File size: 13,239 Bytes
6021dd1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
330
331
332
333
334
import mxnet as mx
from mxnet.rnn import BaseRNNCell
from nowcasting.ops import activation
from nowcasting.operators.common import group_add

class MyBaseRNNCell(BaseRNNCell):
    def __init__(self, prefix="MyBaseRNNCell", params=None):
        super(MyBaseRNNCell, self).__init__(prefix=prefix, params=params)

    def __call__(self, inputs, states, is_initial=False, ret_mid=False):
        raise NotImplementedError()

    def reset(self):
        super(MyBaseRNNCell, self).reset()
        self._curr_states = None

    def get_current_states(self):
        return self._curr_states

    def unroll(self, length, inputs=None, begin_state=None, ret_mid=False,
               input_prefix='', layout='TC', merge_outputs=False):
        """Unroll an RNN cell across time steps.

        Parameters
        ----------
        length : int
            number of steps to unroll
        inputs : Symbol, list of Symbol, or None
            if inputs is a single Symbol (usually the output
            of Embedding symbol), it should have shape
            (batch_size, length, ...) if layout == 'NTC',
            or (length, batch_size, ...) if layout == 'TNC'.

            If inputs is a list of symbols (usually output of
            previous unroll), they should all have shape
            (batch_size, ...).

            If inputs is None, Placeholder variables are
            automatically created.
        begin_state : nested list of Symbol
            input states. Created by begin_state()
            or output state of another cell. Created
            from begin_state() if None.
        input_prefix : str
            prefix for automatically created input
            placehodlers.
        layout : str
            layout of input symbol. Only used if inputs
            is a single Symbol.
        merge_outputs : bool
            if False, return outputs as a list of Symbols.
            If True, concatenate output across time steps
            and return a single symbol with shape
            (batch_size, length, ...) if layout == 'NTC',
            or (length, batch_size, ...) if layout == 'TNC'.

        Returns
        -------
        outputs : list of Symbol
            output symbols.
        states : Symbol or nested list of Symbol
            has the same structure as begin_state()
        mid_info : list of Symbol
        """
        self.reset()
        assert layout == 'TNC' or layout == 'TC'
        if inputs is not None:
            if isinstance(inputs, mx.sym.Symbol):
                assert len(inputs.list_outputs()) == 1, \
                    "unroll doesn't allow grouped symbol as input. Please " \
                    "convert to list first or let unroll handle slicing"
                if 'N' in layout:
                    inputs = mx.sym.SliceChannel(inputs, axis=0, num_outputs=length,
                                                 squeeze_axis=1)
                else:
                    inputs = mx.sym.SliceChannel(inputs, axis=0, num_outputs=length)
            else:
                assert len(inputs) == length
        else:
            inputs = [None] * length
        if begin_state is None:
            states = self.begin_state()
        else:
            states = begin_state
        outputs = []
        mid_infos = []
        for i in range(length):
            output, states, mid_info = self(inputs=inputs[i], states=states,
                                            is_initial=(i == 0 and (begin_state is None)),
                                            ret_mid=True)
            outputs.append(output)
            mid_infos.extend(mid_info)
        if merge_outputs:
            outputs = [mx.sym.expand_dims(i, axis=0) for i in outputs]
            outputs = mx.sym.Concat(*outputs, dim=0)
        if ret_mid:
            return outputs, states, mid_infos
        else:
            return outputs, states


class BaseStackRNN(object):
    def __init__(self, base_rnn_class, stack_num=1,
                 name="BaseStackRNN", residual_connection=True,
                 **kwargs):
        self._base_rnn_class = base_rnn_class
        self._residual_connection = residual_connection
        self._name = name
        self._stack_num = stack_num
        self._prefix = name + "_"
        self._rnns = [base_rnn_class(prefix=self._name + "_%d" %i, **kwargs) for i in range(stack_num)]
        self._init_counter = 0
        self._state_info = None

    def init_state_vars(self):
        """Initial state variable for this cell.

