github/AMD-SHARK-Studio
1
1
Fork 0
mirror of https://github.com/nod-ai/AMD-SHARK-Studio.git synced 2026-04-03 03:00:17 -04:00
Code Issues Packages Projects Releases Wiki Activity

Add EfficientNet B0 and B7 Torch and TF models

Browse Source
This commit is contained in:
mariecwhite
2023-02-20 04:45:23 -08:00
parent c6f44ef1b3
commit f99903e023
5 changed files with 110 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
tank/all_models.csv
View File

@@ -39,3 +39,4 @@ efficientnet_b0,linalg,torch,1e-2,1e-3,default,nhcw-nhwc,False,False,False,"",""
efficientnet_b7,linalg,torch,1e-2,1e-3,default,nhcw-nhwc,False,False,False,"",""
efficientnet_b0,mhlo,tf,1e-2,1e-3,default,None,nhcw-nhwc,False,False,"",""
efficientnet_b7,mhlo,tf,1e-2,1e-3,default,None,nhcw-nhwc,False,False,"",""
gpt2,mhlo,tf,1e-2,1e-3,default,None,False,False,False,"",""
1 resnet50 mhlo tf 1e-2 1e-3 default nhcw-nhwc False False False macos
39 efficientnet_b7 linalg torch 1e-2 1e-3 default nhcw-nhwc False False False
40 efficientnet_b0 mhlo tf 1e-2 1e-3 default None nhcw-nhwc False False
41 efficientnet_b7 mhlo tf 1e-2 1e-3 default None nhcw-nhwc False False
42 gpt2 mhlo tf 1e-2 1e-3 default None False False False

5
tank/generate_sharktank.py
View File

@@ -118,6 +118,7 @@ def save_torch_model(torch_model_list, local_tank_cache):
def save_tf_model(tf_model_list, local_tank_cache):
from tank.model_utils_tf import (
get_causal_image_model,
get_masked_lm_model,
get_causal_lm_model,
get_keras_model,
get_TFhf_model,
@@ -145,13 +146,15 @@ def save_tf_model(tf_model_list, local_tank_cache):
input = None
print(f"Generating artifacts for model {tf_model_name}")
if model_type == "hf":
model, input, _ = get_causal_lm_model(tf_model_name)
model, input, _ = get_masked_lm_model(tf_model_name)
if model_type == "img":
model, input, _ = get_causal_image_model(tf_model_name)
if model_type == "keras":
model, input, _ = get_keras_model(tf_model_name)
if model_type == "TFhf":
model, input, _ = get_TFhf_model(tf_model_name)
if model_type == "hf_causallm":
model, input, _ = get_causal_lm_model(tf_model_name)
tf_model_name = tf_model_name.replace("/", "_")
tf_model_dir = os.path.join(

1
tank/model_metadata.csv
View File

@@ -34,3 +34,4 @@ mnasnet1_0,False,True,-,"cnn, torchvision, mobile, architecture-search","Outperf
bert-large-uncased,True,hf,True,330M,"nlp;bert-variant;transformer-encoder","24 layers, 1024 hidden units, 16 attention heads"
efficientnet_b0,True,False,5.3M,"image-classification;cnn;conv2d;depthwise-conv","Smallest EfficientNet variant with 224x224 input"
efficientnet_b7,True,False,66M,"image-classification;cnn;conv2d;depthwise-conv","Largest EfficientNet variant with 600x600 input"
gpt2,True,False,110M,"nlp;transformer-decoder;auto-regressive","12 layers, 768 hidden units, 12 attention heads"
1 model_name use_tracing dynamic param_count tags notes
34 bert-large-uncased True hf True 330M nlp;bert-variant;transformer-encoder
35 efficientnet_b0 True False 5.3M image-classification;cnn;conv2d;depthwise-conv Smallest EfficientNet variant with 224x224 input
36 efficientnet_b7 True False 66M image-classification;cnn;conv2d;depthwise-conv Largest EfficientNet variant with 600x600 input
37 gpt2 True False 110M nlp;transformer-decoder;auto-regressive 12 layers, 768 hidden units, 12 attention heads

