diff --git a/image_classification/imagenet_dataset.py b/image_classification/imagenet_dataset.py
index 948ac5b8bb4c360bc2ea52d819c2958da52ef68f..158093b3aa9380547490ac5da2386695dd71dd33 100644
--- a/image_classification/imagenet_dataset.py
+++ b/image_classification/imagenet_dataset.py
@@ -19,80 +19,33 @@ import random
 import numpy as np
 
 from datasets.folder import DatasetFolder
-
-
-def center_crop_resize(img):
-    h, w = img.shape[:2]
-    c = int(224 / 256 * min((h, w)))
-    i = (h + 1 - c) // 2
-    j = (w + 1 - c) // 2
-    img = img[i:i + c, j:j + c, :]
-    return cv2.resize(img, (224, 224), 0, 0, cv2.INTER_LINEAR)
-
-
-def random_crop_resize(img):
-    height, width = img.shape[:2]
-    area = height * width
-
-    for attempt in range(10):
-        target_area = random.uniform(0.08, 1.) * area
-        log_ratio = (math.log(3 / 4), math.log(4 / 3))
-        aspect_ratio = math.exp(random.uniform(*log_ratio))
-
-        w = int(round(math.sqrt(target_area * aspect_ratio)))
-        h = int(round(math.sqrt(target_area / aspect_ratio)))
-
-        if w <= width and h <= height:
-            i = random.randint(0, height - h)
-            j = random.randint(0, width - w)
-            img = img[i:i + h, j:j + w, :]
-            return cv2.resize(img, (224, 224), 0, 0, cv2.INTER_LINEAR)
-
-    return center_crop_resize(img)
-
-
-def random_flip(img):
-    if np.random.randint(0, 2) == 1:
-        img = img[:, ::-1, :]
-    return img
-
-
-def normalize_permute(img):
-    # transpose and convert to RGB from BGR
-    img = img.astype(np.float32).transpose((2, 0, 1))[::-1, ...]
-    mean = np.array([123.675, 116.28, 103.53], dtype=np.float32)
-    std = np.array([58.395, 57.120, 57.375], dtype=np.float32)
-    invstd = 1. / std
-    for v, m, s in zip(img, mean, invstd):
-        v.__isub__(m).__imul__(s)
-    return img
-
-
-def compose(functions):
-    def process(sample):
-        img, label = sample
-        for fn in functions:
-            img = fn(img)
-        return img, label
-
-    return process
+from transform import transforms
+from paddle import fluid
 
 
 class ImageNetDataset(DatasetFolder):
     def __init__(self, path, mode='train'):
         super(ImageNetDataset, self).__init__(path)
         self.mode = mode
+
+        normalize = transforms.Normalize(
+            mean=[123.675, 116.28, 103.53], std=[58.395, 57.120, 57.375])
         if self.mode == 'train':
-            self.transform = compose([
-                cv2.imread, random_crop_resize, random_flip, normalize_permute
+            self.transform = transforms.Compose([
+                transforms.RandomResizedCrop(224),
+                transforms.RandomHorizontalFlip(),
+                transforms.Permute(mode='CHW'), normalize
             ])
         else:
-            self.transform = compose(
-                [cv2.imread, center_crop_resize, normalize_permute])
+            self.transform = transforms.Compose([
+                transforms.Resize(256), transforms.CenterCrop(224),
+                transforms.Permute(mode='CHW'), normalize
+            ])
 
     def __getitem__(self, idx):
-        img, label = self.samples[idx]
-        return self.transform((img, [label]))
+        img_path, label = self.samples[idx]
+        img = cv2.imread(img_path).astype(np.float32)
+        return self.transform(img), [label]
 
     def __len__(self):
         return len(self.samples)
diff --git a/transform/__init__.py b/transform/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..7986cdd642998fb0638a81c9ea22615faf8bad0b
--- /dev/null
+++ b/transform/__init__.py
@@ -0,0 +1 @@
+from .transforms import *
diff --git a/transform/functional.py b/transform/functional.py
new file mode 100644
index 0000000000000000000000000000000000000000..b5f6392a2e4ffd06e333ab805ed7e1e2aa5d25c2
--- /dev/null
+++ b/transform/functional.py
@@ -0,0 +1,56 @@
+import sys
+import collections
+import random
+
+import cv2
+import numpy as np
+
+if sys.version_info < (3, 3):
+    Sequence = collections.Sequence
+    Iterable = collections.Iterable
+else:
+    Sequence = collections.abc.Sequence
+    Iterable = collections.abc.Iterable
+
+
+def flip(image, code):
+    """
+    Accordding to the code (the type of flip), flip the input image
+
+    Args:
+        image: Input image, with (H, W, C) shape
+        code: code that indicates the type of flip.
