File size: 10,592 Bytes
3be620b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import numpy as np
import tensorflow as tf
import tensorflow_addons as tfa
from ganime.configs.model_configs import GPTConfig, ModelConfig
from ganime.model.vqgan_clean.vqgan import VQGAN
from ganime.trainer.warmup.cosine import WarmUpCosine
from tensorflow import keras
from tensorflow.keras import Model, layers
from transformers import TFGPT2Model, GPT2Config
from tensorflow.keras import mixed_precision


class Net2Net(Model):
    def __init__(
        self,
        transformer_config: GPTConfig,
        first_stage_config: ModelConfig,
        trainer_config,
        **kwargs,
    ):
        super().__init__(**kwargs)
        self.first_stage_model = VQGAN(**first_stage_config)

        # configuration = GPT2Config(**transformer_config)
        # self.transformer = TFGPT2Model(configuration)#.from_pretrained("gpt2", **self.transformer_config)
        # configuration = GPT2Config(**transformer_config)
        self.transformer = TFGPT2Model.from_pretrained(
            "gpt2-medium"
        )  # , **transformer_config)
        if "checkpoint_path" in transformer_config:
            print(f"Restoring weights from {transformer_config['checkpoint_path']}")
            self.load_weights(transformer_config["checkpoint_path"])

        self.loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(
            from_logits=True, reduction=tf.keras.losses.Reduction.NONE
        )

        self.loss_tracker = keras.metrics.Mean(name="loss")

        self.scheduled_lrs = self.create_warmup_scheduler(trainer_config)

        optimizer = tfa.optimizers.AdamW(
            learning_rate=self.scheduled_lrs, weight_decay=1e-4
        )
        self.compile(
            optimizer=optimizer,
            loss=self.loss_fn,
            # run_eagerly=True,
        )

        # Gradient accumulation
        self.gradient_accumulation = [
            tf.Variable(tf.zeros_like(v, dtype=tf.float32), trainable=False)
            for v in self.transformer.trainable_variables
        ]

    def create_warmup_scheduler(self, trainer_config):
        len_x_train = trainer_config["len_x_train"]
        batch_size = trainer_config["batch_size"]
        n_epochs = trainer_config["n_epochs"]

        total_steps = int(len_x_train / batch_size * n_epochs)
        warmup_epoch_percentage = trainer_config["warmup_epoch_percentage"]
        warmup_steps = int(total_steps * warmup_epoch_percentage)

        scheduled_lrs = WarmUpCosine(
            lr_start=trainer_config["lr_start"],
            lr_max=trainer_config["lr_max"],
            warmup_steps=warmup_steps,
            total_steps=total_steps,
        )

        return scheduled_lrs

    def apply_accu_gradients(self):
        # apply accumulated gradients
        self.optimizer.apply_gradients(
            zip(self.gradient_accumulation, self.transformer.trainable_variables)
        )

        # reset
        for i in range(len(self.gradient_accumulation)):
            self.gradient_accumulation[i].assign(
                tf.zeros_like(self.transformer.trainable_variables[i], dtype=tf.float32)
            )

    @property
    def metrics(self):
        # We list our `Metric` objects here so that `reset_states()` can be
        # called automatically at the start of each epoch
        # or at the start of `evaluate()`.
        # If you don't implement this property, you have to call
        # `reset_states()` yourself at the time of your choosing.
        return [
            self.loss_tracker,
        ]

    @tf.function(
        # input_signature=[
        #     tf.TensorSpec(shape=[None, None, None, 3], dtype=tf.float32),
        # ]
    )
    def encode_to_z(self, x):
        quant_z, indices, quantized_loss = self.first_stage_model.encode(x)

        batch_size = tf.shape(quant_z)[0]

        indices = tf.reshape(indices, shape=(batch_size, -1))
        return quant_z, indices

    def call(self, inputs, training=None, mask=None):

        first_frame = inputs["first_frame"]
        last_frame = inputs["last_frame"]
        n_frames = inputs["n_frames"]

        return self.generate_video(first_frame, last_frame, n_frames)



    @tf.function()
    def predict_next_indices(self, inputs, example_indices):
        logits = self.transformer(inputs)
        logits = logits.last_hidden_state
        logits = tf.cast(logits, dtype=tf.float32)
        # Remove the conditioned part
        logits = logits[
            :, tf.shape(example_indices)[1] - 1 :
        ]  # -1 here 'cause -1 above
        # logits = tf.reshape(logits, shape=(-1, tf.shape(logits)[-1]))
        return logits

    @tf.function()
    def body(self, total_loss, frames, index, last_frame_indices):

        previous_frame_indices = self.encode_to_z(frames[:, index - 1, ...])[1]
        cz_indices = tf.concat((last_frame_indices, previous_frame_indices), axis=1)
        target_indices = self.encode_to_z(frames[:, index, ...])[1]
        # target_indices = tf.reshape(target_indices, shape=(-1,))

        with tf.GradientTape() as tape:
            logits = self.predict_next_indices(
                cz_indices[:, :-1], last_frame_indices
            )  # don't know why -1

            frame_loss = tf.cast(
                tf.reduce_mean(self.loss_fn(target_indices, logits)),
                dtype=tf.float32,
            )

