diff --git a/CNNs/densenet.py b/CNNs/densenet.py
new file mode 100644
index 0000000..8f70424
--- /dev/null
+++ b/CNNs/densenet.py
@@ -0,0 +1,228 @@
+"""
+DenseNet, original: https://github.com/pytorch/vision/blob/master/torchvision/models/densenet.py
+"""
+import re
+from collections import OrderedDict
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.model_zoo as model_zoo
+import torchvision.transforms as transforms
+
+from PIL import Image
+import numpy as np
+
+model_urls = {
+    'densenet121': 'https://download.pytorch.org/models/densenet121-a639ec97.pth',
+    'densenet169': 'https://download.pytorch.org/models/densenet169-b2777c0a.pth',
+    'densenet201': 'https://download.pytorch.org/models/densenet201-c1103571.pth',
+    'densenet161': 'https://download.pytorch.org/models/densenet161-8d451a50.pth',
+}
+
+
+class _DenseLayer(nn.Sequential):
+    """Basic unit of DenseBlock (using bottleneck layer) """
+    def __init__(self, num_input_features, growth_rate, bn_size, drop_rate):
+        super(_DenseLayer, self).__init__()
+        self.add_module("norm1", nn.BatchNorm2d(num_input_features))
+        self.add_module("relu1", nn.ReLU(inplace=True))
+        self.add_module("conv1", nn.Conv2d(num_input_features, bn_size*growth_rate,
+                                           kernel_size=1, stride=1, bias=False))
+        self.add_module("norm2", nn.BatchNorm2d(bn_size*growth_rate))
+        self.add_module("relu2", nn.ReLU(inplace=True))
+        self.add_module("conv2", nn.Conv2d(bn_size*growth_rate, growth_rate,
+                                           kernel_size=3, stride=1, padding=1, bias=False))
+        self.drop_rate = drop_rate
+
+    def forward(self, x):
+        new_features = super(_DenseLayer, self).forward(x)
+        if self.drop_rate > 0:
+            new_features = F.dropout(new_features, p=self.drop_rate, training=self.training)
+        return torch.cat([x, new_features], 1)
+
+class _DenseBlock(nn.Sequential):
+    """DenseBlock"""
+    def __init__(self, num_layers, num_input_features, bn_size, growth_rate, drop_rate):
+        super(_DenseBlock, self).__init__()
+        for i in range(num_layers):
+            layer = _DenseLayer(num_input_features+i*growth_rate, growth_rate, bn_size,
+                                drop_rate)
+            self.add_module("denselayer%d" % (i+1,), layer)
+
+
+class _Transition(nn.Sequential):
+    """Transition layer between two adjacent DenseBlock"""
+    def __init__(self, num_input_feature, num_output_features):
+        super(_Transition, self).__init__()
+        self.add_module("norm", nn.BatchNorm2d(num_input_feature))
+        self.add_module("relu", nn.ReLU(inplace=True))
+        self.add_module("conv", nn.Conv2d(num_input_feature, num_output_features,
+                                          kernel_size=1, stride=1, bias=False))
+        self.add_module("pool", nn.AvgPool2d(2, stride=2))
+
+
+class DenseNet(nn.Module):
+    "DenseNet-BC model"
+    def __init__(self, growth_rate=32, block_config=(6, 12, 24, 16), num_init_features=64,
+                 bn_size=4, compression_rate=0.5, drop_rate=0, num_classes=1000):
+        """
+        :param growth_rate: (int) number of filters used in DenseLayer, `k` in the paper
+        :param block_config: (list of 4 ints) number of layers in each DenseBlock
+        :param num_init_features: (int) number of filters in the first Conv2d
+        :param bn_size: (int) the factor using in the bottleneck layer
+        :param compression_rate: (float) the compression rate used in Transition Layer
+        :param drop_rate: (float) the drop rate after each DenseLayer
+        :param num_classes: (int) number of classes for classification
+        """
+        super(DenseNet, self).__init__()
+        # first Conv2d
+        self.features = nn.Sequential(OrderedDict([
+            ("conv0", nn.Conv2d(3, num_init_features, kernel_size=7, stride=2, padding=3, bias=False)),
+            ("norm0", nn.BatchNorm2d(num_init_features)),
+            ("relu0", nn.ReLU(inplace=True)),
+            ("pool0", nn.MaxPool2d(3, stride=2, padding=1))
+        ]))
+
+        # DenseBlock
+        num_features = num_init_features
+        for i, num_layers in enumerate(block_config):
+            block = _DenseBlock(num_layers, num_features, bn_size, growth_rate, drop_rate)
+            self.features.add_module("denseblock%d" % (i + 1), block)
+            num_features += num_layers*growth_rate
+            if i != len(block_config) - 1:
+                transition = _Transition(num_features, int(num_features*compression_rate))
+                self.features.add_module("transition%d" % (i + 1), transition)
+                num_features = int(num_features * compression_rate)
+
+        # final bn+ReLU
+        self.features.add_module("norm5", nn.BatchNorm2d(num_features))
+        self.features.add_module("relu5", nn.ReLU(inplace=True))
+
+        # classification layer
+        self.classifier = nn.Linear(num_features, num_classes)
+
+        # params initialization
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight)
+            elif isinstance(m, nn.BatchNorm2d):
+                nn.init.constant_(m.bias, 0)
+                nn.init.constant_(m.weight, 1)
+            elif isinstance(m, nn.Linear):
+                nn.init.constant_(m.bias, 0)
+
+    def forward(self, x):
+        features = self.features(x)
+        out = F.avg_pool2d(features, 7, stride=1).view(features.size(0), -1)
+        out = self.classifier(out)
+        return out
+
+class DenseNet_MNIST(nn.Module):
+    """DenseNet for MNIST dataset"""
+    def __init__(self, growth_rate=12, block_config=(6, 6, 6), num_init_features=16,
+                 bn_size=4, compression_rate=0.5, drop_rate=0, num_classes=10):
+        """
+        :param growth_rate: (int) number of filters used in DenseLayer, `k` in the paper
+        :param block_config: (list of 2 ints) number of layers in each DenseBlock
+        :param num_init_features: (int) number of filters in the first Conv2d
+        :param bn_size: (int) the factor using in the bottleneck layer
+        :param compression_rate: (float) the compression rate used in Transition Layer
+        :param drop_rate: (float) the drop rate after each DenseLayer
+        :param num_classes: (int) number of classes for classification
+        """
+        super(DenseNet_MNIST, self).__init__()
+        # first Conv2d
+        self.features = nn.Sequential(OrderedDict([
+            ("conv0", nn.Conv2d(1, num_init_features, kernel_size=3, stride=1, padding=1, bias=False)),
+            ("norm0", nn.BatchNorm2d(num_init_features)),
+            ("relu0", nn.ReLU(inplace=True)),
+        ]))
+
+        # DenseBlock
+        num_features = num_init_features
+        for i, num_layers in enumerate(block_config):
+            block = _DenseBlock(num_layers, num_features, bn_size, growth_rate, drop_rate)
+            self.features.add_module("denseblock%d" % (i + 1), block)
+            num_features += num_layers * growth_rate
+            if i != len(block_config) - 1:
+                transition = _Transition(num_features, int(num_features * compression_rate))
+                self.features.add_module("transition%d" % (i + 1), transition)
+                num_features = int(num_features * compression_rate)
+
+        # final bn+ReLU
+        self.features.add_module("norm5", nn.BatchNorm2d(num_features))
+        self.features.add_module("relu5", nn.ReLU(inplace=True))
+
+        # classification layer
+        self.classifier = nn.Linear(num_features, num_classes)
+
+        # params initialization
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight)
+            elif isinstance(m, nn.BatchNorm2d):
+                nn.init.constant_(m.bias, 0)
+                nn.init.constant_(m.weight, 1)
+            elif isinstance(m, nn.Linear):
+                nn.init.constant_(m.bias, 0)
+
+    def forward(self, x):
+        features = self.features(x)
+        out = F.avg_pool2d(features, 7, stride=1).view(features.size(0), -1)
+        out = self.classifier(out)
+        return out
+
+
+def densenet121(pretrained=False, **kwargs):
+    """DenseNet121"""
+    model = DenseNet(num_init_features=64, growth_rate=32, block_config=(6, 12, 24, 16),
+                     **kwargs)
+
+    if pretrained:
+        # '.'s are no longer allowed in module names, but pervious _DenseLayer
+        # has keys 'norm.1', 'relu.1', 'conv.1', 'norm.2', 'relu.2', 'conv.2'.