+            -1 : Flip horizontally and vertically
+            0 : Flip vertically
+            1 : Flip horizontally
+    """
+    return cv2.flip(image, flipCode=code)
+
+
+def resize(img, size, interpolation=cv2.INTER_LINEAR):
+    """
+    resize the input data to given size
+
+    Args:
+        input: Input data, could be image or masks, with (H, W, C) shape
+        size: Target size of input data, with (height, width) shape.
+        interpolation: Interpolation method.
+    """
+
+    if isinstance(interpolation, Sequence):
+        interpolation = random.choice(interpolation)
+
+    if isinstance(size, int):
+        h, w = img.shape[:2]
+        if (w <= h and w == size) or (h <= w and h == size):
+            return img
+        if w < h:
+            ow = size
+            oh = int(size * h / w)
+            return cv2.resize(img, (ow, oh), interpolation=interpolation)
+        else:
+            oh = size
+            ow = int(size * w / h)
+            return cv2.resize(img, (ow, oh), interpolation=interpolation)
+    else:
+        return cv2.resize(img, size[::-1], interpolation=interpolation)
diff --git a/transform/transforms.py b/transform/transforms.py
new file mode 100644
index 0000000000000000000000000000000000000000..70ae28352628bc74c1184eafcd7975f45bd24eea
--- /dev/null
+++ b/transform/transforms.py
@@ -0,0 +1,437 @@
+from __future__ import division
+
+import math
+import sys
+import random
+import cv2
+
+import numpy as np
+import numbers
+import types
+import collections
+import warnings
+
+from . import functional as F
+
+if sys.version_info < (3, 3):
+    Sequence = collections.Sequence
+    Iterable = collections.Iterable
+else:
+    Sequence = collections.abc.Sequence
+    Iterable = collections.abc.Iterable
+
+
+class Compose(object):
+    """Composes several transforms together.
+
+    Args:
+        transforms (list of ``Transform`` objects): list of transforms to compose.
+
+    """
+
+    def __init__(self, transforms):
+        self.transforms = transforms
+
+    def __call__(self, img):
+        for t in self.transforms:
+            img = t(img)
+        return img
+
+    def __repr__(self):
+        format_string = self.__class__.__name__ + '('
+        for t in self.transforms:
+            format_string += '\n'
+            format_string += '    {0}'.format(t)
+        format_string += '\n)'
+        return format_string
+
+
+class Resize(object):
+    """Resize the input PIL Image to the given size.
+
+    Args:
+        size (int|list|tuple): Desired output size. If size is a sequence like
+            (h, w), output size will be matched to this. If size is an int,
+            smaller edge of the image will be matched to this number.
+            i.e, if height > width, then image will be rescaled to
+            (size * height / width, size)
+        interpolation (int): interpolation mode of resize. Default: cv2.INTER_LINEAR.
+    """
+
+    def __init__(self, size, interpolation=cv2.INTER_LINEAR):
+        assert isinstance(size, int) or (isinstance(size, Iterable) and
+                                         len(size) == 2)
+        self.size = size
+        self.interpolation = interpolation
+
+    def __call__(self, img):
+        """
+        Args:
+            img (PIL Image): Image to be scaled.
+
+        Returns:
+            PIL Image: Rescaled image.
+        """
+        return F.resize(img, self.size, self.interpolation)
+
+
+class RandomResizedCrop(object):
+    """Crop the input data to random size and aspect ratio.
