File size: 8,056 Bytes
6ef31de
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import ml_collections

YOUCOOK_TRAIN_SIZE = 1333  # Number of videos


def get_config(runlocal=''):
    """Returns the base experiment configuration."""

    runlocal = bool(runlocal)

    config = ml_collections.ConfigDict()
    config.token_loss_coef = 1.
    config.runlocal = runlocal
    config.experiment_name = 'youcook'

    config.count_flops = False  # if runlocal else ml_collections.ConfigDict({'count_flops': True})

    # dataset
    config.dataset_name = 'dense_video_captioning'
    config.dataset_configs = ml_collections.ConfigDict()
    config.dataset_configs.corrupt = 0.
    config.dataset_configs.span_len = 3.
    config.dataset_configs.preserve = True
    config.dataset_configs.corrupt_coef = 0.
    config.dataset_configs.proba_corrupt = 0.
    notime = ml_collections.config_dict.FieldReference(False)
    config.dataset_configs.notime = notime
    config.dataset_configs.abs_time_token = False
    config.dataset_configs.random_temporal_crop_proba = 0.5
    config.dataset_configs.time_format = 'se'
    tmp_only = ml_collections.config_dict.FieldReference(False)
    config.dataset_configs.tmp_only = tmp_only
    config.dataset_configs.split = False
    order = ml_collections.config_dict.FieldReference('ld')
    config.dataset_configs.order = order
    config.dataset_configs.from_xm = None

    config.data_dtype_str = 'float32'

    config.dataset_configs.base_dir = '/mnt/petrelfs/wangyiqin/vid_cap/examples'
    config.dataset_configs.tables = {
        'train': 'train.tfrecord.sst@64',
        'validation': 'test@1',
    } 
    config.dataset_configs.examples_per_subset = {
        'train': 0,
        'validation': 1,
    }

    # List of modalities to load, supports `features` only for now.
    # Note that it only specifies which modalities to load, not which to use,
    # which is controlled by config.model.modality_fusion
    config.dataset_configs.modalities = ('features', 'text')
    config.dataset_configs.features_dim = 768
    config.dataset_configs.return_as_dict = True
    num_frames = ml_collections.config_dict.FieldReference(
        100)  # need to change back to 100 in the future -- Yiqin
    config.dataset_configs.num_frames = num_frames
    num_bins = ml_collections.config_dict.FieldReference(100)
    config.dataset_configs.num_bins = num_bins
    config.dataset_configs.one_hot_labels = True
    config.dataset_configs.zero_centering = True
    config.dataset_configs.val_on_test = False
    config.dataset_configs.num_eval_clips = 1
    config.dataset_configs.prefetch_to_device = 2

    # Text params
    config.dataset_configs.max_num_output_words = 256
    config.dataset_configs.max_num_input_words = 1000
    config.dataset_configs.tokenizer = ml_collections.ConfigDict()
    config.dataset_configs.tokenizer.tokenizer_type = 'sentence_piece'
    config.dataset_configs.caption_string = 'caption/string'
    config.dataset_configs.train_caption_string = 'caption/string'
    config.dataset_configs.input_timestamp_name = 'video/timestamps'
    config.dataset_configs.input_duration_name = 'video/duration'
    config.dataset_configs.output_raw_timestamp_name = 'timestamp'
    config.dataset_configs.output_raw_duration_name = 'duration'
    config.dataset_configs.input_feature_name = 'image/clip_embeddings'
    config.dataset_configs.output_raw_feature_name = 'features'
    config.dataset_configs.vocabulary_size = 32128
    config.dataset_configs.max_events = 20
    config.dataset_configs.asr_notime = False
    config.datasets = {'youcook': config.dataset_configs}

    # Decoding
    config.decoding = ml_collections.ConfigDict()
    config.decoding.decoding_method = 'beamsearch'
    # config.decoding.decoding_method = 'temperature_sample'
    config.decoding.num_decodes = 4
    config.decoding.alpha = 1
    config.decoding.temperature = 1.

