File size: 12,109 Bytes
9ce984a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
"""
Title: Text classification with Switch Transformer
Author: [Khalid Salama](https://www.linkedin.com/in/khalid-salama-24403144/)
Date created: 2020/05/10
Last modified: 2021/02/15
Description: Implement a Switch Transformer for text classification.
Accelerator: GPU
"""

"""
## Introduction

This example demonstrates the implementation of the
[Switch Transformer](https://arxiv.org/abs/2101.03961) model for text
classification.

The Switch Transformer replaces the feedforward network (FFN) layer in the standard
Transformer with a Mixture of Expert (MoE) routing layer, where each expert operates
independently on the tokens in the sequence. This allows increasing the model size without
increasing the computation needed to process each example.

Note that, for training the Switch Transformer efficiently, data and model parallelism
need to be applied, so that expert modules can run simultaneously, each on its own accelerator.
While the implementation described in the paper uses the
[TensorFlow Mesh](https://github.com/tensorflow/mesh) framework for distributed training,
this example presents a simple, non-distributed implementation of the Switch Transformer
model for demonstration purposes.
"""

"""
## Setup
"""

import keras
from keras import ops
from keras import layers

"""
## Download and prepare dataset
"""

vocab_size = 20000  # Only consider the top 20k words
num_tokens_per_example = 200  # Only consider the first 200 words of each movie review
(x_train, y_train), (x_val, y_val) = keras.datasets.imdb.load_data(num_words=vocab_size)
print(len(x_train), "Training sequences")
print(len(x_val), "Validation sequences")
x_train = keras.utils.pad_sequences(x_train, maxlen=num_tokens_per_example)
x_val = keras.utils.pad_sequences(x_val, maxlen=num_tokens_per_example)

"""
## Define hyperparameters
"""

embed_dim = 32  # Embedding size for each token.
num_heads = 2  # Number of attention heads
ff_dim = 32  # Hidden layer size in feedforward network.
num_experts = 10  # Number of experts used in the Switch Transformer.
batch_size = 50  # Batch size.
learning_rate = 0.001  # Learning rate.
dropout_rate = 0.25  # Dropout rate.
num_epochs = 3  # Number of epochs.
num_tokens_per_batch = (
    batch_size * num_tokens_per_example
)  # Total number of tokens per batch.
print(f"Number of tokens per batch: {num_tokens_per_batch}")

"""
## Implement token & position embedding layer

It consists of two separate embedding layers, one for tokens, one for token index (positions).
"""


class TokenAndPositionEmbedding(layers.Layer):
    def __init__(self, maxlen, vocab_size, embed_dim):
        super().__init__()
        self.token_emb = layers.Embedding(input_dim=vocab_size, output_dim=embed_dim)
        self.pos_emb = layers.Embedding(input_dim=maxlen, output_dim=embed_dim)

    def call(self, x):
        maxlen = ops.shape(x)[-1]
        positions = ops.arange(start=0, stop=maxlen, step=1)
        positions = self.pos_emb(positions)
        x = self.token_emb(x)
        return x + positions


"""
## Implement the feedforward network

This is used as the Mixture of Experts in the Switch Transformer.
"""


def create_feedforward_network(ff_dim, embed_dim, name=None):
    return keras.Sequential(
        [layers.Dense(ff_dim, activation="relu"), layers.Dense(embed_dim)], name=name
    )


"""
## Implement the load-balanced loss

This is an auxiliary loss to encourage a balanced load across experts.
"""


def load_balanced_loss(router_probs, expert_mask):
    # router_probs [tokens_per_batch, num_experts] is the probability assigned for
    # each expert per token. expert_mask [tokens_per_batch, num_experts] contains
    # the expert with the highest router probability in one−hot format.

