def transfer_eval(runid, module_prep_model, task1, task2, weightsf, c):
if 'model1' in c:
# Support for transfer across models (e.g. from rnn to attn1511)
# XXX: No separate model1_conf
model1_module = importlib.import_module('.'+c['model1'], 'models')
module_prep_model1 = model1_module.prep_model
else:
module_prep_model1 = module_prep_model
# We construct both original and new model, then copy over
# the weights from the original model
print('Model')
model1 = task1.build_model(module_prep_model1, do_compile=False)
model = task2.build_model(module_prep_model)
print('Model (weights)')
model1.load_weights(weightsf)
for n in model1.nodes.keys():
if n in model.nodes:
model.nodes[n].set_weights(model1.nodes[n].get_weights())
else:
print('- skipping (not in target) ' + n)
print('Pre-training Transfer Evaluation')
task2.eval(model)
train_model(runid, model, task2, c)
print('Predict&Eval (best val epoch)')
res = task2.eval(model)
return model, res
def main():
"""
"""
placeholders = ['input', 'label']
train_ops = ['train']
log_ops = ['accuracy']
files = get_data(config.DATA_DIRECTORY)
queue_graph = create_image_queue_graph(files, config.PIXEL_DEPTH,
config.HEIGHT, config.WIDTH,
config.CHANNELS,
config.BATCH_SIZE, config.CAPACITY)
model_graph = create_model_graph(config.HEIGHT, config.WIDTH,
config.CHANNELS, config.NUM_LABELS)
train_model(queue_graph, model_graph, placeholders, train_ops, log_ops)
def transfer_eval(runid, module_prep_model, task1, task2, weightsf, c):
# We construct both original and new model, then copy over
# the weights from the original model
print('Model')
model1 = task1.build_model(module_prep_model, do_compile=False)
model = task2.build_model(module_prep_model)
print('Model (weights)')
model1.load_weights(weightsf)
for n in model1.nodes.keys():
if n in model.nodes:
model.nodes[n].set_weights(model1.nodes[n].get_weights())
else:
print('- skipping (not in target) ' + n)
print('Pre-training Transfer Evaluation')
task2.eval(model)
train_model(runid, model, task2, c)
print('Predict&Eval (best val epoch)')
res = task2.eval(model)
return model, res
def main(file_path, prep_config, aggregated=False, stats=False, stats_config=None, configs=True,
output_file=None):
''' main function!
Inputs:
file_path - string - path to text files or compiled files
prep_config - config - config used to aggregate files
aggregated - boolean - t/f the specified file path is pre aggregated
stats - boolean - t/f to calculate stats for data
configs - bool/config - either config to use or generate configs
'''
if aggregated:
data = load(file_path)
else:
data = prep(file_path, prep_config)
if stats:
data_stats = calc_stats(data, stats_config)
else:
data_stats = 0
if configs == True:
config_list = generate_configs()
else:
config_list = configs
inferences = []
trained_models = []
for config in config_list:
model = model_selection(config, data)
trained_models.append(train_model(config, model, data))
inferences.append(calc_inferences(config, trained_model, data))
if output_file != None:
helper.write_results(output_file, data_stats, trained_models, inferences)
return (data_stats, trained_models, inferences)
type=float,
dest='GRAD_CLIP',
required=False)
args = parser.parse_args()
for k in vars(args):
if getattr(args, k):
setattr(config, k, getattr(args, k))
# print({item:getattr(config, item) for item in dir(config) if not item.startswith("__")})
print("Loading data....")
data, labid_to_id, padding_idx = data_reader.load_data(config.DATA_DIR +
'data.csv')
target_size = len(labid_to_id)
print(f'Using Device: {config.DEVICE}')
print(f'Starting training of {config.MODEL}...')
if not os.path.exists(config.MODEL_SAVE_DIR):
os.makedirs(config.MODEL_SAVE_DIR)
if config.MODEL_NAME_SUFFIX != '':
savename = config.MODEL + '_' + config.MODEL_NAME_SUFFIX
else:
savename = config.MODEL
train_model(data,
get_model(config.MODEL),
target_size,
padding_idx,
n_fold=config.N_FOLDS,
savename=savename)
def main(args):
bert_tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
bert_model = BertModel.from_pretrained('bert-base-uncased')
dataset_descriptors = get_dataset_paths(args.dataset_json)
dataset_types = [k['name'] for k in dataset_descriptors]
# Init Bert layer and Conv
model, conv_model, sent_embedder = init_common(args, bert_model)
task_classifier = TaskClassifierGCDC(conv_model.get_n_blocks() *
args.n_filters)
ep_maker = EpisodeMaker(
bert_tokenizer,
args.max_len,
args.max_sent,
model.cnn.get_max_kernel(),
args.device,
datasets=dataset_descriptors,
sent_embedder=None if args.finetune else sent_embedder)
print(ep_maker.datasets['gcdc'])
task_classifier = task_classifier.to(args.device)
model = model.to(args.device)
optim = torch.optim.Adam(list(model.parameters()) +
list(task_classifier.parameters()),
lr=args.lr)
import random
logging.info('Multitask training starting.')
