peaknet4antfarm

A peaknet API with pytorch backbone

Example

from peaknet import Peaknet
peaknet = Peaknet() # Init a Peaknet instance
peaknet.loadDNWeights() # Load newpeaksv9 network and pretrained weights 
peaknet.model.cuda() # transfer the net to GPU
peaknet.model.print_network() # print network 
next(peaknet.model.parameters()) # a good way to see if the net is on GPU

Setup

Add the following lines to your ~/.bashrc file.

# PEAKNET4ANTFARM
export PYTHONPATH=/reg/neh/home5/liponan/ai/peaknet4antfarm:/reg/neh/home5/liponan/ai/pytorch-yolo2:$PYTHONPATH

I recommend creating an isolated environment for running Peaknet, as it requires an old version of PyTorch.

conda create --name antfarm python=2.7 pytorch=0.1.12 torchvision numpy h5py
conda activate antfarm
conda install --channel lcls-rhel7 psana-conda --force ## adding --force prevents conda from downgrading pytorch...

API

train (for client)

peaknet.train( imgs, labels, box_size = 7 )

imgs is a numpy array with dimensions (n,m,h,w). imgs will be treated as a stack of nxm tiles. labels is a list of tutple of length n. Each item in the list is a tutple of three numpy arrays s, r, c, where s is an array of integers 0~(m-1).

model access (for client)

peaknet.model

returns the current model

update model (for client)

peaknet.updateModel( newModel )

replaces the current model with newModel, including the network and the weights

update grad (for Queen)

peaknet.updateGrad( newModel )

replaces gradients in the current model with that from newModel. newModel must have same network as current model.

optimize (for Queen)

peaknet.optimize()

performs one step of SGD optimization.

liponan/peaknet4antfarm