[test] added transformers models to test model zoo (#3135)

tests/test_fx/test_tracer/test_hf_model/test_hf_albert.py

@@ -1,66 +1,18 @@
-import pytest
-import torch
-import transformers
 from hf_tracer_utils import trace_model_and_compare_output
 
+from tests.kit.model_zoo import model_zoo
+
 BATCH_SIZE = 2
 SEQ_LENGTH = 16
 
 
-def test_single_sentence_albert():
-    MODEL_LIST = [
-        transformers.AlbertModel,
-        transformers.AlbertForPreTraining,
-        transformers.AlbertForMaskedLM,
-        transformers.AlbertForSequenceClassification,
-        transformers.AlbertForTokenClassification,
-    ]
+def test_albert():
+    sub_registry = model_zoo.get_sub_registry('transformers_albert')
 
-    config = transformers.AlbertConfig(embedding_size=128,
-                                       hidden_size=128,
-                                       num_hidden_layers=2,
-                                       num_attention_heads=4,
-                                       intermediate_size=256)
-
-    def data_gen():
-        input_ids = torch.zeros((BATCH_SIZE, SEQ_LENGTH), dtype=torch.int64)
-        token_type_ids = torch.zeros((BATCH_SIZE, SEQ_LENGTH), dtype=torch.int64)
-        attention_mask = torch.zeros((BATCH_SIZE, SEQ_LENGTH), dtype=torch.int64)
-        meta_args = dict(input_ids=input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask)
-        return meta_args
-
-    for model_cls in MODEL_LIST:
-        model = model_cls(config=config)
-        trace_model_and_compare_output(model, data_gen)
-
-
-def test_multi_sentence_albert():
-    config = transformers.AlbertConfig(hidden_size=128,
-                                       num_hidden_layers=2,
-                                       num_attention_heads=4,
-                                       intermediate_size=256)
-    tokenizer = transformers.BertTokenizer.from_pretrained("bert-base-uncased")
-
-    def data_gen_for_qa():
-        question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
-        inputs = tokenizer(question, text, return_tensors="pt")
-        return inputs
-
-    model = transformers.AlbertForQuestionAnswering(config)
-    trace_model_and_compare_output(model, data_gen_for_qa)
-
-    def data_gen_for_mcq():
-        prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
-        choice0 = "It is eaten with a fork and a knife."
-        choice1 = "It is eaten while held in the hand."
-        encoding = tokenizer([prompt, prompt], [choice0, choice1], return_tensors="pt", padding=True)
-        encoding = {k: v.unsqueeze(0) for k, v in encoding.items()}
-        return encoding
-
-    model = transformers.AlbertForMultipleChoice(config)
-    trace_model_and_compare_output(model, data_gen_for_mcq)
+    for name, (model_fn, data_gen_fn, _, _) in sub_registry.items():
+        model = model_fn()
+        trace_model_and_compare_output(model, data_gen_fn)
 
 
 if __name__ == '__main__':
-    test_single_sentence_albert()
-    test_multi_sentence_albert()
+    test_albert()

tests/test_fx/test_tracer/test_hf_model/test_hf_bert.py

@@ -1,69 +1,15 @@
-import pytest
-import torch
-import transformers
 from hf_tracer_utils import trace_model_and_compare_output
 
-BATCH_SIZE = 2
-SEQ_LENGTH = 16
+from tests.kit.model_zoo import model_zoo
 
 
-def test_single_sentence_bert():
-    MODEL_LIST = [
-        transformers.BertModel,
-        transformers.BertForPreTraining,
-        transformers.BertLMHeadModel,
-        transformers.BertForMaskedLM,
-        transformers.BertForSequenceClassification,
-        transformers.BertForTokenClassification,
-    ]
+def test_bert():
+    sub_registry = model_zoo.get_sub_registry('transformers_bert')
 
-    config = transformers.BertConfig(hidden_size=128, num_hidden_layers=2, num_attention_heads=4, intermediate_size=256)
-
-    def data_gen():
-        input_ids = torch.zeros((BATCH_SIZE, SEQ_LENGTH), dtype=torch.int64)
-        token_type_ids = torch.zeros((BATCH_SIZE, SEQ_LENGTH), dtype=torch.int64)
-        attention_mask = torch.zeros((BATCH_SIZE, SEQ_LENGTH), dtype=torch.int64)
-        meta_args = dict(input_ids=input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask)
-        return meta_args
-
-    for model_cls in MODEL_LIST:
-        model = model_cls(config=config)
-        trace_model_and_compare_output(model, data_gen)
-
-
-def test_multi_sentence_bert():
-    config = transformers.BertConfig(hidden_size=128, num_hidden_layers=2, num_attention_heads=4, intermediate_size=256)
-    tokenizer = transformers.BertTokenizer.from_pretrained("bert-base-uncased")
-
-    def data_gen_for_next_sentence():
-        prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
-        next_sentence = "The sky is blue due to the shorter wavelength of blue light."
-        encoding = tokenizer(prompt, next_sentence, return_tensors="pt")
-        return encoding
-
-    model = transformers.BertForNextSentencePrediction(config)
-    trace_model_and_compare_output(model, data_gen_for_next_sentence)
-
-    def data_gen_for_qa():
-        question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
-        inputs = tokenizer(question, text, return_tensors="pt")
-        return inputs
-
-    model = transformers.BertForQuestionAnswering(config)
-    trace_model_and_compare_output(model, data_gen_for_qa)
-
-    def data_gen_for_mcq():
-        prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
-        choice0 = "It is eaten with a fork and a knife."
-        choice1 = "It is eaten while held in the hand."
-        encoding = tokenizer([prompt, prompt], [choice0, choice1], return_tensors="pt", padding=True)
-        encoding = {k: v.unsqueeze(0) for k, v in encoding.items()}
-        return encoding
-
-    model = transformers.BertForMultipleChoice(config)
-    trace_model_and_compare_output(model, data_gen_for_mcq)
+    for name, (model_fn, data_gen_fn, _, _) in sub_registry.items():
+        model = model_fn()
+        trace_model_and_compare_output(model, data_gen_fn)
 
