advanced/dynamic_quantization_tutorial
Run in Google Colab
Colab
Download Notebook
Notebook
View on GitHub
GitHub
Note
Click here to download the full example code
(beta) Dynamic Quantization on an LSTM Word Language Model#
Created On: Oct 07, 2019 | Last Updated: Aug 27, 2024 | Last Verified: Nov 05, 2024
Author: James Reed
Edited by: Seth Weidman
Introduction#
Quantization involves converting the weights and activations of your model from float to int, which can result in smaller model size and faster inference with only a small hit to accuracy.
In this tutorial, we will apply the easiest form of quantization - dynamic quantization - to an LSTM-based next word-prediction model, closely following the word language model from the PyTorch examples.
# imports
import os
from io import open
import time
import torch
import torch.nn as nn
import torch.nn.functional as F
1. Define the model#
Here we define the LSTM model architecture, following the model from the word language model example.
class LSTMModel(nn.Module):
"""Container module with an encoder, a recurrent module, and a decoder."""
def __init__(self, ntoken, ninp, nhid, nlayers, dropout=0.5):
super(LSTMModel, self).__init__()
self.drop = nn.Dropout(dropout)
self.encoder = nn.Embedding(ntoken, ninp)
self.rnn = nn.LSTM(ninp, nhid, nlayers, dropout=dropout)
self.decoder = nn.Linear(nhid, ntoken)
self.init_weights()
self.nhid = nhid
self.nlayers = nlayers
def init_weights(self):
initrange = 0.1
self.encoder.weight.data.uniform_(-initrange, initrange)
self.decoder.bias.data.zero_()
self.decoder.weight.data.uniform_(-initrange, initrange)
def forward(self, input, hidden):
emb = self.drop(self.encoder(input))
output, hidden = self.rnn(emb, hidden)
output = self.drop(output)
decoded = self.decoder(output)
return decoded, hidden
def init_hidden(self, bsz):
weight = next(self.parameters())
return (weight.new_zeros(self.nlayers, bsz, self.nhid),
weight.new_zeros(self.nlayers, bsz, self.nhid))
2. Load in the text data#
Next, we load the Wikitext-2 dataset into a Corpus, again following the preprocessing from the word language model example.
class Dictionary(object):
def __init__(self):
self.word2idx = {}
self.idx2word = []
def add_word(self, word):
if word not in self.word2idx:
self.idx2word.append(word)
self.word2idx[word] = len(self.idx2word) - 1
return self.word2idx[word]
def __len__(self):
return len(self.idx2word)
class Corpus(object):
def __init__(self, path):
self.dictionary = Dictionary()
self.train = self.tokenize(os.path.join(path, 'train.txt'))
self.valid = self.tokenize(os.path.join(path, 'valid.txt'))
self.test = self.tokenize(os.path.join(path, 'test.txt'))
def tokenize(self, path):
"""Tokenizes a text file."""
assert os.path.exists(path)
# Add words to the dictionary
with open(path, 'r', encoding="utf8") as f:
for line in f:
words = line.split() + ['<eos>']
for word in words:
self.dictionary.add_word(word)
# Tokenize file content
with open(path, 'r', encoding="utf8") as f:
idss = []
for line in f:
words = line.split() + ['<eos>']
ids = []
for word in words:
ids.append(self.dictionary.word2idx[word])
idss.append(torch.tensor(ids).type(torch.int64))
ids = torch.cat(idss)
return ids
model_data_filepath = 'data/'
corpus = Corpus(model_data_filepath + 'wikitext-2')
3. Load the pretrained model#
This is a tutorial on dynamic quantization, a quantization technique that is applied after a model has been trained. Therefore, we’ll simply load some pretrained weights into this model architecture; these weights were obtained by training for five epochs using the default settings in the word language model example.