+    A crop of random size (default: of 0.08 to 1.0) of the original size and a random
+    aspect ratio (default: of 3/4 to 1.33) of the original aspect ratio is made.
+    After applying crop transfrom, the input data will be resized to given size.
+
+    Args:
+        output_size (int|list|tuple): Target size of output image, with (height, width) shape.
+        scale (list|tuple): Range of size of the origin size cropped. Default: (0.08, 1.0)
+        ratio (list|tuple): Range of aspect ratio of the origin aspect ratio cropped. Default: (0.75, 1.33)
+    """
+
+    def __init__(self,
+                 output_size,
+                 scale=(0.08, 1.0),
+                 ratio=(3. / 4, 4. / 3),
+                 interpolation=cv2.INTER_LINEAR):
+        if isinstance(output_size, int):
+            self.output_size = (output_size, output_size)
+        else:
+            self.output_size = output_size
+        assert (scale[0] <= scale[1]), "scale should be of kind (min, max)"
+        assert (ratio[0] <= ratio[1]), "ratio should be of kind (min, max)"
+        self.scale = scale
+        self.ratio = ratio
+        self.interpolation = interpolation
+
+    def _get_params(self, image, attempts=10):
+        height, width, _ = image.shape
+        area = height * width
+
+        for _ in range(attempts):
+            target_area = np.random.uniform(*self.scale) * area
+            log_ratio = tuple(math.log(x) for x in self.ratio)
+            aspect_ratio = math.exp(np.random.uniform(*log_ratio))
+
+            w = int(round(math.sqrt(target_area * aspect_ratio)))
+            h = int(round(math.sqrt(target_area / aspect_ratio)))
+
+            if 0 < w <= width and 0 < h <= height:
+                x = np.random.randint(0, width - w + 1)
+                y = np.random.randint(0, height - h + 1)
+                return x, y, w, h
+
+        # Fallback to central crop
+        in_ratio = float(width) / float(height)
+        if in_ratio < min(self.ratio):
+            w = width
+            h = int(round(w / min(self.ratio)))
+        elif in_ratio > max(self.ratio):
+            h = height
+            w = int(round(h * max(self.ratio)))
+        else:  # whole image
+            w = width
+            h = height
+        x = (width - w) // 2
+        y = (height - h) // 2
+        return x, y, w, h
+
+    def __call__(self, img):
+        x, y, w, h = self._get_params(img)
+        cropped_img = img[y:y + h, x:x + w]
+        return F.resize(cropped_img, self.output_size, self.interpolation)
+
+
+class CenterCropResize(object):
+    """Crops to center of image with padding then scales size.
+
+    Args:
+        size (int|list|tuple): Target size of output image, with (height, width) shape.
+        crop_padding (int): center crop with the padding. Default: 32.
+        interpolation (int): interpolation mode of resize. Default: cv2.INTER_LINEAR.
+    """
+
+    def __init__(self, size, crop_padding=32, interpolation=cv2.INTER_LINEAR):
+        if isinstance(size, int):
+            self.size = (size, size)
+        else:
+            self.size = size
+        self.crop_padding = crop_padding
+        self.interpolation = interpolation
+
+    def _get_params(self, img):
+        h, w = img.shape[:2]
+        size = min(self.size)
+        c = int(size / (size + self.crop_padding) * min((h, w)))
+        x = (h + 1 - c) // 2
+        y = (w + 1 - c) // 2
+        return c, x, y
+
+    def __call__(self, img):
+        c, x, y = self._get_params(img)
+        cropped_img = img[x:x + c, y:y + c, :]
+        return F.resize(cropped_img, self.size, self.interpolation)
+
+
+class CenterCrop(object):
+    """Crops the given the input data at the center.
+
+    Args:
+        output_size: Target size of output image, with (height, width) shape.
