Deploying a Seq2Seq Model with TorchScript

Author: Matthew Inkawhich 1.2, this tutorial was updated to work with PyTorch 1.2

This tutorial will walk through the process of transitioning a sequence-to-sequence model to TorchScript using the TorchScript API. The model that we will convert is the chatbot model from the Chatbot tutorial. You can either treat this tutorial as a “Part 2” to the Chatbot tutorial and deploy your own pretrained model, or you can start with this document and use a pretrained model that we host. In the latter case, you can reference the original Chatbot tutorial for details regarding data preprocessing, model theory and definition, and model training.

What is TorchScript?

During the research and development phase of a deep learning-based project, it is advantageous to interact with an eager, imperative interface like PyTorch’s. This gives users the ability to write familiar, idiomatic Python, allowing for the use of Python data structures, control flow operations, print statements, and debugging utilities. Although the eager interface is a beneficial tool for research and experimentation applications, when it comes time to deploy the model in a production environment, having a graph-based model representation is very beneficial. A deferred graph representation allows for optimizations such as out-of-order execution, and the ability to target highly optimized hardware architectures. Also, a graph-based representation enables framework-agnostic model exportation. PyTorch provides mechanisms for incrementally converting eager-mode code into TorchScript, a statically analyzable and optimizable subset of Python that Torch uses to represent deep learning programs independently from the Python runtime.

The API for converting eager-mode PyTorch programs into TorchScript is found in the torch.jit module. This module has two core modalities for converting an eager-mode model to a TorchScript graph representation: tracing and scripting. The torch.jit.trace function takes a module or function and a set of example inputs. It then runs the example input through the function or module while tracing the computational steps that are encountered, and outputs a graph-based function that performs the traced operations. Tracing is great for straightforward modules and functions that do not involve data-dependent control flow, such as standard convolutional neural networks. However, if a function with data-dependent if statements and loops is traced, only the operations called along the execution route taken by the example input will be recorded. In other words, the control flow itself is not captured. To convert modules and functions containing data-dependent control flow, a scripting mechanism is provided. The torch.jit.script function/decorator takes a module or function and does not requires example inputs. Scripting then explicitly converts the module or function code to TorchScript, including all control flows. One caveat with using scripting is that it only supports a subset of Python, so you might need to rewrite the code to make it compatible with the TorchScript syntax.

For all details relating to the supported features, see the TorchScript language reference. To provide the maximum flexibility, you can also mix tracing and scripting modes together to represent your whole program, and these techniques can be applied incrementally.

Acknowledgements

This tutorial was inspired by the following sources:

1. Yuan-Kuei Wu’s pytorch-chatbot implementation: https://github.com/ywk991112/pytorch-chatbot

2. Sean Robertson’s practical-pytorch seq2seq-translation example: https://github.com/spro/practical-pytorch/tree/master/seq2seq-translation

3. FloydHub’s Cornell Movie Corpus preprocessing code: https://github.com/floydhub/textutil-preprocess-cornell-movie-corpus

Prepare Environment

First, we will import the required modules and set some constants. If you are planning on using your own model, be sure that the MAX_LENGTH constant is set correctly. As a reminder, this constant defines the maximum allowed sentence length during training and the maximum length output that the model is capable of producing.

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals

import torch
import torch.nn as nn
import torch.nn.functional as F
import re
import os
import unicodedata
import numpy as np

device = torch.device("cpu")


MAX_LENGTH = 10  # Maximum sentence length

# Default word tokens
PAD_token = 0  # Used for padding short sentences
SOS_token = 1  # Start-of-sentence token
EOS_token = 2  # End-of-sentence token

Model Overview

As mentioned, the model that we are using is a sequence-to-sequence (seq2seq) model. This type of model is used in cases when our input is a variable-length sequence, and our output is also a variable length sequence that is not necessarily a one-to-one mapping of the input. A seq2seq model is comprised of two recurrent neural networks (RNNs) that work cooperatively: an encoder and a decoder.

Image source: https://jeddy92.github.io/JEddy92.github.io/ts_seq2seq_intro/

Encoder

The encoder RNN iterates through the input sentence one token (e.g. word) at a time, at each time step outputting an “output” vector and a “hidden state” vector. The hidden state vector is then passed to the next time step, while the output vector is recorded. The encoder transforms the context it saw at each point in the sequence into a set of points in a high-dimensional space, which the decoder will use to generate a meaningful output for the given task.