time_log = datetime.now().strftime('%y%m%d-%H%M%S')
# writer = SummaryWriter(f'runs/multitaskep_{time_log}')
for batch_nr in range(args.n_epochs):
optim.zero_grad()
dataset_type = 'gcdc'
one_batch_dataset = ep_maker.get_episode(
dataset_type=dataset_type,
n_train=args.train_size_support)['support_set']
binary, loss = loss_task_factory(dataset_type)
train_acc, train_loss = train_model(model,
task_classifier,
one_batch_dataset,
loss,
optim,
binary,
disp_tqdm=False)
# writer.add_scalar(f'Train/{dataset_type}/multi/accuracy', train_acc, batch_nr)
# writer.add_scalar(f'Train/{dataset_type}/multi/loss', train_loss, batch_nr)
logging.info("dataset_type %s, acc %.4f, loss %.4f", dataset_type,
train_acc, train_loss)
logging.debug(
"max of gradients of task_classifier: %f",
max(p.grad.max() for p in task_classifier.parameters()
)) # we take the max because the mean wouldn't be informative
logging.debug(
"avg of gradients of model: %f",
max(p.grad.max() for p in model.parameters()
if p.grad is not None))
for i in range(4):
binary, loss = loss_task_factory('gcdc')
test_set = ep_maker.datasets['gcdc'][i]['test']
test_set.batch_size = 1
test_set.shuffle()
test_set = BertPreprocessor(test_set,
sent_embedder,
conv_model.get_max_kernel(),
device=args.device,
batch_size=8)
acc, loss, _ = eval_model(model,
task_classifier,
test_set,
loss,
binary,
disp_tqdm=False)
logging.info("%s: accuracy %.4f", test_set.file, acc)
def main():
paths = sys.argv[1:]
"""
['./Data/Train/Under_90_min_tuning/data.zip', './Data/Validation/Validation_10_percent/data.zip',
'./tuning_results.txt', './hyperparameter.txt']
"""
with zipfile.ZipFile(paths[0], "r") as zip_ref:
if not os.path.exists("Tmp"):
os.mkdir("Tmp")
if not os.path.exists(os.path.join("Tmp", "train")):
os.mkdir(os.path.join("Tmp", "train"))
zip_ref.extractall(os.path.join("Tmp", "train")) # remove it at last
create_json(os.path.join("Tmp", "train"),
os.path.join('Tmp', 'train_corpus.json'))
with zipfile.ZipFile(paths[1], "r") as zip_ref:
if not os.path.exists("Tmp"):
os.mkdir("Tmp")
if not os.path.exists(os.path.join("Tmp", "validation")):
os.mkdir(os.path.join("Tmp", "validation"))
zip_ref.extractall(os.path.join("Tmp",
"validation")) # remove it at last
create_json(os.path.join("Tmp", "validation"),
os.path.join('Tmp', 'valid_corpus.json'))
number_of_validation_examples = sum(
1 for line in open(os.path.join('Tmp', 'valid_corpus.json')))
print("number of validation examples: " +
str(number_of_validation_examples))
with open(paths[2], 'w') as f: # create file for writing
pass
filters = [150, 200, 250]
kernel_size = [7, 11, 15]
units = [150, 200, 250]
# load the train and test data
data_gen = AudioGenerator(spectrogram=False)
data_gen.load_train_data(desc_file=os.path.join(
'Tmp',
'train_corpus.json')) # necessary to calculate mean and variance
data_gen.load_validation_data(
desc_file=os.path.join('Tmp', 'valid_corpus.json'))
print("validation examples in datagen: " + str(len(data_gen.valid_texts)))
for f in filters:
for k in kernel_size:
for u in units:
model_end = final_model(
input_dim=13,
filters=f, # loaded from hyperparameter
kernel_size=k, # loaded from hyperparameter
conv_stride=2,
conv_border_mode='valid',
units=u, # loaded from hyperparameter
activation='relu',
cell=GRU,
dropout_rate=1,
number_of_layers=2)
train_model(
input_to_softmax=model_end,
#pickle_path=os.path.join('Tmp','model_end.pickle'),
train_json=os.path.join('Tmp', 'train_corpus.json'),
valid_json=os.path.join('Tmp', 'valid_corpus.json'),
save_model_path=os.path.join('Tmp', 'model.h5'),
epochs=10, # changed
spectrogram=False)
wer_sum = 0
for i in range(number_of_validation_examples):
actual, predicted = get_predictions(
data_gen, i, 'validation', model_end,
os.path.join('Tmp', 'model.h5'))
wer_sum += wer(actual.split(), predicted.split())
wer_sum /= number_of_validation_examples
with open(paths[2], 'a') as file: # open file for writing
file.write(
str(f) + " " + str(k) + " " + str(u) + " " +
str(wer_sum) + "\n")
# now take the lowest wer and have the hyperparameters to paths[3]
lines = []
with open(paths[2], 'r') as f:
lines = f.readlines()
min_wer = sys.maxsize
min_filter = None
min_kernel_size = None
min_units = None
for l in lines:
s = l.split()
if float(s[3]) < min_wer:
min_wer = float(s[3])
min_filter = int(s[0])
min_kernel_size = int(s[1])
min_units = int(s[2])
with open(paths[3], 'w') as f:
f.write(str(min_filter) + "\n")
f.write(str(min_kernel_size) + "\n")
f.write(str(min_units) + "\n")
shutil.rmtree("Tmp")
params = {
'path': args.features_dir,
'model_path': args.model_dir,
'batch_size': args.batch_size,
'epochs': args.epochs,
'learning_rate': args.learning_rate,
'n_cnn_filters': [int(x) for x in args.n_filters.split('-')],
'n_cnn_kernels': [int(x) for x in args.n_kernels.split('-')],
'n_fc_units': [int(x) for x in args.n_fc_units.split('-')],
'n_classes': args.n_classes,
'train_val_ratio': args.train_val_ratio,
'baseline_val_loss': args.baseline_val_loss,
}
train_model(params)
print('model training complete')
elif mode == 'prediction':
params = {
'test_data': args.test_speech_dir,
'model_path': os.path.join(args.model_dir, 'vad_model.pt'),
'smoothing': args.smoothing,
'visualize': args.visualize,
'parallel': args.parallel,
'fig_path': args.fig_path
}
prediction(params)
# Intersection over Union (Jaccard) used for scoring
score_func = db_eval_iou
# get the DAVIS 2016 data loaders
loaders = {k: DataLoader(DAVIS(p["/"],p[k], s)) for k in ['train','val']}
# get model and load pre-trained weights
model = load_model( STM(new_arch=new_arch), p["weights"])
# set trainable parameters
select_params( model, contains=weight_hint)
# loss function
criterion = CrossEntropyLoss()
# optimizier
optimizer = Adam(model.parameters(), lr=learning_rate)
# create logger
log = Tracker()
log.create_dir()
log.set_save_string('state_dict_M_{}_E_{}_J_{:5f}_T_{}.pth' )
# train model and validate after each epoch
train_model(loaders, model, criterion, optimizer, log, score_func,
batch_size=batch_size, mode=mode,num_epochs=num_epochs)
# plot log file for statistics
log.plot()
if check_name("INIT", basename):
print("initializing with vgg weights")
model.load_state_dict(init_weights(modelzoo.vgg11(pretrained=True), model))
model = model.to(device)
if adam or finetune:
print("Using adam")
opt = optim.Adam(model.parameters())
else:
opt = optim.SGD(model.parameters(), lr=LR, momentum=0.9)
scheduler = lr_scheduler.StepLR(opt, step_size=BATCH_SIZE//2, gamma=0.1)
train_model(model, loss, metric, opt, scheduler, dataloaders[o], device, dataset_sizes[o],
inepoch_plot=False, loss_name="Batch DICE Loss", metric_name="Batch DICE", model_name=basename + "-" + o, num_epochs=NEPOCHS,
save=True, savefolder=model_folder)
print("Testing train result...")
