indhub commented on a change in pull request #10568: [WIP] [MXNET-325] Model 
parallelism tutorial.
URL: https://github.com/apache/incubator-mxnet/pull/10568#discussion_r181894899
 
 

 ##########
 File path: docs/tutorials/gluon/model_parallelism.md
 ##########
 @@ -0,0 +1,373 @@
+
+# Model parallelism
+- This is model parallelized version of 
http://gluon.mxnet.io/chapter05_recurrent-neural-networks/rnns-gluon.html.
+- Similar to https://mxnet.incubator.apache.org/faq/model_parallel_lstm.html.
+
+
+```python
+import math
+import os
+import time
+import numpy as np
+import mxnet as mx
+from mxnet import gluon, autograd
+from mxnet.gluon import nn, rnn
+import collections
+```
+
+
+```python
+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)
+```
+
+
+```python
+class Corpus(object):
+    def __init__(self, path):
+        self.dictionary = Dictionary()
+        self.train = self.tokenize(path + 'train.txt')
+        self.valid = self.tokenize(path + 'valid.txt')
+        self.test = self.tokenize(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') as f:
+            tokens = 0
+            for line in f:
+                words = line.split() + ['<eos>']
+                tokens += len(words)
+                for word in words:
+                    self.dictionary.add_word(word)
+
+        # Tokenize file content
+        with open(path, 'r') as f:
+            ids = np.zeros((tokens,), dtype='int32')
+            token = 0
+            for line in f:
+                words = line.split() + ['<eos>']
+                for word in words:
+                    ids[token] = self.dictionary.word2idx[word]
+                    token += 1
+
+        return mx.nd.array(ids, dtype='int32')
+```
+
+`MultiGPULSTM` creates stacked LSTM with layers spread across multiple GPUs. 
+For example, `MultiGPULSTM(0, [1, 2, 2, 1], 400, 200, 0.5)` will create a 
stacked LSTM with one layer on GPU(0), two layers on GPU(1), two layers on 
GPU(2), one layer on GPU(3) with a hidden size of 400 embedding size of 200 and 
dropout probability of .5.
+
+
+```python
+class MultiGPULSTM(object):
+    
+    def __init__(self, start_device, num_layers_list, num_hidden, input_size, 
dropout):
+        """Create a MultiGPULSTM. num_layers_list dictates how many layers of 
LSTM
+        gets places in which device. For example, [1, 2, 2, 1] will create a 
stacked LSTM
+        with one layer on GPU(0), two layers on GPU(1), two layers on GPU(2), 
one layer on GPU(3)"""
+        self.lstm_dict = collections.OrderedDict()
+        device_index = start_device
+        self.trainers = []
+        
+        for num_layers in num_layers_list:
+            lstm = gluon.rnn.LSTM(num_hidden, num_layers, dropout=dropout, 
input_size=input_size)
+            input_size = num_hidden
+            self.lstm_dict[device_index] = lstm
+            device_index += 1
+        
+    def begin_state(self, *args, **kwargs):
+        """Return a list of hidden state for each LSTM in the stack"""
+        return [lstm.begin_state(ctx=mx.gpu(gpu_num), *args, **kwargs) 
+                for gpu_num, lstm in self.lstm_dict.items()]
+    
+    def forward(self, inputs, hidden):
+        """Pass the data through all LSTM in the stack
+        copying intermediate outputs to other contexts as necessary"""
+        hidden_indx = 0
+
+        output = inputs
+        for gpu_num, lstm in self.lstm_dict.items():
+            next_input = output.as_in_context(mx.gpu(gpu_num))
+            output, hidden[hidden_indx] = lstm(next_input, hidden[hidden_indx])
+            hidden_indx += 1
+        
+        return output, hidden
+    
+    def init_params(self, init=mx.init.Xavier(), force_reinit=False):
+        """For each LSTM in the stack,
+        initialize its parameters in the context specified by 
num_layers_list"""
+        for gpu_num, lstm in self.lstm_dict.items():
+            lstm.collect_params().initialize(init, ctx=mx.gpu(gpu_num), 
