Tensorflow v1 primitive implementation of ResNet50. Change output activation from ReLU to what you like

resnet50.py

import tensorflow as tf

"""
Thanks to
https://cv-tricks.com/keras/understand-implement-resnets/
https://qiita.com/gucchi0403/items/097d83f0e8a6091c1240
"""

def resnet(inpt, out_units):
    def weight_variable(shape, name=None):
        initial = tf.truncated_normal(shape, stddev=0.1)
        with tf.init_scope() as sc:
            weight = tf.Variable(initial, name=name)
        return weight

    def fc_layer(inpt, shape):
        fc_w = weight_variable(shape)
        with tf.init_scope() as sc:
            fc_b = tf.Variable(tf.constant(0., shape=[shape[1]]))
        fc_h = tf.nn.relu(tf.matmul(inpt, fc_w) + fc_b)
        return fc_h

    def conv_layer(inpt, filter_shape, stride):
        inpt_channels =  inpt.get_shape().as_list()[-1]
        mean, var = tf.nn.moments(inpt, axes=[0,1,2])
        with tf.init_scope() as sc:
            beta = tf.Variable(tf.zeros([inpt_channels]), name="beta")
            gamma = weight_variable([inpt_channels], name="gamma")
        batch_normed = tf.nn.batch_normalization(inpt, mean, var, beta, gamma, tf.constant(1e-6, dtype=tf.float32))
        out_relu = tf.nn.relu(batch_normed)
        filter_ = weight_variable(filter_shape)
        out = tf.nn.conv2d(out_relu, filter=filter_, strides=[1, stride, stride, 1], padding="SAME")
        return out

    def residual_block(inpt, output_depth, first_stride=1, projection=True):
        input_depth = inpt.get_shape().as_list()[-1]

        conv1 = conv_layer(inpt, [1, 1, input_depth, int(output_depth/4)], first_stride)
        conv2 = conv_layer(conv1, [3, 3, int(output_depth/4), int(output_depth/4)], 1)
        conv3 = conv_layer(conv2, [1, 1, int(output_depth/4), output_depth], 1)

        input_layer = inpt
        if input_depth != output_depth:
            if projection:
                # Option B: Projection shortcut
                input_layer = conv_layer(inpt, [1, 1, input_depth, output_depth], first_stride)
            else:
                # Option A: Zero-padding TODO: maxpooling if first_stride > 1
                input_layer = tf.pad(inpt, [[0,0], [0,0], [0,0], [0, output_depth - input_depth]])

        res = conv3 + input_layer
        return res


    with tf.variable_scope('conv0'):
        conv0 = conv_layer(inpt, [7, 7, 3, 64], 2)
        conv0_p = tf.nn.max_pool(conv0, [1, 3, 3, 1], [1, 2, 2, 1], padding="SAME")

    with tf.variable_scope('conv1'):
        conv1_1 = residual_block(conv0_p, 256)
        conv1_2 = residual_block(conv1_1, 256)
        conv1_3 = residual_block(conv1_2, 256)

    with tf.variable_scope('conv2'):
        conv2_1 = residual_block(conv1_3, 512, first_stride=2)
        conv2_2 = residual_block(conv2_1, 512)
        conv2_3 = residual_block(conv2_2, 512)
        conv2_4 = residual_block(conv2_3, 512)

    with tf.variable_scope('conv3'):
        conv3_1 = residual_block(conv2_4, 1024, first_stride=2)
        conv3_2 = residual_block(conv3_1, 1024)
        conv3_3 = residual_block(conv3_2, 1024)
        conv3_4 = residual_block(conv3_3, 1024)
        conv3_5 = residual_block(conv3_4, 1024)
        conv3_6 = residual_block(conv3_5, 1024)

    with tf.variable_scope('conv4'):
        conv4_1 = residual_block(conv3_6, 2048, first_stride=2)
        conv4_2 = residual_block(conv4_1, 2048)
        conv4_3 = residual_block(conv4_2, 2048)

    with tf.variable_scope('fc'):
        global_pool = tf.reduce_mean(conv4_3, [1, 2])
        fc1_1 = fc_layer(global_pool, [2048, out_units])

    return fc1_1

Hi,I'm sorry to disturb you again.In Deep3DFaceReconstruction,my qusetion abou training code
microsoft/Deep3DFaceReconstruction#74
I went to study resnet50 and read your code.But not find the loss functions and regularizations.If I don't see it, please point it out for me.or you write it somewhere else?
Thanks a lot.

@ZJH2233
Sorry I thought you wanted only ResNet code because at that issue I was discussing ResNet. My entire repository for 3D reconstruction is private as a company asked me and payed for it. Official repository has landmark loss and rendering networks from 3D object. I used them and wrote my own photometric loss and perceptual loss.

Oh,So loss function code can't give me because the company contract?What a pity.
About,

Official repository has landmark loss and rendering networks from 3D object.

Could give me this part loss function?
Or,Tell me where I can find this part,In which github repository?In deep3dface original code?
It doesn't matter if you don't remember.Thank you very much for your answer!