test_model(model, loss, metric, dataloaders[o], dataset_sizes[o], device,
loss_name="Batch DICE Loss", metric_name="Batch DICE")
# Only test and show results
else:
print("Loading saved result and testing full volume with {}...".format(basename))
plt.title("Loss in all orientations")
models = get_models(bias, e2d, res, small, bn, dunet, model_folder=model_folder)
for i, o in enumerate(orientations):
path = os.path.join(model_folder, basename + "-" + o + ".pkl")
loaded_results = TrainResults.load(path)
loaded_results.plot(show=False, loss_only=True, o=o)
for epoch in range(num_epochs):
log.log_time('Epoch {}/{}'.format(epoch, num_epochs - 1))
# Iterate over videos.
for video_step, video_loader in enumerate(train_loader):
video_loss = []
# Iterate over frames.
for _, sample in enumerate(video_loader):
n_samples += 1
# Send data to device
y, x = sample['y'].to(device), sample['x'].to(device)
# Train model with sample
loss = train_model(model, {'x':x, 'y':y}, criterion, optimizer)
video_loss.append(loss)
# Logs per video
log.log_time('Video: {}\tTotal Loss: {:.6f}\tAvg Loss: {:.6f}'
.format(n_samples, np.sum(video_loss), np.average(video_loss)))
# Test model
# NOTE: len(train_loader) must be >> len(test_loader)
if video_step % TEST_INTERVAL == 0:
test_loss = []
# Iterate over videos.
for video_step, video_loader in enumerate(test_loader):
# Iterate over frames.
for _, sample in enumerate(video_loader):
dataloaders = data.DataLoader(DataLoader(args),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
device = torch.device("cuda:0")
print("constructing model ....")
model = VSR(args.nframes)
model = nn.DataParallel(model.to(device), gpuids)
if args.resume:
ckpt = torch.load(args.model_path)
new_ckpt = {}
for key in ckpt:
if not key.startswith('module'):
new_key = 'module.' + key
else:
new_key = key
new_ckpt[new_key] = ckpt[key]
model.load_state_dict(new_ckpt)
print("model constructed")
summary_writer = SummaryWriter(args.log_dir)
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
scheduler = ExponentialLR(optimizer, gamma=args.gamma)
train_model(model, optimizer, scheduler, dataloaders, summary_writer,
device, args)
params=model_params,
regularizer=regularizer)
logger.info("Loading the pretrained model from %s",
os.path.join(args.pretrained_dir, args.pretrained_model_name))
try:
pretrained_model_state_path = os.path.join(args.pretrained_dir,
args.pretrained_model_name)
pretrained_model_state = torch.load(pretrained_model_state_path)
model.load_state_dict(state_dict=pretrained_model_state)
except:
raise ConfigurationError(
"It appears that the configuration of the pretrained model and "
"the model to fine-tune are not compatible. "
"Please check the compatibility of the encoders and taggers in the "
"config files.")
### Create multi-task trainer ###
multi_task_trainer_params = params.pop("multi_task_trainer")
trainer = MultiTaskTrainer.from_params(model=model,
task_list=task_list,
serialization_dir=serialization_dir,
params=multi_task_trainer_params)
### Launch training ###
metrics = train_model(multi_task_trainer=trainer, recover=False)
if metrics is not None:
logging.info(
"Fine-tuning is finished ! Let's have a drink. It's on the house !"
)
]
test_data = [
np.expand_dims(test_data[:, :, :, i], axis=3)
for i in range(data_chn_num)
]
''' Train Networks'''
train_config = TrainConfig(args)
# U-Net (only FLAIR)
train_dat = [train_data[0], train_trgt]
test_dat = [test_data[0], test_trgt]
train_model(train_config,
START_TIME,
net_depth=args.depth,
SALIENCY=False,
DILATION=False,
restore_dir=None,
net_type='FLAIR',
train_dat=train_dat,
test_dat=test_dat)
# U-Net (only IAM)
train_dat = [train_data[1], train_trgt]
test_dat = [test_data[1], test_trgt]
train_model(train_config,
START_TIME,
net_depth=args.depth,
SALIENCY=False,
DILATION=False,
restore_dir=None,
net_type='IAM',
if (vars.mode == 'test'): #test
#model.load_state_dict(torch.load("D:\\Projects\\Python\\Zeitoon Detection\"))
model.load_state_dict(torch.load(vars.test_model))
model = model.to(vars.device)
test_model(model, vars.criterion, 'test')
else:
optimizer = optim.SGD(model.parameters(),
lr=0.001 if vars.pre_trained else 0.06,
momentum=0.9)
#optimizer = optim.Adam(model.parameters(), lr=0.05)
# Decay LR by a factor of 0.6 every 6 epochs
exp_lr_scheduler = lr_scheduler.StepLR(
optimizer, step_size=10 if vars.pre_trained else 6, gamma=0.6)
model = model.to(vars.device)
model = train_model(model, vars.criterion, optimizer, exp_lr_scheduler,
vars.num_epochs)
visualize_model(model)
# ######################################################################
# # ConvNet as fixed feature extractor
# # Here, we need to freeze all the network except the final layer. We need
# # to set ``requires_grad == False`` to freeze the parameters so that the
# # gradients are not computed in ``backward()``.
# # You can read more about this in the documentation
# # `here <https://pytorch.org/docs/notes/autograd.html#excluding-subgraphs-from-backward>`__.
# #
# model_conv = torchvision.models.resnet18(pretrained=True)
# for param in model_conv.parameters():
# param.requires_grad = False
# try:
# s = sys.argv[1]
# except IndexError:
# s = ""
# create_interactions(s)
epochs = 1
for i in range(0, 10):
# Generate new data
create_interactions(str(i))
# Load
model = load_model(name="model")
model.compile(loss='mse',
optimizer=RMSprop())
# Train
train_filename = "/ssd/train_extra.csv{}".format(i)
model, losses = train_model(model, epochs, train_filename, nb_epoch=2)
# Test
print("MSE", losses[-1])
test_filename = "/ssd/test_extra.csv{}".format(i)
m = test_model(model, test_filename)
# Save model
save_model(model, name="model")
# if m > 0.93:
# break
if args.phase == 'train':
dataloader = data.DataLoader(DataLoader(args),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
device = torch.device("cuda:0")
print("constructing model ....")