    num_experts = ops.shape(expert_mask)[-1]
    # Get the fraction of tokens routed to each expert.
    # density is a vector of length num experts that sums to 1.
    density = ops.mean(expert_mask, axis=0)
    # Get fraction of probability mass assigned to each expert from the router
    # across all tokens. density_proxy is a vector of length num experts that sums to 1.
    density_proxy = ops.mean(router_probs, axis=0)
    # Want both vectors to have uniform allocation (1/num experts) across all
    # num_expert elements. The two vectors will be pushed towards uniform allocation
    # when the dot product is minimized.
    loss = ops.mean(density_proxy * density) * ops.cast((num_experts**2), "float32")
    return loss


"""
### Implement the router as a layer
"""


class Router(layers.Layer):
    def __init__(self, num_experts, expert_capacity):
        self.num_experts = num_experts
        self.route = layers.Dense(units=num_experts)
        self.expert_capacity = expert_capacity
        super().__init__()

    def call(self, inputs, training=False):
        # inputs shape: [tokens_per_batch, embed_dim]
        # router_logits shape: [tokens_per_batch, num_experts]
        router_logits = self.route(inputs)

        if training:
            # Add noise for exploration across experts.
            router_logits += keras.random.uniform(
                shape=router_logits.shape, minval=0.9, maxval=1.1
            )
        # Probabilities for each token of what expert it should be sent to.
        router_probs = keras.activations.softmax(router_logits, axis=-1)
        # Get the top−1 expert for each token. expert_gate is the top−1 probability
        # from the router for each token. expert_index is what expert each token
        # is going to be routed to.
        expert_gate, expert_index = ops.top_k(router_probs, k=1)
        # expert_mask shape: [tokens_per_batch, num_experts]
        expert_mask = ops.one_hot(expert_index, self.num_experts)
        # Compute load balancing loss.
        aux_loss = load_balanced_loss(router_probs, expert_mask)
        self.add_loss(aux_loss)
        # Experts have a fixed capacity, ensure we do not exceed it. Construct
        # the batch indices, to each expert, with position in expert make sure that
        # not more that expert capacity examples can be routed to each expert.
        position_in_expert = ops.cast(
            ops.cumsum(expert_mask, axis=0) * expert_mask, "int32"
        )
        # Keep only tokens that fit within expert capacity.
        expert_mask *= ops.cast(
            ops.less(ops.cast(position_in_expert, "int32"), self.expert_capacity),
            "float32",
        )
        expert_mask_flat = ops.sum(expert_mask, axis=-1)
        # Mask out the experts that have overflowed the expert capacity.
        expert_gate *= expert_mask_flat
        # Combine expert outputs and scaling with router probability.
        # combine_tensor shape: [tokens_per_batch, num_experts, expert_capacity]
        combined_tensor = ops.expand_dims(
            expert_gate
            * expert_mask_flat
            * ops.squeeze(ops.one_hot(expert_index, self.num_experts), 1),
            -1,
        ) * ops.squeeze(ops.one_hot(position_in_expert, self.expert_capacity), 1)
        # Create binary dispatch_tensor [tokens_per_batch, num_experts, expert_capacity]
        # that is 1 if the token gets routed to the corresponding expert.
        dispatch_tensor = ops.cast(combined_tensor, "float32")

        return dispatch_tensor, combined_tensor


"""
### Implement a Switch layer
"""


class Switch(layers.Layer):
    def __init__(
        self, num_experts, embed_dim, ff_dim, num_tokens_per_batch, capacity_factor=1
    ):
        self.num_experts = num_experts
        self.embed_dim = embed_dim
        self.experts = [
            create_feedforward_network(ff_dim, embed_dim) for _ in range(num_experts)
        ]

        self.expert_capacity = num_tokens_per_batch // self.num_experts
        self.router = Router(self.num_experts, self.expert_capacity)
        super().__init__()

    def call(self, inputs):
        batch_size = ops.shape(inputs)[0]
        num_tokens_per_example = ops.shape(inputs)[1]