Decoder

The decoder RNN generates the response sentence in a token-by-token fashion. It uses the encoder’s context vectors, and internal hidden states to generate the next word in the sequence. It continues generating words until it outputs an EOS_token, representing the end of the sentence. We use an attention mechanism in our decoder to help it to “pay attention” to certain parts of the input when generating the output. For our model, we implement Luong et al.’s “Global attention” module, and use it as a submodule in our decode model.

Data Handling

Although our models conceptually deal with sequences of tokens, in reality, they deal with numbers like all machine learning models do. In this case, every word in the model’s vocabulary, which was established before training, is mapped to an integer index. We use a Voc object to contain the mappings from word to index, as well as the total number of words in the vocabulary. We will load the object later before we run the model.

Also, in order for us to be able to run evaluations, we must provide a tool for processing our string inputs. The normalizeString function converts all characters in a string to lowercase and removes all non-letter characters. The indexesFromSentence function takes a sentence of words and returns the corresponding sequence of word indexes.

class Voc:
    def __init__(self, name):
        self.name = name
        self.trimmed = False
        self.word2index = {}
        self.word2count = {}
        self.index2word = {PAD_token: "PAD", SOS_token: "SOS", EOS_token: "EOS"}
        self.num_words = 3  # Count SOS, EOS, PAD

    def addSentence(self, sentence):
        for word in sentence.split(' '):
            self.addWord(word)

    def addWord(self, word):
        if word not in self.word2index:
            self.word2index[word] = self.num_words
            self.word2count[word] = 1
            self.index2word[self.num_words] = word
            self.num_words += 1
        else:
            self.word2count[word] += 1

    # Remove words below a certain count threshold
    def trim(self, min_count):
        if self.trimmed:
            return
        self.trimmed = True
        keep_words = []
        for k, v in self.word2count.items():
            if v >= min_count:
                keep_words.append(k)

        print('keep_words {} / {} = {:.4f}'.format(
            len(keep_words), len(self.word2index), len(keep_words) / len(self.word2index)
        ))
        # Reinitialize dictionaries
        self.word2index = {}
        self.word2count = {}
        self.index2word = {PAD_token: "PAD", SOS_token: "SOS", EOS_token: "EOS"}
        self.num_words = 3 # Count default tokens
        for word in keep_words:
            self.addWord(word)


# Lowercase and remove non-letter characters
def normalizeString(s):
    s = s.lower()
    s = re.sub(r"([.!?])", r" \1", s)
    s = re.sub(r"[^a-zA-Z.!?]+", r" ", s)
    return s


# Takes string sentence, returns sentence of word indexes
def indexesFromSentence(voc, sentence):
    return [voc.word2index[word] for word in sentence.split(' ')] + [EOS_token]

Define Encoder

We implement our encoder’s RNN with the torch.nn.GRU module which we feed a batch of sentences (vectors of word embeddings) and it internally iterates through the sentences one token at a time calculating the hidden states. We initialize this module to be bidirectional, meaning that we have two independent GRUs: one that iterates through the sequences in chronological order, and another that iterates in reverse order. We ultimately return the sum of these two GRUs’ outputs. Since our model was trained using batching, our EncoderRNN model’s forward function expects a padded input batch. To batch variable-length sentences, we allow a maximum of MAX_LENGTH tokens in a sentence, and all sentences in the batch that have less than MAX_LENGTH tokens are padded at the end with our dedicated PAD_token tokens. To use padded batches with a PyTorch RNN module, we must wrap the forward pass call with torch.nn.utils.rnn.pack_padded_sequence and torch.nn.utils.rnn.pad_packed_sequence data transformations. Note that the forward function also takes an input_lengths list, which contains the length of each sentence in the batch. This input is used by the torch.nn.utils.rnn.pack_padded_sequence function when padding.

