Skip to content

Instantly share code, notes, and snippets.

@piyush2896

piyush2896/resnet50.py

Created January 4, 2018 08:31
Show Gist options
  • Save piyush2896/6e8055da6a60a3585ed4835c76603131 to your computer and use it in GitHub Desktop.
Save piyush2896/6e8055da6a60a3585ed4835c76603131 to your computer and use it in GitHub Desktop.
Download ZIP
ResNet50 using TensorFlow.
Raw
resnet50.py
import tensorflow as tf
from pprint import pprint
def get_weights(shape, name):
return tf.get_variable(name, shape=shape)
def get_bias(shape, name):
return tf.zeros(shape=shape, name=name)
def zero_padding(X, pad=(3, 3)):
paddings = tf.constant([[0, 0], [pad[0], pad[0]],
[pad[1], pad[1]], [0, 0]])
return tf.pad(X, paddings, 'CONSTANT')
def flatten(X):
return tf.contrib.layers.flatten(X)
def dense(X, out, name):
in_prev = X.shape.as_list()[1]
W = get_weights((in_prev, out), name=name+'_W')
b = get_bias((1, out), name=name+'_b')
Z = tf.add(tf.matmul(X, W), b, name=name+'_Z')
A = tf.nn.softmax(Z, name=name)
params = {'W':W, 'b':b, 'Z':Z, 'A':A}
return A, params
def conv2D(A_prev, filters, k_size, strides, padding, name):
m, in_H, in_W, in_C = A_prev.shape.as_list()
w_shape = (k_size[0], k_size[1], in_C, filters)
b_shape = (1, 1, 1, filters)
W = get_weights(shape=w_shape, name=name+'_W')
b = get_bias(shape=b_shape, name=name+'_b')
strides = [1, strides[0], strides[1], 1]
A = tf.nn.conv2d(A_prev, W, strides=strides, padding=padding, name=name)
params = {'W':W, 'b':b, 'A':A}
return A, params
def batch_norm(X, name):
m_, v_ = tf.nn.moments(X, axes=[0, 1, 2], keep_dims=False)
beta_ = tf.zeros(X.shape.as_list()[3])
gamma_ = tf.ones(X.shape.as_list()[3])
bn = tf.nn.batch_normalization(X, mean=m_, variance=v_,
offset=beta_, scale=gamma_,
variance_epsilon=1e-4)
return bn
def identity_block(X, f, filters, stage, block):
"""
Implementing a ResNet identity block with shortcut path
passing over 3 Conv Layers
@params
X - input tensor of shape (m, in_H, in_W, in_C)
f - size of middle layer filter
filters - tuple of number of filters in 3 layers
stage - used to name the layers
block - used to name the layers
@returns
A - Output of identity_block
params - Params used in identity block
"""
conv_name = 'res' + str(stage) + block + '_branch'
bn_name = 'bn' + str(stage) + block + '_branch'
l1_f, l2_f, l3_f = filters
params = {}
A1, params[conv_name+'2a'] = conv2D(X, filters=l1_f, k_size=(1, 1), strides=(1, 1),
padding='VALID', name=conv_name+'2a')
A1_bn = batch_norm(A1, name=bn_name+'2a')
A1_act = tf.nn.relu(A1_bn)
params[conv_name+'2a']['bn'] = A1_bn
params[conv_name+'2a']['act'] = A1_bn
A2, params[conv_name+'2b'] = conv2D(A1_act, filters=l2_f, k_size=(f, f), strides=(1, 1),
padding='SAME', name=conv_name+'2b')
A2_bn = batch_norm(A2, name=bn_name+'2b')
A2_act = tf.nn.relu(A2_bn)
params[conv_name+'2b']['bn'] = A2_bn
params[conv_name+'2b']['act'] = A2_act
A3, params[conv_name+'2c'] = conv2D(A2_act, filters=l3_f, k_size=(1, 1), strides=(1, 1),
padding='VALID', name=conv_name+'2c')
A3_bn=batch_norm(A3, name=bn_name+'2c')
A3_add = tf.add(A3_bn, X)
A = tf.nn.relu(A3_add)
params[conv_name+'2c']['bn'] = A3_bn
params[conv_name+'2c']['add'] = A3_add
params['out'] = A
return A, params
def convolutional_block(X, f, filters, stage, block, s=2):
"""
Implementing a ResNet convolutional block with shortcut path
passing over 3 Conv Layers having different sizes
@params
X - input tensor of shape (m, in_H, in_W, in_C)
f - size of middle layer filter
filters - tuple of number of filters in 3 layers
stage - used to name the layers
block - used to name the layers
s - strides used in first layer of convolutional block
@returns
A - Output of convolutional_block
params - Params used in convolutional block
"""
conv_name = 'res' + str(stage) + block + '_branch'
bn_name = 'bn' + str(stage) + block + '_branch'
l1_f, l2_f, l3_f = filters
params = {}
A1, params[conv_name+'2a'] = conv2D(X, filters=l1_f, k_size=(1, 1), strides=(s, s),
