Esempio n. 1
0
def test_2Dexperiment():
    c = Config()

    c.batch_size = 200
    c.n_epochs = 40
    c.learning_rate = 0.001
    if torch.cuda.is_available():
        c.use_cuda = True
    else:
        c.use_cuda = False
    c.rnd_seed = 1
    c.log_interval = 200
    # model-specific
    c.n_coupling = 8
    c.prior = 'gauss'

    exp = SmileyExperiment(
        c,
        name='gauss',
        n_epochs=c.n_epochs,
        seed=42,
        base_dir='experiment_dir',
        loggers={'visdom': ['visdom', {
            "exp_name": "myenv"
        }]})

    exp.run()

    # sampling
    samples = exp.model.sample(1000).cpu().numpy()
    sns.jointplot(samples[:, 0], samples[:, 1])
    plt.show()
Esempio n. 2
0
def get_config():
    c = Config()
    # cli flags using trixi, overwrite using e.g. --learning_rate=0.001
    c.txt_file = 'assets/001ssb.txt'  # Path to a .txt file to train on
    c.seq_length = 30        # Length of an input sequence
    c.gen_length = 250       # Length of the generated sequence
    c.lstm_num_hidden = 128  # Number of hidden units in the LSTM
    c.lstm_num_layers = 2    # Number of LSTM layers in the model

    # Training params
    c.batch_size = 64        # Number of examples to process in a batch
    c.learning_rate = 2e-3   # Learning rate

    # It is not necessary to implement the following three params, but it may help training.
    c.learning_rate_decay = 0.96  # Learning rate decay fraction
    c.learning_rate_step = 5000  # Learning rate step
    c.dropout_keep_prob = 1.0  # Dropout keep probability

    c.train_steps = 1e6      # Number of training steps
    c.max_norm = 5.0

    # Misc params
    c.summary_path = './summaries/'  # Output path for summaries
    c.print_every = 5        # How often to print training progress
    c.sample_every = 100     # How often to sample from the model
    c.device = 'cuda:0'      # Training device 'cpu' or 'cuda:0'
    c.temperature = 0.5      # balances the sampling strategy between fully-greedy (near 0) and fully-random (higher). e.g. 0.5, 1.0, 2.0.

    return c
def get_config():
    c = Config()
    # cli flags using trixi, overwrite using e.g. --learning_rate=0.001
    c.model_type = 'RNN'    # Model type, should be 'RNN' or 'LSTM'
    c.input_length = 10     # Length of an input sequence
    c.input_dim = 1         # Dimensionality of input sequence
    c.num_classes = 10      # Dimensionality of output sequence
    c.num_hidden = 128      # Number of hidden units in the model
    c.batch_size = 128      # Number of examples to process in a batch
    c.learning_rate = 0.001 # Learning rate
    c.train_steps = 10000   # Number of training steps
    c.max_norm = 10.0
    c.device = 'cuda:0'     # Training device 'cpu' or 'cuda:0'
    return c
Esempio n. 4
0
def test_MNIST_experiment():
    c = Config()

    c.batch_size = 64
    c.n_epochs = 50
    c.learning_rate = 0.001
    c.weight_decay = 5e-5
    if torch.cuda.is_available():
        c.use_cuda = True
    else:
        c.use_cuda = False
    c.rnd_seed = 1
    c.log_interval = 100
    c.subset_size = 10
    # model-specific
    c.n_coupling = 8
    c.n_filters = 64

    exp = MNISTExperiment(
        c,
        name='mnist_test',
        n_epochs=c.n_epochs,
        seed=42,
        base_dir='experiment_dir',
        loggers={'visdom': ['visdom', {
            "exp_name": "myenv"
        }]})

    exp.run()

    exp.model.eval()
    exp.model.to('cpu')
    with torch.no_grad():
        samples = exp.model.sample(16, device='cpu')
        img_grid = make_grid(samples).permute((1, 2, 0))
    plt.imshow(img_grid)
    plt.show()
    return exp.model
Esempio n. 5
0
def test_Resnet():
    c = Config()

    c.batch_size = 64
    c.batch_size_test = 1000
    c.n_epochs = 10
    c.learning_rate = 0.01
    c.momentum = 0.9
    if torch.cuda.is_available():
        c.use_cuda = True
    else:
        c.use_cuda = False
    c.rnd_seed = 1
    c.log_interval = 200

    exp = MNIST_classification(config=c,
                               name='experiment',
                               n_epochs=c.n_epochs,
                               seed=42,
                               base_dir='./experiment_dir',
                               loggers={"visdom": "visdom"})

    exp.run()
def get_config():
    c = Config()

    c.batch_size = 6
    c.patch_size = 512
    c.n_epochs = 20
    c.learning_rate = 0.0002
    c.do_ce_weighting = True
    c.do_batchnorm = True
    if torch.cuda.is_available():
        c.use_cuda = True
    else:
        c.use_cuda = False
    c.rnd_seed = 1
    c.log_interval = 200
    c.base_dir='/media/kleina/Data2/output/meddec'
    c.data_dir='/media/kleina/Data2/Data/meddec/Task07_Pancreas_expert_preprocessed'
    c.split_dir='/media/kleina/Data2/Data/meddec/Task07_Pancreas_preprocessed'
    c.data_file = 'C:/dev/data/Endoviz2018/GIANA/polyp_detection_segmentation/image_gt_data_file_list_all_640x640.csv'
    c.additional_slices=5
    c.name=''

    print(c)
    return c
Esempio n. 7
0
import numpy as np
import torch
from trixi.util import Config
from experiment import MNISTexperiment
from util import plot_dependency_map
import matplotlib.pyplot as plt

c = Config()
c.batch_size = 128
c.n_epochs = 10
c.learning_rate = 0.001
if torch.cuda.is_available():
    c.use_cuda = True
else:
    c.use_cuda = False
c.rnd_seed = 1
c.log_interval = 100

exp = MNISTexperiment(config=c,
                      name='test',
                      n_epochs=c.n_epochs,
                      seed=c.rnd_seed,
                      base_dir='./experiment_dir',
                      loggers={"visdom": ["visdom", {
                          "exp_name": "myenv"
                      }]})

# # run backpropagation for each dimension to compute what other
# # dimensions it depends on.
# exp.setup()
# d = 28