Counting 2,987 Big Data & Machine Learning Frameworks, Toolsets, and Examples...
Suggestion? Feedback? Tweet @stkim1

Author
Contributors
Last Commit
Jul. 20, 2018
Created
Nov. 26, 2017

receptivefield

Gradient based receptive field estimation for Convolutional Neural Networks. receptivefield uses backpropagation of the gradients from output of selected feature maps to the input image in order to estimate the size (width, height), stride and offset of resulting receptive field. Numerical estimation of receptive field can be useful when dealing with more complicated neural networks like ResNet, Inception (see notebooks) where analytical approach of computing receptive fields cannot be used.

Build Status

Installation

  • Requires: python (in version >= 3.6), numpy, matplotlib, pillow (check requirements.txt)

  • Depending on the selected API this library requires also:

    • for Keras API:
      • keras>=2.1.6
      • tensorflow-gpu>=1.8.0
    • for Tensorflow API:
      • tensorflow-gpu>=1.8.0
  • pip install receptivefield

Some remarks

  • In order to get better results or avoid possible NaNs in the estimated receptive field parameters, it is suggested to use Linear (instead Relu) activation and AvgPool2D instead of MaxPool2D. This improves gradient flow in the network and hence better signal in the input image. Note, that this is required only for RF estimation.

  • Additionally, one may even initialize network with constant positive values in all weights (positive if max pooling is used) and set biases to zero. In the case of Keras API this can be obtained by setting init_weight=True in the KerasReceptiveField(init_weight=True) constructor.

Limitations

  • Numerical approach cannot be used when RF is larger that input image, however one may try to increase the input image size, since RF parameters depend on the architecture not input image size.

Supported APIs

Currently only Keras and Tensorflow API are supported. However it should be possible to extend receptivefield functionality by deriving abstract class ReceptiveField in base.py file.

  • Keras with KerasReceptiveField, example usage in notebooks/keras_api.ipynb
  • Tensorflow with TFReceptiveField or TFFeatureMapsReceptiveField example usage in notebooks/tensorflow_api.ipynb

How does it work?

This is description of a general approach which is almost the same for other APIs.

  1. Define build_function which returns model (here Keras model)

    def model_build_func(input_shape=[224, 224, 3]):
        ...
        return Model(input, output)
  2. Compute receptive field parameters with KerasReceptiveField

    from receptivefield.keras import KerasReceptiveField
    rf_params = KerasReceptiveField(model_build_func).compute(
        input_shape=[224, 224, 3], # this will be passed to model_build_func
        input_layer='input_image', # must exist - usually input image layer
        output_layers=['feature_map'] # for example last conv layer
    )
  3. The rf_params is a list of FeatureMapDescription. Here we selected only single feature map output_layers=['feature_map'] and the rf_params will contain

    rf_params = [
       FeatureMapDescription(
           size=Size(w=60, h=60), 
           rf=ReceptiveFieldDescription(
               offset=(2.5, 2.5), 
               stride=(1.0, 1.0), 
               size=Size(w=9, h=9))
           )
    ]

    Explanation of FeatureMapDescription fields:

    • size - defined the spatial dimensions of the feature map i.e. the width and height of the feature map grid.
    • rf is an instance of ReceptiveFieldDescription:
      • offset - defines location of the first left-top anchor in the image coordinates (defined in pixels).
      • stride - defines how much RF of the network moves w.r.t unit displacement in the feature_map tensor.
      • size - defines the effective area in the input image which one point in the feature_map tensor is seeing.