+        # They are also in the checkpoints in model_urls. This pattern is used
+        # to find such keys.
+        pattern = re.compile(
+            r'^(.*denselayer\d+\.(?:norm|relu|conv))\.((?:[12])\.(?:weight|bias|running_mean|running_var))$')
+        state_dict = model_zoo.load_url(http://github.com/model_urls['densenet121'])
+        for key in list(state_dict.keys()):
+            res = pattern.match(key)
+            if res:
+                new_key = res.group(1) + res.group(2)
+                state_dict[new_key] = state_dict[key]
+                del state_dict[key]
+        model.load_state_dict(state_dict)
+    return model
+
+if __name__ == "__main__":
+    densenet = densenet121(pretrained=True)
+    densenet.eval()
+
+    img = Image.open("./images/cat.jpg")
+
+    trans_ops = transforms.Compose([
+        transforms.Resize(256),
+        transforms.CenterCrop(224),
+        transforms.ToTensor(),
+        transforms.Normalize(mean=[0.485, 0.456, 0.406],
+                             std=[0.229, 0.224, 0.225])
+    ])
+
+    images = trans_ops(img).view(-1, 3, 224, 224)
+    print(images)
+    outputs = densenet(images)
+
+    _, predictions = outputs.topk(5, dim=1)
+
+    labels = list(map(lambda s: s.strip(), open("./data/imagenet/synset_words.txt").readlines()))
+    for idx in predictions.numpy()[0]:
+        print("Predicted labels:", labels[idx])
+
+
+
+
+
+
+
diff --git a/CNNs/mobilenet_v2.py b/CNNs/mobilenet_v2.py
new file mode 100644
index 0000000..23d6c0f
--- /dev/null
+++ b/CNNs/mobilenet_v2.py
@@ -0,0 +1,349 @@
+"""
+2018-11-24
+"""
+
+from collections import namedtuple
+import copy
+
+import tensorflow as tf
+
+slim = tf.contrib.slim
+
+def _make_divisible(v, divisor, min_value=None):
+    """make `v` is divided exactly by `divisor`, but keep the min_value"""
+    if min_value is None:
+        min_value = divisor
+    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
+    # Make sure that round down does not go down by more than 10%.
+    if new_v < 0.9 * v:
+        new_v += divisor
+    return new_v
+
+
+@slim.add_arg_scope
+def _depth_multiplier_func(params,
+                           multiplier,
+                           divisible_by=8,
+                           min_depth=8):
+    """get the new channles"""
+    if 'num_outputs' not in params:
+        return
+    d = params['num_outputs']
+    params['num_outputs'] = _make_divisible(d * multiplier, divisible_by,
+                                                   min_depth)
+
+def _fixed_padding(inputs, kernel_size, rate=1):
+    """Pads the input along the spatial dimensions independently of input size.
+      Pads the input such that if it was used in a convolution with 'VALID' padding,
+      the output would have the same dimensions as if the unpadded input was used
+      in a convolution with 'SAME' padding.
+      Args:
+        inputs: A tensor of size [batch, height_in, width_in, channels].
+        kernel_size: The kernel to be used in the conv2d or max_pool2d operation.
+        rate: An integer, rate for atrous convolution.
+      Returns:
+        output: A tensor of size [batch, height_out, width_out, channels] with the
+        input, either intact (if kernel_size == 1) or padded (if kernel_size > 1).
+    """
+    kernel_size_effective = [kernel_size[0] + (kernel_size[0] - 1) * (rate - 1),
+                               kernel_size[0] + (kernel_size[0] - 1) * (rate - 1)]
+    pad_total = [kernel_size_effective[0] - 1, kernel_size_effective[1] - 1]
+    pad_beg = [pad_total[0] // 2, pad_total[1] // 2]
+    pad_end = [pad_total[0] - pad_beg[0], pad_total[1] - pad_beg[1]]
+    padded_inputs = tf.pad(inputs, [[0, 0], [pad_beg[0], pad_end[0]],
+                                      [pad_beg[1], pad_end[1]], [0, 0]])
+    return padded_inputs
+
+
+@slim.add_arg_scope
+def expanded_conv(x,
+                  num_outputs,
+                  expansion=6,
+                  stride=1,
+                  rate=1,
+                  normalizer_fn=slim.batch_norm,
+                  project_activation_fn=tf.identity,
+                  padding="SAME",
+                  scope=None):
+    """The expand conv op in MobileNetv2
+        1x1 conv -> depthwise 3x3 conv -> 1x1 linear conv
+    """
+    with tf.variable_scope(scope, default_name="expanded_conv") as s, \
+       tf.name_scope(s.original_name_scope):
+        prev_depth = x.get_shape().as_list()[3]
+        # the filters of expanded conv
+        inner_size = prev_depth * expansion
+        net = x
+        # only inner_size > prev_depth, use expanded conv
+        if inner_size > prev_depth:
+            net = slim.conv2d(net, inner_size, 1, normalizer_fn=normalizer_fn,
+                              scope="expand")
+        # depthwise conv
+        net = slim.separable_conv2d(net, num_outputs=None, kernel_size=3,
+                                    depth_multiplier=1, stride=stride,
+                                    rate=rate, normalizer_fn=normalizer_fn,
+                                    padding=padding, scope="depthwise")
+        # projection
+        net = slim.conv2d(net, num_outputs, 1, normalizer_fn=normalizer_fn,
+                          activation_fn=project_activation_fn, scope="project")
+
+        # residual connection
+        if stride == 1 and net.get_shape().as_list()[-1] == prev_depth:
+            net += x
+
+        return net
+
+def global_pool(x, pool_op=tf.nn.avg_pool):
+    """Applies avg pool to produce 1x1 output.
+    NOTE: This function is funcitonally equivalenet to reduce_mean, but it has
+        baked in average pool which has better support across hardware.
+    Args:
+        input_tensor: input tensor
+        pool_op: pooling op (avg pool is default)
+    Returns:
+        a tensor batch_size x 1 x 1 x depth.
+    """
+    shape = x.get_shape().as_list()
+    if shape[1] is None or shape[2] is None:
+        kernel_size = tf.convert_to_tensor(
+            [1, tf.shape(x)[1], tf.shape(x)[2], 1])
+    else:
+        kernel_size = [1, shape[1], shape[2], 1]
+    output = pool_op(x, ksize=kernel_size, strides=[1, 1, 1, 1], padding='VALID')
+    # Recover output shape, for unknown shape.
+    output.set_shape([None, 1, 1, None])
+    return output
+
+
+_Op = namedtuple("Op", ['op', 'params', 'multiplier_func'])
+
+def op(op_func, **params):
+    return _Op(op=op_func, params=params,
+               multiplier_func=_depth_multiplier_func)
+
+
+CONV_DEF = [op(slim.conv2d, num_outputs=32, stride=2, kernel_size=3),
+            op(expanded_conv, num_outputs=16, expansion=1),
+            op(expanded_conv, num_outputs=24, stride=2),
+            op(expanded_conv, num_outputs=24, stride=1),
+            op(expanded_conv, num_outputs=32, stride=2),
+            op(expanded_conv, num_outputs=32, stride=1),
+            op(expanded_conv, num_outputs=32, stride=1),
+            op(expanded_conv, num_outputs=64, stride=2),
+            op(expanded_conv, num_outputs=64, stride=1),
+            op(expanded_conv, num_outputs=64, stride=1),
+            op(expanded_conv, num_outputs=64, stride=1),
+            op(expanded_conv, num_outputs=96, stride=1),
+            op(expanded_conv, num_outputs=96, stride=1),
+            op(expanded_conv, num_outputs=96, stride=1),
+            op(expanded_conv, num_outputs=160, stride=2),
+            op(expanded_conv, num_outputs=160, stride=1),
+            op(expanded_conv, num_outputs=160, stride=1),
+            op(expanded_conv, num_outputs=320, stride=1),
+            op(slim.conv2d, num_outputs=1280, stride=1, kernel_size=1),
+            ]
+
+
+def mobilenet_arg_scope(is_training=True,
+                        weight_decay=0.00004,
+                        stddev=0.09,
+                        dropout_keep_prob=0.8,
+                        bn_decay=0.997):
+    """Defines Mobilenet default arg scope.
+    Usage:
+     with tf.contrib.slim.arg_scope(mobilenet.training_scope()):
+       logits, endpoints = mobilenet_v2.mobilenet(input_tensor)
+     # the network created will be trainble with dropout/batch norm
+     # initialized appropriately.
