import generative_playground except: import sys, os, inspect my_location = os.path.dirname(os.path.abspath(inspect.getfile(inspect.currentframe()))) sys.path.append('../../..') sys.path.append('../../../../../transformer_pytorch') from generative_playground.molecules.train.vae.main_train_vae import train_vae from generative_playground.molecules.model_settings import get_settings molecules = True grammar = True settings = get_settings(molecules,grammar) save_file =settings['filename_stub'] + 'dr0.2_rnn__.h5' model, fitter, main_dataset = train_vae(molecules=molecules, BATCH_SIZE=150, # 250 max for p2.xlarge drop_rate=0.2, save_file=save_file, sample_z=False, encoder_type='rnn', decoder_type='action', lr=1e-4, plot_prefix='rnn do0.2 no_sam 1e-4', preload_weights=False) while True: next(fitter)
import sys, os, inspect my_location = os.path.dirname(os.path.abspath(inspect.getfile(inspect.currentframe()))) sys.path.append('../../..') sys.path.append('../../../../../transformer_pytorch') from generative_playground.molecules.train.vae.main_train_vae import train_vae from generative_playground.molecules.model_settings import get_settings molecules = True grammar = True settings = get_settings(molecules,grammar) save_file =settings['filename_stub'] + 'baseline__.h5' model, fitter, _ = train_vae(molecules=molecules, grammar=grammar, BATCH_SIZE=50, # max 500 on a p2.xlarge save_file=save_file, sample_z=True, encoder_type='cnn', decoder_type='step', lr=5e-4, plot_prefix='baseline lr 5e-4 KLW 0.01', reg_weight= 1, epsilon_std = 0.01, dashboard='main', preload_weights=False) while True: next(fitter)
import generative_playground except: import sys, os, inspect my_location = os.path.dirname( os.path.abspath(inspect.getfile(inspect.currentframe()))) sys.path.append('../../..') sys.path.append('../../../../../transformer_pytorch') from generative_playground.molecules.train.vae.main_train_vae import train_vae from generative_playground.molecules.model_settings import get_settings molecules = True grammar = True settings = get_settings(molecules, grammar) save_file = settings['filename_stub'] + 'dr0.4_rnn_resnet__.h5' model, fitter, main_dataset = train_vae( molecules=molecules, BATCH_SIZE=100, # max 200 on p2.xlarge drop_rate=0.2, save_file=save_file, sample_z=False, encoder_type='rnn', decoder_type='action_resnet', lr=1e-4, plot_prefix='rnn_resnet do=0.4 1e-4') while True: next(fitter)
from generative_playground.models.simple_models import DenseHead import numpy as np from generative_playground.utils.gpu_utils import to_gpu molecules = True grammar = True settings = get_settings(molecules, grammar) dash_name = 'test' visdom = Dashboard(dash_name) model, fitter, main_dataset = train_vae(molecules=True, grammar=True, BATCH_SIZE=150, drop_rate=0.3, sample_z=True, save_file='next_gen.h5', encoder_type=False, lr=5e-4, plot_prefix='RNN enc lr 1e-4', dashboard=dash_name, preload_weights=False) # this is a wrapper for encoding/decodng grammar_model = ZincGrammarModel(model=model) validity_model = to_gpu( DenseHead(model.encoder, body_out_dim=settings['z_size'], drop_rate=0.3)) valid_smile_ds = IncrementingHDF5Dataset('valid_smiles.h5', valid_frac=0.1) invalid_smile_ds = IncrementingHDF5Dataset('invalid_smiles.h5', valid_frac=0.1) valid_fitter = train_validity(grammar=grammar, model=validity_model,
except: import sys, os, inspect my_location = os.path.dirname( os.path.abspath(inspect.getfile(inspect.currentframe()))) sys.path.append('../../..') sys.path.append('../../../../../transformer_pytorch') from generative_playground.molecules.train.vae.main_train_vae import train_vae from generative_playground.molecules.model_settings import get_settings molecules = True grammar = True settings = get_settings(molecules, grammar) save_file = settings['filename_stub'] + 'dr0.2_attn__.h5' model, fitter, main_dataset = train_vae( molecules=molecules, BATCH_SIZE=10, # it's a bit of a GPU RAM hog drop_rate=0.2, save_file=save_file, sample_z=False, reg_weight=1, # with 0.01 and do 0.1 had real trouble generalizing? encoder_type='attention', decoder_type='attention', lr=1e-4, plot_prefix='attn do=0.2 no_sam 1e-4', preload_weights=False) while True: next(fitter)