Keras minimal - copy/paste example

  • Python code:

    from keras.layers import Conv2D, Input, AvgPool2D
    from keras.models import Model
    from receptivefield.image import get_default_image
    from receptivefield.keras import KerasReceptiveField
    
    # define model function
    def model_build_func(input_shape):
        act = 'linear' # see Remarks
        inp = Input(shape=input_shape, name='input_image')
        x = Conv2D(32, (7, 7), activation=act)(inp)
        x = Conv2D(32, (5, 5), activation=act)(x)
        x = AvgPool2D()(x)
        x = Conv2D(64, (5, 5), activation=act, name='feature_grid')(x)
        x = AvgPool2D()(x)
        model = Model(inp, x)
        return model
    
    shape = [64, 64, 3]
    # compute receptive field
    rf = KerasReceptiveField(model_build_func, init_weights=True)
    rf_params = rf.compute(shape, 'input_image', ['feature_grid'])
    # debug receptive field
    rf.plot_rf_grids(get_default_image(shape, name='doge'))
  • Logger output + example RF grid

      [ INFO][keras.py]::Feature maps shape: [[None, 23, 23, 1]]
      [ INFO][keras.py]::Input shape       : (None, 64, 64, 3)
      [ INFO][base.py]::Estimated receptive field for feature map [0]: ReceptiveFieldDescription(offset=(10.0, 10.0), stride=(2.0, 2.0), size=Size(w=20, h=20))
    

Tensorflow minimal - copy/paste example

  • Python code with multiple feature maps selected and using secondary API with TFFeatureMapsReceptiveField. Same results can be obtained using TFReceptiveField (check notebooks for example usage).

    from typing import List
    from tensorflow.contrib import slim
    import tensorflow as tf
    from receptivefield.image import get_default_image
    from receptivefield.tensorflow import TFFeatureMapsReceptiveField
    
    
    def vgg_feature_extractor(input_image: tf.Tensor) -> List[tf.Tensor]:
        """
        A function which accepts image tensor [1, width, height, num_channels] and 
        returns a list of feature maps of shape [1, fm_width, fm_height, fm_channels]
        """
        
        def linear(x):
            return x
        
        with slim.arg_scope([slim.conv2d],
                              activation_fn=linear,
                              weights_initializer=tf.constant_initializer(0.001),
                              biases_initializer=tf.constant_initializer(0.0)):
            
            net = slim.repeat(input_image, 2, slim.conv2d, 64, [3, 3], scope='conv1')
            net = slim.avg_pool2d(net, [2, 2], scope='pool1')
            net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
            net = slim.avg_pool2d(net, [2, 2], scope='pool2')
            net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
            net = slim.avg_pool2d(net, [2, 2], scope='pool3')
            net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
            fm0 = net # first feature map
            net = slim.avg_pool2d(net, [2, 2], scope='pool4')
            net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
            fm1 = net # second feature map
        
        return [fm0, fm1]
      
    
    image = get_default_image(shape=(96, 96), tile_factor=1, name="cat")
    rf = TFFeatureMapsReceptiveField(vgg_feature_extractor)
    rf_params = rf.compute(input_shape=image.shape)
    # debug receptive field
    rf.plot_rf_grids(get_default_image(shape, name='doge'))
  • Logger output + example RF grid

    [ INFO][tensorflow.py]::Feature maps shape: [(1, 36, 36, 1), (1, 18, 18, 1)]
    [ INFO][tensorflow.py]::Input shape       : [1, 288, 288, 3]
    [ INFO][base.py]::Estimated receptive field for feature map [0]: ReceptiveFieldDescription(offset=(4.0, 4.0), stride=(8.0, 8.0), size=Size(w=92, h=92))
    [ INFO][base.py]::Estimated receptive field for feature map [1]: ReceptiveFieldDescription(offset=(8.0, 8.0), stride=(16.0, 16.0), size=Size(w=196, h=196))
    

Keras more detailed example

Here we show, how to estimate effective receptive field of any Keras model.

