require 'torch'
require 'nn'
require 'optim'
require 'model'
require 'image'
require 'cutorch'
require 'cunn'
require('gnuplot')
require('lfs')
require('data')
data = {}
dtype = 'torch.CudaTensor'
-- specify directories
model_root = 'models/'
data_root = 'data/deepbind/'
viz_dir = 'visualization_results/'
-- ****************************************************************** --
-- ****************** CHANGE THESE FIELDS *************************** --
TFs = {'ATF1_K562_ATF1_-06-325-_Harvard'}
cnn_model_name = 'model=CNN,cnn_size=128,cnn_filters=9-5-3,dropout=0.5,learning_rate=0.01,batch_size=256'
rnn_model_name = 'model=RNN,rnn_size=32,rnn_layers=1,dropout=0.5,learning_rate=0.01,batch_size=256'
cnnrnn_model_name = 'model=CNN-RNN,cnn_size=128,cnn_filter=9,rnn_size=32,rnn_layers=1,dropout=0.5,learning_rate=0.01,batch_size=256'
model_names = {rnn_model_name,cnn_model_name,cnnrnn_model_name}
-- ****************************************************************** --
-- ****************************************************************** --
alphabet = 'ACGT'
OneHot = OneHot(#alphabet):type(dtype)
crit = nn.ClassNLLCriterion():type(dtype) --c
start_pos = 1
end_pos = start_pos + 0
lambda = 0.009
config = {learningRate=.05,momentum=0.9}
iterations = 1000
for _,TF in pairs(TFs) do
print(TF)
save_path = viz_dir..TF..'/'
os.execute('mkdir '..save_path..' > /dev/null 2>&1')
-- Load Models
models = {}
for _,model_name in pairs(model_names) do
load_path = model_root..model_name..'/'..TF..'/'
model = torch.load(load_path..'best_model.t7')
model:evaluate()
model.model:type(dtype)
models[model_name] = model
end
--#######################################################################--
--######################### CLASS OPTIMIZATION ##########################--
--#######################################################################--
for model_name, model in pairs(models) do
print('\n ****** Optimizing '..model_name..' *******\n')
print(model.model)
model.model:remove(1)
model:resetStates()
motif = torch.rand(1,101,4):type(dtype)
target = torch.Tensor({1}):type(dtype)
-- motif weight update
feval = function(X)
local output = model:forward(X)
local loss = crit:forward(output[1], target)
local df_do = crit:backward(output[1], target)
local inputGrads = model:backward(motif, df_do)
return (loss + lambda*(X:norm())^2), (inputGrads + X*2*lambda)
end
-- SGD Loop
for i = 1,iterations do
motif,f = optim.rmsprop(feval,motif,config)
print(f[1])
end
-- resize
motif = motif[1]:type(dtype)
-- clamp to values in (0,1)
motif:clamp(0,1)
max = motif:max()
for i = 1,101 do
sum = motif[i]:sum()
if sum == 0 then
motif[i] = torch.zeros(4)
else
for j = 1,4 do motif[i][j] = motif[i][j]/max end
end
end
for i = 1,101 do
--add smoothing constant
for j = 1,4 do motif[i][j] = motif[i][j]+0.01 end
--normalize
sum = motif[i]:sum()
for j = 1,4 do motif[i][j] = motif[i][j]/sum end
end
s2l_filename = save_path..model_name..'_optimization.txt'
optimization_file = io.open(s2l_filename, 'w')
optimization_file:write('PO ')
alphabet:gsub(".",function(c) optimization_file:write(tostring(c)..' ') end)
optimization_file:write('\n')
for i=1,motif:size(1) do
optimization_file:write(tostring(i)..' ')
for j=1,motif:size(2) do
optimization_file:write(tostring(motif[i][j])..' ')
end
optimization_file:write('\n')
end
optimization_file:close()
cmd = "weblogo -D transfac -F png -o "..save_path..model_name.."_optimization.png --errorbars NO --show-xaxis NO --show-yaxis NO -A dna --composition none -n 101 --color '#00CC00' 'A' 'A' --color '#0000CC' 'C' 'C' --color '#FFB300' 'G' 'G' --color '#CC0000' 'T' 'T' < "..s2l_filename
os.execute(cmd)
end
print('')
print(lfs.currentdir()..'/'..save_path)
os.execute('rm '..save_path..'/*.csv > /dev/null 2>&1')
os.execute('rm '..save_path..'/*.txt > /dev/null 2>&1')
end
Datasets
Models