131
tank/model_utils_tf.py
View File

@@ -7,7 +7,6 @@ from transformers import (
)
BATCH_SIZE = 1
MAX_SEQUENCE_LENGTH = 128
################################## MHLO/TF models #########################################
# TODO : Generate these lists or fetch model source from tank/tf/tf_model_list.csv
@@ -37,6 +36,9 @@ maskedlm_models = [
"hf-internal-testing/tiny-random-flaubert",
"xlm-roberta",
]
causallm_models = [
"gpt2",
]
tfhf_models = [
"microsoft/MiniLM-L12-H384-uncased",
]
@@ -50,6 +52,8 @@ def get_tf_model(name):
if name in keras_models:
return get_keras_model(name)
elif name in maskedlm_models:
return get_masked_lm_model(name)
elif name in causallm_models:
return get_causal_lm_model(name)
elif name in tfhf_models:
return get_TFhf_model(name)
@@ -61,12 +65,20 @@ def get_tf_model(name):
)
##################### Tensorflow Hugging Face LM Models ###################################
##################### Tensorflow Hugging Face Bert Models ###################################
BERT_MAX_SEQUENCE_LENGTH = 128
# Create a set of 2-dimensional inputs
tf_bert_input = [
tf.TensorSpec(shape=[BATCH_SIZE, MAX_SEQUENCE_LENGTH], dtype=tf.int32),
tf.TensorSpec(shape=[BATCH_SIZE, MAX_SEQUENCE_LENGTH], dtype=tf.int32),
tf.TensorSpec(shape=[BATCH_SIZE, MAX_SEQUENCE_LENGTH], dtype=tf.int32),
tf.TensorSpec(
shape=[BATCH_SIZE, BERT_MAX_SEQUENCE_LENGTH], dtype=tf.int32
),
tf.TensorSpec(
shape=[BATCH_SIZE, BERT_MAX_SEQUENCE_LENGTH], dtype=tf.int32
),
tf.TensorSpec(
shape=[BATCH_SIZE, BERT_MAX_SEQUENCE_LENGTH], dtype=tf.int32
),
]
@@ -92,28 +104,29 @@ def get_TFhf_model(name):
"microsoft/MiniLM-L12-H384-uncased"
)
text = "Replace me by any text you'd like."
text = [text] * BATCH_SIZE
encoded_input = tokenizer(
text,
padding="max_length",
truncation=True,
max_length=MAX_SEQUENCE_LENGTH,
max_length=BERT_MAX_SEQUENCE_LENGTH,
)
test_input = [
tf.reshape(
tf.convert_to_tensor(encoded_input["input_ids"], dtype=tf.int32),
[BATCH_SIZE, MAX_SEQUENCE_LENGTH],
[BATCH_SIZE, BERT_MAX_SEQUENCE_LENGTH],
),
tf.reshape(
tf.convert_to_tensor(
encoded_input["attention_mask"], dtype=tf.int32
),
[BATCH_SIZE, MAX_SEQUENCE_LENGTH],
[BATCH_SIZE, BERT_MAX_SEQUENCE_LENGTH],
),
tf.reshape(
tf.convert_to_tensor(
encoded_input["token_type_ids"], dtype=tf.int32
),
[BATCH_SIZE, MAX_SEQUENCE_LENGTH],
[BATCH_SIZE, BERT_MAX_SEQUENCE_LENGTH],
),
]
actual_out = model.forward(*test_input)
@@ -129,35 +142,41 @@ def compare_tensors_tf(tf_tensor, numpy_tensor):
return np.allclose(tf_to_numpy, numpy_tensor, rtol, atol)
##################### Tensorflow Hugging Face Masked LM Models ###################################
from transformers import TFAutoModelForMaskedLM, AutoTokenizer
import tensorflow as tf
# Create a set of input signature.
input_signature_maskedlm = [
tf.TensorSpec(shape=[BATCH_SIZE, MAX_SEQUENCE_LENGTH], dtype=tf.int32),
tf.TensorSpec(shape=[BATCH_SIZE, MAX_SEQUENCE_LENGTH], dtype=tf.int32),
]
# For supported models please see here:
# https://huggingface.co/docs/transformers/model_doc/auto#transformers.TFAutoModelForCasualLM
# Tokenizer for language models
def preprocess_input(
model_name, text="This is just used to compile the model"
model_name, max_length, text="This is just used to compile the model"
):
tokenizer = AutoTokenizer.from_pretrained(model_name)
text = [text] * BATCH_SIZE
inputs = tokenizer(
text,
padding="max_length",
return_tensors="tf",