ntokens = len(corpus.dictionary)
model = LSTMModel(
ntoken = ntokens,
ninp = 512,
nhid = 256,
nlayers = 5,
)
model.load_state_dict(
torch.load(
model_data_filepath + 'word_language_model_quantize.pth',
map_location=torch.device('cpu'),
weights_only=True
)
)
model.eval()
print(model)
LSTMModel(
(drop): Dropout(p=0.5, inplace=False)
(encoder): Embedding(33278, 512)
(rnn): LSTM(512, 256, num_layers=5, dropout=0.5)
(decoder): Linear(in_features=256, out_features=33278, bias=True)
)
Now let’s generate some text to ensure that the pretrained model is working properly - similarly to before, we follow here
input_ = torch.randint(ntokens, (1, 1), dtype=torch.long)
hidden = model.init_hidden(1)
temperature = 1.0
num_words = 1000
with open(model_data_filepath + 'out.txt', 'w') as outf:
with torch.no_grad(): # no tracking history
for i in range(num_words):
output, hidden = model(input_, hidden)
word_weights = output.squeeze().div(temperature).exp().cpu()
word_idx = torch.multinomial(word_weights, 1)[0]
input_.fill_(word_idx)
word = corpus.dictionary.idx2word[word_idx]
outf.write(str(word.encode('utf-8')) + ('\n' if i % 20 == 19 else ' '))
if i % 100 == 0:
print('| Generated {}/{} words'.format(i, 1000))
with open(model_data_filepath + 'out.txt', 'r') as outf:
all_output = outf.read()
print(all_output)
| Generated 0/1000 words
| Generated 100/1000 words
| Generated 200/1000 words
| Generated 300/1000 words
| Generated 400/1000 words
| Generated 500/1000 words
| Generated 600/1000 words
| Generated 700/1000 words
| Generated 800/1000 words
| Generated 900/1000 words
b'surfaced' b'esteemed' b'out' b'center' b'condemned' b'a' b'hole' b'to' b'Zack' b'Brooks' b',' b'believes' b'for' b'casemated' b'performance' b'conical' b'including' b'(' b'Vendetta' b')'
b'.' b'Edwin' b'Pareil' b'offers' b'that' b'it' b'would' b'be' b'suitable' b'by' b'kingship' b'individuals' b']' b'below' b'three' b'months' b'and' b'were' b'sought' b'to'
b'be' b'.' b'Despite' b'numerous' b'respects' b'for' b'those' b',' b'it' b'was' b'both' b'reported' b'to' b'be' b'made' b'and' b'do' b'not' b'be' b'inappropriate'
b'a' b'area' b'Far' b'Power' b'<unk>' b'.' b'By' b'Holden' b',' b'the' b'village' b'underwent' b'mostly' b'only' b'two' b'at' b'the' b'Joyce' b'Council' b'('
b'<unk>' b')' b',' b'as' b'he' b'became' b'delicate' b'gameplay' b',' b'and' b'priory' b'was' b'to' b'control' b'for' b'all' b'over' b'ten' b'times' b','
b'saying' b'he' b'were' b'possible' b'to' b'solve' b'the' b'"' b'Jenny' b'System' b'"' b',' b'a' b'"' b'distinction' b'Creation' b'with' b'victim' b'or' b'mortar'
b'parties' b'"' b'.' b'Conservation' b'under' b'R' b"'Malley" b'overlap' b'amongst' b'<unk>' b'in' b'Africa' b',' b'Cantrell' b'and' b'Jupiter' b'in' b'2009' b'.' b'This'
b'forces' b'was' b'also' b'reduced' b'with' b'over' b'50' b'%' b'of' b'a' b'civilian' b'invasion' b',' b'which' b'are' b'reminiscent' b'of' b'<unk>' b'.' b'Caroline'
b'Evans' b'(' b'Christian' b'uniforms' b',' b'was' b'mentioned' b'to' b'improve' b'their' b'NSB' b'Inca' b'never' b'78' b'\xe2\x80\x93' b'16' b'years' b',' b'as' b'a'
b'sale' b')' b'was' b'a' b'mere' b'experts' b'at' b'"' b'a' b'subtropical' b'@-@' b'century' b',' b'perspective' b'"' b'for' b'speaking' b'teams' b'in' b'the'