force_reinit=force_reinit)
+    
+    def init_trainer(self, optimizer, optimizer_params=None, kvstore='device'):
+        """Create seperate trainer for each LSTM
+        since one trainer cannot have parameters from multiple contexts"""
+        for gpu_num, lstm in self.lstm_dict.items():
+            self.trainers.append(gluon.Trainer(lstm.collect_params(), 
optimizer, optimizer_params, kvstore))
+
+    def step(self, batch_size, ignore_stale_grad=False):
+        """Call step on each LSTM's trainer"""
+        for trainer in self.trainers:
+            trainer.step(batch_size, ignore_stale_grad)
+
+    def clip_global_norm(self, max_norm):
+        """Clip gradients for each LSTM"""
+        for gpu_num, lstm in self.lstm_dict.items():
+            grads = [i.grad(mx.gpu(gpu_num)) for i in 
lstm.collect_params().values()]
+            gluon.utils.clip_global_norm(grads, max_norm)
+            
+    def reset_optimizer(self, optimizer, optimizer_params=None):
+        """Used to change learning rate. Not used tight now."""
+        for trainer in self.trainers:
+            trainer._init_optimizer(optimizer, optimizer_params)
+```
+
+`LSTMModel` adds an encoder in the beginning and decoder at the end to a 
`MultiGPULSTM`
+
+
+```python
+class LSTMModel():
+    def __init__(self, vocab_size, embedding_size, num_hidden,
+                 num_layers_list, dropout=0.5, **kwargs):
+        self.encoder = nn.Embedding(vocab_size, embedding_size,
+                                    weight_initializer = mx.init.Uniform(0.1))
+        self.lstm = MultiGPULSTM(0, num_layers_list, num_hidden, 
embedding_size, dropout)
+        self.decoder = nn.Dense(vocab_size, in_units = num_hidden)
+        self.dropout = nn.Dropout(dropout)
+        self.num_hidden = num_hidden
+        self.num_layers_list = num_layers_list
+        
+    def forward(self, inputs, hidden):
+        embedding = self.encoder(inputs)
+        embedding = self.dropout(embedding)
+        
+        output, hidden = self.lstm.forward(embedding, hidden)
+        output = self.dropout(output)
+
+        decoded = self.decoder(output.reshape((-1, self.num_hidden)))
+        return decoded, hidden
+    
+    def begin_state(self, *args, **kwargs):
+        return self.lstm.begin_state(*args, **kwargs)
+    
+    def init_params(self, init=mx.init.Xavier(), force_reinit=False):
+        self.encoder.collect_params().initialize(init, ctx=mx.gpu(0), 
force_reinit=force_reinit)
+        self.lstm.init_params(init, force_reinit)
+        last_gpu = len(self.num_layers_list) - 1
+        self.decoder.collect_params().initialize(init, ctx=mx.gpu(last_gpu), 
force_reinit=force_reinit)
+    
+    def init_trainer(self, optimizer, optimizer_params=None, kvstore='device'):
+        self.encoder_trainer = gluon.Trainer(self.encoder.collect_params(), 
optimizer, optimizer_params, kvstore)
+        self.decoder_trainer = gluon.Trainer(self.decoder.collect_params(), 
optimizer, optimizer_params, kvstore)
+        self.lstm.init_trainer(optimizer, optimizer_params, kvstore)
+
+    def step(self, batch_size, ignore_stale_grad=False):
+        self.encoder_trainer.step(batch_size, ignore_stale_grad)
+        self.decoder_trainer.step(batch_size, ignore_stale_grad)
+        self.lstm.step(batch_size, ignore_stale_grad)
+
+    def clip_global_norm(self, max_norm):
+        self.lstm.clip_global_norm(max_norm)
+    
+    def reset_optimizer(self, optimizer, optimizer_params=None):
+        self.encoder_trainer._init_optimizer(optimizer, optimizer_params)
+        self.decoder_trainer._init_optimizer(optimizer, optimizer_params)
+        self.lstm.reset_optimizer(optimizer, optimizer_params)
+```
+
+
+```python
+args_data = 'data/ptb.'
 
 Review comment:
   Sure

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
us...@infra.apache.org


With regards,
Apache Git Services

Reply via email to