model1 = EDVR(128, args.nframes, 8, 5, 40)
model1 = nn.DataParallel(model1.to(device), gpuids)
model2 = D()
model2 = nn.DataParallel(model2.to(device), gpuids)
if args.resume:
model1 = load_state_dict(model1, args.model_path)
model2 = load_state_dict(model2, args.dis_model_path)
print("model constructed")
optimizer1 = torch.optim.Adam(model1.parameters(), lr=args.lr)
scheduler1 = ExponentialLR(optimizer1, gamma=args.gamma)
optimizer2 = torch.optim.Adam(model2.parameters(), lr=args.lr)
scheduler2 = ExponentialLR(optimizer2, gamma=args.gamma)
summery_writer = SummaryWriter(args.log_dir)
train_model(model1, model2, optimizer1, optimizer2, scheduler1, scheduler2,
dataloader, summery_writer, device, args)
def train_or_eval(do_train=True,
weights_type_bits=None,
hidden_sizes=None,
use_pca=None,
use_old_pca=None,
pca_d=None):
weights_type, bits = ask_weight_info() if weights_type_bits is None \
else weights_type_bits
hidden_sizes = ask_layer_info() if hidden_sizes is None else hidden_sizes
use_pca = ask_use_pca() if use_pca is None else use_pca
# TODO: need to ask for the number of hidden layers and their
# sizes. When compiling / evaluating, we should be able to get
# that information by looking at the weights, but that might mess
# with our current method of determining whether the weights are
# quantized or not. An easy fix is probably to actually check for
# shift/scale values (maybe by looking at the shape of where they
# should be) instead of only looking at the length of the weights
# tuple.
if use_pca:
# Check if any PCA data exists.
if Path('tf/emnist/test_pca.pkl').is_file():
data = load_pickled('tf/emnist/test_pca.pkl')
if use_old_pca:
print('Using existing PCA data.')
else:
if use_old_pca is None:
questions = [
inquirer.Confirm(
'old_pca',
message=
'Use existing PCA data (%d principal components)?'
% data.images.shape[1])
]
if inquirer.prompt(questions)['old_pca']:
print('Using existing PCA data.')
else:
gen_pca(pca_d)
else:
if use_old_pca: print('Using existing PCA data.')
else: gen_pca(pca_d)
else:
if use_old_pca:
print('train_or_eval error: use_old_pca is set but there \
is no existing PCA data')
exit(-1)
else:
gen_pca(pca_d)
# Write PCA flag to disk so we can easily check later whether PCA
# is being used or not (so we use the appropriate number of
# batches during evaluation).
save_pickled('pca.pkl', use_pca)
# Choose between float and quantized models.
model = float_model if weights_type == 'float' else quantized_model
# Load parameters and data for the chosen dataset.
_, load_data \
= choose_dataset('emnist' + ('_pca' if use_pca else ''))
print('Loading data...')
train_data, validation_data, test_data = load_data('tf/')
# Choose which set to use (train, test, etc.).
# We pass 3 for combined train+validation if PCA isn't being used,
# otherwise we pass 2 to use only training data.
images, labels = choose_images_labels(train_data, validation_data,
test_data, 2 if use_pca else 3)
example_shape = images.shape[1:]
input_size = reduce(mul, example_shape, 1)
with tf.Graph().as_default():
# Build all of the ops.
print("Initializing...")
# Choose between a few different values for learning rate and
# stuff depending on the options chosen by the user.
learning_rate, max_steps, decay_step, decay_factor = choose_hypers(
weights_type, bits)
batch_size = 100
x, y, weights, logits, loss_op, pred_op, train_op = build_ops(
100, bits, learning_rate, decay_step, decay_factor, model,
input_size, hidden_sizes)
# Create session and initialize variables.
sess = init_session()
# Go.
if do_train:
print('Training model...')
seq = train_model(sess,
model,
x,
y,
train_op,
loss_op,
pred_op,
weights,
images,
labels,
batch_size,
max_steps,
'tf/models/default/',
bits,
1.0,
log=False)
i = 0
for _, acc in seq:
i += 1
render_progress_bar(i, max_steps,
(colored('Accuracy', 'yellow') + ': %0.02f%%') \
% (acc * 100.0))
clear_bottom_bar()
print('Done training model.')
print('Selecting weights with best accuracy (%.02f%%).' %
(acc * 100.0))
print(str(acc))
else:
print('Evaluating model...')
model.load_weights(sess,
'tf/models/default/',
num_bits=bits,
pca_d=pca_d)
acc = evaluate(sess, x, y, pred_op, images, labels, 100, log=False)
print('acc: %.02f' % acc)
def train_and_evaluate(num_epochs, model, optimizer, loss_fn, train_dataloader,
val_dataloader, early_stopping_criteria, directory,
use_bert, use_mongo):
"""Train on training set and evaluate on evaluation set
Args:
num_epochs: Number of epochs to run the training and evaluation
model: Model
optimizer: Optimizer
loss_fn: Loss function
dataloader: Dataloader for the training set
val_dataloader: Dataloader for the validation set
scheduler: Scheduler
directory: Directory path name to story the logging files
Returns train and evaluation metrics with epoch, loss, accuracy, recall, precision and f1-score
"""
train_metrics = pd.DataFrame(
columns=['epoch', 'loss', 'accuracy', 'recall', 'precision', 'f1'])
val_metrics = pd.DataFrame(
columns=['epoch', 'loss', 'accuracy', 'recall', 'precision', 'f1'])
best_val_loss = float("inf")
early_stop_step = 0
for epoch in trange(num_epochs, desc="Epoch"):
### TRAINING ###
train_results = train_model(model, optimizer, loss_fn,
train_dataloader, device, use_bert)
train_metrics.loc[len(train_metrics)] = {
'epoch': epoch,
'loss': train_results['loss'],
'accuracy': train_results['accuracy'],
'recall': train_results['recall'],
'precision': train_results['precision'],
'f1': train_results['f1']
}
if use_mongo: log_scalars(train_results, "Train")
### EVALUATION ###
val_results = evaluate_model(model, optimizer, loss_fn, val_dataloader,
device, use_bert)
val_metrics.loc[len(val_metrics)] = {
'epoch': epoch,
'loss': val_results['loss'],
'accuracy': val_results['accuracy'],
'recall': val_results['recall'],
'precision': val_results['precision'],
'f1': val_results['f1']
}
if use_mongo: log_scalars(val_results, "Validation")
#Save best and latest state
best_model = val_results['loss'] < best_val_loss
#last_model = epoch == num_epochs-1
if best_model:
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'optim_dict': optimizer.state_dict()
},
directory=directory,
checkpoint='best_model.pth.tar')
#Early stopping
if val_results['loss'] >= best_val_loss:
early_stop_step += 1
print("Early stop step:", early_stop_step)
else:
best_val_loss = val_results['loss']
early_stop_step = 0
stop_early = early_stop_step >= early_stopping_criteria
if stop_early:
print("Stopping early at epoch {}".format(epoch))
return train_metrics, val_metrics
print('\n')
print('Train Loss: {} | Train Acc: {}'.format(
train_results['loss'], train_results['accuracy']))
print('Valid Loss: {} | Valid Acc: {}'.format(val_results['loss'],
val_results['accuracy']))
print('Train recall: {} | Train precision: {} | Train f1: {}'.format(
train_results['recall'], train_results['precision'],
train_results['f1']))
print('Valid recall: {} | Valid precision: {} | Valid f1 {}'.format(
val_results['recall'], val_results['precision'],
val_results['f1']))
return train_metrics, val_metrics
if not Path(file + '.json').is_file():
print("File not found")
load = 0
else:
model = load_model(file)
if load == 0:
print("Creating neural network...")
