDict()
+# NETWORK.OPTIMIZER_CLASS = tf.train.MomentumOptimizer
+# NETWORK.OPTIMIZER_KWARGS = {"momentum": 0.9}
+# NETWORK.LEARNING_RATE_FUNC = tf.train.piecewise_constant
+# NETWORK.LEARNING_RATE_KWARGS = {
+#     "values": [0.1, 0.01, 0.001, 0.0001],
+#     "boundaries": [40000, 60000, 80000],
+# }
+NETWORK.IMAGE_SIZE = IMAGE_SIZE
+NETWORK.BATCH_SIZE = BATCH_SIZE
+NETWORK.DATA_FORMAT = DATA_FORMAT
+# NETWORK.WEIGHT_DECAY_RATE = 0.0001
+NETWORK.ACTIVATION_QUANTIZER = linear_mid_tread_half_quantizer
+NETWORK.ACTIVATION_QUANTIZER_KWARGS = {
+    'bit': 2,
+    'max_value': 2
+}
+NETWORK.WEIGHT_QUANTIZER = binary_mean_scaling_quantizer
+NETWORK.WEIGHT_QUANTIZER_KWARGS = {}
+
+# dataset
+DATASET = EasyDict()
+DATASET.BATCH_SIZE = BATCH_SIZE
+DATASET.DATA_FORMAT = DATA_FORMAT
+DATASET.PRE_PROCESSOR = PRE_PROCESSOR
+DATASET.AUGMENTOR = Sequence([
+    # Pad(2),
+    Crop(size=IMAGE_SIZE),
+    # FlipLeftRight(),
+    # Crop(size=IMAGE_SIZE, resize=256),
+    FlipLeftRight(),
+    # Brightness((0.75, 1.25)),
+    # Color((0.75, 1.25)),
+    # Contrast((0.75, 1.25)),
+    # Hue((-10, 10)),
+])
+DATASET.ENABLE_PREFETCH = True

lmnet/configs/example/classification.py

@@ -9,79 +9,77 @@
 #
 # Unless required by applicable law or agreed to in writing, software
 # distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# WITH WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # =============================================================================
 from easydict import EasyDict
 import tensorflow as tf
 
 from lmnet.common import Tasks
+from lmnet.networks.classification.lmnet_v1 import LmnetV1Quantize
 from lmnet.datasets.cifar10 import Cifar10
-from lmnet.networks.classification.lmnet_v0 import LmnetV0Quantize
 from lmnet.data_processor import Sequence
 from lmnet.pre_processor import (
     Resize,
-    PerImageStandardization,
+    DivideBy255,
 )
 from lmnet.data_augmentor import (
-    Brightness,
-    Color,
-    Contrast,
+    Crop,
     FlipLeftRight,
-    Hue,
+    Pad,
 )
 from lmnet.quantizations import (
     binary_mean_scaling_quantizer,
     linear_mid_tread_half_quantizer,
 )
 
-IS_DEBUG = True
+IS_DEBUG = False
 
-NETWORK_CLASS = LmnetV0Quantize
+NETWORK_CLASS = LmnetV1Quantize
 DATASET_CLASS = Cifar10
 
-IMAGE_SIZE = [28, 28]
-BATCH_SIZE = 32
+IMAGE_SIZE = [32, 32]
+BATCH_SIZE = 100
 DATA_FORMAT = "NHWC"
 TASK = Tasks.CLASSIFICATION
 CLASSES = DATASET_CLASS.classes
 
+MAX_STEPS = 100000
+SAVE_CHECKPOINT_STEPS = 1000
 KEEP_CHECKPOINT_MAX = 5
-MAX_EPOCHS = 1  # MAX_STEPS = 1561
-SAVE_CHECKPOINT_STEPS = 100
-TEST_STEPS = 100
-SUMMARISE_STEPS = 10
-
-
+TEST_STEPS = 1000
+SUMMARISE_STEPS = 100
 # pretrain
 IS_PRETRAIN = False
-PRETRAIN_VARS = [
-    "conv1/kernel:",
-    "conv1/bias:",
-    "conv2/kernel:",
-    "conv2/bias:",
-    "conv3/kernel:",
-    "conv3/bias:",
-    "conv4/kernel:",
-    "conv4/bias:",
-    "conv5/kernel:",
-    "conv5/bias:",
-    "conv6/kernel:",
-    "conv6/bias:",
-]
-PRETRAIN_DIR = "saved/lmnet_0.01_caltech101/checkpoints"
-PRETRAIN_FILE = "save.ckpt-99001"
+PRETRAIN_VARS = []
+PRETRAIN_DIR = ""
+PRETRAIN_FILE = ""
+
+
+# for debug
+# MAX_STEPS = 10
+# BATCH_SIZE = 31
+# SAVE_CHECKPOINT_STEPS = 2
+# TEST_STEPS = 10
+# SUMMARISE_STEPS = 2
+# IS_DEBUG = True
 
