Add evolla rebase main (#36232)
* add evolla * adding protein encoder part * add initial processing test * save processor * add docstring * add evolla processor * add two test * change vision to protein * change resampler to sequence_compressor * change vision to protein * initial update for llama * add initial update for llamaForCausalLM * add `test_processor`, `test_saprot_output`, `test_protein_encoder_output` * change evolla, but still working on it * add test_single_forward * pass test_attention_outputs * pass test_hidden_states_output * pass test_save_load and test_from_pretrained_no_checkpoint * pass test_cpu_offload * skip some tests * update new progress * skip test_model_is_small * pass test_model_weights_reload_no_missing_tied_weights * pass test_model_get_set_embeddings * pass test_cpu_offload * skip test_resize_embeddings * add pipeline_model_mapping * remote old setUp * pass processor save_pretrained and load_pretrained * remove pooling layer * pass test_inputs_embeds_matches_input_ids * pass test_model_is_small * pass test_attention_outputs * pass test_initialization * pass test_model_get_set_embeddings * pass test_single_forward * skip test_disk_offload_bin and test_disk_offload_safetensors * fix most tests * pass test_protein_encoder_output * remove useless code * add EvollaForProteinText2Text * pass test_saprot_output * pass all EvollaModelTest test and remove processor test * add processor test to its own file * skip is_training since esm skipped it and the saprot code causes error when setting is_training True * pass processor tests * solve all except config * pass most cases * change init * add doc to `configuration_evolla.py` * remove image_processing test * remove extra processor test * remove extra modules * remove extra modules * change all configs into one config * pass all evolla test * pass `make fixup` * update short summary * update Evolla-10B-hf * pass check_dummies.py and check_code_quality * fix `tests/models/auto/test_tokenization_auto.py::AutoTokenizerTest::test_model_name_edge_cases_in_mappings` * remove dummy codes * change format * fix llava issue * update format * update to solve llama3 access issue * update to make forward right * solve processor save load problem from instructblip solution * remove unexpected file * skip `test_generation_tester_mixin_inheritance` * add `test_single_forward_correct` and `test_inference_natural_language_protein_reasoning` * add `modular_evolla.py` * solved issue #36362 * run `make fixup` * update modular * solve float32 training * add fix * solve `utils/check_docstrings.py` * update * update * update * remove other files and replace sequential and einsum * add use case in document * update the models * update model * change some wrong code * Update src/transformers/models/evolla/modular_evolla.py Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com> * Update src/transformers/models/evolla/modular_evolla.py Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com> * Update src/transformers/models/evolla/modular_evolla.py Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com> * Update src/transformers/models/evolla/modular_evolla.py Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com> * fix issues mentioned in PR * update style and rearrange the placement * fix return_dict argument issue * solve SaProtConfig issue * Solve EvollaSaProtRotaryEmbedding issue * solve attention_mask issue * solve almosst all issues * make style * update config * remove unrelated pickle file * delete pickle files * fix config * simplify a lot * remove past k-v from encoder * continue work * style * skip it from init * fix init * fix init * simplify more * fill in docstrings * change test for generation * skip test * fix style --------- Co-authored-by: Chenchen Han <13980209828@163.com> Co-authored-by: Cyril Vallez <cyril.vallez@huggingface.co> Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com>
This commit is contained in:
Xibin Bayes Zhou
295
tests/models/evolla/test_processor_evolla.py
Normal file
295
tests/models/evolla/test_processor_evolla.py
Normal file
@@ -0,0 +1,295 @@
|
||||
# Copyright 2025 The HuggingFace Team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
import random
|
||||
import shutil
|
||||
import tempfile
|
||||
import unittest
|
||||
|
||||
from transformers import (
|
||||
AutoProcessor,
|
||||
EvollaProcessor,
|
||||
)
|
||||
from transformers.testing_utils import require_torch
|
||||
from transformers.utils import is_torch_available
|
||||
|
||||
from ...test_processing_common import ProcessorTesterMixin
|
||||
|
||||
|
||||
if is_torch_available():
|
||||
import torch
|
||||
|
||||
|
||||
EVOLLA_VALID_AA = list("ACDEFGHIKLMNPQRSTVWY#")
|
||||
EVOLLA_VALID_FS = list("pynwrqhgdlvtmfsaeikc#")
|
||||
|
||||
|
||||
@require_torch
|
||||
class EvollaProcessorTest(ProcessorTesterMixin, unittest.TestCase):
|
||||
processor_class = EvollaProcessor
|
||||
|
||||
def setUp(self):
|
||||
self.tmpdirname = tempfile.mkdtemp()
|
||||
|
||||
processor = EvollaProcessor.from_pretrained("westlake-repl/Evolla-10B-hf")
|
||||
|
||||
processor.save_pretrained(self.tmpdirname)
|
||||
|
||||
self.input_keys = ["protein_input_ids", "protein_attention_mask", "input_ids", "attention_mask"]
|
||||
|
||||
def prepare_input_and_expected_output(self):
|
||||
amino_acid_sequence = "AAAA"
|
||||
foldseek_sequence = "dddd"
|
||||
question = "What is the function of this protein?"