file = input("Name: ")
epochs = int(input("Epcohs: "))
batch = int(input("Batch Size: "))
model = train_model(X_train, y_train, batch, epochs, file)
more_train = int(input("More training (0/1): "))
if more_train == 1:
more_epochs = int(input("How much epochs: "))
batch = int(input("Batch Size: "))
model = train_model(X_train, y_train, batch, more_epochs, file)
predict = int(input("Predict y_test (0/1): "))
y_pred = model.predict(X_test)
y_pred_bool = (y_pred > 0.5)
if predict == 1:
from train import train_model
from data_loader import load
from examples.NIPS.MNIST.mnist import test_MNIST, MNIST_Net, neural_predicate
from model import Model
from optimizer import Optimizer
from network import Network
import torch
queries = load('train_data.txt')
test_queries = load('test_data.txt')
with open('addition.pl') as f:
problog_string = f.read()
network = MNIST_Net()
net = Network(network, 'mnist_net', neural_predicate)
net.optimizer = torch.optim.Adam(network.parameters(), lr=0.001)
model = Model(problog_string, [net], caching=False)
optimizer = Optimizer(model, 2)
train_model(model,
queries,
1,
optimizer,
test_iter=1000,
test=lambda x: x.accuracy(test_queries, test=True),
snapshot_iter=10000)
if use_cuda:
print('Using GPU')
model.cuda()
else:
print('Using CPU')
optimizer = optim.SGD(model.parameters(),
lr=config["lr"],
momentum=config["momentum"])
criterion = nn.CrossEntropyLoss()
# Run the functions and save the best model in the function model_ft.
model_ft, losses_train, accuracy_train, losses_val, accuracy_val = train_model(
model,
criterion,
optimizer,
dataloders,
dataset_sizes,
use_cuda,
num_epochs=config["epochs"])
plt.figure(1)
plt.subplot(221)
plt.plot(losses_train)
plt.ylabel('Losses Train')
plt.subplot(222)
plt.plot(accuracy_train)
plt.ylabel('Accuracy Train')
plt.subplot(223)
plt.plot(losses_val)
plt.ylabel('Losses Eval')
plt.subplot(224)
print_partition_index(val_mask, "Val", y_val)
print_partition_index(test_mask, "Test", y_test)
# Some preprocessing
features = preprocess_features(features)
known_percent = show_data_stats(adj, features, y_train, y_val, y_test,
train_mask, val_mask, test_mask,
SHOW_TEST_VAL_DATASET_STATS)
model_func, support, sub_sampled_support, num_supports = get_model_and_support(
'k-nn', adj, initial_train_mask, train_mask, val_mask, test_mask,
WITH_TEST)
test_acc, list_node_correctly_classified = train_model(model_func,
num_supports,
support,
features,
y_train,
y_val,
y_test,
train_mask,
val_mask,
test_mask,
sub_sampled_support,
VERBOSE_TRAINING,
seed=seed,
list_adj=list_adj)
print(str(test_acc) + "%")
correct_paths_to_known, incorrect_paths_to_known = get_classification_stats(
list_node_correctly_classified, get_list_from_mask(test_mask),
paths_to_known_list)
print_classification_stats(correct_paths_to_known, incorrect_paths_to_known)
smoothing_parameter = 0.13
pos_prior = 0.78
response_list = []
context_list = []
label_response_dict = {}
word_frequency = {}
tokenizer = TweetTokenizer()
classified = []
sarcasm_tokens, non_sarcasm_tokens, total_sarcasm_tokens, total_non_sarcasm_tokens = train_model(
)
with open("data/test.jsonl", encoding="utf-8") as json_file:
data = json.loads("[" + json_file.read().replace("}\n{", "},\n{") + "]")
for p in data:
response_list.append(p["response"].replace("@USER", ""))
context_list.append(p["context"])
for line in response_list:
words = tokenizer.tokenize(line.lower())
sarcasm_prob = 0
non_sarcasm_prob = 0
def main(current_path, args):
print(args)
# Give a folder path as an argument with '--log_dir' to save
# TensorBoard summaries. Default is a log folder in current directory.
# Create the directory for TensorBoard variables if there is not.
if not os.path.exists(args.log_dir):
os.makedirs(args.log_dir)
filename = args.file
# preprocessing raw data and building vocabulary
vocabulary = read_data(filename)
print('Data size', len(vocabulary))
# Step 2: Build the dictionary and replace rare words with UNK token
# Filling 4 global variables:
# data - list of codes (integers from 0 to vocabulary_size-1).
# This is the original text but words are replaced by their codes
# count - map of words(strings) to count of occurrences
# dictionary - map of words(strings) to their codes(integers)
# reverse_dictionary - maps codes(integers) to words(strings)
data, count, dictionary, reverse_dictionary = build_dataset(vocabulary, 0)
del vocabulary # Hint to reduce memory.
vocabulary_size = len(dictionary)
num_sampled = args.neg_samples # Number of negative examples to sample.
train_args = {
'batch_size': args.train_bs,
'embedding_size': args.emb_size, # Dimension of the embedding vector.
'skip_window':
args.skip_window, # How many words to consider left and right.
'num_skips': args.
num_skips, # How many times to reuse an input to generate a label.
