-
Notifications
You must be signed in to change notification settings - Fork 12
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
update example_script in test folder
- Loading branch information
Showing
1 changed file
with
32 additions
and
123 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -1,127 +1,36 @@ | ||
| # Copyright (c) 2021 Venkata K. Ramaswamy, Samuel C. Musson, Chris G. Willcocks, Matteo T. Degiacomi | ||
| # | ||
| # Molearn is free software ; | ||
| # you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation ; | ||
| # either version 2 of the License, or (at your option) any later version. | ||
| # molearn is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY ; | ||
| # without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. | ||
| # See the GNU General Public License for more details. | ||
| # You should have received a copy of the GNU General Public License along with molearn ; | ||
| # if not, write to the Free Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA. | ||
|
|
||
| import sys | ||
| import os | ||
| import shutil | ||
| import numpy as np | ||
| import time | ||
| import torch | ||
| import torch.nn as nn | ||
| import torch.nn.functional as F | ||
| import torch.optim | ||
| from copy import deepcopy | ||
| import biobox | ||
|
|
||
| from molearn import load_data | ||
| from molearn import Auto_potential | ||
| from molearn import Autoencoder | ||
|
|
||
| device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') | ||
|
|
||
| floc = ["MurD_test.pdb"] # test protein (contains only 16 conformations of the MurD protein) | ||
| batch_size = 4 # if this is too small, gpu utilization goes down | ||
| epoch = 0 | ||
| iter_per_epoch = 5 #use higher iter_per_epoch = 1000 for smoother plots (iter_per_epoch = smoothness of statistics) | ||
| method = 'roll' # 3 method for available in Auto_potential: 'roll', 'convolutional', 'indexing' | ||
|
|
||
| # load multiPDB and create loss function | ||
| dataset, meanval, stdval, atom_names, mol, test0, test1 = load_data(floc[0], atoms = ["CA", "C", "N", "CB", "O"], device=device) | ||
| lf = Auto_potential(frame=dataset[0]*stdval, pdb_atom_names=atom_names, method = method, device=torch.device('cpu')) | ||
|
|
||
| # Saving test structures (the most extreme conformations in terms of RMSD) | ||
| # Remember to rescale with stdval, permute axis from [3,N] to [N,3] | ||
| # unsqueeze to [1, N, 3], send back to cpu, and convert to numpy array. | ||
| crds = (test0*stdval).permute(1,0).unsqueeze(0).data.cpu().numpy() | ||
| mol.coordinates = crds | ||
| mol.write_pdb("TEST0.pdb") | ||
|
|
||
| crds = (test1*stdval).permute(1,0).unsqueeze(0).data.cpu().numpy() | ||
| mol.coordinates = crds | ||
| mol.write_pdb("TEST1.pdb") | ||
|
|
||
| # helper function to make getting another batch of data easier | ||
| def cycle(iterable): | ||
| while True: | ||
| for x in iterable: | ||
| yield x | ||
|
|
||
| ######################################################################## | ||
| train_loader = torch.utils.data.DataLoader(torch.utils.data.TensorDataset(dataset.float()), | ||
| batch_size=batch_size, shuffle=True, drop_last=True, num_workers=6) | ||
| iterator = iter(cycle(train_loader)) | ||
|
|
||
| # define networks | ||
| network = Autoencoder(m=2.0, latent_z=2, r=2).to(device) | ||
| print("> Network parameters: ", len(torch.nn.utils.parameters_to_vector(network.parameters()))) | ||
| # define optimisers | ||
| optimiser = torch.optim.Adam(network.parameters(), lr=0.001, amsgrad=True) | ||
|
|
||
| #training loop | ||
| while (epoch<20): | ||
| print("> epoch: ", epoch) | ||
| for i in range(iter_per_epoch): | ||
| # get two batches of training data | ||