|
||||
|
||||
expected_output = {
|
||||
"protein_input_ids": torch.tensor([[0, 13, 13, 13, 13, 2]]),
|
||||
"protein_attention_mask": torch.tensor([[1, 1, 1, 1, 1, 1]]),
|
||||
"input_ids": torch.tensor(
|
||||
[
|
||||
[
|
||||
128000,
|
||||
128006,
|
||||
9125,
|
||||
128007,
|
||||
271,
|
||||
2675,
|
||||
527,
|
||||
459,
|
||||
15592,
|
||||
6335,
|
||||
430,
|
||||
649,
|
||||
4320,
|
||||
904,
|
||||
4860,
|
||||
922,
|
||||
13128,
|
||||
13,
|
||||
128009,
|
||||
128006,
|
||||
882,
|
||||
128007,
|
||||
271,
|
||||
3923,
|
||||
374,
|
||||
279,
|
||||
734,
|
||||
315,
|
||||
420,
|
||||
13128,
|
||||
30,
|
||||
128009,
|
||||
128006,
|
||||
78191,
|
||||
128007,
|
||||
271,
|
||||
]
|
||||
]
|
||||
),
|
||||
"attention_mask": torch.tensor(
|
||||
[
|
||||
[
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
1,
|
||||
]
|
||||
]
|
||||
),
|
||||
}
|
||||
protein_dict = {"aa_seq": amino_acid_sequence, "foldseek": foldseek_sequence}
|
||||
message = [
|
||||
{"role": "system", "content": "You are an AI expert that can answer any questions about protein."},
|
||||
{"role": "user", "content": question},
|
||||
]
|
||||
return protein_dict, message, expected_output
|
||||
|
||||
def test_processor(self):
|
||||
protein_tokenizer = self.get_protein_tokenizer()
|
||||
tokenizer = self.get_tokenizer()
|
||||
|
||||
processor = EvollaProcessor(protein_tokenizer, tokenizer)
|
||||
|
||||
protein_dict, message, expected_output = self.prepare_input_and_expected_output()
|
||||
inputs = processor(proteins=[protein_dict], messages_list=[message])
|
||||
|
||||
# check if the input is correct
|
||||
for key, value in expected_output.items():
|
||||
self.assertTrue(
|
||||
torch.equal(inputs[key], value),
|
||||
f"inputs[key] is {inputs[key]} and expected_output[key] is {expected_output[key]}",
|
||||
)
|
||||
|
||||
def get_tokenizer(self, **kwargs):
|
||||
return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).tokenizer
|
||||
|
||||
def get_protein_tokenizer(self, **kwargs):
|
||||
return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).protein_tokenizer
|
||||
|
||||
def tearDown(self):
|
||||
shutil.rmtree(self.tmpdirname)
|
||||
|
||||
def prepare_inputs_single(self):
|
||||
proteins = {
|
||||
"aa_seq": "".join(random.choices(EVOLLA_VALID_AA, k=100)),
|
||||
"foldseek": "".join(random.choices(EVOLLA_VALID_FS, k=100)),
|
||||
}
|
||||
return proteins
|
||||
|
||||
def prepare_inputs_pair(self):
|
||||
proteins = [
|
||||
{
|
||||
"aa_seq": "".join(random.choices(EVOLLA_VALID_AA, k=100)),
|
||||
"foldseek": "".join(random.choices(EVOLLA_VALID_FS, k=100)),
|
||||
},
|
||||
{
|
||||
"aa_seq": "".join(random.choices(EVOLLA_VALID_AA, k=100)),
|
||||
"foldseek": "".join(random.choices(EVOLLA_VALID_FS, k=100)),
|
||||
},
|
||||
]
|
||||
return proteins
|
||||
|
||||
def prepare_inputs_long(self):
|
||||
proteins = [
|
||||
{
|
||||
"aa_seq": "".join(random.choices(EVOLLA_VALID_AA, k=100)),
|
||||
"foldseek": "".join(random.choices(EVOLLA_VALID_FS, k=100)),
|
||||
},
|
||||
{
|
||||
"aa_seq": "".join(random.choices(EVOLLA_VALID_AA, k=2000)),
|
||||
"foldseek": "".join(random.choices(EVOLLA_VALID_FS, k=2000)),
|
||||
},
|
||||
]
|
||||
return proteins
|
||||
|
||||
def prepare_inputs_short(self):
|
||||
proteins = [
|
||||
{
|
||||