TorchScript Notes:

Since the encoder’s forward function does not contain any data-dependent control flow, we will use tracing to convert it to script mode. When tracing a module, we can leave the module definition as-is. We will initialize all models towards the end of this document before we run evaluations.

class EncoderRNN(nn.Module):
    def __init__(self, hidden_size, embedding, n_layers=1, dropout=0):
        super(EncoderRNN, self).__init__()
        self.n_layers = n_layers
        self.hidden_size = hidden_size
        self.embedding = embedding

        # Initialize GRU; the input_size and hidden_size params are both set to 'hidden_size'
        #   because our input size is a word embedding with number of features == hidden_size
        self.gru = nn.GRU(hidden_size, hidden_size, n_layers,
                          dropout=(0 if n_layers == 1 else dropout), bidirectional=True)

    def forward(self, input_seq, input_lengths, hidden=None):
        # type: (Tensor, Tensor, Optional[Tensor]) -> Tuple[Tensor, Tensor]
        # Convert word indexes to embeddings
        embedded = self.embedding(input_seq)
        # Pack padded batch of sequences for RNN module
        packed = torch.nn.utils.rnn.pack_padded_sequence(embedded, input_lengths)
        # Forward pass through GRU
        outputs, hidden = self.gru(packed, hidden)
        # Unpack padding
        outputs, _ = torch.nn.utils.rnn.pad_packed_sequence(outputs)
        # Sum bidirectional GRU outputs
        outputs = outputs[:, :, :self.hidden_size] + outputs[:, : ,self.hidden_size:]
        # Return output and final hidden state
        return outputs, hidden

Define Decoder’s Attention Module

Next, we’ll define our attention module (Attn). Note that this module will be used as a submodule in our decoder model. Luong et al. consider various “score functions”, which take the current decoder RNN output and the entire encoder output, and return attention “energies”. This attention energies tensor is the same size as the encoder output, and the two are ultimately multiplied, resulting in a weighted tensor whose largest values represent the most important parts of the query sentence at a particular time-step of decoding.

# Luong attention layer
class Attn(nn.Module):
    def __init__(self, method, hidden_size):
        super(Attn, self).__init__()
        self.method = method
        if self.method not in ['dot', 'general', 'concat']:
            raise ValueError(self.method, "is not an appropriate attention method.")
        self.hidden_size = hidden_size
        if self.method == 'general':
            self.attn = nn.Linear(self.hidden_size, hidden_size)
        elif self.method == 'concat':
            self.attn = nn.Linear(self.hidden_size * 2, hidden_size)
            self.v = nn.Parameter(torch.FloatTensor(hidden_size))

    def dot_score(self, hidden, encoder_output):
        return torch.sum(hidden * encoder_output, dim=2)

    def general_score(self, hidden, encoder_output):
        energy = self.attn(encoder_output)
        return torch.sum(hidden * energy, dim=2)

    def concat_score(self, hidden, encoder_output):
        energy = self.attn(torch.cat((hidden.expand(encoder_output.size(0), -1, -1), encoder_output), 2)).tanh()
        return torch.sum(self.v * energy, dim=2)

    def forward(self, hidden, encoder_outputs):
        # Calculate the attention weights (energies) based on the given method
        if self.method == 'general':
            attn_energies = self.general_score(hidden, encoder_outputs)
        elif self.method == 'concat':
            attn_energies = self.concat_score(hidden, encoder_outputs)
        elif self.method == 'dot':
            attn_energies = self.dot_score(hidden, encoder_outputs)

        # Transpose max_length and batch_size dimensions
        attn_energies = attn_energies.t()

        # Return the softmax normalized probability scores (with added dimension)
        return F.softmax(attn_energies, dim=1).unsqueeze(1)

Define Decoder

Similarly to the EncoderRNN, we use the torch.nn.GRU module for our decoder’s RNN. This time, however, we use a unidirectional GRU. It is important to note that unlike the encoder, we will feed the decoder RNN one word at a time. We start by getting the embedding of the current word and applying a dropout. Next, we forward the embedding and the last hidden state to the GRU and obtain a current GRU output and hidden state. We then use our Attn module as a layer to obtain the attention weights, which we multiply by the encoder’s output to obtain our attended encoder output. We use this attended encoder output as our context tensor, which represents a weighted sum indicating what parts of the encoder’s output to pay attention to. From here, we use a linear layer and softmax normalization to select the next word in the output sequence.