'num_steps': args.epochs, # Number of epochs
'log_dir': args.log_dir
}
model = Word2Vec(args.train_bs, args.emb_size, vocabulary_size,
num_sampled)
graph = model._graph
embeddings = None
with tf.Session(graph=graph) as session:
# Open a writer to write summaries.
writer = tf.summary.FileWriter(args.log_dir, session.graph)
embeddings = train_model(session, model, reverse_dictionary, writer,
data, **train_args)
#similarity = model.similarity.eval()
np.savetxt(args.emb_file, embeddings)
# Save the model for checkpoints.
model.savemodel.save(session, os.path.join(args.log_dir, 'model.ckpt'))
# Create a configuration for visualizing embeddings with the labels in TensorBoard.
config = projector.ProjectorConfig()
embedding_conf = config.embeddings.add()
embedding_conf.tensor_name = model.embeddings.name
embedding_conf.metadata_path = os.path.join(args.log_dir,
'metadata.tsv')
projector.visualize_embeddings(writer, config)
# Closing writer
writer.close()
print("Plotting items on TSNE plot...")
runTSNE(embeddings, reverse_dictionary, args.tsne_img_file)
print('\n')
with open('data2.bin', 'wb') as fp:
pickle.dump(data, fp)
def createWord2Vec(modelFile, dataset):
data = pickle.load(open(dataset, 'rb'))
# train model
model = Word2Vec(data, size=300, window=3, min_count=5, sg=1, workers=8)
# save model
model.save(modelFile)
def process_word2Vec(modelFile, words):
model = Word2Vec.load(modelFile)
for word in words:
print(word)
print(model.most_similar(word))
if __name__ == '__main__':
process_embedding()
tokenizer, max_len = train_model()
# uncomment and change csvpath to fill a file with prediction from our model
# since train_model save our best model in './model7.h5'we use this path to get it
# fill_csv(csvpath, './model7.h5', tokenizer, max_len)
transform=composed_for_valset,
train=False)
datasets = {'train': train_dataset, 'val': val_dataset}
# pdb.set_trace()
dataloaders = {
x: torch.utils.data.DataLoader(datasets[x],
batch_size=batch_size,
shuffle=True,
num_workers=4)
for x in ['train', 'val']
}
dataset_sizes = {x: len(datasets[x]) for x in ['train', 'val']}
class_names = datasets['train'].classes
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# pdb.set_trace()
track_sample = siss[91]
train_model(net.to(device),
optimizer,
scheduler,
dataloaders,
dataset_sizes,
track_sample,
batch_size,
num_epochs=epochs)
from build_model import build_model
from train import train_model
from utils import parse_args
from dataset import get_streams
if __name__ == "__main__":
args = parse_args()
cost = build_model(args)
train_stream, valid_stream = get_streams(args)
train_model(cost, train_stream, valid_stream, args)
def main():
print("Running train_aml.py")
parser = argparse.ArgumentParser("train")
parser.add_argument(
"--model_name",
type=str,
help="Name of the Model",
default="COVID19Articles_model.pkl",
)
parser.add_argument("--step_output",
type=str,
help=("output for passing data to next step"))
parser.add_argument("--dataset_version",
type=str,
help=("dataset version"))
parser.add_argument("--data_file_path",
type=str,
help=("data file path, if specified,\
a new version of the dataset will be registered"))
parser.add_argument(
"--caller_run_id",
type=str,
help=("caller run id, for example ADF pipeline run id"))
parser.add_argument("--dataset_name",
type=str,
help=("Dataset name. Dataset must be passed by name\
to always get the desired dataset version\
rather than the one used while the pipeline creation"))
args = parser.parse_args()
print("Argument [model_name]: %s" % args.model_name)
print("Argument [step_output]: %s" % args.step_output)
print("Argument [dataset_version]: %s" % args.dataset_version)
print("Argument [data_file_path]: %s" % args.data_file_path)
print("Argument [caller_run_id]: %s" % args.caller_run_id)
print("Argument [dataset_name]: %s" % args.dataset_name)
model_name = args.model_name
step_output_path = args.step_output
dataset_version = args.dataset_version
data_file_path = args.data_file_path
dataset_name = args.dataset_name
run = Run.get_context()
print("Getting training parameters")
# Load the training parameters from the parameters file
with open("parameters.json") as f:
pars = json.load(f)
try:
train_args = pars["training"]
except KeyError:
print("Could not load training values from file")
train_args = {}
# Log the training parameters
print(f"Parameters: {train_args}")
for (k, v) in train_args.items():
run.log(k, v)
run.parent.log(k, v)
# Get the dataset
if (dataset_name):
if (data_file_path == 'none'):
dataset = Dataset.get_by_name(run.experiment.workspace,
dataset_name,
dataset_version) # NOQA: E402, E501
else:
dataset = register_dataset(run.experiment.workspace, dataset_name,
os.environ.get("DATASTORE_NAME"),
data_file_path)
else:
e = ("No dataset provided")
print(e)
raise Exception(e)
# Link dataset to the step run so it is trackable in the UI
run.input_datasets['training_data'] = dataset
run.parent.tag("dataset_id", value=dataset.id)
# Split the data into test/train
df = dataset.to_pandas_dataframe()
data = split_data(df)
# Train the model
model = train_model(data, train_args)
# Evaluate and log the metrics returned from the train function
metrics = get_model_metrics(model, data)
for (k, v) in metrics.items():
run.log(k, v)
run.parent.log(k, v)
# Pass model file to next step
os.makedirs(step_output_path, exist_ok=True)
model_output_path = os.path.join(step_output_path, model_name)
joblib.dump(value=model, filename=model_output_path)
# Also upload model file to run outputs for history
os.makedirs('outputs', exist_ok=True)
output_path = os.path.join('outputs', model_name)
joblib.dump(value=model, filename=output_path)
run.tag("run_type", value="train")
print(f"tags now present for run: {run.tags}")
run.complete()
def main():
args = parser.parse_args()
logger = setup_logger()
## Use below for slurm setting.
# slurm_job_id = os.getenv('SLURM_JOB_ID', 'nojobid')
# slurm_proc_id = os.getenv('SLURM_PROC_ID', None)
# unique_identifier = str(slurm_job_id)
# if slurm_proc_id is not None:
# unique_identifier = unique_identifier + "_" + str(slurm_proc_id)
unique_identifier = ''
# Load config
config_path = args.config
episode_path = args.episode
if args.resume_ckpt and not args.config:
base_dir = os.path.dirname(os.path.dirname(args.resume_ckpt))
config_path = os.path.join(base_dir, 'config.yaml')
episode_path = os.path.join(base_dir, 'episode.yaml')
config = yaml.load(open(config_path), Loader=yaml.FullLoader)
episode = yaml.load(open(episode_path), Loader=yaml.FullLoader)
config['data_schedule'] = episode
# Override options
for option in args.override.split('|'):
if not option:
continue
address, value = option.split('=')
keys = address.split('.')