 
 if __name__ == '__main__':
-    test_single_sentence_bert()
-    test_multi_sentence_bert()
+    test_bert()

tests/test_fx/test_tracer/test_hf_model/test_hf_gpt.py

@@ -1,35 +1,17 @@
 import pytest
-import torch
-import transformers
 from hf_tracer_utils import trace_model_and_compare_output
 
-BATCH_SIZE = 1
-SEQ_LENGTH = 16
+from tests.kit.model_zoo import model_zoo
 
 
 # TODO: remove this skip once we handle the latest gpt model
 @pytest.mark.skip
 def test_gpt():
-    MODEL_LIST = [
-        transformers.GPT2Model,
-        transformers.GPT2LMHeadModel,
-        transformers.GPT2DoubleHeadsModel,
-        transformers.GPT2ForTokenClassification,
-    # transformers.GPT2ForSequenceClassification, # not supported yet
-    ]
+    sub_registry = model_zoo.get_sub_registry('transformers_gpt')
 
-    config = transformers.GPT2Config(n_position=64, n_layer=2, n_head=4)
-
-    def data_gen():
-        input_ids = torch.zeros((BATCH_SIZE, SEQ_LENGTH), dtype=torch.int64)
-        token_type_ids = torch.zeros((BATCH_SIZE, SEQ_LENGTH), dtype=torch.int64)
-        attention_mask = torch.zeros((BATCH_SIZE, SEQ_LENGTH), dtype=torch.int64)
-        kwargs = dict(input_ids=input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask)
-        return kwargs
-
-    for model_cls in MODEL_LIST:
-        model = model_cls(config=config)
-        trace_model_and_compare_output(model, data_gen)
+    for name, (model_fn, data_gen_fn, _, _) in sub_registry.items():
+        model = model_fn()
+        trace_model_and_compare_output(model, data_gen_fn)
 
 
 if __name__ == '__main__':

tests/test_fx/test_tracer/test_hf_model/test_hf_opt.py

@@ -1,29 +1,14 @@
-import pytest
-import torch
-import transformers
 from hf_tracer_utils import trace_model_and_compare_output
 
-BATCH_SIZE = 1
-SEQ_LENGTH = 16
+from tests.kit.model_zoo import model_zoo
 
 
 def test_opt():
-    MODEL_LIST = [
-        transformers.OPTModel,
-        transformers.OPTForCausalLM,
-    ]
+    sub_registry = model_zoo.get_sub_registry('transformers_opt')
 
-    config = transformers.OPTConfig(hidden_size=128, num_hidden_layers=2, num_attention_heads=4)
-
-    def data_gen():
-        input_ids = torch.zeros((BATCH_SIZE, SEQ_LENGTH), dtype=torch.int64)
-        attention_mask = torch.zeros((BATCH_SIZE, SEQ_LENGTH), dtype=torch.int64)
-        kwargs = dict(input_ids=input_ids, attention_mask=attention_mask)
-        return kwargs
-
-    for model_cls in MODEL_LIST:
-        model = model_cls(config=config)
-        trace_model_and_compare_output(model, data_gen)
+    for name, (model_fn, data_gen_fn, _, _) in sub_registry.items():
+        model = model_fn()
+        trace_model_and_compare_output(model, data_gen_fn)
 
 
 if __name__ == '__main__':

tests/test_fx/test_tracer/test_hf_model/test_hf_t5.py

@@ -1,41 +1,14 @@
-import pytest
-import torch
-import transformers
 from hf_tracer_utils import trace_model_and_compare_output
 
-BATCH_SIZE = 1
-SEQ_LENGTH = 16
+from tests.kit.model_zoo import model_zoo
 
 
 def test_t5():
-    MODEL_LIST = [
-        transformers.T5Model,
-        transformers.T5ForConditionalGeneration,
-        transformers.T5EncoderModel,
-    ]
+    sub_registry = model_zoo.get_sub_registry('transformers_t5')
 
-    config = transformers.T5Config(d_model=128, num_layers=2)
-
-    def data_gen():
-        input_ids = torch.zeros((BATCH_SIZE, SEQ_LENGTH), dtype=torch.int64)
-        decoder_input_ids = torch.zeros((BATCH_SIZE, SEQ_LENGTH), dtype=torch.int64)
-        kwargs = dict(input_ids=input_ids, decoder_input_ids=decoder_input_ids)
-        return kwargs
-
-    def data_gen_for_encoder_only():
-        input_ids = torch.zeros((BATCH_SIZE, SEQ_LENGTH), dtype=torch.int64)
-        kwargs = dict(input_ids=input_ids)
-        return kwargs
-
-    for model_cls in MODEL_LIST:
-        model = model_cls(config=config)
-
-        if isinstance(model, transformers.T5EncoderModel):
-            data_gen_func = data_gen_for_encoder_only
-        else:
-            data_gen_func = data_gen
-
-        trace_model_and_compare_output(model, data_gen_func)
+    for name, (model_fn, data_gen_fn, _, _) in sub_registry.items():
+        model = model_fn()
+        trace_model_and_compare_output(model, data_gen_fn)
 
 
 if __name__ == '__main__':