        Parameters
        ----------

        Returns
        -------
        state_vars : nested list of Symbol
            starting states for first RNN step
        """
        state_vars = []
        for i, info in enumerate(self.state_info):
            state = mx.sym.var(name='%s_begin_state_%s' % (self._name, self.state_postfix[i]), **info)
            state_vars.append(state)
        return state_vars

    def concat_to_split(self, concat_states):
        assert len(concat_states) == len(self.state_info)
        split_states = [[] for i in range(self._stack_num)]
        for i in range(len(self.state_info)):
            channel_axis = self.state_info[i]['__layout__'].lower().find('c')
            ele = mx.sym.split(concat_states[i], num_outputs=self._stack_num, axis=channel_axis)
            for j in range(self._stack_num):
                split_states[j].append(ele[j])
        return split_states

    def split_to_concat(self, split_states):
        # Concat the states together
        concat_states = []
        for i in range(len(self.state_info)):
            channel_axis = self.state_info[i]['__layout__'].lower().find('c')
            concat_states.append(mx.sym.concat(*[ele[i] for ele in split_states],
                                               dim=channel_axis))
        return concat_states

    def check_concat(self, states):
        ret = not isinstance(states[0], list)
        return ret

    def to_concat(self, states):
        if not self.check_concat(states):
            states = self.split_to_concat(states)
        return states

    def to_split(self, states):
        if self.check_concat(states):
            states = self.concat_to_split(states)
        return states

    @property
    def state_postfix(self):
        return self._rnns[0].state_postfix

    @property
    def state_info(self):
        if self._state_info is None:
            info = []
            for i in range(len(self._rnns[0].state_info)):
                ele = {}
                for rnn in self._rnns:
                    if 'shape' not in ele:
                        ele['shape'] = list(rnn.state_info[i]['shape'])
                    else:
                        channel_dim = rnn.state_info[i]['__layout__'].lower().find('c')
                        ele['shape'][channel_dim] += rnn.state_info[i]['shape'][channel_dim]
                    if '__layout__' not in ele:
                        ele['__layout__'] = rnn.state_info[i]['__layout__'].upper()
                    else:
                        assert rnn.state_info[i]['__layout__'] == ele['__layout__'].upper()
                ele['shape'] = tuple(ele['shape'])
                info.append(ele)
            self._state_info = info
            return info
        else:
            return self._state_info

    def flatten_add_layout(self, states, blocked=False):
        """
        
        Parameters
        ----------
        states : list of list or list

        Returns
        -------
        ret : list
        """
        states = self.to_concat(states)
        assert self.check_concat(states)
        ret = []
        for i, ele in enumerate(states):
            if blocked:
                ret.append(mx.sym.BlockGrad(ele, __layout__=self.state_info[i]['__layout__']))
            else:
                ele._set_attr(__layout__=self.state_info[i]['__layout__'])
                ret.append(ele)
        return ret

    def reset(self):
        for i in range(len(self._rnns)):
            self._rnns[i].reset()

    def unroll(self, length, inputs=None, begin_states=None, ret_mid=False):
        if begin_states is None:
            begin_states = self.init_state_vars()
        begin_states = self.to_split(begin_states)
        assert len(begin_states) == self._stack_num, len(begin_states)
        for ele in begin_states:
            assert len(ele) == len(self.state_info)
        outputs = []
        final_states = []
        mid_infos = []
        for i in range(len(self._rnns)):
            rnn_out_list, rnn_final_states, rnn_mid_infos =\
                self._rnns[i].unroll(length=length, inputs=inputs,
                                     begin_state=begin_states[i],
                                     layout="TC",
                                     ret_mid=True)
            if self._residual_connection and i > 0:
                # Use residual connections
                rnn_out_list = group_add(lhs=rnn_out_list, rhs=inputs)
            inputs = rnn_out_list
            outputs.append(rnn_out_list)
            final_states.append(rnn_final_states)
            mid_infos.append(rnn_mid_infos)
        if ret_mid:
            return outputs, final_states, mid_infos
        else:
            return outputs, final_states


class MyGRU(MyBaseRNNCell):
    """GRU cell.