        # Calculate batch gradients
        gradients = tape.gradient(frame_loss, self.transformer.trainable_variables)

        # Accumulate batch gradients
        for i in range(len(self.gradient_accumulation)):
            self.gradient_accumulation[i].assign_add(tf.cast(gradients[i], tf.float32))

        index = tf.add(index, 1)
        total_loss = tf.add(total_loss, frame_loss)
        return total_loss, frames, index, last_frame_indices

    def cond(self, total_loss, frames, index, last_frame_indices):
        return tf.less(index, tf.shape(frames)[1])

    def train_step(self, data):
        first_frame = data["first_frame"]
        last_frame = data["last_frame"]
        frames = data["y"]
        n_frames = data["n_frames"]

        last_frame_indices = self.encode_to_z(last_frame)[1]
        total_loss = 0.0

        total_loss, _, _, _ = tf.while_loop(
            cond=self.cond,
            body=self.body,
            loop_vars=(tf.constant(0.0), frames, tf.constant(1), last_frame_indices),
        )

        self.apply_accu_gradients()
        self.loss_tracker.update_state(total_loss)
        return {m.name: m.result() for m in self.metrics}

    def cond_test_step(self, total_loss, frames, index, last_frame_indices):
        return tf.less(index, tf.shape(frames)[1])

    @tf.function()
    def body_test_step(self, total_loss, frames, index, predicted_logits):
        target_indices = self.encode_to_z(frames[:, index, ...])[1]
        # target_indices = tf.reshape(target_indices, shape=(-1,))
        logits = predicted_logits[index]

        frame_loss = tf.cast(
            tf.reduce_mean(self.loss_fn(target_indices, logits)),
            dtype=tf.float32,
        )

        index = tf.add(index, 1)
        total_loss = tf.add(total_loss, frame_loss)
        return total_loss, frames, index, predicted_logits

    def test_step(self, data):
        first_frame = data["first_frame"]
        last_frame = data["last_frame"]
        frames = data["y"]
        n_frames = data["n_frames"]

        predicted_logits, _, _ = self.predict_logits(first_frame, last_frame, n_frames)

        total_loss, _, _, _ = tf.while_loop(
            cond=self.cond_test_step,
            body=self.body_test_step,
            loop_vars=(tf.constant(0.0), frames, tf.constant(1), predicted_logits),
        )
       

        self.loss_tracker.update_state(total_loss)
        return {m.name: m.result() for m in self.metrics}

    @tf.function()
    def convert_logits_to_indices(self, logits, shape):
        probs = tf.keras.activations.softmax(logits)
        _, generated_indices = tf.math.top_k(probs)
        generated_indices = tf.reshape(
            generated_indices,
            shape,  # , self.first_stage_model.quantize.num_embeddings)
        )
        return generated_indices
        # quant = self.first_stage_model.quantize.get_codebook_entry(
        #     generated_indices, shape=shape
        # )

        # return self.first_stage_model.decode(quant)

    @tf.function()
    def predict_logits(self, first_frame, last_frame, n_frames):
        quant_first, indices_first = self.encode_to_z(first_frame)
        quant_last, indices_last = self.encode_to_z(last_frame)

        indices_previous = indices_first

        predicted_logits = tf.TensorArray(
            tf.float32, size=0, dynamic_size=True, clear_after_read=False
        )

        index = tf.constant(1)
        while tf.less(index, tf.reduce_max(n_frames)):
            tf.autograph.experimental.set_loop_options(
                shape_invariants=[(indices_previous, tf.TensorShape([None, None]))]
            )
            cz_indices = tf.concat((indices_last, indices_previous), axis=1)
            logits = self.predict_next_indices(cz_indices[:, :-1], indices_last)

            # generated_indices = self.convert_logits_to_indices(
            #     logits, tf.shape(indices_last)
            # )
            predicted_logits = predicted_logits.write(index, logits)
            indices_previous = self.convert_logits_to_indices(
                logits, tf.shape(indices_first)
            )
            index = tf.add(index, 1)

        return predicted_logits.stack(), tf.shape(quant_first), tf.shape(indices_first)

    @tf.function()
    def generate_video(self, first_frame, last_frame, n_frames):
        predicted_logits, quant_shape, indices_shape = self.predict_logits(
            first_frame, last_frame, n_frames
        )

        generated_images = tf.TensorArray(
            tf.float32, size=0, dynamic_size=True, clear_after_read=False
        )
        generated_images = generated_images.write(0, first_frame)

        index = tf.constant(1)
        while tf.less(index, tf.reduce_max(n_frames)):
            indices = self.convert_logits_to_indices(predicted_logits[index], indices_shape)
            quant = self.first_stage_model.quantize.get_codebook_entry(
                indices,
                shape=quant_shape,
            )
            decoded = self.first_stage_model.decode(quant)
            generated_images = generated_images.write(index, decoded)
            index = tf.add(index, 1)

        stacked_images = generated_images.stack()
        videos = tf.transpose(stacked_images, (1, 0, 2, 3, 4))
        return videos