padding="max_length",
truncation=True,
max_length=MAX_SEQUENCE_LENGTH,
max_length=max_length,
)
return inputs
##################### Tensorflow Hugging Face Masked LM Models ###################################
from transformers import TFAutoModelForMaskedLM, AutoTokenizer
import tensorflow as tf
MASKED_LM_MAX_SEQUENCE_LENGTH = 128
# Create a set of input signature.
input_signature_maskedlm = [
tf.TensorSpec(
shape=[BATCH_SIZE, MASKED_LM_MAX_SEQUENCE_LENGTH], dtype=tf.int32
),
tf.TensorSpec(
shape=[BATCH_SIZE, MASKED_LM_MAX_SEQUENCE_LENGTH], dtype=tf.int32
),
]
# For supported models please see here:
# https://huggingface.co/docs/transformers/model_doc/auto#transformers.TFAutoModelForMaskedLM
class MaskedLM(tf.Module):
def __init__(self, model_name):
super(MaskedLM, self).__init__()
@@ -171,9 +190,65 @@ class MaskedLM(tf.Module):
return self.m.predict(input_ids, attention_mask)
def get_causal_lm_model(hf_name, text="Hello, this is the default text."):
def get_masked_lm_model(hf_name, text="Hello, this is the default text."):
model = MaskedLM(hf_name)
encoded_input = preprocess_input(hf_name, text)
encoded_input = preprocess_input(
hf_name, MASKED_LM_MAX_SEQUENCE_LENGTH, text
)
test_input = (encoded_input["input_ids"], encoded_input["attention_mask"])
actual_out = model.forward(*test_input)
return model, test_input, actual_out
##################### Tensorflow Hugging Face Causal LM Models ###################################
from transformers import AutoConfig, TFAutoModelForCausalLM, TFGPT2Model
CAUSAL_LM_MAX_SEQUENCE_LENGTH = 1024
input_signature_causallm = [
tf.TensorSpec(
shape=[BATCH_SIZE, CAUSAL_LM_MAX_SEQUENCE_LENGTH], dtype=tf.int32
),
tf.TensorSpec(
shape=[BATCH_SIZE, CAUSAL_LM_MAX_SEQUENCE_LENGTH], dtype=tf.int32
),
]
# For supported models please see here:
# https://huggingface.co/docs/transformers/model_doc/auto#transformers.TFAutoModelForCausalLM
# For more background, see:
# https://huggingface.co/blog/tf-xla-generate
class CausalLM(tf.Module):
def __init__(self, model_name):
super(CausalLM, self).__init__()
# Decoder-only models need left padding.
self.tokenizer = AutoTokenizer.from_pretrained(
model_name, padding_side="left", pad_token="</s>"
)
self.tokenization_kwargs = {
"pad_to_multiple_of": CAUSAL_LM_MAX_SEQUENCE_LENGTH,
"padding": True,
"return_tensors": "tf",
}
self.model = TFGPT2Model.from_pretrained(model_name, return_dict=True)
self.model.predict = lambda x, y: self.model(
input_ids=x, attention_mask=y
)[0]
def preprocess_input(self, text):
return self.tokenizer(text, **self.tokenization_kwargs)
@tf.function(input_signature=input_signature_causallm, jit_compile=True)
def forward(self, input_ids, attention_mask):
return self.model.predict(input_ids, attention_mask)
def get_causal_lm_model(hf_name, text="Hello, this is the default text."):
model = CausalLM(hf_name)
batched_text = [text] * BATCH_SIZE
encoded_input = model.preprocess_input(batched_text)
test_input = (encoded_input["input_ids"], encoded_input["attention_mask"])
actual_out = model.forward(*test_input)
return model, test_input, actual_out

1
tank/tf_model_list.csv
View File

@@ -21,3 +21,4 @@ efficientnet-v2-s,keras
bert-large-uncased,hf
efficientnet_b0,keras
efficientnet_b7,keras
gpt2,hf_causallm
1 model_name model_type
21 bert-large-uncased hf
22 efficientnet_b0 keras
23 efficientnet_b7 keras
24 gpt2 hf_causallm