    # Model
    config.model_name = 'vid2seq'
    config.model = ml_collections.ConfigDict()
    config.model.from_xm = None

    # Encoder configs
    config.model.encoder = ml_collections.ConfigDict()
    config.model.encoder.share_encoder = True
    config.model.encoder.encoder_type = 'cat_encoder'
    config.model.encoder.cat_encoder = ml_collections.ConfigDict()
    config.model.encoder.cat_encoder.dim = 2048
    config.model.encoder.cat_encoder.layers = 12
    config.model.encoder.cat_encoder.heads = 12
    config.model.encoder.cat_encoder.pos_embed = 'learned_1d'
    config.model.encoder.cat_encoder.dropout_rate = 0.
    config.model.encoder.cat_encoder.t5_dropout_rate = 0.1
    config.model.encoder.cat_encoder.stochastic_depth = 0.
    config.model.encoder.cat_encoder.pretrained_config = 't5_1_1_base'
    config.model.encoder.from_xm = None

    # Decoder configs
    config.model.decoder_type = 't5_decoder'
    config.model.decoder = ml_collections.ConfigDict()
    config.model.decoder.order = order
    config.model.decoder.t5_decoder = ml_collections.ConfigDict()
    config.model.decoder.t5_decoder.logits_via_embedding = False
    config.model.decoder.t5_decoder.dropout_rate = 0.1
    config.model.decoder.t5_decoder.num_frames = num_frames
    config.model.decoder.notime = notime
    config.model.decoder.num_bins = num_bins
    config.model.decoder.tmp_only = tmp_only
    config.model.decoder.t5_decoder.pretrained_config = 't5_1_1_base'

    # Initalisation configs
    config.init_from = ml_collections.ConfigDict()
    # Replace with your checkpoint pretrained on YT-temporal-1bn, assuming it has
    # been trained for 200K iterations
    config.init_from.checkpoint_path = '/mnt/petrelfs/wangyiqin/vid_cap/vid2seq_model'
    # config.init_from.model_config = '/mnt/petrelfs/wangyiqin/vid_cap/scenic/scenic/projects/vid2seq/configs/yttemporal.py'
    config.init_from.step = 200001 # ytt 200000, anet 200001

    config.init_from.encoder = ml_collections.ConfigDict()
    config.init_from.encoder.checkpoint_path = None
    config.init_from.encoder.init_from_vit = False
    config.init_from.encoder = ml_collections.ConfigDict()
    config.init_from.encoder.load_pretrained_weights = True

    config.init_from.decoder = ml_collections.ConfigDict()
    config.init_from.decoder.load_pretrained_weights = True

    config.init_from.t5 = ml_collections.ConfigDict()
    config.init_from.t5.load_pretrained_weights = True

    # Training
    config.trainer_name = 'densevidcap_trainer'
    config.optimizer = 'adam'
    config.optimizer_configs = ml_collections.ConfigDict()
    config.optimizer_configs.weight_decay = 0.
    config.l2_decay_factor = 0.
    config.max_grad_norm = 1.
    config.label_smoothing = 0.1
    epochs = ml_collections.config_dict.FieldReference(0)  ### add
    config.num_training_epochs = 0
    batch_size = ml_collections.config_dict.FieldReference(1)
    config.batch_size = 1  #if runlocal else batch_size  # 128  # Minimum is num_devices = 32
    config.eval_batch_size = 1  #if runlocal else 32  # Needs to be num_local_devices
    config.rng_seed = 0

    # Learning schedule.
    steps_per_epoch = 3 if runlocal else YOUCOOK_TRAIN_SIZE // batch_size
    total_steps = epochs * steps_per_epoch
    config.lr_configs = ml_collections.ConfigDict()
    config.lr_configs.learning_rate_schedule = 'compound'
    config.lr_configs.factors = 'constant * cosine_decay * linear_warmup'
    config.lr_configs.warmup_steps = total_steps // 10
    config.lr_configs.steps_per_cycle = total_steps
    config.lr_configs.total_steps = total_steps
    config.lr_configs.base_learning_rate = 3e-4

    config.eval_metrics = ['cider', 'meteor', 'soda']

    # Logging
    config.log_eval_steps = steps_per_epoch  # write TB and/or XM summary
    config.log_summary_steps = steps_per_epoch  # write TB and/or XM summary
    config.write_summary = True  # write TB and/or XM summary
    config.write_xm_measurements = True  # write XM measurements
    config.xprof = True  # Profile using xprof
    config.checkpoint = True  # do checkpointing
    config.debug_train = False  # debug mode during training
    config.debug_eval = True  # debug mode during eval
    return config