+    """
+
+    def __init__(self, output_size):
+        if isinstance(output_size, int):
+            self.output_size = (output_size, output_size)
+        else:
+            self.output_size = output_size
+
+    def _get_params(self, img):
+        th, tw = self.output_size
+        h, w, _ = img.shape
+        assert th <= h and tw <= w, "output size is bigger than image size"
+        x = int(round((w - tw) / 2.0))
+        y = int(round((h - th) / 2.0))
+        return x, y
+
+    def __call__(self, img):
+        x, y = self._get_params(img)
+        th, tw = self.output_size
+        return img[y:y + th, x:x + tw]
+
+
+class RandomHorizontalFlip(object):
+    """Horizontally flip the input data randomly with a given probability.
+
+    Args:
+        prob (float): probability of the input data being flipped. Default: 0.5
+    """
+
+    def __init__(self, prob=0.5):
+        self.prob = prob
+
+    def __call__(self, img):
+        if np.random.random() < self.prob:
+            return F.flip(img, code=1)
+        return img
+
+
+class RandomVerticalFlip(object):
+    """Vertically flip the input data randomly with a given probability.
+
+    Args:
+        prob (float): probability of the input data being flipped. Default: 0.5
+    """
+
+    def __init__(self, prob=0.5):
+        self.prob = prob
+
+    def __call__(self, img):
+        if np.random.random() < self.prob:
+            return F.flip(img, code=0)
+        return img
+
+
+class Normalize(object):
+    """Normalize the input data with mean and standard deviation.
+    Given mean: ``(M1,...,Mn)`` and std: ``(S1,..,Sn)`` for ``n`` channels,
+    this transform will normalize each channel of the input data.
+    ``output[channel] = (input[channel] - mean[channel]) / std[channel]``
+
+    Args:
+        mean (int|float|list): Sequence of means for each channel.
+        std (int|float|list): Sequence of standard deviations for each channel.
+    
+    """
+
+    def __init__(self, mean=0.0, std=1.0):
+        if isinstance(mean, numbers.Number):
+            mean = [mean, mean, mean]
+
+        if isinstance(std, numbers.Number):
+            mean = [std, std, std]
+
+        self.mean = np.array(mean, dtype=np.float32).reshape(len(mean), 1, 1)
+        self.std = np.array(std, dtype=np.float32).reshape(len(std), 1, 1)
+
+    def __call__(self, img):
+        return (img - self.mean) / self.std
+
+
+class Permute(object):
+    """Change input data to a target mode.
+    For example, most transforms use HWC mode image,
+    while the Neural Network might use CHW mode input tensor
+
+    Args:
+        mode: Output mode of input. Use "CHW" mode by default.
+    """
+
+    def __init__(self, mode="CHW"):
+        assert mode in ["CHW"], "unsupported mode: {}".format(mode)
+        self.mode = mode
+
+    def __call__(self, img):
+        if self.mode == "CHW":
+            return img.transpose((2, 0, 1))[::-1, ...]
+        return img
+
+
+class GaussianNoise(object):
+    """Add random gaussian noise to the input data.
+    Gaussian noise is generated with given mean and std.
+
+    Args:
+        mean: Gaussian mean used to generate noise.
+        std: Gaussian standard deviation used to generate noise.
+    """
+
+    def __init__(self, mean=0.0, std=1.0):
+        self.mean = np.array(mean, dtype=np.float32)
+        self.std = np.array(std, dtype=np.float32)
+
+    def __call__(self, img):
+        dtype = img.dtype
+        noise = np.random.normal(self.mean, self.std, img.shape) * 255
+        img = img + noise.astype(np.float32)
+        return np.clip(img, 0, 255).astype(dtype)
+
+
+class BrightnessTransform(object):
+    """Adjust brightness of the image.
+
+    Args:
+        value: How much to adjust the brightness. Can be any
+            non negative number. 0 gives the original image
+    """
+
+    def __init__(self, value):
+        if value < 0:
+            raise ValueError("brightness value should be non-negative")
+        self.value = value
+
+    def __call__(self, img):
+        if self.value == 0:
+            return img
+
+        dtype = img.dtype
+        img = img.astype(np.float32)
+        alpha = np.random.uniform(max(0, 1 - self.value), 1 + self.value)
+        img = img * alpha
+        return img.clip(0, 255).astype(dtype)
+
+
+class ContrastTransform(object):
+    """Adjust contrast of the image.
