Converting original T5 to be used in Transformers

I want to use original T5 checkpoint in Transformers library. I found multiple answers referring to convert_t5_original_tf_checkpoint_to_pytorch.py which does not seem to exist. Any other way? Or where can I find a (currently working) version of that file?

[marton-avrios]

…actually the link you sent for the example config file proved to be extremely useful! Starting from there I’ve found all related files. Here is everything (including the config file) for T5 Small: https://huggingface.co/t5-small. Also an example workflow for future reference:

mkdir t5
gsutil -m cp -r gs://t5-data/pretrained_models/small t5
python ~/transformers/src/transformers/models/t5/convert_t5_original_tf_checkpoint_to_pytorch.py --tf_checkpoint_path t5/small/model.ckpt-1000000 --pytorch_dump_path t5-small-pt --config_file t5/small_config.json

[NielsRogge]

I see that they store configurations in .gin files, like this one: https://console.cloud.google.com/storage/browser/_details/t5-data/pretrained_models/small/operative_config.gin

When opening this on my laptop in Notepad, this looks like this:

import t5.models.mesh_transformer
import t5.data.sentencepiece_vocabulary
import mesh_tensorflow.optimize
import mesh_tensorflow.transformer.dataset
import mesh_tensorflow.transformer.learning_rate_schedules
import mesh_tensorflow.transformer.t2t_vocabulary
import mesh_tensorflow.transformer.transformer_layers
import mesh_tensorflow.transformer.utils

# Macros:
# ==============================================================================
d_ff = 2048
d_kv = 64
d_model = 512
dropout_rate = 0.1
inputs_length = 512
mean_noise_span_length = 3.0
MIXTURE_NAME = 'all_mix'
noise_density = 0.15
num_heads = 8
num_layers = 6
targets_length = 512
init_checkpoint = "gs://t5-data/pretrained_models/small/model.ckpt-1000000"
tokens_per_batch = 1048576

# Parameters for AdafactorOptimizer:
# ==============================================================================
AdafactorOptimizer.beta1 = 0.0
AdafactorOptimizer.clipping_threshold = 1.0
AdafactorOptimizer.decay_rate = None
AdafactorOptimizer.epsilon1 = 1e-30
AdafactorOptimizer.epsilon2 = 0.001
AdafactorOptimizer.factored = True
AdafactorOptimizer.min_dim_size_to_factor = 128
AdafactorOptimizer.multiply_by_parameter_scale = True

# Parameters for Bitransformer:
# ==============================================================================
Bitransformer.shared_embedding = True

# Parameters for denoise:
# ==============================================================================
denoise.inputs_fn = @preprocessors.noise_span_to_unique_sentinel
denoise.noise_density = %noise_density
denoise.noise_mask_fn = @preprocessors.random_spans_noise_mask
denoise.targets_fn = @preprocessors.nonnoise_span_to_unique_sentinel

# Parameters for decoder/DenseReluDense:
# ==============================================================================
decoder/DenseReluDense.dropout_rate = %dropout_rate
decoder/DenseReluDense.hidden_size = %d_ff

# Parameters for encoder/DenseReluDense:
# ==============================================================================
encoder/DenseReluDense.dropout_rate = %dropout_rate
encoder/DenseReluDense.hidden_size = %d_ff

# Parameters for decoder/EncDecAttention:
# ==============================================================================
# None.

# Parameters for get_sentencepiece_model_path:
# ==============================================================================
get_sentencepiece_model_path.mixture_or_task_name = %MIXTURE_NAME

# Parameters for get_variable_dtype:
# ==============================================================================
get_variable_dtype.activation_dtype = 'bfloat16'

# Parameters for decoder/LayerStack:
# ==============================================================================
decoder/LayerStack.dropout_rate = %dropout_rate
decoder/LayerStack.norm_epsilon = 1e-06

# Parameters for encoder/LayerStack:
# ==============================================================================
encoder/LayerStack.dropout_rate = %dropout_rate
encoder/LayerStack.norm_epsilon = 1e-06

# Parameters for learning_rate_schedule_noam:
# ==============================================================================
learning_rate_schedule_noam.linear_decay_fraction = 0.1
learning_rate_schedule_noam.multiplier = 1.0
learning_rate_schedule_noam.offset = 0
learning_rate_schedule_noam.warmup_steps = 10000

# Parameters for make_bitransformer:
# ==============================================================================
make_bitransformer.decoder_name = 'decoder'
make_bitransformer.encoder_name = 'encoder'