+    Args:
+        is_training: if set to False this will ensure that all customizations are
+            set to non-training mode. This might be helpful for code that is reused
+        across both training/evaluation, but most of the time training_scope with
+        value False is not needed. If this is set to None, the parameters is not
+        added to the batch_norm arg_scope.
+        weight_decay: The weight decay to use for regularizing the model.
+        stddev: Standard deviation for initialization, if negative uses xavier.
+        dropout_keep_prob: dropout keep probability (not set if equals to None).
+        bn_decay: decay for the batch norm moving averages (not set if equals to
+            None).
+    Returns:
+        An argument scope to use via arg_scope.
+    """
+    # Note: do not introduce parameters that would change the inference
+    # model here (for example whether to use bias), modify conv_def instead.
+    batch_norm_params = {
+        'center': True,
+        'scale': True,
+        'decay': bn_decay,
+        'is_training': is_training
+    }
+    if stddev < 0:
+        weight_intitializer = slim.initializers.xavier_initializer()
+    else:
+        weight_intitializer = tf.truncated_normal_initializer(stddev=stddev)
+
+    # Set weight_decay for weights in Conv and FC layers.
+    with slim.arg_scope(
+        [slim.conv2d, slim.fully_connected, slim.separable_conv2d],
+        weights_initializer=weight_intitializer,
+        normalizer_fn=slim.batch_norm,
+        activation_fn=tf.nn.relu6), \
+        slim.arg_scope([slim.batch_norm], **batch_norm_params), \
+        slim.arg_scope([slim.dropout], is_training=is_training,
+                     keep_prob=dropout_keep_prob), \
+        slim.arg_scope([slim.conv2d, slim.separable_conv2d],
+                       biases_initializer=None,
+                       padding="SAME"), \
+        slim.arg_scope([slim.conv2d],
+                     weights_regularizer=slim.l2_regularizer(weight_decay)), \
+        slim.arg_scope([slim.separable_conv2d], weights_regularizer=None) as s:
+        return s
+
+
+def mobilenetv2(x,
+                num_classes=1001,
+                depth_multiplier=1.0,
+                scope='MobilenetV2',
+                finegrain_classification_mode=False,
+                min_depth=8,
+                divisible_by=8,
+                output_stride=None,
+                ):
+    """Mobilenet v2
+    Args:
+        x: The input tensor
+        num_classes: number of classes
+        depth_multiplier: The multiplier applied to scale number of
+            channels in each layer. Note: this is called depth multiplier in the
+            paper but the name is kept for consistency with slim's model builder.
+        scope: Scope of the operator
+        finegrain_classification_mode: When set to True, the model
+            will keep the last layer large even for small multipliers.
+            The paper suggests that it improves performance for ImageNet-type of problems.
+        min_depth: If provided, will ensure that all layers will have that
+          many channels after application of depth multiplier.
+       divisible_by: If provided will ensure that all layers # channels
+          will be divisible by this number.
+    """
+    conv_defs = CONV_DEF
+
+    # keep the last conv layer very larger channel
+    if finegrain_classification_mode:
+        conv_defs = copy.deepcopy(conv_defs)
+        if depth_multiplier < 1:
+            conv_defs[-1].params['num_outputs'] /= depth_multiplier
+
+    depth_args = {}
+    # NB: do not set depth_args unless they are provided to avoid overriding
+    # whatever default depth_multiplier might have thanks to arg_scope.
+    if min_depth is not None:
+        depth_args['min_depth'] = min_depth
+    if divisible_by is not None:
+        depth_args['divisible_by'] = divisible_by
+
+    with slim.arg_scope([_depth_multiplier_func], **depth_args):
+        with tf.variable_scope(scope, default_name='Mobilenet'):
+            # The current_stride variable keeps track of the output stride of the
+            # activations, i.e., the running product of convolution strides up to the
+            # current network layer. This allows us to invoke atrous convolution
+            # whenever applying the next convolution would result in the activations
+            # having output stride larger than the target output_stride.
+            current_stride = 1
+
+            # The atrous convolution rate parameter.
+            rate = 1
+
+            net = x
+            # Insert default parameters before the base scope which includes
+            # any custom overrides set in mobilenet.
+            end_points = {}
+            scopes = {}
+            for i, opdef in enumerate(conv_defs):
+                params = dict(opdef.params)
+                opdef.multiplier_func(params, depth_multiplier)
+                stride = params.get('stride', 1)
+                if output_stride is not None and current_stride == output_stride:
+                    # If we have reached the target output_stride, then we need to employ
+                    # atrous convolution with stride=1 and multiply the atrous rate by the
+                    # current unit's stride for use in subsequent layers.
+                    layer_stride = 1
+                    layer_rate = rate
+                    rate *= stride
+                else:
+                    layer_stride = stride
+                    layer_rate = 1
+                    current_stride *= stride
+                # Update params.
+                params['stride'] = layer_stride
+                # Only insert rate to params if rate > 1.
+                if layer_rate > 1:
+                    params['rate'] = layer_rate
+
+                try:
+                    net = opdef.op(net, **params)
+                except Exception:
+                    raise ValueError('Failed to create op %i: %r params: %r' % (i, opdef, params))
+
+            with tf.variable_scope('Logits'):
+                net = global_pool(net)
+                end_points['global_pool'] = net
+                if not num_classes:
+                    return net, end_points
+                net = slim.dropout(net, scope='Dropout')
+                # 1 x 1 x num_classes
+                # Note: legacy scope name.
+                logits = slim.conv2d(
+                    net,
+                    num_classes, [1, 1],
+                    activation_fn=None,
+                    normalizer_fn=None,
+                    biases_initializer=tf.zeros_initializer(),
+                    scope='Conv2d_1c_1x1')
+
+                logits = tf.squeeze(logits, [1, 2])
+
+                return logits
+
+
+if __name__ == "__main__":
+    import cv2
+    import numpy as np
+
+    inputs = tf.placeholder(tf.uint8, [None, None, 3])
+    images = tf.expand_dims(inputs, 0)
+    images = tf.cast(images, tf.float32) / 128. - 1
+    images.set_shape((None, None, None, 3))
+    images = tf.image.resize_images(images, (224, 224))
+
+    with slim.arg_scope(mobilenet_arg_scope(is_training=False)):
+        logits = mobilenetv2(images)
+
+    # Restore using exponential moving average since it produces (1.5-2%) higher
+    # accuracy
+    ema = tf.train.ExponentialMovingAverage(0.999)
+    vars = ema.variables_to_restore()
+
+    saver = tf.train.Saver(vars)
+
+    print(len(tf.global_variables()))
+    for var in tf.global_variables():
+        print(var)
