likesum / bpn Goto Github PK

Hello, this paper is excellent work. But I have a problem with the flops of network. I count with my own code is vastly different from what is described in the paper. Is this reasonable? or is it a matter of my code statistics?
resolution: 1024x768，K=15，B=90, Burst=8, grasacale. The flops of my code output is as follows:

==================Model Analysis Report======================
Doc:
op: The nodes are operation kernel type, such as MatMul, Conv2D. Graph nodes belonging to the same type are aggregated together.
flops: Number of float operations. Note: Please read the implementation for the math behind it.

Profile:
node name | # float_ops
Conv2D 1286.04b float_ops (100.00%, 99.91%)
BiasAdd 640.54m float_ops (0.09%, 0.05%)
Softmax 354.70m float_ops (0.04%, 0.03%)
MaxPool 97.52m float_ops (0.01%, 0.01%)
Mean 47.97m float_ops (0.00%, 0.00%)
Mul 26 float_ops (0.00%, 0.00%)
Sub 2 float_ops (0.00%, 0.00%)
======================End of Report==========================
The FLOPs is:1287180783100

The flops in the paper are 29.9 GB. Is that correct?

the code to calculate the flops as follows:
def get_flops(model):
concrete = tf.function(lambda inputs: model(inputs))
concrete_func = concrete.get_concrete_function(
[tf.TensorSpec([1, *inputs.shape[1:]]) for inputs in model.inputs])
frozen_func, graph_def = convert_variables_to_constants_v2_as_graph(concrete_func)
with tf.Graph().as_default() as graph:
tf.graph_util.import_graph_def(graph_def, name='')
run_meta = tf.compat.v1.RunMetadata()
opts = tf.compat.v1.profiler.ProfileOptionBuilder.float_operation()
flops = tf.compat.v1.profiler.profile(graph=graph, run_meta=run_meta, cmd="op", options=opts)
return flops.total_float_ops

Hi,
When I was processing the 1024*1536 image, the system indicated OOM. What is the size of the feature map you processed before?
The problem might be multiplying basis and coeff, but I haven't been able to solve it yet....

This works attracts me a lot. I am trying to reproduce the result and find the pretrained model link is failed. Hopely, the author can update the link.

It seems that the training set is extremely huge. Could you tell me the hardware and time consumption for training on your experiments?