        # inputs shape: [num_tokens_per_batch, embed_dim]
        inputs = ops.reshape(inputs, [num_tokens_per_batch, self.embed_dim])
        # dispatch_tensor shape: [expert_capacity, num_experts, tokens_per_batch]
        # combine_tensor shape: [tokens_per_batch, num_experts, expert_capacity]
        dispatch_tensor, combine_tensor = self.router(inputs)
        # expert_inputs shape: [num_experts, expert_capacity, embed_dim]
        expert_inputs = ops.einsum("ab,acd->cdb", inputs, dispatch_tensor)
        expert_inputs = ops.reshape(
            expert_inputs, [self.num_experts, self.expert_capacity, self.embed_dim]
        )
        # Dispatch to experts
        expert_input_list = ops.unstack(expert_inputs, axis=0)
        expert_output_list = [
            self.experts[idx](expert_input)
            for idx, expert_input in enumerate(expert_input_list)
        ]
        # expert_outputs shape: [expert_capacity, num_experts, embed_dim]
        expert_outputs = ops.stack(expert_output_list, axis=1)
        # expert_outputs_combined shape: [tokens_per_batch, embed_dim]
        expert_outputs_combined = ops.einsum(
            "abc,xba->xc", expert_outputs, combine_tensor
        )
        # output shape: [batch_size, num_tokens_per_example, embed_dim]
        outputs = ops.reshape(
            expert_outputs_combined,
            [batch_size, num_tokens_per_example, self.embed_dim],
        )
        return outputs


"""
## Implement a Transformer block layer
"""


class TransformerBlock(layers.Layer):
    def __init__(self, embed_dim, num_heads, ffn, dropout_rate=0.1):
        super().__init__()
        self.att = layers.MultiHeadAttention(num_heads=num_heads, key_dim=embed_dim)
        # The ffn can be either a standard feedforward network or a switch
        # layer with a Mixture of Experts.
        self.ffn = ffn
        self.layernorm1 = layers.LayerNormalization(epsilon=1e-6)
        self.layernorm2 = layers.LayerNormalization(epsilon=1e-6)
        self.dropout1 = layers.Dropout(dropout_rate)
        self.dropout2 = layers.Dropout(dropout_rate)

    def call(self, inputs, training=False):
        attn_output = self.att(inputs, inputs)
        attn_output = self.dropout1(attn_output, training=training)
        out1 = self.layernorm1(inputs + attn_output)
        ffn_output = self.ffn(out1)
        ffn_output = self.dropout2(ffn_output, training=training)
        return self.layernorm2(out1 + ffn_output)


"""
## Implement the classifier

The `TransformerBlock` layer outputs one vector for each time step of our input sequence.
Here, we take the mean across all time steps and use a feedforward network on top
of it to classify text.
"""


def create_classifier():
    switch = Switch(num_experts, embed_dim, ff_dim, num_tokens_per_batch)
    transformer_block = TransformerBlock(embed_dim // num_heads, num_heads, switch)

    inputs = layers.Input(shape=(num_tokens_per_example,))
    embedding_layer = TokenAndPositionEmbedding(
        num_tokens_per_example, vocab_size, embed_dim
    )
    x = embedding_layer(inputs)
    x = transformer_block(x)
    x = layers.GlobalAveragePooling1D()(x)
    x = layers.Dropout(dropout_rate)(x)
    x = layers.Dense(ff_dim, activation="relu")(x)
    x = layers.Dropout(dropout_rate)(x)
    outputs = layers.Dense(2, activation="softmax")(x)

    classifier = keras.Model(inputs=inputs, outputs=outputs)
    return classifier


"""
## Train and evaluate the model
"""


def run_experiment(classifier):
    classifier.compile(
        optimizer=keras.optimizers.Adam(learning_rate),
        loss="sparse_categorical_crossentropy",
        metrics=["accuracy"],
    )
    history = classifier.fit(
        x_train,
        y_train,
        batch_size=batch_size,
        epochs=num_epochs,
        validation_data=(x_val, y_val),
    )
    return history


classifier = create_classifier()
run_experiment(classifier)


"""
## Conclusion

Compared to the standard Transformer architecture, the Switch Transformer can have a much
larger number of parameters, leading to increased model
capacity, while maintaining a reasonable computational cost.
"""