# Parameters for decoder/make_layer_stack:
# ==============================================================================
decoder/make_layer_stack.block_scope = True
decoder/make_layer_stack.layers = \
    [@mesh_tensorflow.transformer.transformer_layers.SelfAttention,
     @mesh_tensorflow.transformer.transformer_layers.EncDecAttention,
     @mesh_tensorflow.transformer.transformer_layers.DenseReluDense]
decoder/make_layer_stack.num_layers = %num_layers

# Parameters for encoder/make_layer_stack:
# ==============================================================================
encoder/make_layer_stack.block_scope = True
encoder/make_layer_stack.layers = \
    [@mesh_tensorflow.transformer.transformer_layers.SelfAttention,
     @mesh_tensorflow.transformer.transformer_layers.DenseReluDense]
encoder/make_layer_stack.num_layers = %num_layers

# Parameters for mesh_train_dataset_fn:
# ==============================================================================
mesh_train_dataset_fn.mixture_or_task_name = %MIXTURE_NAME


# Parameters for noise_span_to_unique_sentinel:
# ==============================================================================
# None.

# Parameters for nonnoise_span_to_unique_sentinel:
# ==============================================================================
# None.

# Parameters for pack_dataset:
# ==============================================================================


# Parameters for pack_or_pad:
# ==============================================================================
# None.

# Parameters for random_spans_helper:
# ==============================================================================
random_spans_helper.extra_tokens_per_span_inputs = 1
random_spans_helper.extra_tokens_per_span_targets = 1
random_spans_helper.inputs_length = %inputs_length
random_spans_helper.mean_noise_span_length = %mean_noise_span_length
random_spans_helper.noise_density = %noise_density

# Parameters for targets_length/random_spans_helper:
# ==============================================================================
targets_length/random_spans_helper.extra_tokens_per_span_inputs = 1
targets_length/random_spans_helper.extra_tokens_per_span_targets = 1
targets_length/random_spans_helper.inputs_length = %inputs_length
targets_length/random_spans_helper.mean_noise_span_length = %mean_noise_span_length
targets_length/random_spans_helper.noise_density = %noise_density

# Parameters for random_spans_noise_mask:
# ==============================================================================
random_spans_noise_mask.mean_noise_span_length = %mean_noise_span_length

# Parameters for targets_length/random_spans_targets_length:
# ==============================================================================
# None.

# Parameters for random_spans_tokens_length:
# ==============================================================================
# None.

# Parameters for rate_num_examples:
# ==============================================================================
rate_num_examples.maximum = 1000000.0
rate_num_examples.scale = 1.0
rate_num_examples.temperature = 1.0

# Parameters for rate_unsupervised:
# ==============================================================================
rate_unsupervised.value = 710000.0

# Parameters for reduce_concat_tokens:
# ==============================================================================
reduce_concat_tokens.batch_size = 128
reduce_concat_tokens.feature_key = 'targets'

# Parameters for run:
# ==============================================================================
run.autostack = True
run.batch_size = ('tokens_per_batch', %tokens_per_batch)
run.dataset_split = 'train'
run.ensemble_inputs = None
run.eval_checkpoint_step = None
run.eval_dataset_fn = None
run.eval_summary_dir = None
run.export_path = ''
run.iterations_per_loop = 100
run.keep_checkpoint_max = None
run.layout_rules = \
    'ensemble:ensemble,batch:batch,d_ff:model,heads:model,vocab:model,experts:batch'
run.learning_rate_schedule = @learning_rate_schedules.learning_rate_schedule_noam
run.mesh_shape = @mesh_tensorflow.transformer.utils.tpu_mesh_shape()
run.mode = 'train' 
run.init_checkpoint = %init_checkpoint
run.model_type = 'bitransformer'
run.optimizer = @optimize.AdafactorOptimizer
run.perplexity_eval_steps = 10
run.predict_fn = None
run.save_checkpoints_steps = 2400
run.sequence_length = {'inputs': %inputs_length, 'targets': %targets_length}
run.train_dataset_fn = \
    @t5.models.mesh_transformer.mesh_train_dataset_fn
run.train_steps = 1000000000
run.variable_filter = None
run.vocabulary = \
    @t5.data.sentencepiece_vocabulary.SentencePieceVocabulary()

# Parameters for select_random_chunk:
# ==============================================================================
select_random_chunk.feature_key = 'targets'
select_random_chunk.max_length = 65536