# TorchScript Notes:
# ~~~~~~~~~~~~~~~~~~~~~~
#
# Similarly to the ``EncoderRNN``, this module does not contain any
# data-dependent control flow. Therefore, we can once again use
# **tracing** to convert this model to TorchScript after it
# is initialized and its parameters are loaded.
#

class LuongAttnDecoderRNN(nn.Module):
    def __init__(self, attn_model, embedding, hidden_size, output_size, n_layers=1, dropout=0.1):
        super(LuongAttnDecoderRNN, self).__init__()

        # Keep for reference
        self.attn_model = attn_model
        self.hidden_size = hidden_size
        self.output_size = output_size
        self.n_layers = n_layers
        self.dropout = dropout

        # Define layers
        self.embedding = embedding
        self.embedding_dropout = nn.Dropout(dropout)
        self.gru = nn.GRU(hidden_size, hidden_size, n_layers, dropout=(0 if n_layers == 1 else dropout))
        self.concat = nn.Linear(hidden_size * 2, hidden_size)
        self.out = nn.Linear(hidden_size, output_size)

        self.attn = Attn(attn_model, hidden_size)

    def forward(self, input_step, last_hidden, encoder_outputs):
        # Note: we run this one step (word) at a time
        # Get embedding of current input word
        embedded = self.embedding(input_step)
        embedded = self.embedding_dropout(embedded)
        # Forward through unidirectional GRU
        rnn_output, hidden = self.gru(embedded, last_hidden)
        # Calculate attention weights from the current GRU output
        attn_weights = self.attn(rnn_output, encoder_outputs)
        # Multiply attention weights to encoder outputs to get new "weighted sum" context vector
        context = attn_weights.bmm(encoder_outputs.transpose(0, 1))
        # Concatenate weighted context vector and GRU output using Luong eq. 5
        rnn_output = rnn_output.squeeze(0)
        context = context.squeeze(1)
        concat_input = torch.cat((rnn_output, context), 1)
        concat_output = torch.tanh(self.concat(concat_input))
        # Predict next word using Luong eq. 6
        output = self.out(concat_output)
        output = F.softmax(output, dim=1)
        # Return output and final hidden state
        return output, hidden

Define Evaluation

Greedy Search Decoder

As in the chatbot tutorial, we use a GreedySearchDecoder module to facilitate the actual decoding process. This module has the trained encoder and decoder models as attributes, and drives the process of encoding an input sentence (a vector of word indexes), and iteratively decoding an output response sequence one word (word index) at a time.

Encoding the input sequence is straightforward: simply forward the entire sequence tensor and its corresponding lengths vector to the encoder. It is important to note that this module only deals with one input sequence at a time, NOT batches of sequences. Therefore, when the constant 1 is used for declaring tensor sizes, this corresponds to a batch size of 1. To decode a given decoder output, we must iteratively run forward passes through our decoder model, which outputs softmax scores corresponding to the probability of each word being the correct next word in the decoded sequence. We initialize the decoder_input to a tensor containing an SOS_token. After each pass through the decoder, we greedily append the word with the highest softmax probability to the decoded_words list. We also use this word as the decoder_input for the next iteration. The decoding process terminates either if the decoded_words list has reached a length of MAX_LENGTH or if the predicted word is the EOS_token.

TorchScript Notes:

The forward method of this module involves iterating over the range of \([0, max\_length)\) when decoding an output sequence one word at a time. Because of this, we should use scripting to convert this module to TorchScript. Unlike with our encoder and decoder models, which we can trace, we must make some necessary changes to the GreedySearchDecoder module in order to initialize an object without error. In other words, we must ensure that our module adheres to the rules of the TorchScript mechanism, and does not utilize any language features outside of the subset of Python that TorchScript includes.

To get an idea of some manipulations that may be required, we will go over the diffs between the GreedySearchDecoder implementation from the chatbot tutorial and the implementation that we use in the cell below. Note that the lines highlighted in red are lines removed from the original implementation and the lines highlighted in green are new.

Changes:

• Added decoder_n_layers to the constructor arguments

  • This change stems from the fact that the encoder and decoder models that we pass to this module will be a child of TracedModule (not Module). Therefore, we cannot access the decoder’s number of layers with decoder.n_layers. Instead, we plan for this, and pass this value in during module construction.

