I'm currently implement the sequantial deep matching model (https://arxiv.org/abs/1909.00385) using tensorflow 2.3. And I included the preprocessing layer as part of the model via subclassing keras.layers.Layer.
The preprocessing part of code is listed below
class Preprocessing(keras.layers.Layer):
def __init__(self, str_columns, hash_bins, float_columns, float_buckets, embedding_dim, user_columns, short_seq_columns, prefer_seq_columns, item_key_feats,
item_key_hash_bucket_size, series_feats, series_feats_hash_bucket_size, deviceid_num, device_list, **kwargs):
super(Preprocessing, self).__init__(**kwargs)
self.str_columns = str_columns
self.hash_bins = hash_bins
self.float_columns = float_columns
self.float_buckets = float_buckets
self.embedding_dim = embedding_dim
self.user_columns = user_columns
self.short_seq_columns = short_seq_columns
self.prefer_seq_columns = prefer_seq_columns
self.item_key_feats = item_key_feats
self.item_key_hash_bucket_size = item_key_hash_bucket_size
self.series_feats = series_feats
self.series_feats_hash_bucket_size = series_feats_hash_bucket_size
self.deviceid_num = deviceid_num
self.device_list = device_list
self.user_outputs = {}
self.short_outputs = {}
self.prefer_outputs = {}
deviceid_lookup = keras.layers.experimental.preprocessing.StringLookup(vocabulary=device_list, mask_token=None, oov_token="-1")
deviceid_embedding = keras.layers.Embedding(input_dim=deviceid_num, output_dim=embedding_dim)
item_key_hashing = keras.layers.experimental.preprocessing.Hashing(num_bins=item_key_hash_bucket_size)
item_key_embedding = keras.layers.Embedding(input_dim=item_key_hash_bucket_size, output_dim=embedding_dim)
series_hashing = keras.layers.experimental.preprocessing.Hashing(num_bins=series_feats_hash_bucket_size)
series_embedding = keras.layers.Embedding(input_dim=series_feats_hash_bucket_size, output_dim=embedding_dim)
for i in str_columns:
if i == "device_id":
process = [deviceid_lookup, deviceid_embedding]
elif i in item_key_feats:
process = [item_key_hashing, item_key_embedding]
elif i in series_feats:
process = [series_hashing, series_embedding]
else:
hashing = keras.layers.experimental.preprocessing.Hashing(num_bins=hash_bins[i])
embedding = keras.layers.Embedding(input_dim=hash_bins[i], output_dim=embedding_dim)
process = [hashing, embedding]
if i in user_columns:
self.user_outputs[i] = process
if i in short_seq_columns:
self.short_outputs[i] = process
if i in prefer_seq_columns:
self.prefer_outputs[i] = process
for l in float_columns:
discrete = keras.layers.experimental.preprocessing.Discretization(bins=float_buckets[l])
embedding = keras.layers.Embedding(input_dim=len(float_buckets[l]) + 1, output_dim=embedding_dim)
if l in user_columns:
self.user_outputs[l] = [discrete, embedding]
if l in short_seq_columns:
self.short_outputs[l] = [discrete, embedding]
if l in prefer_seq_columns:
self.prefer_outputs[l] = [discrete, embedding]
@staticmethod
def get_embedding(input_tmp, name, embed_dict):
func = embed_dict[name]
if len(func) < 2:
print(func)
raise Exception('Not enough function to retrieve embedding')
output = func[0](input_tmp)
output = func[1](output)
return output
def call(self, inputs):
user_embedding = tf.concat([tf.reduce_mean(self.get_embedding(inputs[i], i, self.user_outputs), axis=[1, 2]) for i in self.user_columns], axis=-1)
short_embedding = tf.concat([tf.squeeze(self.get_embedding(inputs[l], l, self.short_outputs), axis=1).to_tensor() for l in self.short_seq_columns], axis=-1)
prefer_embedding = {k: tf.squeeze(self.get_embedding(inputs[k], k, self.prefer_outputs).to_tensor(), axis=1) for k in self.prefer_seq_columns}
return user_embedding, short_embedding, prefer_embedding
And also my input code:
def read_row(csv_row):
record_defaults = [[0.]] * numeric_feature_size + [['']] * category_feature_size + [['0-0']] + [['0']]
row = tf.io.decode_csv(csv_row, record_defaults=record_defaults, field_delim='', use_quote_delim=False)
features = []
for i, feature in enumerate(row):
if i < numeric_feature_size:
features.append(feature)
elif i < numeric_feature_size + category_feature_size:
tmp_tf = tf.strings.split([feature], ";")
features.append(tmp_tf)
res = OrderedDict(zip(numeric_columns + category_columns, features))
res['target'] = [tf.cast(row[-2], tf.string)]
return res
The other part of code is not giving here, cause I believe it's right, and might be too much to list here.