# Parameters for decoder/SelfAttention:
# ==============================================================================
decoder/SelfAttention.attention_kwargs = None
decoder/SelfAttention.dropout_rate = %dropout_rate
decoder/SelfAttention.key_value_size = %d_kv
decoder/SelfAttention.num_heads = %num_heads
decoder/SelfAttention.num_memory_heads = 0
decoder/SelfAttention.relative_attention_num_buckets = 32
decoder/SelfAttention.relative_attention_type = 'bias_shared'
decoder/SelfAttention.shared_kv = False

# Parameters for encoder/SelfAttention:
# ==============================================================================
encoder/SelfAttention.attention_kwargs = None
encoder/SelfAttention.dropout_rate = %dropout_rate
encoder/SelfAttention.key_value_size = %d_kv
encoder/SelfAttention.num_heads = %num_heads
encoder/SelfAttention.num_memory_heads = 0
encoder/SelfAttention.relative_attention_num_buckets = 32
encoder/SelfAttention.relative_attention_type = 'bias_shared'
encoder/SelfAttention.shared_kv = False

# Parameters for SentencePieceVocabulary:
# ==============================================================================
SentencePieceVocabulary.extra_ids = 100
SentencePieceVocabulary.sentencepiece_model_file = \
    @t5.models.mesh_transformer.get_sentencepiece_model_path()

# Parameters for serialize_num_microbatches:
# ==============================================================================
serialize_num_microbatches.tokens_per_microbatch_per_replica = 8192

# Parameters for split_tokens:
# ==============================================================================
split_tokens.feature_key = 'targets'
split_tokens.max_tokens_per_segment = @preprocessors.random_spans_tokens_length()
split_tokens.min_tokens_per_segment = None

# Parameters for tpu_estimator_model_fn:
# ==============================================================================
tpu_estimator_model_fn.init_checkpoint = %init_checkpoint
tpu_estimator_model_fn.outer_batch_size = 1
tpu_estimator_model_fn.tpu_summaries = False

# Parameters for tpu_mesh_shape:
# ==============================================================================
tpu_mesh_shape.ensemble_parallelism = None
tpu_mesh_shape.model_parallelism = 1
tpu_mesh_shape.tpu_topology = '8x8'

# Parameters for decoder/Unitransformer:
# ==============================================================================
decoder/Unitransformer.d_model = %d_model
decoder/Unitransformer.ensemble = None
decoder/Unitransformer.input_full_attention = False
decoder/Unitransformer.label_smoothing = 0.0
decoder/Unitransformer.loss_denominator = None
decoder/Unitransformer.loss_fn = None
decoder/Unitransformer.loss_on_targets_only = False
decoder/Unitransformer.max_length = 512
decoder/Unitransformer.positional_embedding = False
decoder/Unitransformer.shared_embedding_and_softmax_weights = True
decoder/Unitransformer.vocab_divisor = 128
decoder/Unitransformer.z_loss = 0.0001
decoder/Unitransformer.loss_denominator = 233472

# Parameters for encoder/Unitransformer:
# ==============================================================================
encoder/Unitransformer.d_model = %d_model
encoder/Unitransformer.ensemble = None
encoder/Unitransformer.input_full_attention = False
encoder/Unitransformer.label_smoothing = 0.0
encoder/Unitransformer.loss_denominator = None
encoder/Unitransformer.loss_fn = None
encoder/Unitransformer.loss_on_targets_only = False
encoder/Unitransformer.max_length = 512
encoder/Unitransformer.positional_embedding = False
encoder/Unitransformer.shared_embedding_and_softmax_weights = True
encoder/Unitransformer.vocab_divisor = 128
encoder/Unitransformer.z_loss = 0.0001

# Parameters for unsupervised:
# ==============================================================================
unsupervised.preprocessors = \
    [@preprocessors.select_random_chunk,
     @preprocessors.reduce_concat_tokens,
     @preprocessors.split_tokens,
     @preprocessors.denoise]

=> the relevant part here seems to be only the model hyperparameters:

d_ff = 2048
d_kv = 64
d_model = 512
dropout_rate = 0.1
inputs_length = 512
mean_noise_span_length = 3.0
MIXTURE_NAME = 'all_mix'
noise_density = 0.15
num_heads = 8
num_layers = 6
targets_length = 512
init_checkpoint = "gs://t5-data/pretrained_models/small/model.ckpt-1000000"
tokens_per_batch = 1048576

So maybe you can create a config.json based on those?

Thanks! (you are a lifesaver by the way with these response times 😃).

And happy to hear this 😃 you’re welcome