  • Create model build_function which returns model. This function should accept one parameter input_shape.

    from keras.layers import Conv2D, Input
    from keras.layers import AvgPool2D
    from keras.models import Model
    
    def model_build_func(input_shape):
        activation = 'linear'
        padding='valid'
        
        inp = Input(shape=input_shape, name='input_image')
        x = Conv2D(32, (5, 5), padding=padding, activation=activation)(inp)
        x = Conv2D(32, (3, 3), padding=padding, activation=activation)(x)
        x = AvgPool2D()(x)
        x = Conv2D(64, (3, 3), activation=activation, padding=padding)(x)
        x = Conv2D(64, (3, 3), activation=activation, padding=padding)(x)
        x = AvgPool2D()(x)
        x = Conv2D(128, (3, 3), activation=activation, padding=padding)(x)
        x = Conv2D(128, (3, 3), activation=activation, padding=padding, name='feature_grid')(x)
    
        model = Model(inp, x)
        return model
  • Check if model is building properly:

    model = model_build_func(input_shape=(96, 96, 3))
    model.summary()
    _________________________________________________________________
    Layer (type)                 Output Shape              Param #   
    =================================================================
    input_image (InputLayer)     (None, 96, 96, 3)         0         
    _________________________________________________________________
    conv2d_1 (Conv2D)            (None, 92, 92, 32)        2432      
    _________________________________________________________________
    conv2d_2 (Conv2D)            (None, 90, 90, 32)        9248      
    _________________________________________________________________
    average_pooling2d_1 (Average (None, 45, 45, 32)        0         
    _________________________________________________________________
    conv2d_3 (Conv2D)            (None, 43, 43, 64)        18496     
    _________________________________________________________________
    conv2d_4 (Conv2D)            (None, 41, 41, 64)        36928     
    _________________________________________________________________
    average_pooling2d_2 (Average (None, 20, 20, 64)        0         
    _________________________________________________________________
    conv2d_5 (Conv2D)            (None, 18, 18, 128)       73856     
    _________________________________________________________________
    feature_grid (Conv2D)        (None, 16, 16, 128)       147584    
    =================================================================
    Total params: 288,544
    Trainable params: 288,544
    Non-trainable params: 0
    
  • This step is not required but it is useful to plot results in the example image. For instance, you would like to see what is the size of network receptive field in comparision to some objects you wish detect (or localize) by this network.

    from receptivefield.image import get_default_image
    import matplotlib.pyplot as plt
    # Load sample image of `Lena`.
    image = get_default_image(shape=(32, 32), tile_factor=1)
    plt.imshow(image)
  • Compute receptive field of the network by calling rf.compute

    from receptivefield.keras import KerasReceptiveField
    
    rf = KerasReceptiveField(model_build_func, init_weights=False)
    
    rf_params = rf.compute(
        input_shape=image.shape, 
        input_layer='input_image', 
        output_layers=['feature_grid']
    )
    print(rf_params)
  • The resulting receptive field is:

    [FeatureMapDescription(
       size=Size(w=16, h=16), 
       rf=ReceptiveFieldDescription(
           offset=(17, 17), 
           stride=(4.0, 4.0), 
           size=Size(w=34, h=34))
    )]
    
  • Input shape: rf.input_shape==ImageShape(w=96, h=96, c=3)

  • Output feature map spatial dimensions: rf.output_shapes==[Size(w=16, h=16)].

  • You may want to see how gradients backpropagate to the input image. Here point=(8, 8) refers to the (W, H) position of the source signal from the output grid. Here we select the first feature map with fm_id=0.

    rf.plot_gradient_at(fm_id=0, point=(8, 8), image=None, figsize=(7, 7))
  • Or even plot whole receptive field grid:

    rf.plot_rf_grids(custom_image=image, figsize=(6, 6))
  • In the above, the red rectangle corresponds to the area which top-left grid point is seeing in the input image. Blue rectangle corresponds to the central grid point, green to the bottom-right point. Green dots show the position of the centers of the grid anchors in the source image.

  • For more plotting functions check the implementation of the base class ReceptiveField defined in base.py file.

Latest Releases
0.2
 Nov. 28 2017
0.19
 Nov. 27 2017
0.1.0
 Nov. 27 2017
Beta version
 Nov. 27 2017