b'Atlantic' b'language' b'.' b'(' b'Kevin' b'atrox' b')' b'was' b'often' b'keen' b'to' b'make' b'a' b'kakapo' b'to' b'move' b'the' b'Florida' b'situation' b'received'
b'at' b'tomb' b'points' b'.' b'With' b'a' b'third' b'total' b'down' b',' b'it' b'has' b'high' b',' b'with' b'each' b'of' b'them' b'and' b'such'
b'Los' b',' b'indicating' b'over' b'equal' b'dead' b'signature' b'.' b'Wiley' b'Cramp' b'claimed' b'that' b'"' b'it' b'was' b'a' b'<unk>' b'in' b'the' b'Roman'
b'obsession' b'to' b'said' b'that' b'extravagant' b'power' b'or' b'@-@' b'iron' b'drug' b'<unk>' b'and' b'absolute' b'features' b'progress' b'is' b'left' b'to' b'enhance' b','
b'spacing' b',' b'Transits' b',' b'and' b'Bangor' b'"' b'.' b'As' b'part' b'of' b'specific' b'occupation' b',' b'they' b'are' b'to' b'engage' b'at' b'a'
b'maximum' b'level' b'.' b'<eos>' b'<eos>' b'=' b'=' b'Development' b'=' b'=' b'<eos>' b'<eos>' b'The' b'titles' b'used' b'written' b'by' b'<unk>' b'Cartwright' b','
b'between' b'<unk>' b'and' b'<unk>' b'5' b'or' b'he' b'agreed' b'a' b'entrance' b'to' b'doors' b',' b'Sciences' b',' b'ninjas' b',' b'G.' b'Cecil' b','
b'<unk>' b'and' b'a' b'absence' b'to' b'16' b'points' b'of' b'the' b'form' b';' b'instead' b'of' b'embryos' b'gives' b'seek' b'.' b'\xc3\x8dmar' b'arrives' b'on'
b'the' b'people' b'of' b'the' b'high' b'directional' b'name' b'as' b'none' b'of' b'the' b'oldest' b'ice' b'viewing' b'series' b'.' b'<eos>' b'The' b'first' b'@-@'
b'season' b'defense' b'(' b'which' b'could' b'have' b'since' b'manage' b'value' b')' b',' b'glider' b',' b'ZX' b'Ministry' b'(' b'dividing' b'Union' b')' b','
b'are' b'also' b'eldest' b'.' b'Occasionally' b',' b'constructive' b'for' b'a' b'average' b'scaled' b'has' b'a' b'link' b'<unk>' b'daily' b'off' b'as' b'the' b'clock'
b'effect' b'themselves' b'<unk>' b'down' b',' b'a' b'<unk>' b'reason' b'on' b'many' b'pads' b'who' b'died' b'around' b'50' b'%' b',' b'and' b'such' b'as'
b'its' b'<unk>' b',' b'2' b'@.@' b'8' b'cm' b'(' b'5' b'@.@' b'4' b'lb' b';' b'6' b'@.@' b'44' b'mph' b')' b'at' b'a'
b'deep' b'fastball' b'on' b'a' b'role' b'consisting' b'of' b'8' b'honors' b',' b'and' b'18' b'birds' b'.' b'The' b'female' b'is' b'a' b'definitive' b'game'
b',' b'while' b'Korean' b'forests' b'helicopters' b'that' b'only' b'largest' b'hair' b'shorter' b'denominations' b'sea' b'.' b'It' b'aggressor' b'on' b'30' b'October' b'1922' b'('
b'PHP' b')' b'.' b'<eos>' b'<eos>' b'=' b'=' b'Charts' b'=' b'=' b'<eos>' b'<eos>' b'The' b'kakapo' b'sits' b'on' b'4' b'April' b'1801' b'in'
b',' b'Blaine' b"'s" b'clerical' b'producer' b'claims' b'that' b'they' b'were' b'thought' b'or' b'<unk>' b'on' b'1' b'January' b'2015' b',' b'though' b'it' b'is'
b'apparent' b'by' b'Christians' b'.' b'Once' b'the' b'fungus' b'becomes' b'large' b',' b'poised' b'homes' b',' b'improved' b'derived' b',' b'as' b'conferred' b'by' b'symbols'
b',' b'including' b'common' b'telephone' b',' b'Wife' b',' b'objects' b',' b'spruce' b'and' b'other' b'are' b'one' b'of' b'its' b'studies' b'and' b'its' b'abundance'
b'only' b'as' b'being' b'under' b'dark' b'superficially' b',' b'spoken' b'Gedney' b',' b'on' b'a' b'woodcut' b'with' b'supplement' b',' b'carrying' b'plurality' b'as' b'its'
b'eyes' b'prepared' b'from' b'<unk>' b'else' b'and' b'unrelated' b'.' b'Travelling' b'areas' b',' b'redundant' b',' b'quote' b',' b'the' b'<unk>' b',' b'rivers' b'and'