• Store away new attributes as constants

  • In the original implementation, we were free to use variables from the surrounding (global) scope in our GreedySearchDecoder’s forward method. However, now that we are using scripting, we do not have this freedom, as the assumption with scripting is that we cannot necessarily hold on to Python objects, especially when exporting. An easy solution to this is to store these values from the global scope as attributes to the module in the constructor, and add them to a special list called __constants__ so that they can be used as literal values when constructing the graph in the forward method. An example of this usage is on NEW line 19, where instead of using the device and SOS_token global values, we use our constant attributes self._device and self._SOS_token.

• Enforce types of forward method arguments

  • By default, all parameters to a TorchScript function are assumed to be Tensor. If we need to pass an argument of a different type, we can use function type annotations as introduced in PEP 3107. In addition, it is possible to declare arguments of different types using MyPy-style type annotations (see doc).

• Change initialization of decoder_input

  • In the original implementation, we initialized our decoder_input tensor with torch.LongTensor([[SOS_token]]). When scripting, we are not allowed to initialize tensors in a literal fashion like this. Instead, we can initialize our tensor with an explicit torch function such as torch.ones. In this case, we can easily replicate the scalar decoder_input tensor by multiplying 1 by our SOS_token value stored in the constant self._SOS_token.

class GreedySearchDecoder(nn.Module):
    def __init__(self, encoder, decoder, decoder_n_layers):
        super(GreedySearchDecoder, self).__init__()
        self.encoder = encoder
        self.decoder = decoder
        self._device = device
        self._SOS_token = SOS_token
        self._decoder_n_layers = decoder_n_layers

    __constants__ = ['_device', '_SOS_token', '_decoder_n_layers']

    def forward(self, input_seq : torch.Tensor, input_length : torch.Tensor, max_length : int):
        # Forward input through encoder model
        encoder_outputs, encoder_hidden = self.encoder(input_seq, input_length)
        # Prepare encoder's final hidden layer to be first hidden input to the decoder
        decoder_hidden = encoder_hidden[:self._decoder_n_layers]
        # Initialize decoder input with SOS_token
        decoder_input = torch.ones(1, 1, device=self._device, dtype=torch.long) * self._SOS_token
        # Initialize tensors to append decoded words to
        all_tokens = torch.zeros([0], device=self._device, dtype=torch.long)
        all_scores = torch.zeros([0], device=self._device)
        # Iteratively decode one word token at a time
        for _ in range(max_length):
            # Forward pass through decoder
            decoder_output, decoder_hidden = self.decoder(decoder_input, decoder_hidden, encoder_outputs)
            # Obtain most likely word token and its softmax score
            decoder_scores, decoder_input = torch.max(decoder_output, dim=1)
            # Record token and score
            all_tokens = torch.cat((all_tokens, decoder_input), dim=0)
            all_scores = torch.cat((all_scores, decoder_scores), dim=0)
            # Prepare current token to be next decoder input (add a dimension)
            decoder_input = torch.unsqueeze(decoder_input, 0)
        # Return collections of word tokens and scores
        return all_tokens, all_scores

Evaluating an Input

Next, we define some functions for evaluating an input. The evaluate function takes a normalized string sentence, processes it to a tensor of its corresponding word indexes (with batch size of 1), and passes this tensor to a GreedySearchDecoder instance called searcher to handle the encoding/decoding process. The searcher returns the output word index vector and a scores tensor corresponding to the softmax scores for each decoded word token. The final step is to convert each word index back to its string representation using voc.index2word.

We also define two functions for evaluating an input sentence. The evaluateInput function prompts a user for an input, and evaluates it. It will continue to ask for another input until the user enters ‘q’ or ‘quit’.