+    checkpoint_path = r"C:\Users\xiaoh\Desktop\temp\mobilenet_v2_1.0_224\mobilenet_v2_1.0_224.ckpt"
+    image_file = "C:/Users/xiaoh/Desktop/temp/pandas.jpg"
+    with tf.Session() as sess:
+        saver.restore(sess, checkpoint_path)
+
+        img = cv2.imread(image_file)
+        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+
+        print(np.argmax(sess.run(logits, feed_dict={inputs: img})[0]))
+
+
+
+
+
+
+
+
+
+
diff --git a/CNNs/shufflenet_v2.py b/CNNs/shufflenet_v2.py
new file mode 100644
index 0000000..dbb25c9
--- /dev/null
+++ b/CNNs/shufflenet_v2.py
@@ -0,0 +1,243 @@
+"""
+The implement of shufflenet_v2 by Keras
+"""
+
+import tensorflow as tf
+from tensorflow.keras.layers import Conv2D, DepthwiseConv2D
+from tensorflow.keras.layers import MaxPool2D, GlobalAveragePooling2D, Dense
+from tensorflow.keras.layers import BatchNormalization, Activation
+
+
+def channle_shuffle(inputs, group):
+    """Shuffle the channel
+    Args:
+        inputs: 4D Tensor
+        group: int, number of groups
+    Returns:
+        Shuffled 4D Tensor
+    """
+    in_shape = inputs.get_shape().as_list()
+    h, w, in_channel = in_shape[1:]
+    assert in_channel % group == 0
+    l = tf.reshape(inputs, [-1, h, w, in_channel // group, group])
+    l = tf.transpose(l, [0, 1, 2, 4, 3])
+    l = tf.reshape(l, [-1, h, w, in_channel])
+
+    return l
+
+class Conv2D_BN_ReLU(tf.keras.Model):
+    """Conv2D -> BN -> ReLU"""
+    def __init__(self, channel, kernel_size=1, stride=1):
+        super(Conv2D_BN_ReLU, self).__init__()
+
+        self.conv = Conv2D(channel, kernel_size, strides=stride,
+                            padding="SAME", use_bias=False)
+        self.bn = BatchNormalization(axis=-1, momentum=0.9, epsilon=1e-5)
+        self.relu = Activation("relu")
+
+    def call(self, inputs, training=True):
+        x = self.conv(inputs)
+        x = self.bn(x, training=training)
+        x = self.relu(x)
+        return x
+
+class DepthwiseConv2D_BN(tf.keras.Model):
+    """DepthwiseConv2D -> BN"""
+    def __init__(self, kernel_size=3, stride=1):
+        super(DepthwiseConv2D_BN, self).__init__()
+
+        self.dconv = DepthwiseConv2D(kernel_size, strides=stride,
+                                     depth_multiplier=1,
+                                     padding="SAME", use_bias=False)
+        self.bn = BatchNormalization(axis=-1, momentum=0.9, epsilon=1e-5)
+
+    def call(self, inputs, training=True):
+        x = self.dconv(inputs)
+        x = self.bn(x, training=training)
+        return x
+
+
+class ShufflenetUnit1(tf.keras.Model):
+    def __init__(self, out_channel):
+        """The unit of shufflenetv2 for stride=1
+        Args:
+            out_channel: int, number of channels
+        """
+        super(ShufflenetUnit1, self).__init__()
+
+        assert out_channel % 2 == 0
+        self.out_channel = out_channel
+
+        self.conv1_bn_relu = Conv2D_BN_ReLU(out_channel // 2, 1, 1)
+        self.dconv_bn = DepthwiseConv2D_BN(3, 1)
+        self.conv2_bn_relu = Conv2D_BN_ReLU(out_channel // 2, 1, 1)
+
+    def call(self, inputs, training=False):
+        # split the channel
+        shortcut, x = tf.split(inputs, 2, axis=3)
+
+        x = self.conv1_bn_relu(x, training=training)
+        x = self.dconv_bn(x, training=training)
+        x = self.conv2_bn_relu(x, training=training)
+
+        x = tf.concat([shortcut, x], axis=3)
+        x = channle_shuffle(x, 2)
+        return x
+
+class ShufflenetUnit2(tf.keras.Model):
+    """The unit of shufflenetv2 for stride=2"""
+    def __init__(self, in_channel, out_channel):
+        super(ShufflenetUnit2, self).__init__()
+
+        assert out_channel % 2 == 0
+        self.in_channel = in_channel
+        self.out_channel = out_channel
+
+        self.conv1_bn_relu = Conv2D_BN_ReLU(out_channel // 2, 1, 1)
+        self.dconv_bn = DepthwiseConv2D_BN(3, 2)
+        self.conv2_bn_relu = Conv2D_BN_ReLU(out_channel - in_channel, 1, 1)
+
+        # for shortcut
+        self.shortcut_dconv_bn = DepthwiseConv2D_BN(3, 2)
+        self.shortcut_conv_bn_relu = Conv2D_BN_ReLU(in_channel, 1, 1)
+
+    def call(self, inputs, training=False):
+        shortcut, x = inputs, inputs
+
+        x = self.conv1_bn_relu(x, training=training)
+        x = self.dconv_bn(x, training=training)
+        x = self.conv2_bn_relu(x, training=training)
+
+        shortcut = self.shortcut_dconv_bn(shortcut, training=training)
+        shortcut = self.shortcut_conv_bn_relu(shortcut, training=training)
+
+        x = tf.concat([shortcut, x], axis=3)
+        x = channle_shuffle(x, 2)
+        return x
+
+class ShufflenetStage(tf.keras.Model):
+    """The stage of shufflenet"""
+    def __init__(self, in_channel, out_channel, num_blocks):
+        super(ShufflenetStage, self).__init__()
+
+        self.in_channel = in_channel
+        self.out_channel = out_channel
+
+        self.ops = []
+        for i in range(num_blocks):
+            if i == 0:
+                op = ShufflenetUnit2(in_channel, out_channel)
+            else:
+                op = ShufflenetUnit1(out_channel)
+            self.ops.append(op)
+
+    def call(self, inputs, training=False):
+        x = inputs
+        for op in self.ops:
+            x = op(x, training=training)
+        return x
+
+
+class ShuffleNetv2(tf.keras.Model):
+    """Shufflenetv2"""
+    def __init__(self, num_classes, first_channel=24, channels_per_stage=(116, 232, 464)):
+        super(ShuffleNetv2, self).__init__()
+
+        self.num_classes = num_classes
+
+        self.conv1_bn_relu = Conv2D_BN_ReLU(first_channel, 3, 2)
+        self.pool1 = MaxPool2D(3, strides=2, padding="SAME")
+        self.stage2 = ShufflenetStage(first_channel, channels_per_stage[0], 4)
+        self.stage3 = ShufflenetStage(channels_per_stage[0], channels_per_stage[1], 8)
+        self.stage4 = ShufflenetStage(channels_per_stage[1], channels_per_stage[2], 4)
+        self.conv5_bn_relu = Conv2D_BN_ReLU(1024, 1, 1)
+        self.gap = GlobalAveragePooling2D()
+        self.linear = Dense(num_classes)
+
+    def call(self, inputs, training=False):
+        x = self.conv1_bn_relu(inputs, training=training)
+        x = self.pool1(x)
+        x = self.stage2(x, training=training)
+        x = self.stage3(x, training=training)
+        x = self.stage4(x, training=training)
+        x = self.conv5_bn_relu(x, training=training)
+        x = self.gap(x)
+        x = self.linear(x)
+        return x
+
+
+if __name__ =="__main__":
+    """
+    inputs = tf.placeholder(tf.float32, [None, 224, 224, 3])
+
+    model = ShuffleNetv2(1000)
+    outputs = model(inputs)
+
+    print(model.summary())
+
+    with tf.Session() as sess:
+        pass
+    
+
+    vars = []
+    for v in tf.global_variables():
+
+        vars.append((v.name, v))
+        print(v.name)
+    print(len(vars))
+
+
+    import numpy as np
+
+    path = "C:/models/ShuffleNetV2-1x.npz"
+    weights = np.load(path)
+    np_vars = []
+    for k in weights:
+        k_ = k.replace("beta", "gbeta")
+        k_ = k_.replace("/dconv", "/conv10_dconv")
+        k_ = k_.replace("shortcut_dconv", "shortcut_a_dconv")
+        k_ = k_.replace("conv5", "su_conv5")
+        k_ = k_.replace("linear", "t_linear")
+        np_vars.append((k_, weights[k]))
+    np_vars.sort(key=lambda x: x[0])
+
+    for k, _ in np_vars:
+        print(k)
+
+    saver = tf.train.Saver(tf.global_variables())
+    with tf.Session() as sess:
+        sess.run(tf.global_variables_initializer())
+
+        assign_ops = []
+        for id in range(len(vars)):
+            print(vars[id][0], np_vars[id][0])
