kimoktm / u-net Goto Github PK

You've done great job!

But I face so many troubles with training the model, could you please provide link (Google Disk for example) to your model (it could be not fully trainedl).

I just want to look at the model structure, to figure out, if it fits my needs :)

Thank you for your work!

Hi! I'm new at this. Can somebody help me how to import my data images in dataset_to_tfrecords?

Hello I am trying to use unet on a custom data set of labels that have the standard 3 RGB channels. My question is how would I configure the mask for each image given that there are three different classifications I want?

I can only get training dataset of 1 channel from the internet,so where can I find the color dataset with RGB channels?

It will be very kind if someone can share the color dataset!

ValueError: Shape of a new variable (conv6_1/weights) must be fully defined, but instead was (3, 3, ?, 512).

how to resolve it

def deconv_upsample(inputs, factor, name, padding = 'SAME', activation_fn = None):
"""
Convolution Transpose upsampling layer with bilinear interpolation weights:
ISSUE: problems with odd scaling factors
----------
Args:
inputs: Tensor, [batch_size, height, width, channels]
factor: Integer, upsampling factor
name: String, scope name
padding: String, input padding
activation_fn: Tensor fn, activation function on output (can be None)

Returns:
    outputs: Tensor, [batch_size, height * factor, width * factor, num_filters_in]
"""

with tf.variable_scope(name):
    stride_shape   = [1, factor, factor, 1]
    input_shape    = tf.shape(inputs)# [1,14,14,1024]
    num_filters_in = inputs.get_shape()[-1].value
    output_shape   = tf.stack([input_shape[0], input_shape[1] * factor, input_shape[2] * factor, num_filters_in])# [   1   28   28 1024]

    weights = bilinear_upsample_weights(factor, num_filters_in)# [   4    4 1024 1024]
    outputs = tf.nn.conv2d_transpose(inputs, weights, output_shape, stride_shape, padding = padding)#shape=(?,?,?,?)!!!

    if activation_fn is not None:
        outputs = activation_fn(outputs)

    return outputs