 PRE_PROCESSOR = Sequence([
     Resize(size=IMAGE_SIZE),
-    PerImageStandardization()
+    DivideBy255()
 ])
 POST_PROCESSOR = None
 
 NETWORK = EasyDict()
-NETWORK.OPTIMIZER_CLASS = tf.train.AdamOptimizer
-NETWORK.OPTIMIZER_KWARGS = {"learning_rate": 0.001}
+NETWORK.OPTIMIZER_CLASS = tf.train.MomentumOptimizer
+NETWORK.OPTIMIZER_KWARGS = {"momentum": 0.9}
+NETWORK.LEARNING_RATE_FUNC = tf.train.piecewise_constant
+step_per_epoch = int(50000 / BATCH_SIZE)
+NETWORK.LEARNING_RATE_KWARGS = {
+    "values": [0.01, 0.001, 0.0001, 0.00001],
+    "boundaries": [step_per_epoch * 50, step_per_epoch * 100, step_per_epoch * 150],
+}
 NETWORK.IMAGE_SIZE = IMAGE_SIZE
 NETWORK.BATCH_SIZE = BATCH_SIZE
 NETWORK.DATA_FORMAT = DATA_FORMAT
@@ -100,9 +98,8 @@
 DATASET.DATA_FORMAT = DATA_FORMAT
 DATASET.PRE_PROCESSOR = PRE_PROCESSOR
 DATASET.AUGMENTOR = Sequence([
+    Pad(2),
+    Crop(size=IMAGE_SIZE),
     FlipLeftRight(),
-    Brightness((0.75, 1.25)),
-    Color((0.75, 1.25)),
-    Contrast((0.75, 1.25)),
-    Hue((-10, 10)),
 ])
+DATASET.TRAIN_VALIDATION_SAVING_SIZE = 5000

lmnet/executor/train.py

@@ -156,14 +156,20 @@ def start_training(config):
             )
         )
     else:
+        session_config = tf.ConfigProto(
+            gpu_options=tf.GPUOptions(
+                allow_growth=True
+                # visible_device_list=str(hvd.local_rank())
+            )
+        )
         # TODO(wakisaka): For debug.
         # session_config = tf.ConfigProto(
         #     gpu_options=tf.GPUOptions(
         #         allow_growth=True,
         #         per_process_gpu_memory_fraction=0.1
         #     )
         # )
-        session_config = tf.ConfigProto()  # tf.ConfigProto(log_device_placement=True)
+        # session_config = tf.ConfigProto()  # tf.ConfigProto(log_device_placement=True)
     # TODO(wakisaka): XLA JIT
     # session_config.graph_options.optimizer_options.global_jit_level = tf.OptimizerOptions.ON_1
 

lmnet/get-img-from-cifar.py

@@ -0,0 +1,22 @@
+from six.moves import cPickle as pickle
+from  PIL import Image
+import numpy as np
+import scipy.misc
+
+f = open('dataset/CIFAR_10/cifar-10-batches-py/data_batch_1', 'rb')
+
+tupled_data= pickle.load(f, encoding='bytes')
+
+f.close()
+
+img = tupled_data[b'data']
+label = tupled_data[b'labels']
+
+for i in range(10):
+    single_img = np.array(img[i])
+    single_label = str(label[i])
+    single_img_reshaped = np.transpose(np.reshape(single_img,(3, 32,32)), (1,2,0))
+    scipy.misc.imsave(f'images/{i}-{single_label}.jpg',single_img_reshaped)
+    #single_img_reshaped = single_img.reshape(32,32,3)
+
+#plt.imshow(single_img_reshaped)