The evaluateExample function simply takes a string input sentence as an argument, normalizes it, evaluates it, and prints the response.

def evaluate(searcher, voc, sentence, max_length=MAX_LENGTH):
    ### Format input sentence as a batch
    # words -> indexes
    indexes_batch = [indexesFromSentence(voc, sentence)]
    # Create lengths tensor
    lengths = torch.tensor([len(indexes) for indexes in indexes_batch])
    # Transpose dimensions of batch to match models' expectations
    input_batch = torch.LongTensor(indexes_batch).transpose(0, 1)
    # Use appropriate device
    input_batch = input_batch.to(device)
    lengths = lengths.to(device)
    # Decode sentence with searcher
    tokens, scores = searcher(input_batch, lengths, max_length)
    # indexes -> words
    decoded_words = [voc.index2word[token.item()] for token in tokens]
    return decoded_words


# Evaluate inputs from user input (stdin)
def evaluateInput(searcher, voc):
    input_sentence = ''
    while(1):
        try:
            # Get input sentence
            input_sentence = input('> ')
            # Check if it is quit case
            if input_sentence == 'q' or input_sentence == 'quit': break
            # Normalize sentence
            input_sentence = normalizeString(input_sentence)
            # Evaluate sentence
            output_words = evaluate(searcher, voc, input_sentence)
            # Format and print response sentence
            output_words[:] = [x for x in output_words if not (x == 'EOS' or x == 'PAD')]
            print('Bot:', ' '.join(output_words))

        except KeyError:
            print("Error: Encountered unknown word.")

# Normalize input sentence and call evaluate()
def evaluateExample(sentence, searcher, voc):
    print("> " + sentence)
    # Normalize sentence
    input_sentence = normalizeString(sentence)
    # Evaluate sentence
    output_words = evaluate(searcher, voc, input_sentence)
    output_words[:] = [x for x in output_words if not (x == 'EOS' or x == 'PAD')]
    print('Bot:', ' '.join(output_words))

Load Pretrained Parameters

Ok, its time to load our model!

Use hosted model

To load the hosted model:

1. Download the model here.

2. Set the loadFilename variable to the path to the downloaded checkpoint file.

3. Leave the checkpoint = torch.load(loadFilename) line uncommented, as the hosted model was trained on CPU.

Use your own model

To load your own pre-trained model:

1. Set the loadFilename variable to the path to the checkpoint file that you wish to load. Note that if you followed the convention for saving the model from the chatbot tutorial, this may involve changing the model_name, encoder_n_layers, decoder_n_layers, hidden_size, and checkpoint_iter (as these values are used in the model path).

2. If you trained the model on a CPU, make sure that you are opening the checkpoint with the checkpoint = torch.load(loadFilename) line. If you trained the model on a GPU and are running this tutorial on a CPU, uncomment the checkpoint = torch.load(loadFilename, map_location=torch.device('cpu')) line.

TorchScript Notes:

Notice that we initialize and load parameters into our encoder and decoder models as usual. If you are using tracing mode(torch.jit.trace) for some part of your models, you must call .to(device) to set the device options of the models and .eval() to set the dropout layers to test mode before tracing the models. TracedModule objects do not inherit the to or eval methods. Since in this tutorial we are only using scripting instead of tracing, we only need to do this before we do evaluation (which is the same as we normally do in eager mode).

save_dir = os.path.join("data", "save")
corpus_name = "cornell movie-dialogs corpus"

# Configure models
model_name = 'cb_model'
attn_model = 'dot'
#attn_model = 'general'
#attn_model = 'concat'
hidden_size = 500
encoder_n_layers = 2
decoder_n_layers = 2
dropout = 0.1
batch_size = 64

# If you're loading your own model
# Set checkpoint to load from
checkpoint_iter = 4000
# loadFilename = os.path.join(save_dir, model_name, corpus_name,
#                             '{}-{}_{}'.format(encoder_n_layers, decoder_n_layers, hidden_size),
#                             '{}_checkpoint.tar'.format(checkpoint_iter))

# If you're loading the hosted model
loadFilename = 'data/4000_checkpoint.tar'

# Load model
# Force CPU device options (to match tensors in this tutorial)
checkpoint = torch.load(loadFilename, map_location=torch.device('cpu'))
encoder_sd = checkpoint['en']
decoder_sd = checkpoint['de']
encoder_optimizer_sd = checkpoint['en_opt']
decoder_optimizer_sd = checkpoint['de_opt']
embedding_sd = checkpoint['embedding']
voc = Voc(corpus_name)
voc.__dict__ = checkpoint['voc_dict']


print('Building encoder and decoder ...')
# Initialize word embeddings
embedding = nn.Embedding(voc.num_words, hidden_size)
embedding.load_state_dict(embedding_sd)
# Initialize encoder & decoder models
encoder = EncoderRNN(hidden_size, embedding, encoder_n_layers, dropout)
decoder = LuongAttnDecoderRNN(attn_model, embedding, hidden_size, voc.num_words, decoder_n_layers, dropout)
# Load trained model params
encoder.load_state_dict(encoder_sd)
decoder.load_state_dict(decoder_sd)
# Use appropriate device
encoder = encoder.to(device)
decoder = decoder.to(device)
# Set dropout layers to eval mode
encoder.eval()
decoder.eval()
print('Models built and ready to go!')