b'restrict' b'spiracles' b'probably' b'act' b',' b'and' b'can' b'engage' b'a' b'white' b'authority' b',' b'improve' b'only' b'so' b'in' b'which' b'old' b',' b'could'
b'be' b'lost' b'.' b'There' b'are' b'no' b'general' b'stories' b'in' b'oath' b'\xe2\x80\x94' b'perch' b',' b'and' b'they' b'eruptions' b'in' b'the' b'transient' b'translation'
b'.' b'A' b'combination' b'of' b'justify' b'these' b'radiation' b'types' b'will' b'plaintive' b'short' b'types' b'or' b'pale' b'parents' b'.' b'<unk>' b'that' b'bear' b'the'
b'NHA' b'that' b'gives' b'about' b'difficult' b'water' b'because' b'km2' b'are' b'in' b'1965' b',' b'have' b'a' b'surface' b'"' b'Christ' b'"' b'eyes' b'during'
b'"' b'while' b'most' b'observed' b'.' b'"' b'Similarly' b',' b'in' b'less' b'cases' b'they' b'they' b'were' b'<unk>' b',' b'no' b'skill' b'of' b'certain'
b'eggs' b'to' b'be' b'given' b',' b'if' b'they' b'are' b'lower' b'today' b'of' b'<unk>' b'.' b'AML' b'Have' b'Goldstein' b'Enemy' b'Will' b'G.' b'Jie'
b'assert' b'was' b'weren' b'Minister' b'with' b'a' b'native' b'group' b'of' b'6' b'@.@' b'2' b'million' b'Arabs' b'the' b'same' b'beam' b'and' b',' b'bullied'
b'a' b'outgassing' b'before' b'Bat' b'activity' b',' b'fairly' b'or' b'twice' b',' b'harmonic' b'ones' b'and' b'legs' b'greatly' b'the' b'presence' b'of' b'wings' b'business'
b'.' b'Between' b'fertilized' b'and' b'12' b'birds' b',' b'they' b'recorded' b'reduced' b'territorial' b'females' b',' b'either' b'taking' b'in' b'high' b'situations' b'particularly' b'feeds'
b'toward' b'the' b'National' b'chapels' b'.' b'The' b'males' b'rarely' b'find' b'observed' b'outflow' b';' b'the' b'other' b'trees' b'would' b'be' b'folded' b'to' b'Moscow'
b'meter' b'and' b'rarely' b'recognise' b'the' b'ball' b'.' b'Like' b'modern' b'females' b',' b'they' b'were' b'recognized' b'by' b'Method' b',' b'employed' b'on' b'35'
b'to' b'11' b'AU' b'.' b'These' b'cats' b'<unk>' b',' b'in' b'which' b'they' b'impact' b'and' b'eaten' b'.' b'<eos>' b'Muhammed' b'nucleus' b'is' b'suggested'
b'to' b'only' b'be' b'transported' b'to' b'bellied' b',' b'or' b'Herbert' b',' b'which' b'may' b'act' b'if' b'their' b'mature' b'descriptions' b'may' b'be' b'frustrated'
b'in' b'both' b'some' b'or' b'herding' b'prosperity' b'.' b'bird' b'a' b'complete' b'litter' b'eventually' b'<unk>' b'below' b'grey' b'planet' b'.' b'These' b'birds' b','
b'however' b',' b'as' b'they' b'can' b'be' b'tightly' b'inferred' b'.' b'The' b'kakapo' b'are' b'commonly' b'rank' b'of' b'hotels' b'on' b'stroke' b'antibodies' b','
b'both' b'day' b',' b'and' b'several' b'years' b'old' b'sites' b'.' b'This' b'behaviour' b'of' b'this' b'time' b'may' b'be' b'sulfur' b'.' b'High' b'Irish'
b'Therefore' b',' b'Indian' b',' b'Babyface' b'trajectory' b',' b'capability' b',' b'parking' b',' b'any' b'ballistic' b'volunteer' b',' b'someday' b'furniture' b',' b'were' b'pioneered'
b'by' b'humanity' b'(' b'derives' b'and' b'foliage' b',' b'as' b'of' b'GAR' b')' b',' b'and' b'also' b'yellowish' b'but' b'lays' b'organized' b'@-@' b'cause'
b'or' b'wolves' b'from' b'a' b'pair' b'of' b'permit' b'cloud' b'across' b'Pongolapoort' b'and' b'264' b'water' b'.' b'New' b'330' b',' b'degenerate' b'after' b'being'
It’s no GPT-2, but it looks like the model has started to learn the structure of language!