padding='VALID', name=conv_name+'2a')
A1_bn = batch_norm(A1, name=bn_name+'2a')
A1_act = tf.nn.relu(A1_bn)
params[conv_name+'2a']['bn'] = A1_bn
params[conv_name+'2a']['act'] = A1_bn
A2, params[conv_name+'2b'] = conv2D(A1_act, filters=l2_f, k_size=(f, f), strides=(1, 1),
padding='SAME', name=conv_name+'2b')
A2_bn = batch_norm(A2, name=bn_name+'2b')
A2_act = tf.nn.relu(A2_bn)
params[conv_name+'2b']['bn'] = A2_bn
params[conv_name+'2b']['act'] = A2_act
A3, params[conv_name+'2c'] = conv2D(A2_act, filters=l3_f, k_size=(1, 1), strides=(1, 1),
padding='VALID', name=conv_name+'2c')
A3_bn=batch_norm(A3, name=bn_name+'2c')
params[conv_name+'2c']['bn'] = A3_bn
A_, params[conv_name+'1'] = conv2D(X, filters=l3_f, k_size=(1, 1), strides=(s, s),
padding='VALID', name=conv_name+'1')
A_bn_ = batch_norm(A_, name=bn_name+'1')
A3_add = tf.add(A3_bn, A_bn_)
A = tf.nn.relu(A3_add)
params[conv_name+'2c']['add'] = A3_add
params[conv_name+'1']['bn'] = A_bn_
params['out'] = A
return A, params
def ResNet50(input_shape=[64, 64, 3], classes=2):
input_shape=[None]+ input_shape
params={}
X_input = tf.placeholder(tf.float32, shape=input_shape, name='input_layer')
X = zero_padding(X_input, (3, 3))
params['input'] = X_input
params['zero_pad'] = X
# Stage 1
params['stage1'] = {}
A_1, params['stage1']['conv'] = conv2D(X, filters=64, k_size=(7, 7), strides=(2, 2),
padding='VALID', name='conv1')
A_1_bn = batch_norm(A_1, name='bn_conv1')
A_1_act = tf.nn.relu(A_1_bn)
A_1_pool = tf.nn.max_pool(A_1_act, ksize=(1, 3, 3, 1), strides=(1, 2, 2, 1),
padding='VALID')
params['stage1']['bn'] = A_1_bn
params['stage1']['act'] = A_1_act
params['stage1']['pool'] = A_1_pool
# Stage 2
params['stage2'] = {}
A_2_cb, params['stage2']['cb'] = convolutional_block(A_1_pool, f=3, filters=[64, 64, 256],
stage=2, block='a', s=1)
A_2_ib1, params['stage2']['ib1'] = identity_block(A_2_cb, f=3, filters=[64, 64, 256],
stage=2, block='b')
A_2_ib2, params['stage2']['ib2'] = identity_block(A_2_ib1, f=3, filters=[64, 64, 256],
stage=2, block='c')
# Stage 3
params['stage3'] = {}
A_3_cb, params['stage3']['cb'] = convolutional_block(A_2_ib2, 3, [128, 128, 512],
stage=3, block='a', s=2)
A_3_ib1, params['stage3']['ib1'] = identity_block(A_3_cb, 3, [128, 128, 512],
stage=3, block='b')
A_3_ib2, params['stage3']['ib2'] = identity_block(A_3_ib1, 3, [128, 128, 512],
stage=3, block='c')
A_3_ib3, params['stage3']['ib3'] = identity_block(A_3_ib2, 3, [128, 128, 512],
stage=3, block='d')
# Stage 4
params['stage4'] = {}
A_4_cb, params['stage4']['cb'] = convolutional_block(A_3_ib3, 3, [256, 256, 1024],
stage=4, block='a', s=2)
A_4_ib1, params['stage4']['ib1'] = identity_block(A_4_cb, 3, [256, 256, 1024],
stage=4, block='b')
A_4_ib2, params['stage4']['ib2'] = identity_block(A_4_ib1, 3, [256, 256, 1024],
stage=4, block='c')
A_4_ib3, params['stage4']['ib3'] = identity_block(A_4_ib2, 3, [256, 256, 1024],
stage=4, block='d')
A_4_ib4, params['stage4']['ib4'] = identity_block(A_4_ib3, 3, [256, 256, 1024],
stage=4, block='e')
A_4_ib5, params['stage4']['ib5'] = identity_block(A_4_ib4, 3, [256, 256, 1024],
stage=4, block='f')
# Stage 5
params['stage5'] = {}
A_5_cb, params['stage5']['cb'] = convolutional_block(A_4_ib5, 3, [512, 512, 2048],
stage=5, block='a', s=2)
A_5_ib1, params['stage5']['ib1'] = identity_block(A_5_cb, 3, [512, 512, 2048],
stage=5, block='b')
A_5_ib2, params['stage5']['ib2'] = identity_block(A_5_ib1, 3, [512, 512, 2048],
stage=5, block='c')
# Average Pooling
A_avg_pool = tf.nn.avg_pool(A_5_ib2, ksize=(1, 2, 2, 1), strides=(1, 2, 2, 1),
padding='VALID', name='avg_pool')
params['avg_pool'] = A_avg_pool
# Output Layer
A_flat = flatten(A_avg_pool)
params['flatten'] = A_flat
A_out, params['out'] = dense(A_flat, classes, name='fc'+str(classes))
return A_out, params
if __name__ == '__main__':
A, params = ResNet50()
pprint(params, stream=open('ResNet50.json', 'w'), indent=2)
pprint(params, indent=2)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment