New model

[hhexiy]

• Separate original graph model into Graph and GraphEmbed
• Message passing in GraphEmbed

Still some TODOs:

• Encoder should run over tokens instead of entities
• Move generate() from the model class to NeuralSession

[hhexiy]

On a separate note, here's the plan for implementing mini-batching:

Instead of process a whole dialogue, we will input a pair of (partner utterance, agent utterance). The model encodes the partner utterance, updates the graph embedding, decode the agent utterance and update the graph embedding again.

Within a dialogue, we will initialize the hidden state, utterance, graph from the previous example but won't BP through it.

[percyliang]

Sounds good.

On Thu, Oct 6, 2016 at 10:29 PM, hhexiy notifications@github.com wrote:

On a separate note, here's the plan for implementing mini-batching:

Instead of process a whole dialogue, we will input a pair of (partner utterance, agent utterance) and the model encodes the partner utterance, updates the graph embedding, and decode the agent utterance.

Within a dialogue, we will initialize the hidden state, utterance, graph from the previous example but won't BP through it.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/stanfordnlp/game-dialogue/pull/14#issuecomment-252156399, or mute the thread https://github.com/notifications/unsubscribe-auth/AAakuANYLHN3QYoFi96Tv3Bqfl2BvuA9ks5qxdi7gaJpZM4KPFFh .

[hhexiy]

Some updates: graph.py

• GraphMetadata: schema, config (e.g. padding, feature size)
• GraphBatch: takes a batch of graphs and output normalized (padded) minibatches
• Graph: construct initial graph from KB and update graph structure given new utterances

graph_embedder.py

• Take data from GraphBatch
• Update utterance matrices
• Compute context (attention slots) given utterances

learn.py

• _run_batch(): take a sequence of minibatches where each batch continues from the previous batch, run truncated rnn over them.

encdec.py

• BasicEncoder/Decoder: rnn w/o graph or attention
• GraphEncoder: rnn, update graph at the end
• GraphDecoder: rnn + attention, update graph at the end
• CopyGraphDecoder: rnn + attention, update graph at the end, softmax over both vocab and attention scores

Put all test code under model/test.