here = config
for key in keys[:-1]:
if key not in here:
raise ValueError('{} is not defined in config file. '
'Failed to override.'.format(address))
here = here[key]
if keys[-1] not in here:
raise ValueError('{} is not defined in config file. '
'Failed to override.'.format(address))
here[keys[-1]] = yaml.load(value, Loader=yaml.FullLoader)
# Set log directory
config['log_dir'] = os.path.join(args.log_dir, unique_identifier)
if not args.resume_ckpt and os.path.exists(config['log_dir']):
logger.warning('%s already exists' % config['log_dir'])
input('Press enter to continue')
# print the configuration
print(colorful.bold_white("configuration:").styled_string)
pprint(config)
print(colorful.bold_white("configuration end").styled_string)
if args.resume_ckpt and not args.log_dir:
config['log_dir'] = os.path.dirname(
os.path.dirname(args.resume_ckpt)
)
# Save config
os.makedirs(config['log_dir'], mode=0o755, exist_ok=True)
if not args.resume_ckpt or args.config:
config_save_path = os.path.join(config['log_dir'], 'config.yaml')
episode_save_path = os.path.join(config['log_dir'], 'episode.yaml')
yaml.dump(config, open(config_save_path, 'w'))
yaml.dump(episode, open(episode_save_path, 'w'))
print(colorful.bold_yellow('config & episode saved to {}'.format(config['log_dir'])).styled_string)
# Build components
data_scheduler = DataScheduler(config)
writer = SummaryWriter(config['log_dir'])
model = MODEL[config['model_name']](config, writer)
if args.resume_ckpt:
model.load_state_dict(torch.load(args.resume_ckpt))
model.to(config['device'])
train_model(config, model, data_scheduler, writer)
print(colorful.bold_white("\nThank you and Good Job Computer").styled_string)
from train import train_model
from data_loader import load
from examples.NIPS.MNIST.mnist import test_MNIST, MNIST_Net, neural_predicate
from model import Model
from optimizer import Optimizer
from network import Network
import torch
queries = load('train_data.txt')
with open('addition.pl') as f:
problog_string = f.read()
network = MNIST_Net()
net = Network(network, 'mnist_net', neural_predicate)
net.optimizer = torch.optim.Adam(network.parameters(), lr=0.001)
model = Model(problog_string, [net], caching=True)
optimizer = Optimizer(model, 2)
train_model(model,
queries,
1,
optimizer,
test_iter=1000,
test=test_MNIST,
snapshot_iter=10000)
parser.add_argument(
'--format',
'-f',
required=True,
choices=['10x_h5', '10x_mtx'],
help=
'The format of input files. It has two options, \'10x_h5\' or \'10x_mtx\'.'
)
args = parser.parse_args()
#Argv 1: Pathway of the input file
if args.task == 'train':
train.train_model(input_file=args.input_file,
weight_file=args.weight_file,
output_path=args.output_path,
gene_file=args.gene_file,
hidden_size=args.latent_size,
batch_size=args.batch_size,
optimizer=args.algorithm,
format_type=args.format)
#Argv 2: Pathway of the load_weight
elif args.task == 'prediction':
load_model.load_weight(input_file=args.input_file,
load_weight_file=args.load_file,
weight_file=args.weight_file,
gene_file=args.gene_file,
hidden_size=args.latent_size,
output_path=args.output_path,
filtered=args.filtered,
optimizer=args.algorithm,
mode=args.mode,
format_type=args.format)
def main(opt):
setup_seed(opt.seed)
if torch.cuda.is_available():
device = torch.device('cuda')
torch.cuda.set_device(opt.gpu_id)
else:
device = torch.device('cpu')
log_dir = opt.log_dir + '/' + opt.network + '-' + str(opt.layers)
utils.mkdir(log_dir)
model = get_model(opt)
# model = nn.DataParallel(model, device_ids=[1, 2, 3])
# model = nn.DataParallel(model, device_ids=[0, 1, 2, 3])
# model = nn.DataParallel(model, device_ids=[4, 5, 6, 7])
model = nn.DataParallel(model, device_ids=[0, 1, 2, 3, 4, 5, 6, 7])
# model = convert_model(model)
model = model.to(device)
summary_writer = SummaryWriter(logdir=log_dir)
weight = None
if opt.classes == 9:
weight = torch.tensor([1.8, 1, 1, 1.2, 1, 1.6, 1.2, 1.4, 1],
device=device)
elif opt.classes == 8:
weight = torch.tensor([1.8, 1, 1.2, 1.6, 1, 1.2, 1.8, 1],
device=device)
elif opt.classes == 2:
weight = torch.tensor([1., 1.5], device=device)
if opt.criterion == 'lsr':
criterion = LabelSmoothSoftmaxCE(weight=weight,
use_focal_loss=opt.use_focal,
reduction='sum').cuda()
elif opt.criterion == 'focal':
# criterion = FocalLoss(alpha=1, gamma=2, reduction='sum')
criterion = FocalLoss2()
elif opt.criterion == 'ce':
criterion = nn.CrossEntropyLoss(weight=weight, reduction='sum').cuda()
elif opt.criterion == 'bce':
criterion = nn.BCEWithLogitsLoss(weight=weight, reduction='sum').cuda()
if opt.classes > 2:
# all data
images, labels = utils.read_data(
os.path.join(opt.root_dir, opt.train_dir),
os.path.join(opt.root_dir, opt.train_label), opt.train_less,
opt.clean_data)
elif opt.classes == 2:
# 2 categories
images, labels = utils.read_ice_snow_data(
os.path.join(opt.root_dir, opt.train_dir),
os.path.join(opt.root_dir, opt.train_label))
# 7 categories
# images, labels = utils.read_non_ice_snow_data(
# os.path.join(opt.root_dir, opt.train_dir),
# os.path.join(opt.root_dir, opt.train_label))
################ devide set #################
if opt.fore:
train_im, train_label = images[opt.num_val:], labels[opt.num_val:]
val_im, val_label = images[:opt.num_val], labels[:opt.num_val]
else:
train_im, train_label = images[:-opt.num_val], labels[:-opt.num_val]
val_im, val_label = images[-opt.num_val:], labels[-opt.num_val:]
if opt.cu_mode:
train_data_1 = train_im[:4439], train_label[:4439]
train_data_2 = train_im[:5385], train_label[:5385]
train_data_3 = train_im, train_label
# train_datas = [train_data_1, train_data_2, train_data_3]
train_datas = [train_data_2, train_data_3]
opt.num_epochs //= len(train_datas)
else:
train_datas = [(train_im, train_label)]
val_data = val_im, val_label
#########################################
if opt.retrain:
state_dict = torch.load(opt.model_dir + '/' + opt.network + '-' +
str(opt.layers) + '-' + str(opt.crop_size) +