"aa_seq": "".join(random.choices(EVOLLA_VALID_AA, k=1)),
|
||||
"foldseek": "".join(random.choices(EVOLLA_VALID_FS, k=1)),
|
||||
},
|
||||
{
|
||||
"aa_seq": "".join(random.choices(EVOLLA_VALID_AA, k=100)),
|
||||
"foldseek": "".join(random.choices(EVOLLA_VALID_FS, k=100)),
|
||||
},
|
||||
]
|
||||
return proteins
|
||||
|
||||
def prepare_inputs_empty(self):
|
||||
proteins = [
|
||||
{
|
||||
"aa_seq": "",
|
||||
"foldseek": "",
|
||||
},
|
||||
{
|
||||
"aa_seq": "".join(random.choices(EVOLLA_VALID_AA, k=100)),
|
||||
"foldseek": "".join(random.choices(EVOLLA_VALID_FS, k=100)),
|
||||
},
|
||||
]
|
||||
return proteins
|
||||
|
||||
def prepare_inputs(self, protein_types="pair"):
|
||||
r"""
|
||||
Prepare inputs for the test.
|
||||
|
||||
Args:
|
||||
protein_types (`str`): the types of proteins to prepare.
|
||||
- "single": a single correct protein.
|
||||
- "pair": a pair of correct proteins.
|
||||
- "long": a long sequence of correct proteins and a correct protein.
|
||||
- "short": a short sequence of correct proteins (only have 1 aa) and a correct protein.
|
||||
- "empty": an empty sequence of proteins and a correct protein.
|
||||
"""
|
||||
if protein_types == "single":
|
||||
proteins = self.prepare_inputs_single()
|
||||
elif protein_types == "pair":
|
||||
proteins = self.prepare_inputs_pair()
|
||||
elif protein_types == "long":
|
||||
proteins = self.prepare_inputs_long()
|
||||
elif protein_types == "short":
|
||||
proteins = self.prepare_inputs_short()
|
||||
elif protein_types == "empty":
|
||||
proteins = self.prepare_inputs_empty()
|
||||
else:
|
||||
raise ValueError(
|
||||
f"protein_types should be one of 'single', 'pair', 'long','short', 'empty', but got {protein_types}"
|
||||
)
|
||||
|
||||
questions = ["What is the function of the protein?"] * len(proteins)
|
||||
messages_list = []
|
||||
for question in questions:
|
||||
messages = [
|
||||
{"role": "system", "content": "You are an AI expert that can answer any questions about protein."},
|
||||
{"role": "user", "content": question},
|
||||
]
|
||||
messages_list.append(messages)
|
||||
return proteins, messages_list
|
||||
|
||||
def test_tokenizer_decode(self):
|
||||
protein_tokenizer = self.get_protein_tokenizer()
|
||||
tokenizer = self.get_tokenizer()
|
||||
|
||||
processor = EvollaProcessor(tokenizer=tokenizer, protein_tokenizer=protein_tokenizer, return_tensors="pt")
|
||||
|
||||
predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]]
|
||||
|
||||
decoded_processor = processor.batch_decode(predicted_ids)
|
||||
decoded_tok = tokenizer.batch_decode(predicted_ids)
|
||||
|
||||
self.assertListEqual(decoded_tok, decoded_processor)
|
||||
|
||||
def test_model_input_names(self):
|
||||
protein_tokenizer = self.get_protein_tokenizer()
|
||||
tokenizer = self.get_tokenizer()
|
||||
|
||||
processor = EvollaProcessor(tokenizer=tokenizer, protein_tokenizer=protein_tokenizer)
|
||||
proteins, messages_list = self.prepare_inputs()
|
||||
|
||||
inputs = processor(messages_list=messages_list, proteins=proteins, padding="longest", return_tensors="pt")
|
||||
|
||||
# For now the processor supports only ['pixel_values', 'input_ids', 'attention_mask']
|
||||
self.assertSetEqual(set(inputs.keys()), set(self.input_keys))
|
||||
Reference in New Issue
Block a user