+            assign_ops.append(tf.assign(vars[id][1], np_vars[id][1]))
+
+        sess.run(assign_ops)
+        saver.save(sess, "./models/shufflene_v2_1.0.ckpt")
+
+        model.save("./models/shufflenet_v2_1.0.hdf5")
+    
+    """
+
+    import numpy as np
+    from tensorflow.keras.preprocessing import image
+    from tensorflow.keras.applications.densenet import preprocess_input, decode_predictions
+
+    img_path = './images/cat.jpg'
+    img = image.load_img(img_path, target_size=(224, 224))
+    x = image.img_to_array(img)
+    x = np.expand_dims(x, axis=0)
+    x = preprocess_input(x)
+
+    inputs = tf.placeholder(tf.float32, [None, 224, 224, 3])
+    model = ShuffleNetv2(1000)
+    outputs = model(inputs, training=False)
+    outputs = tf.nn.softmax(outputs)
+
+    saver = tf.train.Saver()
+    with tf.Session() as sess:
+        saver.restore(sess, "./models/shufflene_v2_1.0.ckpt")
+        preds = sess.run(outputs, feed_dict={inputs: x})
+        print(decode_predictions(preds, top=3)[0])
+
diff --git a/ObjectDetections/yolo2/config.py b/ObjectDetections/yolo2/config.py
new file mode 100644
index 0000000..ad7fa91
--- /dev/null
+++ b/ObjectDetections/yolo2/config.py
@@ -0,0 +1,25 @@
+"""
+Yolov2 anchors and coco classes
+"""
+
+"""
+anchors = [[0.738768, 0.874946],
+           [2.42204, 2.65704],
+           [4.30971, 7.04493],
+           [10.246, 4.59428],
+           [12.6868, 11.8741]]
+"""
+anchors = [[0.57273, 0.677385],
+           [1.87446, 2.06253],
+           [3.33843, 5.47434],
+           [7.88282, 3.52778],
+           [9.77052, 9.16828]]
+
+def read_coco_labels():
+    f = open("./data/coco_classes.txt")
+    class_names = []
+    for l in f.readlines():
+        class_names.append(l[:-1])
+    return class_names
+
+class_names = read_coco_labels()
\ No newline at end of file
diff --git a/ObjectDetections/yolo2/data/coco_classes.txt b/ObjectDetections/yolo2/data/coco_classes.txt
new file mode 100644
index 0000000..ca76c80
--- /dev/null
+++ b/ObjectDetections/yolo2/data/coco_classes.txt
@@ -0,0 +1,80 @@
+person
+bicycle
+car
+motorbike
+aeroplane
+bus
+train
+truck
+boat
+traffic light
+fire hydrant
+stop sign
+parking meter
+bench
+bird
+cat
+dog
+horse
+sheep
+cow
+elephant
+bear
+zebra
+giraffe
+backpack
+umbrella
+handbag
+tie
+suitcase
+frisbee
+skis
+snowboard
+sports ball
+kite
+baseball bat
+baseball glove
+skateboard
+surfboard
+tennis racket
+bottle
+wine glass
+cup
+fork
+knife
+spoon
+bowl
+banana
+apple
+sandwich
+orange
+broccoli
+carrot
+hot dog
+pizza
+donut
+cake
+chair
+sofa
+pottedplant
+bed
+diningtable
+toilet
+tvmonitor
+laptop
+mouse
+remote
+keyboard
+cell phone
+microwave
+oven
+toaster
+sink
+refrigerator
+book
+clock
+vase
+scissors
+teddy bear
+hair drier
+toothbrush
diff --git a/ObjectDetections/yolo2/demo.py b/ObjectDetections/yolo2/demo.py
new file mode 100644
index 0000000..4a7183c
--- /dev/null
+++ b/ObjectDetections/yolo2/demo.py
@@ -0,0 +1,50 @@
+"""
+Demo for yolov2
+"""
+
+import numpy as np
+import tensorflow as tf
+import cv2
+from PIL import Image
+
+from model import darknet
+from detect_ops import decode
+from utils import preprocess_image, postprocess, draw_detection
+from config import anchors, class_names
+
+
+input_size = (416, 416)
+image_file = "./images/car.jpg"
+image = cv2.imread(image_file)
+image_shape = image.shape[:2]
+image_cp = preprocess_image(image, input_size)
+"""
+image = Image.open(image_file)
+image_cp = image.resize(input_size, Image.BICUBIC)
+image_cp = np.array(image_cp, dtype=np.float32)/255.0
+image_cp = np.expand_dims(image_cp, 0)
+#print(image_cp)
+"""
+
+
+images = tf.placeholder(tf.float32, [1, input_size[0], input_size[1], 3])
+detection_feat = darknet(images)
+feat_sizes = input_size[0] // 32, input_size[1] // 32
+detection_results = decode(detection_feat, feat_sizes, len(class_names), anchors)
+
+checkpoint_path = "./checkpoint_dir/yolo2_coco.ckpt"
+saver = tf.train.Saver()
+with tf.Session() as sess:
+    saver.restore(sess, checkpoint_path)
+    bboxes, obj_probs, class_probs = sess.run(detection_results, feed_dict={images: image_cp})
+
+bboxes, scores, class_inds = postprocess(bboxes, obj_probs, class_probs,
+                                         image_shape=image_shape)
+img_detection = draw_detection(image, bboxes, scores, class_inds, class_names)
+cv2.imwrite("detection.jpg", img_detection)
+cv2.imshow("detection results", img_detection)
+
+cv2.waitKey(0)
+
+
+
diff --git a/ObjectDetections/yolo2/detect_ops.py b/ObjectDetections/yolo2/detect_ops.py
new file mode 100644
index 0000000..6060ece
--- /dev/null
+++ b/ObjectDetections/yolo2/detect_ops.py
@@ -0,0 +1,39 @@
+"""
+Detection ops for Yolov2
+"""
+
+import tensorflow as tf
+import numpy as np
+
+
+def decode(detection_feat, feat_sizes=(13, 13), num_classes=80,
+           anchors=None):
+    """decode from the detection feature"""
+    H, W = feat_sizes
+    num_anchors = len(anchors)
+    detetion_results = tf.reshape(detection_feat, [-1, H * W, num_anchors,
+                                        num_classes + 5])
+
+    bbox_xy = tf.nn.sigmoid(detetion_results[:, :, :, 0:2])
+    bbox_wh = tf.exp(detetion_results[:, :, :, 2:4])
+    obj_probs = tf.nn.sigmoid(detetion_results[:, :, :, 4])
+    class_probs = tf.nn.softmax(detetion_results[:, :, :, 5:])
+
+    anchors = tf.constant(anchors, dtype=tf.float32)
+
+    height_ind = tf.range(H, dtype=tf.float32)
+    width_ind = tf.range(W, dtype=tf.float32)
+    x_offset, y_offset = tf.meshgrid(height_ind, width_ind)
+    x_offset = tf.reshape(x_offset, [1, -1, 1])
+    y_offset = tf.reshape(y_offset, [1, -1, 1])
+
+    # decode
+    bbox_x = (bbox_xy[:, :, :, 0] + x_offset) / W
+    bbox_y = (bbox_xy[:, :, :, 1] + y_offset) / H
+    bbox_w = bbox_wh[:, :, :, 0] * anchors[:, 0] / W * 0.5
+    bbox_h = bbox_wh[:, :, :, 1] * anchors[:, 1] / H * 0.5
+
+    bboxes = tf.stack([bbox_x - bbox_w, bbox_y - bbox_h,
+                       bbox_x + bbox_w, bbox_y + bbox_h], axis=3)
+
+    return bboxes, obj_probs, class_probs
diff --git a/ObjectDetections/yolo2/loss.py b/ObjectDetections/yolo2/loss.py
new file mode 100644
index 0000000..931359f
--- /dev/null
+++ b/ObjectDetections/yolo2/loss.py
@@ -0,0 +1,86 @@
+"""
+Loss function for YOLOv2
+"""
+
+import numpy as np
+import tensorflow as tf
+
+def compute_loss(predictions, targets, anchors, scales, num_classes=20, feat_sizes=(13, 13)):
+    """
+    Compute the loss of Yolov2 for training
+    """
+    H, W = feat_sizes
+    C = num_classes
+    B = len(anchors)
+    anchors = tf.constant(anchors, dtype=tf.float32)
+    anchors = tf.reshape(anchors, [1, 1, B, 2])
+
+    sprob, sconf, snoob, scoor = scales  # the scales for different parts
+
+    _coords = targets["coords"]  # ground truth [-1, H*W, B, 4]
+    _probs = targets["probs"]    # class probability [-1, H*W, B, C] one hot
+    _confs = targets["confs"]    # 1 for object, 0 for background, [-1, H*W, B]
+
+    # decode the net output
+    predictions = tf.reshape(predictions, [-1, H, W, B, (5 + C)])
+    coords = predictions[:, :, :, :, 0:4]   # t_x, t_y, t_w, t_h
+    coords = tf.reshape(coords, [-1, H*W, B, 4])
+    coords_xy = tf.nn.sigmoid(coords[:, :, :, 0:2])  # (0, 1) relative cell top left
+    coords_wh = tf.sqrt(tf.exp(coords[:, :, :, 2:4]) * anchors /
+                        np.reshape([W, H], [1, 1, 1, 2])) # sqrt of w, h (0, 1)
+    coords = tf.concat([coords_xy, coords_wh], axis=3)  # [batch_size, H*W, B, 4]
+
+    confs = tf.nn.sigmoid(predictions[:, :, :, :, 4])  # object confidence