The model is working correctly during training using model.compile then model.fit, however, after I saved it with model.save(path), the resulting Graph gets many unknown inputs and none of the inputs name is saved.
saved_model_cli show --dir ./ --tag_set serve --signature_def serving_default
The given SavedModel SignatureDef contains the following input(s):
inputs['args_0'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: serving_default_args_0:0
inputs['args_0_1'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_1:0
inputs['args_0_10'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_10:0
inputs['args_0_11'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_11:0
inputs['args_0_12'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: serving_default_args_0_12:0
inputs['args_0_13'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_13:0
inputs['args_0_14'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_14:0
inputs['args_0_15'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: serving_default_args_0_15:0
inputs['args_0_16'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_16:0
inputs['args_0_17'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_17:0
inputs['args_0_18'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: serving_default_args_0_18:0
inputs['args_0_19'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_19:0
inputs['args_0_2'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_2:0
inputs['args_0_20'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_20:0
inputs['args_0_21'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: serving_default_args_0_21:0
inputs['args_0_22'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_22:0
inputs['args_0_23'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_23:0
inputs['args_0_24'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: serving_default_args_0_24:0
inputs['args_0_25'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_25:0
inputs['args_0_26'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_26:0
inputs['args_0_27'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: serving_default_args_0_27:0
inputs['args_0_28'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_28:0
inputs['args_0_29'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_29:0
inputs['args_0_3'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: serving_default_args_0_3:0
inputs['args_0_30'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: serving_default_args_0_30:0
inputs['args_0_31'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_31:0
inputs['args_0_32'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_32:0
inputs['args_0_33'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: serving_default_args_0_33:0
inputs['args_0_34'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_34:0
inputs['args_0_35'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_35:0
inputs['args_0_36'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: serving_default_args_0_36:0
inputs['args_0_37'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_37:0
inputs['args_0_38'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_38:0
inputs['args_0_39'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: serving_default_args_0_39:0
inputs['args_0_4'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_4:0
inputs['args_0_40'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_40:0
inputs['args_0_41'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_41:0
inputs['args_0_42'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: serving_default_args_0_42:0
inputs['args_0_43'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_43:0
inputs['args_0_44'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_44:0
inputs['args_0_45'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: serving_default_args_0_45:0
inputs['args_0_46'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_46:0
inputs['args_0_47'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_47:0
inputs['args_0_48'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: serving_default_args_0_48:0
inputs['args_0_49'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_49:0
inputs['args_0_5'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_5:0
inputs['args_0_50'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_50:0
inputs['args_0_6'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: serving_default_args_0_6:0
inputs['args_0_7'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_7:0
inputs['args_0_8'] tensor_info:
dtype: DT_INT64
shape: (-1)
name: serving_default_args_0_8:0
inputs['args_0_9'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: serving_default_args_0_9:0
The given SavedModel SignatureDef contains the following output(s):
outputs['output_1'] tensor_info:
dtype: DT_FLOAT
shape: (-1, 64)
name: StatefulPartitionedCall:0
In this model, I only used the categorical features with dtype as tf.string, so all the inputs with dtype of DT_INT64 is not part of my model inputs.
Can anyone help me with this?
I finally got this work.
The error came from my tf.io.decode_csv method. tf.strings.split()will return a RaggedTensor by default, which will also contain two int variables. That's why the input signatures contains 17 string type and 34 int type.
The RaggedTensor will also harm the keras serialization, which is why all of the input names were missing.
I transformed all the RaggedTensor to EagerTensor, and all things worked.
However, that's not the only error I've encountered when I tried to load the model.
I also encountered
The same saveable will be restored with two names error, which cost me tons of time to resolved it. And it turns out to be as bug of keras.layers.experimental.preprocessing module, the same function in it cannot be used twice, cause variables will be recorded as the same name, and result in a non-loadable savedmodel.