We’re almost ready to demonstrate dynamic quantization. We just need to define a few more helper functions:
bptt = 25
criterion = nn.CrossEntropyLoss()
eval_batch_size = 1
# create test data set
def batchify(data, bsz):
# Work out how cleanly we can divide the dataset into ``bsz`` parts.
nbatch = data.size(0) // bsz
# Trim off any extra elements that wouldn't cleanly fit (remainders).
data = data.narrow(0, 0, nbatch * bsz)
# Evenly divide the data across the ``bsz`` batches.
return data.view(bsz, -1).t().contiguous()
test_data = batchify(corpus.test, eval_batch_size)
# Evaluation functions
def get_batch(source, i):
seq_len = min(bptt, len(source) - 1 - i)
data = source[i:i+seq_len]
target = source[i+1:i+1+seq_len].reshape(-1)
return data, target
def repackage_hidden(h):
"""Wraps hidden states in new Tensors, to detach them from their history."""
if isinstance(h, torch.Tensor):
return h.detach()
else:
return tuple(repackage_hidden(v) for v in h)
def evaluate(model_, data_source):
# Turn on evaluation mode which disables dropout.
model_.eval()
total_loss = 0.
hidden = model_.init_hidden(eval_batch_size)
with torch.no_grad():
for i in range(0, data_source.size(0) - 1, bptt):
data, targets = get_batch(data_source, i)
output, hidden = model_(data, hidden)
hidden = repackage_hidden(hidden)
output_flat = output.view(-1, ntokens)
total_loss += len(data) * criterion(output_flat, targets).item()
return total_loss / (len(data_source) - 1)
4. Test dynamic quantization#
Finally, we can call torch.quantization.quantize_dynamic on the model!
Specifically,
We specify that we want the
nn.LSTMandnn.Linearmodules in our model to be quantizedWe specify that we want weights to be converted to
int8values
import torch.quantization
quantized_model = torch.quantization.quantize_dynamic(
model, {nn.LSTM, nn.Linear}, dtype=torch.qint8
)
print(quantized_model)
LSTMModel(
(drop): Dropout(p=0.5, inplace=False)
(encoder): Embedding(33278, 512)
(rnn): DynamicQuantizedLSTM(512, 256, num_layers=5, dropout=0.5)
(decoder): DynamicQuantizedLinear(in_features=256, out_features=33278, dtype=torch.qint8, qscheme=torch.per_tensor_affine)
)
The model looks the same; how has this benefited us? First, we see a significant reduction in model size:
def print_size_of_model(model):
torch.save(model.state_dict(), "temp.p")
print('Size (MB):', os.path.getsize("temp.p")/1e6)
os.remove('temp.p')
print_size_of_model(model)
print_size_of_model(quantized_model)
Size (MB): 113.944455
Size (MB): 79.738939
Second, we see faster inference time, with no difference in evaluation loss:
Note: we set the number of threads to one for single threaded comparison, since quantized models run single threaded.
torch.set_num_threads(1)
def time_model_evaluation(model, test_data):
s = time.time()
loss = evaluate(model, test_data)
elapsed = time.time() - s
print('''loss: {0:.3f}\nelapsed time (seconds): {1:.1f}'''.format(loss, elapsed))
time_model_evaluation(model, test_data)
time_model_evaluation(quantized_model, test_data)
loss: 5.167
elapsed time (seconds): 198.8
loss: 5.168
elapsed time (seconds): 113.1
Running this locally on a MacBook Pro, without quantization, inference takes about 200 seconds, and with quantization it takes just about 100 seconds.