+    confs = tf.reshape(confs, [-1, H*W, B, 1])
+
+    probs = tf.nn.softmax(predictions[:, :, :, :, 5:])  # class probability
+    probs = tf.reshape(probs, [-1, H*W, B, C])
+
+    preds = tf.concat([coords, confs, probs], axis=3)  # [-1, H*W, B, (4+1+C)]
+
+    # match ground truths with anchors (predictions in fact)
+    # assign ground truths to the predictions with the best IOU (select 1 among 5 anchors)
+    wh = tf.pow(coords[:, :, :, 2:4], 2) * np.reshape([W, H], [1, 1, 1, 2])
+    areas = wh[:, :, :, 0] * wh[:, :, :, 1]
+    centers = coords[:, :, :, 0:2]
+    up_left, down_right = centers - (wh * 0.5), centers + (wh * 0.5)
+
+    # the ground truth
+    _wh = tf.pow(_coords[:, :, :, 2:4], 2) * np.reshape([W, H], [1, 1, 1, 2])
+    _areas = _wh[:, :, :, 0] * _wh[:, :, :, 1]
+    _centers = _coords[:, :, :, 0:2]
+    _up_left, _down_right = _centers - (_wh * 0.5), _centers + (_wh * 0.5)
+
+    # compute IOU
+    inter_upleft = tf.maximum(up_left, _up_left)
+    inter_downright = tf.minimum(down_right, _down_right)
+    inter_wh = tf.maximum(inter_downright - inter_upleft, 0.0)
+    intersects = inter_wh[:, :, :, 0] * inter_wh[:, :, :, 1]
+    ious = tf.truediv(intersects, areas + _areas - intersects)
+
+    best_iou_mask = tf.equal(ious, tf.reduce_max(ious, axis=2, keep_dims=True))
+    best_iou_mask = tf.cast(best_iou_mask, tf.float32)
+    mask = best_iou_mask * _confs  # [-1, H*W, B]
+    mask = tf.expand_dims(mask, -1)  # [-1, H*W, B, 1]
+
+    # compute weight terms
+    confs_w = snoob * (1 - mask) + sconf * mask
+    coords_w = scoor * mask
+    probs_w = sprob * mask
+    weights = tf.concat([coords_w, confs_w, probs_w], axis=3)
+
+    truths = tf.concat([_coords, tf.expand_dims(_confs, -1), _probs], 3)
+
+    loss = tf.pow(preds - truths, 2) * weights
+    loss = tf.reduce_sum(loss, axis=[1, 2, 3])
+    loss = 0.5 * tf.reduce_mean(loss)
+    return loss
+
+
+
+
+
+
+
+
+
diff --git a/ObjectDetections/yolo2/model.png b/ObjectDetections/yolo2/model.png
new file mode 100644
index 0000000..07ab142
Binary files /dev/null and b/ObjectDetections/yolo2/model.png differ
diff --git a/ObjectDetections/yolo2/model.py b/ObjectDetections/yolo2/model.py
new file mode 100644
index 0000000..697dd37
--- /dev/null
+++ b/ObjectDetections/yolo2/model.py
@@ -0,0 +1,89 @@
+"""
+YOLOv2 implemented by Tensorflow, only for predicting
+"""
+import os
+
+import numpy as np
+import tensorflow as tf
+
+
+
+######## basic layers #######
+
+def leaky_relu(x):
+    return tf.nn.leaky_relu(x, alpha=0.1, name="leaky_relu")
+
+# Conv2d
+def conv2d(x, filters, size, pad=0, stride=1, batch_normalize=1,
+           activation=leaky_relu, use_bias=False, name="conv2d"):
+    if pad > 0:
+        x = tf.pad(x, [[0, 0], [pad, pad], [pad, pad], [0, 0]])
+    out = tf.layers.conv2d(x, filters, size, strides=stride, padding="VALID",
+                           activation=None, use_bias=use_bias, name=name)
+    if batch_normalize == 1:
+        out = tf.layers.batch_normalization(out, axis=-1, momentum=0.9,
+                                            training=False, name=name+"_bn")
+    if activation:
+        out = activation(out)
+    return out
+
+# maxpool2d
+def maxpool(x, size=2, stride=2, name="maxpool"):
+    return tf.layers.max_pooling2d(x, size, stride)
+
+# reorg layer
+def reorg(x, stride):
+    return tf.extract_image_patches(x, [1, stride, stride, 1],
+                        [1, stride, stride, 1], [1,1,1,1], padding="VALID")
+
+
+def darknet(images, n_last_channels=425):
+    """Darknet19 for YOLOv2"""
+    net = conv2d(images, 32, 3, 1, name="conv1")
+    net = maxpool(net, name="pool1")
+    net = conv2d(net, 64, 3, 1, name="conv2")
+    net = maxpool(net, name="pool2")
+    net = conv2d(net, 128, 3, 1, name="conv3_1")
+    net = conv2d(net, 64, 1, name="conv3_2")
+    net = conv2d(net, 128, 3, 1, name="conv3_3")
+    net = maxpool(net, name="pool3")
+    net = conv2d(net, 256, 3, 1, name="conv4_1")
+    net = conv2d(net, 128, 1, name="conv4_2")
+    net = conv2d(net, 256, 3, 1, name="conv4_3")
+    net = maxpool(net, name="pool4")
+    net = conv2d(net, 512, 3, 1, name="conv5_1")
+    net = conv2d(net, 256, 1, name="conv5_2")
+    net = conv2d(net, 512, 3, 1, name="conv5_3")
+    net = conv2d(net, 256, 1, name="conv5_4")
+    net = conv2d(net, 512, 3, 1, name="conv5_5")
+    shortcut = net
+    net = maxpool(net, name="pool5")
+    net = conv2d(net, 1024, 3, 1, name="conv6_1")
+    net = conv2d(net, 512, 1, name="conv6_2")
+    net = conv2d(net, 1024, 3, 1, name="conv6_3")
+    net = conv2d(net, 512, 1, name="conv6_4")
+    net = conv2d(net, 1024, 3, 1, name="conv6_5")
+    # ---------
+    net = conv2d(net, 1024, 3, 1, name="conv7_1")
+    net = conv2d(net, 1024, 3, 1, name="conv7_2")
+    # shortcut
+    shortcut = conv2d(shortcut, 64, 1, name="conv_shortcut")
+    shortcut = reorg(shortcut, 2)
+    net = tf.concat([shortcut, net], axis=-1)
+    net = conv2d(net, 1024, 3, 1, name="conv8")
+    # detection layer
+    net = conv2d(net, n_last_channels, 1, batch_normalize=0,
+                 activation=None, use_bias=True, name="conv_dec")
+    return net
+
+
+
+if __name__ == "__main__":
+    x = tf.random_normal([1, 416, 416, 3])
+    model = darknet(x)
+
+    saver = tf.train.Saver()
+    with tf.Session() as sess:
+        saver.restore(sess, "./checkpoint_dir/yolo2_coco.ckpt")
+        print(sess.run(model).shape)
+
diff --git a/ObjectDetections/yolo2/utils.py b/ObjectDetections/yolo2/utils.py
new file mode 100644
index 0000000..5821a3b
--- /dev/null
+++ b/ObjectDetections/yolo2/utils.py
@@ -0,0 +1,163 @@
+"""
+Help functions for YOLOv2
+"""
+import random
+import colorsys
+
+import cv2
+import numpy as np
+
+
+
+############## preprocess image ##################
+
+
+def preprocess_image(image, image_size=(416, 416)):
+    """Preprocess a image to inference"""
+    image_cp = np.copy(image).astype(np.float32)
+    # resize the image
+    image_rgb = cv2.cvtColor(image_cp, cv2.COLOR_BGR2RGB)
+    image_resized = cv2.resize(image_rgb, image_size)
+    # normalize
+    image_normalized = image_resized.astype(np.float32) / 255.0
+    # expand the batch_size dim
+    image_expanded = np.expand_dims(image_normalized, axis=0)
+    return image_expanded
+
+def postprocess(bboxes, obj_probs, class_probs, image_shape=(416, 416),
+                threshold=0.5):
+    """post process the detection results"""
+    bboxes = np.reshape(bboxes, [-1, 4])
+    bboxes[:, 0::2] *= float(image_shape[1])
+    bboxes[:, 1::2] *= float(image_shape[0])
+    bboxes = bboxes.astype(np.int32)
+
+    # clip the bboxs
+    bbox_ref = [0, 0, image_shape[1] - 1, image_shape[0] - 1]
+    bboxes = bboxes_clip(bbox_ref, bboxes)
+
+    obj_probs = np.reshape(obj_probs, [-1])
+    class_probs = np.reshape(class_probs, [len(obj_probs), -1])
+    class_inds = np.argmax(class_probs, axis=1)
+    class_probs = class_probs[np.arange(len(obj_probs)), class_inds]
+    scores = obj_probs * class_probs
+
+    # filter bboxes with scores > threshold
+    keep_inds = scores > threshold
+    bboxes = bboxes[keep_inds]
+    scores = scores[keep_inds]
+    class_inds = class_inds[keep_inds]
+
+    # sort top K
+    class_inds, scores, bboxes = bboxes_sort(class_inds, scores, bboxes)
+    # nms
+    class_inds, scores, bboxes = bboxes_nms(class_inds, scores, bboxes)
+
+    return bboxes, scores, class_inds
+
+def draw_detection(im, bboxes, scores, cls_inds, labels, thr=0.3):
+    # for display
+    ############################
+    # Generate colors for drawing bounding boxes.
