chmxu / ecm-tf

implementation of ECM with tensorflow

Emotional Chatting Machine: Emotional Conversation Generation with Internal and External Memory

Introduction

This project is a tensorflow implement of Emotional Conversation Generation. We revise the interface of the seq2seq model on tensorflow to realize our model.

requirements

• Python 3.5
• Tensorflow 1.0.1
• Numpy

Usage

1. train

python train.py -m [model name]

We provide three models, including isolated emotion embedding('embedding') and internal memory('internal') and the mixed one('ECM'). You can set the batch size and the learning rate in this script. As long as we know, no optimizers other than SGD can perform well on this model, so there is no option for optimizers.

2. test

python test.py -m [model name]

This script will generate a pickle file named "test_result_"+model_chosen+".pkl" consisting of posts, ground truth and the generated responses.

3. apply

python apply.py -m [model name] -p [model path]

We provide a simple application script. When running this script, you need to enter what you want to say to the machine(which should be in Chinese and segmented) and the emotion category(1 for like, 2 for sadness, 3 for disgust, 4 for anger, 5 for happiness and 0 for other), then it will return you a response. We provide three pretrained models, and you can use your own trained model, too.

TODO

• Add external memory to the model.
• Speed up the application.

ecm-tf's Issues

can you give a data sample

i didnt find data_origin.pkl and other data sample;
can you give a data sample,then we can make our train data and test data

数据集 KeyError: '<START>'

请问train_data.json 是下面这个数据集吗？

而且我找不到data_medium.pkl and word2idx_mini.pkl,
希望您能提供帮助，谢谢！

Dimension Mismatch in MatMul in model_revise.py

decoder_outputs, decoder_states = tf.contrib.legacy_seq2seq.attention_decoder(decoder_inputs, h, attention_states, cell, num_heads=1, output_size=self.vocab_num)

is throwing following error

ValueError: Dimensions must be equal, but are 456 and 356 for 'attention_seq2seq/rnn/rnn/multi_rnn_cell/cell_0/cell_0/gru_cell/MatMul_68' (op: 'MatMul') with input shapes: [?,456], [356,512].

Error: No such file or directory: 'data_train.pkl'

Hello
When I run
python train.py -m 'ECM'
Why do I get error
No such file or directory: 'data_train.pkl'

Thanks in advance

where is train_data.json?

I think there is no data file in this repository,
How can I get data file ?