Deep Learning Library based on Theano

Design Philosophy

Totally speaking, in every task, _the operation flow/process is unchanged_ and _the only changing thing is just the specific operation in each operation_. (业务流程是不变的，变化的只是 具体的业务。)

Every artificial intelligence(AI) task involves four components: Model, Data, Execution and Evaluation.

The process of AI task is as follows:

So, by extracting the unchanged process, we got our deep learning framework.

Thdl framework contains four modules:

• data: for processing AND providing data;
• model: for building neural models;
• exeval: for executing and evaluating models using provided data;
• task: for building a task.

Features

• Support commonly used models, including convnets, RNNs and LSTMS.
• Support visualization of evaluation results.

Documentation

Preparing.

Requirements

• Python3
• Theano >= 0.9
• scipy
• Numpy
• Matplotlib
• xlwt
• nltk

Example

from thdl.data import text_classification as corpus
from thdl import model
from thdl.exeval import ClassifyExeEval
from thdl.task import ClassificationTask


# data
data_getter = corpus.SentenceGetter('./files/handled/trec.data')
data_processor = corpus.SentenceProcessor()
data_module = corpus.SentenceProvider(shuffle=True)
data_module.set_getter(data_getter)
data_module.set_processor(data_processor)
data_module.build()

# model
model_module = model.Model()
model_module.set_input_tensor(model.tensors.imatrix())
model_module.add_layer(model.layers.Embedding(data_module.get_embedding()))
model_module.add_layer(model.layers.Dropout(0.5))
model_module.add_layer(model.layers.LSTM(n_out=200))
model_module.add_layer(model.layers.Mean(axis=1))
model_module.add_layer(model.layers.Softmax(n_out=len(data_module.index_to_tag)))
model_module.set_output_tensor(model.tensors.fmatrix())
model_module.set_objective(model.objective.CategoricalCrossEntropy())
model_module.set_optimizer(model.optimizer.Adam(learning_rate=0.0001))
model_module.set_metrics([model.metrics.CategoricalAccuracy(),
                          model.metrics.Loss()],
                         model.metrics.Regularizer())
model_module.build()

# execution and evaluation
exeval_module = ClassifyExeEval(batch_size=50)
exeval_module.set_aspects('training', 'valid', 'test')

# task
task = ClassificationTask()
task.set_model(model_module)
task.set_data(data_module)
task.set_exeval(exeval_module)
task.set_logfile("temp.log")
task.hold_out_validation()