Convert Model to TorchScript

Encoder

As previously mentioned, to convert the encoder model to TorchScript, we use scripting. The encoder model takes an input sequence and a corresponding lengths tensor. Therefore, we create an example input sequence tensor test_seq, which is of appropriate size (MAX_LENGTH, 1), contains numbers in the appropriate range \([0, voc.num\_words)\), and is of the appropriate type (int64). We also create a test_seq_length scalar which realistically contains the value corresponding to how many words are in the test_seq. The next step is to use the torch.jit.trace function to trace the model. Notice that the first argument we pass is the module that we want to trace, and the second is a tuple of arguments to the module’s forward method.

Decoder

We perform the same process for tracing the decoder as we did for the encoder. Notice that we call forward on a set of random inputs to the traced_encoder to get the output that we need for the decoder. This is not required, as we could also simply manufacture a tensor of the correct shape, type, and value range. This method is possible because in our case we do not have any constraints on the values of the tensors because we do not have any operations that could fault on out-of-range inputs.

GreedySearchDecoder

Recall that we scripted our searcher module due to the presence of data-dependent control flow. In the case of scripting, we do necessary language changes to make sure the implementation complies with TorchScript. We initialize the scripted searcher the same way that we would initialize an un-scripted variant.

### Compile the whole greedy search model to TorchScript model
# Create artificial inputs
test_seq = torch.LongTensor(MAX_LENGTH, 1).random_(0, voc.num_words).to(device)
test_seq_length = torch.LongTensor([test_seq.size()[0]]).to(device)
# Trace the model
traced_encoder = torch.jit.trace(encoder, (test_seq, test_seq_length))

### Convert decoder model
# Create and generate artificial inputs
test_encoder_outputs, test_encoder_hidden = traced_encoder(test_seq, test_seq_length)
test_decoder_hidden = test_encoder_hidden[:decoder.n_layers]
test_decoder_input = torch.LongTensor(1, 1).random_(0, voc.num_words)
# Trace the model
traced_decoder = torch.jit.trace(decoder, (test_decoder_input, test_decoder_hidden, test_encoder_outputs))

### Initialize searcher module by wrapping ``torch.jit.script`` call
scripted_searcher = torch.jit.script(GreedySearchDecoder(traced_encoder, traced_decoder, decoder.n_layers))

Print Graphs

Now that our models are in TorchScript form, we can print the graphs of each to ensure that we captured the computational graph appropriately. Since TorchScript allow us to recursively compile the whole model hierarchy and inline the encoder and decoder graph into a single graph, we just need to print the scripted_searcher graph

print('scripted_searcher graph:\n', scripted_searcher.graph)

Run Evaluation

Finally, we will run evaluation of the chatbot model using the TorchScript models. If converted correctly, the models will behave exactly as they would in their eager-mode representation.

By default, we evaluate a few common query sentences. If you want to chat with the bot yourself, uncomment the evaluateInput line and give it a spin.

# Use appropriate device
scripted_searcher.to(device)
# Set dropout layers to eval mode
scripted_searcher.eval()

# Evaluate examples
sentences = ["hello", "what's up?", "who are you?", "where am I?", "where are you from?"]
for s in sentences:
    evaluateExample(s, scripted_searcher, voc)

# Evaluate your input
#evaluateInput(traced_encoder, traced_decoder, scripted_searcher, voc)

Save Model

Now that we have successfully converted our model to TorchScript, we will serialize it for use in a non-Python deployment environment. To do this, we can simply save our scripted_searcher module, as this is the user-facing interface for running inference against the chatbot model. When saving a Script module, use script_module.save(PATH) instead of torch.save(model, PATH).

scripted_searcher.save("scripted_chatbot.pth")