+    hsv_tuples = [(x / float(len(labels)), 1., 1.)
+                  for x in range(len(labels))]
+    colors = list(map(lambda x: colorsys.hsv_to_rgb(*x), hsv_tuples))
+    colors = list(
+        map(lambda x: (int(x[0] * 255), int(x[1] * 255), int(x[2] * 255)),
+            colors))
+    random.seed(10101)  # Fixed seed for consistent colors across runs.
+    random.shuffle(colors)  # Shuffle colors to decorrelate adjacent classes.
+    random.seed(None)  # Reset seed to default.
+    # draw image
+    imgcv = np.copy(im)
+    h, w, _ = imgcv.shape
+    for i, box in enumerate(bboxes):
+        if scores[i] < thr:
+            continue
+        cls_indx = cls_inds[i]
+
+        thick = int((h + w) / 300)
+        cv2.rectangle(imgcv,
+                      (box[0], box[1]), (box[2], box[3]),
+                      colors[cls_indx], thick)
+        mess = '%s: %.3f' % (labels[cls_indx], scores[i])
+        if box[1] < 20:
+            text_loc = (box[0] + 2, box[1] + 15)
+        else:
+            text_loc = (box[0], box[1] - 10)
+        cv2.putText(imgcv, mess, text_loc,
+                    cv2.FONT_HERSHEY_SIMPLEX, 1e-3 * h, colors[cls_indx], thick // 3)
+
+    return imgcv
+
+
+############## process bboxes ##################
+def bboxes_clip(bbox_ref, bboxes):
+    """Clip bounding boxes with respect to reference bbox.
+    """
+    bboxes = np.copy(bboxes)
+    bboxes = np.transpose(bboxes)
+    bbox_ref = np.transpose(bbox_ref)
+    bboxes[0] = np.maximum(bboxes[0], bbox_ref[0])
+    bboxes[1] = np.maximum(bboxes[1], bbox_ref[1])
+    bboxes[2] = np.minimum(bboxes[2], bbox_ref[2])
+    bboxes[3] = np.minimum(bboxes[3], bbox_ref[3])
+    bboxes = np.transpose(bboxes)
+    return bboxes
+
+def bboxes_sort(classes, scores, bboxes, top_k=400):
+    """Sort bounding boxes by decreasing order and keep only the top_k
+    """
+    # if priority_inside:
+    #     inside = (bboxes[:, 0] > margin) & (bboxes[:, 1] > margin) & \
+    #         (bboxes[:, 2] < 1-margin) & (bboxes[:, 3] < 1-margin)
+    #     idxes = np.argsort(-scores)
+    #     inside = inside[idxes]
+    #     idxes = np.concatenate([idxes[inside], idxes[~inside]])
+    idxes = np.argsort(-scores)
+    classes = classes[idxes][:top_k]
+    scores = scores[idxes][:top_k]
+    bboxes = bboxes[idxes][:top_k]
+    return classes, scores, bboxes
+
+def bboxes_iou(bboxes1, bboxes2):
+    """Computing iou between bboxes1 and bboxes2.
+    Note: bboxes1 and bboxes2 can be multi-dimensional, but should broacastable.
+    """
+    bboxes1 = np.transpose(bboxes1)
+    bboxes2 = np.transpose(bboxes2)
+    # Intersection bbox and volume.
+    int_ymin = np.maximum(bboxes1[0], bboxes2[0])
+    int_xmin = np.maximum(bboxes1[1], bboxes2[1])
+    int_ymax = np.minimum(bboxes1[2], bboxes2[2])
+    int_xmax = np.minimum(bboxes1[3], bboxes2[3])
+
+    int_h = np.maximum(int_ymax - int_ymin, 0.)
+    int_w = np.maximum(int_xmax - int_xmin, 0.)
+    int_vol = int_h * int_w
+    # Union volume.
+    vol1 = (bboxes1[2] - bboxes1[0]) * (bboxes1[3] - bboxes1[1])
+    vol2 = (bboxes2[2] - bboxes2[0]) * (bboxes2[3] - bboxes2[1])
+    iou = int_vol / (vol1 + vol2 - int_vol)
+    return iou
+
+def bboxes_nms(classes, scores, bboxes, nms_threshold=0.5):
+    """Apply non-maximum selection to bounding boxes.
+    """
+    keep_bboxes = np.ones(scores.shape, dtype=np.bool)
+    for i in range(scores.size-1):
+        if keep_bboxes[i]:
+            # Computer overlap with bboxes which are following.
+            overlap = bboxes_iou(bboxes[i], bboxes[(i+1):])
+            # Overlap threshold for keeping + checking part of the same class
+            keep_overlap = np.logical_or(overlap < nms_threshold, classes[(i+1):] != classes[i])
+            keep_bboxes[(i+1):] = np.logical_and(keep_bboxes[(i+1):], keep_overlap)
+
+    idxes = np.where(keep_bboxes)
+    return classes[idxes], scores[idxes], bboxes[idxes]
+
+
+
+
+
+
diff --git a/README.md b/README.md
index bd61406..b1ebefe 100644
--- a/README.md
+++ b/README.md
@@ -18,13 +18,17 @@ Note: the project aims at imitating the well-implemented algorithms in [Deep Lea
 
 ### CNN Models
 - MobileNet [[self](https://github.com/xiaohu2015/DeepLearning_tutorials/blob/master/CNNs/MobileNet.py) [paper](https://arxiv.org/abs/1704.04861) [ref](https://github.com/Zehaos/MobileNet/blob/master/nets/mobilenet.py)]
+- MobileNetv2 [[self](https://github.com/xiaohu2015/DeepLearning_tutorials/blob/master/CNNs/mobilenet_v2.py) [paper](https://arxiv.org/pdf/1801.04381.pdf) [ref](https://github.com/tensorflow/models/tree/master/research/slim/nets/mobilenet)]
 - SqueezeNet [[self](https://github.com/xiaohu2015/DeepLearning_tutorials/blob/master/CNNs/SqueezeNet.py) [paper](https://arxiv.org/abs/1602.07360)]
 - ResNet [[self](https://github.com/xiaohu2015/DeepLearning_tutorials/blob/master/CNNs/ResNet50.py) [caffe ref](https://github.com/KaimingHe/deep-residual-networks) [paper1](https://arxiv.org/abs/1512.03385) [paper2](https://arxiv.org/abs/1603.05027)]
 - ShuffleNet [[self](https://github.com/xiaohu2015/DeepLearning_tutorials/blob/master/CNNs/ShuffleNet.py) by pytorch [paper](http://cn.arxiv.org/pdf/1707.01083v2)]
+- ShuffleNetv2 [[self](https://github.com/xiaohu2015/DeepLearning_tutorials/blob/master/CNNs/shufflenet_v2.py) [ref](https://github.com/tensorpack/tensorpack/blob/master/examples/ImageNetModels/shufflenet.py) [paper](https://arxiv.org/abs/1807.11164)] 
+- DenseNet [[self](https://github.com/xiaohu2015/DeepLearning_tutorials/blob/master/CNNs/densenet.py) [pytorch_ref](https://github.com/pytorch/vision/blob/master/torchvision/models/densenet.py) [paper](https://arxiv.org/abs/1608.06993)]
 
 ### Object detection
 - YOLOv1 [[self](https://github.com/xiaohu2015/DeepLearning_tutorials/blob/master/ObjectDetections/yolo/yolo_tf.py) [paper](https://arxiv.org/abs/1506.02640) [ref](https://github.com/gliese581gg/YOLO_tensorflow)]
 - SSD [[self](https://github.com/xiaohu2015/DeepLearning_tutorials/blob/master/ObjectDetections/SSD/SSD_demo.py) [paper](https://arxiv.org/pdf/1611.10012.pdf) [slides](http://www.cs.unc.edu/~wliu/papers/ssd_eccv2016_slide.pdf) [cafe](https://github.com/weiliu89/caffe/tree/ssd) [TF](https://arxiv.org/abs/1512.02325) [pytorch](https://github.com/amdegroot/ssd.pytorch) ]
+- YOLOv2 [[self](https://github.com/xiaohu2015/DeepLearning_tutorials/tree/master/ObjectDetections/yolo2) [paper](https://arxiv.org/abs/1612.08242) [ref](https://github.com/yhcc/yolo2)]
 
 ### Practical examples
 You can find more practical examples with tensorflow here:
diff --git a/examples/cnn_setence_classification/text_cnn.py b/examples/cnn_setence_classification/text_cnn.py
index f186faf..3518a5c 100644
--- a/examples/cnn_setence_classification/text_cnn.py
+++ b/examples/cnn_setence_classification/text_cnn.py
@@ -81,7 +81,7 @@ def __init__(self, seq_len, vocab_size, embedding_size, filter_sizes, num_filter