| x0 = next(iterator)[0].to(device) | ||
| x1 = next(iterator)[0].to(device) | ||
| optimiser.zero_grad() | ||
|
|
||
| #encode | ||
| z0 = network.encode(x0) | ||
| z1 = network.encode(x1) | ||
|
|
||
| #interpolate | ||
| alpha = torch.rand(x0.size(0), 1, 1).to(device) | ||
| z_interpolated = (1-alpha)*z0 + alpha*z1 | ||
|
|
||
| #decode | ||
| out = network.decode(z0)[:,:,:x0.size(2)] | ||
| out_interpolated = network.decode(z_interpolated)[:,:,:x0.size(2)] | ||
|
|
||
| #calculate MSE | ||
| mse_loss = ((x0-out)**2).mean() # reconstructive loss (Mean square error) | ||
| out *= stdval | ||
| out_interpolated *= stdval | ||
|
|
||
| #calculate physics for interpolated samples | ||
| bond_energy, angle_energy, torsion_energy, NB_energy = lf.get_loss(out_interpolated) | ||
| #by being enclosed in torch.no_grad() torch autograd cannot see where this scaling | ||
| #factor came from and hence although mathematically the physics cancels, no gradients | ||
| #are found and the scale is simply redefined at each step | ||
| with torch.no_grad(): | ||
| scale = 0.1*mse_loss.item()/(bond_energy.item()+angle_energy.item()+torsion_energy.item()+NB_energy.item()) | ||
|
|
||
| network_loss = mse_loss + scale*(bond_energy + angle_energy + torsion_energy + NB_energy) | ||
|
|
||
| #determine gradients | ||
| network_loss.backward() | ||
|
|
||
| #advance the network weights | ||
| optimiser.step() | ||
|
|
||
| #save interpolations between test0 and test1 every 5 epochs | ||
| epoch+=1 | ||
| if epoch%5 == 0: | ||
| interpolation_out = torch.zeros(20, x0.size(2), 3) | ||
| #encode test with each network | ||
| #Not training so switch to eval mode | ||
| network.eval() | ||
| with torch.no_grad(): # don't need gradients for this bit | ||
| test0_z = network.encode(test0.unsqueeze(0).float()) | ||
| test1_z = network.encode(test1.unsqueeze(0).float()) | ||
|
|
||
| #interpolate between the encoded Z space for each network between test0 and test1 | ||
| for idx, t in enumerate(np.linspace(0, 1, 20)): | ||
| interpolation_out[idx] = network.decode(float(t)*test0_z + (1-float(t))*test1_z)[:,:,:x0.size(2)].squeeze(0).permute(1,0).cpu().data | ||
| interpolation_out *= stdval | ||
| import biobox as bb | ||
| import molearn | ||
|
|
||
|
|
||
| if __name__=='__main__': | ||
| trainer = molearn.Molearn_Trainer(device=torch.device('cpu')) | ||
| data = molearn.PDBData() | ||
| #data.import_pdb('/home/wppj21/Workshop/proteins/MurD-Degiacomi/MurD_closed_open.pdb') | ||
| data.import_pdb('MurD_test.pdb') | ||
| #data.atomselect(atoms='no_hydrogen') | ||
| data.atomselect(atoms = ['CA', 'C', 'N', 'CB', 'O']) | ||
| #trainer.set_dataloader(*data.get_dataloader(batch_size=8, validation_split=0.1, dataset_sample_size=80)) | ||
| trainer.set_data(data, batch_size=8, validation_split=0.1, dataset_sample_size=80) | ||
| #trainer.get_dataset(filename='/home/wppj21/Workshop/proteins/MurD-Degiacomi/MurD_closed_open.pdb', dataset_sample_size=40) | ||
| trainer.get_network(autoencoder_kwargs= | ||
| { 'init_z': 32, | ||
| 'latent_z': 2, | ||
| 'depth': 3, | ||
| 'm': 2.0, | ||
| 'r': 2, | ||
| 'use_spectral_norm': True, | ||
| 'use_group_norm': False, | ||
| 'num_groups': 8, | ||
| 'init_n': 206 | ||
| }) | ||
| trainer.get_optimiser(dict(lr=1e-3, momentum=0.9, weight_decay=1e-5)) | ||
| runkwargs = dict(log_filename='log_file.dat', checkpoint_folder='checkpoints', max_epochs=100, checkpoint_frequency=1) | ||
| trainer.run(**runkwargs) | ||
|
|
||
| # Remember to switch back to train mode when you are done | ||
| network.train() | ||
|
|
||
| #save interpolations | ||
| mol.coordinates = interpolation_out.numpy() | ||
| mol.write_pdb("epoch_%s_interpolation.pdb"%epoch) |