'_model.ckpt')
model.load_state_dict(state_dict)
################ optimizer #################
if opt.retrain and not opt.teacher_mode:
if opt.network in ['effnet']:
for param in model.module.parameters():
param.requires_grad = False
for param in model.module._fc.parameters():
param.requires_grad = True
# for param in model.module._swish.parameters():
# param.requires_grad = True
for param in model.module.model._bn1.parameters():
param.requires_grad = True
elif opt.network in ['resnet', 'resnext', \
'resnext_wsl_32x8d', 'resnext_wsl_32x16d', 'resnext_wsl_32x32d', \
'resnext_swsl']:
for param in model.parameters():
param.requires_grad = False
for param in model.module.fc.parameters():
param.requires_grad = True
for param in model.module.layer4[2].bn3.parameters():
# for param in model.module.layer4[2].bn2.parameters():
param.requires_grad = True
elif opt.network in ['pnasnet_m', 'senet_m']:
for param in model.module.parameters():
param.requires_grad = False
for param in model.module.classifier.parameters():
param.requires_grad = True
if opt.network == 'senet_m':
for param in model.module.features.layer4.parameters():
# for param in model.module.features.layer4[2].bn3.parameters():
param.requires_grad = True
elif opt.network in ['inception_v3']:
for param in model.parameters():
param.requires_grad = False
for param in model.fc.parameters():
param.requires_grad = True
for param in model.Mixed_7c.parameters():
param.requires_grad = True
else:
for param in model.module.parameters():
param.requires_grad = False
for param in model.module.last_linear.parameters():
param.requires_grad = True
if opt.network in ['se_resnext50_32x4d', 'se_resnext101_32x4d']:
for param in model.module.layer4[2].bn3.parameters():
param.requires_grad = True
elif opt.network in ['senet154']:
for param in model.module.layer4.parameters():
param.requires_grad = True
elif opt.network in ['xception']:
for param in model.module.bn4.parameters():
param.requires_grad = True
elif opt.network in ['inceptionresnetv2']:
for param in model.module.conv2d_7b.bn.parameters():
param.requires_grad = True
elif opt.network in ['inceptionv4']:
for param in model.module.features[-1].branch3.parameters():
param.requires_grad = True
elif opt.network in ['fixpnas']:
for param in model.module.cell_11.parameters():
param.requires_grad = True
for param in model.module.cell_10.parameters():
param.requires_grad = True
for param in model.module.cell_9.parameters():
param.requires_grad = True
params = filter(lambda p: p.requires_grad, model.parameters())
else:
if opt.network in ['effnet'] and not opt.retrain:
params = utils.add_weight_decay(model.module.model, 1e-4)
params.append({
'params': model.module._fc.parameters(),
'lr': opt.lr * 10
})
else:
params = utils.add_weight_decay(model, 1e-4)
################ optimizer #################
optimizer = get_optimizer(opt, params, weight_decay=1e-4)
if opt.scheduler in ['step', 'multistep', 'plateau', 'exponential']:
scheduler = get_schedule(opt, optimizer)
############################################
crop_size = opt.crop_size - 128
val_transforms = my_transform(False, crop_size)
val_dataset = WeatherDataset(val_data[0], val_data[1], val_transforms)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=opt.batch_size,
shuffle=False,
num_workers=8,
pin_memory=True)
for train_data in train_datas:
val_dis = np.bincount(val_label) + 1e-20
train_dis = np.bincount(train_data[1])
print(val_dis, opt.num_val)
print(train_dis, len(train_data[1]))
train_transforms = my_transform(True, crop_size, opt.cutout,
opt.n_holes, opt.length, opt.auto_aug,
opt.rand_aug)
train_dataset = WeatherDataset(train_data[0], train_data[1],
train_transforms)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=opt.batch_size,
shuffle=True,
num_workers=8,
drop_last=True,
pin_memory=True)
loader = {'train': train_loader, 'val': val_loader}
################ scheduler #################
if opt.scheduler in ['warmup', 'cycle', 'cos', 'cosw', 'sgdr']:
scheduler = get_schedule(opt, optimizer, len(train_loader))
############################################
model, acc = train_model(loader,
model,
criterion,
optimizer,
summary_writer,
scheduler=scheduler,
scheduler_name=opt.scheduler,
num_epochs=opt.num_epochs,
device=device,
is_inception=opt.is_inception,
mixup=opt.mixup,
cutmix=opt.cutmix,
alpha=opt.alpha,
val_dis=val_dis)
utils.mkdir(opt.model_dir)
torch.save(
model.state_dict(), opt.model_dir + '/' + opt.network + '-' +
str(opt.layers) + '-' + str(crop_size) + '_model.ckpt')
if args.dataset != "mscoco" and (not "ro" in args.dataset
or "predict" in args.trg_len_option
or "average" in args.trg_len_option):
#if "predict" in args.trg_len_option or "average" in args.trg_len_option:
#trg_len_dic = torch.load(os.path.join(data_path(args.dataset), "trg_len"))
trg_len_dic = torch.load(
str(args.data_prefix / "trg_len_dic" / args.dataset[-4:]))
trg_len_dic = organise_trg_len_dic(trg_len_dic)
if args.mode == 'train':
logger.info('starting training')
if args.dataset != "mscoco":
train_model(args,
model,
train_real,
dev_real,
src=SRC,
trg=TRG,
trg_len_dic=trg_len_dic)
else:
train_model(args,
model,
train_real,
dev_real,
src=None,
trg=mscoco_dataset,
trg_len_dic=trg_len_dic)
elif args.mode == 'test':
logger.info(
'starting decoding from the pre-trained model, on the test set...')
def launch_training():
train.train_model()
sequences_per_batch = 16
sequence_length = 16
learning_rate = 1e-3
number_epochs = 32
model = Char2Char(
dictionary_size,
embedding_size,
lstm_hidden_size,
number_lstm_layers,
dropout_probability,
output_size,
device,
).to(device)
print(model)
model = train_model(
model,
dataset,
device,
sequences_per_batch=sequences_per_batch,
sequence_length=sequence_length,
epochs=number_epochs,
learning_rate=learning_rate,
show_loss_plot=True,
)
output = sample(model, dataset, device, 1000, "int adxl_decode(")
print(output)