Problems.txt

Learning Rate / batch size issues

Tensorflow poolAllocator for so long：
	entropy estimate contains too much tensors

    config = tf.ConfigProto()
    config.gpu_options.per_process_gpu_memory_fraction = 0.9

	LD_PRELOAD=..tcmalloc
	
	use BFCAllocator: run with environment variable TF_CPU_ALLOCATOR_USE_BFC=true
	
Low GPU usage and high CPU usage:
	Copy too much data from gpu to cpu due to py_func in quantization, use stop_gradient trick
	
	Replace subpixel with depth to space
	
	Use input pipeline
	
	Distributed training, in-graph
	
Hard implementation of GSM entropy rate, instead use GMM
	tf.log may cause Nan error
	
	Arithmetic coding algorithm should be applied per channel as entropy estimation does.
	
	Problems related to matlab arithmetic coding code.
	
Variable shape:
	build tensor, op, variables
	
	while_loop