                 pooled_outputs.append(pool_output)   # [None, 1, 1, num_filters]
         # Combine all pooled features
         num_filters_total = num_filters * len(filter_sizes)
-        self.h_pool = tf.concat(3, pooled_outputs)  # [None, 1, 1, num_filters_total]
+        self.h_pool = tf.concat( pooled_outputs,3)  # [None, 1, 1, num_filters_total]
         self.h_pool_flat = tf.reshape(self.h_pool, shape=[-1, num_filters_total]) # [None, num_filters_total]
 
         # The dropout layer
@@ -100,7 +100,7 @@ def __init__(self, seq_len, vocab_size, embedding_size, filter_sizes, num_filter
         
         # The loss
         with tf.name_scope("loss"):
-            losses = tf.nn.softmax_cross_entropy_with_logits(self.scores, self.y)
+            losses = tf.nn.softmax_cross_entropy_with_logits(logits=self.scores, labels=self.y)
             self.loss = tf.reduce_mean(losses) + L2_loss * l2_reg_lambda
         
         # Accuracy
diff --git a/examples/cnn_setence_classification/train_cnn.py b/examples/cnn_setence_classification/train_cnn.py
index aefa2b5..bc58dc0 100644
--- a/examples/cnn_setence_classification/train_cnn.py
+++ b/examples/cnn_setence_classification/train_cnn.py
@@ -1,74 +1,74 @@
-"""
-Test the TextRNN class 
-2016/12/22
-"""
-import os
-import sys
-import numpy as np
-import tensorflow as tf
-from sklearn.model_selection import train_test_split
-from tensorflow.contrib import learn
-
-from data_helpers import load_data_and_labels, batch_iter
-from text_cnn import TextCNN
-
-
-# Load original data
-path = sys.path[0]
-pos_filename = path + "/data/rt-polarity.pos"
-neg_filename = path + "/data/rt-polarity.neg"
-
-X_data, y_data = load_data_and_labels(pos_filename, neg_filename)
-max_document_length = max([len(sen.split(" ")) for sen in X_data])
-print("Max_document_length:,", max_document_length)
-# Create the vacabulary
-vocab_processor = learn.preprocessing.VocabularyProcessor(max_document_length)
-# The idx data
-x = np.array(list(vocab_processor.fit_transform(X_data)), dtype=np.float32)
-y = np.array(y_data, dtype=np.int32)
-vocabulary_size = len(vocab_processor.vocabulary_)
-print("The size of vocabulary:", vocabulary_size)
-# Split the data
-X_train, X_test, y_train, y_test = train_test_split(x, y, test_size=0.1, random_state=1111)
-print("X_train shape {0}, y_train shape {1}".format(X_train.shape, y_train.shape))
-print("X_test shape {0}, y_test shape {1}".format(X_test.shape, y_test.shape))
-
-# The parameters of RNN
-seq_len = X_train.shape[1]
-vocab_size = vocabulary_size
-embedding_size = 128
-filter_sizes = [2, 3, 4]
-num_filters = 128
-num_classes = y_train.shape[1]
-l2_reg_lambda = 0.0
-
-# Construct RNN model
-text_rnn_model = TextCNN(seq_len=seq_len, vocab_size=vocab_size, embedding_size=embedding_size, filter_sizes=
-                        filter_sizes, num_filters=num_filters, num_classes=num_classes)
-loss = text_rnn_model.loss
-train_op = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)
-accuracy = text_rnn_model.accuracy
-# The parameters for training
-batch_size = 64
-training_epochs = 10
-dispaly_every = 1
-dropout_keep_prob = 0.5
-
-batch_num = int(X_train.shape[0]/batch_size)
-
-sess = tf.Session()
-sess.run(tf.global_variables_initializer())
-print("Starting training...")
-for epoch in range(training_epochs):
-    avg_cost = 0
-    for batch in range(batch_num):
-        _, cost = sess.run([train_op, loss], feed_dict={text_rnn_model.x: X_train[batch*batch_size:(batch+1)*batch_size],
-                                    text_rnn_model.y: y_train[batch*batch_size:(batch+1)*batch_size],
-                                    text_rnn_model.dropout_keep_prob:dropout_keep_prob})
-        avg_cost += cost
-    if epoch % dispaly_every == 0:
-        cost, acc = sess.run([loss, accuracy], feed_dict={text_rnn_model.x: X_test,
-                                    text_rnn_model.y: y_test,
-                                    text_rnn_model.dropout_keep_prob: 1.0})
-        print("\nEpoch {0} : loss {1}, accuracy {2}".format(epoch, cost, acc))
-
+"""
+Test the TextRNN class
+2016/12/22
+"""
+import os
+import sys
+import numpy as np
+import tensorflow as tf
+from sklearn.model_selection import train_test_split
+from tensorflow.contrib import learn
+
+from data_helpers import load_data_and_labels, batch_iter
+from text_cnn import TextCNN
+import pudb;pu.db
+
+# Load original data
+path = sys.path[0]
+pos_filename = path + "/data/rt-polarity.pos"
+neg_filename = path + "/data/rt-polarity.neg"
+
+X_data, y_data = load_data_and_labels(pos_filename, neg_filename)
+max_document_length = max([len(sen.split(" ")) for sen in X_data])
+print("Max_document_length:,", max_document_length)
+# Create the vacabulary
+vocab_processor = learn.preprocessing.VocabularyProcessor(max_document_length)
+# The idx data
+x = np.array(list(vocab_processor.fit_transform(X_data)), dtype=np.float32)
+y = np.array(y_data, dtype=np.int32)
+vocabulary_size = len(vocab_processor.vocabulary_)
+print("The size of vocabulary:", vocabulary_size)
+# Split the data
+X_train, X_test, y_train, y_test = train_test_split(x, y, test_size=0.1, random_state=1111)
+print("X_train shape {0}, y_train shape {1}".format(X_train.shape, y_train.shape))
+print("X_test shape {0}, y_test shape {1}".format(X_test.shape, y_test.shape))
+
+# The parameters of RNN
+seq_len = X_train.shape[1]
+vocab_size = vocabulary_size
+embedding_size = 128
+filter_sizes = [2, 3, 4]
+num_filters = 128
+num_classes = y_train.shape[1]
+l2_reg_lambda = 0.0
+
+# Construct RNN model
+text_rnn_model = TextCNN(seq_len=seq_len, vocab_size=vocab_size, embedding_size=embedding_size, filter_sizes=
+                        filter_sizes, num_filters=num_filters, num_classes=num_classes)
+loss = text_rnn_model.loss
+train_op = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)
+accuracy = text_rnn_model.accuracy
+# The parameters for training
+batch_size = 64
+training_epochs = 10
+dispaly_every = 1
+dropout_keep_prob = 0.5
+
+batch_num = int(X_train.shape[0]/batch_size)
+
+sess = tf.Session()
+sess.run(tf.global_variables_initializer())
+print("Starting training...")
+for epoch in range(training_epochs):
+    avg_cost = 0
+    for batch in range(batch_num):
+        _, cost = sess.run([train_op, loss], feed_dict={text_rnn_model.x: X_train[batch*batch_size:(batch+1)*batch_size],
+                                    text_rnn_model.y: y_train[batch*batch_size:(batch+1)*batch_size],
+                                    text_rnn_model.dropout_keep_prob:dropout_keep_prob})
+        avg_cost += cost
+    if epoch % dispaly_every == 0:
+        cost, acc = sess.run([loss, accuracy], feed_dict={text_rnn_model.x: X_test,
+                                    text_rnn_model.y: y_test,
+                                    text_rnn_model.dropout_keep_prob: 1.0})
+        print("\nEpoch {0} : loss {1}, accuracy {2}".format(epoch, cost, acc))
+