    Parameters
    ----------
    num_hidden : int
        number of units in output symbol
    prefix : str, default 'rnn_'
        prefix for name of layers
        (and name of weight if params is None)
    params : RNNParams or None
        container for weight sharing between cells.
        created if None.
    """
    def __init__(self, num_hidden, zoneout=0.0, act_type="tanh", prefix='gru_', params=None):
        super(MyGRU, self).__init__(prefix=prefix, params=params)
        self._num_hidden = num_hidden
        self._act_type = act_type
        self._zoneout = zoneout
        self._i2h_weight = self.params.get('i2h_weight')
        self._i2h_bias = self.params.get('i2h_bias')
        self._h2h_weight = self.params.get('h2h_weight')
        self._h2h_bias = self.params.get('h2h_bias')

    @property
    def state_info(self):
        """shape(s) of states"""
        return [{'shape': (0, self._num_hidden), '__layout__': "NC"}]

    def __call__(self, inputs, states=None, is_initial=False, ret_mid=False):
        name = '%s_t%d' % (self._prefix, self._counter)
        self._counter += 1
        if is_initial:
            prev_h = self.begin_state()[0]
        else:
            prev_h = states[0]
        assert states is not None
        if inputs is not None:
            inputs = mx.sym.reshape(inputs, shape=(0, -1))
            i2h = mx.sym.FullyConnected(data=inputs,
                                        num_hidden=self._num_hidden * 3,
                                        weight=self._i2h_weight,
                                        bias=self._i2h_bias,
                                        name="%s_i2h" %name)
            i2h_slice = mx.sym.SliceChannel(i2h, num_outputs=3, axis=1)
        else:
            i2h_slice = None
        h2h = mx.sym.FullyConnected(data=prev_h,
                                    num_hidden=self._num_hidden * 3,
                                    weight=self._h2h_weight,
                                    bias=self._h2h_bias,
                                    name="%s_h2h" %name)
        h2h_slice = mx.sym.SliceChannel(h2h, num_outputs=3, axis=1)
        if i2h_slice is not None:
            reset_gate = activation(i2h_slice[0] + h2h_slice[0], act_type="sigmoid",
                                    name=name + "_r")
            update_gate = activation(i2h_slice[1] + h2h_slice[1], act_type="sigmoid",
                                     name=name + "_u")
            new_mem = activation(i2h_slice[2] + reset_gate * h2h_slice[2],
                                 act_type=self._act_type,
                                 name=name + "_h")
        else:
            reset_gate = activation(h2h_slice[0], act_type="sigmoid",
                                    name=name + "_r")
            update_gate = activation(h2h_slice[1], act_type="sigmoid",
                                     name=name + "_u")
            new_mem = activation(reset_gate * h2h_slice[2],
                                 act_type=self._act_type,
                                 name=name + "_h")
        next_h = update_gate * prev_h + (1 - update_gate) * new_mem
        if self._zoneout > 0.0:
            mask = mx.sym.Dropout(mx.sym.ones_like(prev_h), p=self._zoneout)
            next_h = mx.sym.where(mask, next_h, prev_h)
        self._curr_states = [next_h]
        if not ret_mid:
            return next_h, [next_h]
        else:
            return next_h, [next_h], []

if __name__ == '__main__':
    from nowcasting.operators.conv_rnn import ConvGRU
    brnn1 = BaseStackRNN(base_rnn_class=ConvGRU, stack_num=5,
                         b_h_w=(4, 32, 32), num_filter=32)
    print(brnn1.state_info)
    inputs = mx.sym.var(name="inputs", shape=(8, 4, 16, 32, 32))
    outputs, final_states, mid_infos = brnn1.unroll(length=8, inputs=inputs, ret_mid=True)
    print(len(outputs), len(outputs[0]))
    print(len(final_states), len(final_states[0]))