+
+    Args:
+        value: How much to adjust the contrast. Can be any
+            non negative number. 0 gives the original image
+    """
+
+    def __init__(self, value):
+        if value < 0:
+            raise ValueError("contrast value should be non-negative")
+        self.value = value
+
+    def __call__(self, img):
+        if self.value == 0:
+            return img
+
+        dtype = img.dtype
+        img = img.astype(np.float32)
+        alpha = np.random.uniform(max(0, 1 - self.value), 1 + self.value)
+        img = img * alpha + cv2.cvtColor(img, cv2.COLOR_BGR2GRAY).mean() * (
+            1 - alpha)
+        return img.clip(0, 255).astype(dtype)
+
+
+class SaturationTransform(object):
+    """Adjust saturation of the image.
+
+    Args:
+        value: How much to adjust the saturation. Can be any
+            non negative number. 0 gives the original image
+    """
+
+    def __init__(self, value):
+        if value < 0:
+            raise ValueError("saturation value should be non-negative")
+        self.value = value
+
+    def __call__(self, img):
+        if self.value == 0:
+            return img
+
+        dtype = img.dtype
+        img = img.astype(np.float32)
+        alpha = np.random.uniform(max(0, 1 - self.value), 1 + self.value)
+        gray_img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
+        gray_img = gray_img[..., np.newaxis]
+        img = img * alpha + gray_img * (1 - alpha)
+        return img.clip(0, 255).astype(dtype)
+
+
+class HueTransform(object):
+    """Adjust hue of the image.
+
+    Args:
+        value: How much to adjust the hue. Can be any number
+            between 0 and 0.5, 0 gives the original image
+    """
+
+    def __init__(self, value):
+        if value < 0 or value > 0.5:
+            raise ValueError("hue value should be in [0.0, 0.5]")
+        self.value = value
+
+    def __call__(self, img):
+        if self.value == 0:
+            return img
+
+        dtype = img.dtype
+        img = img.astype(np.uint8)
+        hsv_img = cv2.cvtColor(img, cv2.COLOR_BGR2HSV_FULL)
+        h, s, v = cv2.split(hsv_img)
+
+        alpha = np.random.uniform(-self.value, self.value)
+        h = h.astype(np.uint8)
+        # uint8 addition take cares of rotation across boundaries
+        with np.errstate(over="ignore"):
+            h += np.uint8(alpha * 255)
+        hsv_img = cv2.merge([h, s, v])
+        return cv2.cvtColor(hsv_img, cv2.COLOR_HSV2BGR_FULL).astype(dtype)
+
+
+class ColorJitter(object):
+    """Randomly change the brightness, contrast, saturation and hue of an image.
+
+    Args:
+        brightness: How much to jitter brightness.
+            Chosen uniformly from [max(0, 1 - brightness), 1 + brightness]
+            or the given [min, max]. Should be non negative numbers.
+        contrast: How much to jitter contrast.
+            Chosen uniformly from [max(0, 1 - contrast), 1 + contrast]
+            or the given [min, max]. Should be non negative numbers.
+        saturation: How much to jitter saturation.
+            Chosen uniformly from [max(0, 1 - saturation), 1 + saturation]
+            or the given [min, max]. Should be non negative numbers.
+        hue: How much to jitter hue.
+            Chosen uniformly from [-hue, hue] or the given [min, max].
+            Should have 0<= hue <= 0.5 or -0.5 <= min <= max <= 0.5.
+    """
+
+    def __init__(self, brightness=0, contrast=0, saturation=0, hue=0):
+        transforms = []
+        if brightness != 0:
+            transforms.append(BrightnessTransform(brightness))
+        if contrast != 0:
+            transforms.append(ContrastTransform(contrast))
+        if saturation != 0:
+            transforms.append(SaturationTransform(saturation))
+        if hue != 0:
+            transforms.append(HueTransform(hue))
+
+        random.shuffle(transforms)
+        self.transforms = Compose(transforms)
+
+    def __call__(self, img):
+        return self.transforms(img)