Skip to content

Latest commit

 

History

History
88 lines (68 loc) · 4.07 KB

tf-resnet.md

File metadata and controls

88 lines (68 loc) · 4.07 KB
Raw

Reset in Tensorflow

Tensorflow includes some of the models proposed in the Residual Network paper under tf.keras.applications.

The Models included are:-

There are also the 2nd version for each of them, proposed by another paper one year later in 2016, so we also have

Each model can be instanciated by a function under tf.keras.applications. you can optionally choose to have the trained weights on imagenet dataset or not, also you can somewhat change in the network structure.

Custtomization options

The functions presented here have few parameters to control the returned model instance.

# all of them have the same signature
tf.keras.applications.ResNet50(
    include_top=True, weights='imagenet', input_tensor=None,
    input_shape=None, pooling=None, classes=1000, **kwargs
)
  1. include_top: Bool used to control keeping the first input layer (True:Default) or not (False).
  2. weights: it can either be:
    1. None: for random initialization.
    2. 'imagenet': Default for the weights trained on imagenet.
    3. path: path to file that includes the weights to be loaded.
  3. input_shape: The shape of the InputLayer to be used, the default is (224, 224, 3).
    The shape must have exactly 3 Dimensions, with width and height no less than 32. NOTE: You can only specify a different size if include_top=False
  4. input_tensor: an Optional input layer to be used. NOTE: You can only specify a different layer if include_top=False
  5. classes: The number of output classes to be used, 1000 by default (for imagenet).
  6. pooling: specifies if Pooling should be used or not. NOTE: You can only specify a different value if include_top=True and weights argument is 'not specified'.

How to use

You can import the required function from tf.keras.applications.

import tensorflow as tf
from tensorflow.keras.applications import ResNet50

Then pass in the parameters you desire to get an instance of the model.

# in order to create the model you need only import the function
# and pass in the required parameters.
model = ResNet50()

preprocessing

You can use this model but not this fast, you need to apply some preprocessing to the input images before it's passed to the model.

# in order to use the model you need to first apply some preprocessing to input image
# the function for that is included here
from tensorflow.keras.applications.resnet import preprocess_input

# and it can be used as simple as
image = preprocess_input(image)
# them pass the output to the model
labels = model(image)

This kind of preprocessing is done by multiple models included in Tensorflow, and done in pretty much the same way, you can check it here

The code below builds a new model from the MobileNet model that first applies the required preprocessing, then passes the result to the imported model, it may not be as simple as in previous cell but it's the same concept.

i = tf.keras.layers.Input([None, None, 3], dtype = tf.uint8)
x = tf.cast(i, tf.float32)
x = tf.keras.applications.mobilenet.preprocess_input(x)
core = tf.keras.applications.MobileNet()
x = core(x)
model = tf.keras.Model(inputs=[i], outputs=[x])

image = tf.image.decode_png(tf.io.read_file('file.png'))
result = model(image)