421 lines
17 KiB
Python
421 lines
17 KiB
Python
# coding=utf-8
|
|
# Copyright 2023 The HuggingFace Inc. 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.
|
|
""" Testing suite for the PyTorch VipLlava model. """
|
|
|
|
import copy
|
|
import gc
|
|
import unittest
|
|
|
|
import requests
|
|
|
|
from transformers import (
|
|
AutoProcessor,
|
|
VipLlavaConfig,
|
|
VipLlavaForConditionalGeneration,
|
|
is_torch_available,
|
|
is_vision_available,
|
|
)
|
|
from transformers.testing_utils import require_bitsandbytes, require_torch, require_torch_gpu, slow, torch_device
|
|
|
|
from ...test_configuration_common import ConfigTester
|
|
from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
|
|
|
|
|
|
if is_torch_available():
|
|
import torch
|
|
else:
|
|
is_torch_greater_or_equal_than_2_0 = False
|
|
|
|
if is_vision_available():
|
|
from PIL import Image
|
|
|
|
|
|
# Copied from transformers.tests.models.llava.test_modeling_llava.LlavaVisionText2TextModelTester with Llava->VipLlava
|
|
class VipLlavaVisionText2TextModelTester:
|
|
# Ignore copy
|
|
def __init__(
|
|
self,
|
|
parent,
|
|
ignore_index=-100,
|
|
image_token_index=0,
|
|
projector_hidden_act="gelu",
|
|
seq_length=7,
|
|
vision_feature_layers=[0, 0, 1, 1, 0],
|
|
text_config={
|
|
"model_type": "llama",
|
|
"seq_length": 7,
|
|
"is_training": True,
|
|
"use_input_mask": True,
|
|
"use_token_type_ids": False,
|
|
"use_labels": True,
|
|
"vocab_size": 99,
|
|
"hidden_size": 32,
|
|
"num_hidden_layers": 2,
|
|
"num_attention_heads": 4,
|
|
"intermediate_size": 37,
|
|
"hidden_act": "gelu",
|
|
"hidden_dropout_prob": 0.1,
|
|
"attention_probs_dropout_prob": 0.1,
|
|
"max_position_embeddings": 512,
|
|
"type_vocab_size": 16,
|
|
"type_sequence_label_size": 2,
|
|
"initializer_range": 0.02,
|
|
"num_labels": 3,
|
|
"num_choices": 4,
|
|
"pad_token_id": 0,
|
|
},
|
|
is_training=True,
|
|
vision_config={
|
|
"batch_size": 12,
|
|
"image_size": 30,
|
|
"patch_size": 2,
|
|
"num_channels": 3,
|
|
"is_training": True,
|
|
"hidden_size": 32,
|
|
"projection_dim": 32,
|
|
"num_hidden_layers": 2,
|
|
"num_attention_heads": 4,
|
|
"intermediate_size": 37,
|
|
"dropout": 0.1,
|
|
"attention_dropout": 0.1,
|
|
"initializer_range": 0.02,
|
|
},
|
|
):
|
|
self.parent = parent
|
|
self.ignore_index = ignore_index
|
|
self.image_token_index = image_token_index
|
|
self.projector_hidden_act = projector_hidden_act
|
|
self.vision_feature_layers = vision_feature_layers
|
|
self.text_config = text_config
|
|
self.vision_config = vision_config
|
|
self.seq_length = seq_length
|
|
|
|
self.num_hidden_layers = text_config["num_hidden_layers"]
|
|
self.vocab_size = text_config["vocab_size"]
|
|
self.hidden_size = text_config["hidden_size"]
|
|
self.num_attention_heads = text_config["num_attention_heads"]
|
|
self.is_training = is_training
|
|
|
|
self.batch_size = 3
|
|
self.num_channels = 3
|
|
self.image_size = 336
|
|
self.encoder_seq_length = 231
|
|
|
|
def get_config(self):
|
|
return VipLlavaConfig(
|
|
text_config=self.text_config,
|
|
vision_config=self.vision_config,
|
|
ignore_index=self.ignore_index,
|
|
image_token_index=self.image_token_index,
|
|
projector_hidden_act=self.projector_hidden_act,
|
|
vision_feature_layers=self.vision_feature_layers,
|
|
)
|
|
|
|
def prepare_config_and_inputs(self):
|
|
pixel_values = floats_tensor(
|
|
[
|
|
self.batch_size,
|
|
self.vision_config["num_channels"],
|
|
self.vision_config["image_size"],
|
|
self.vision_config["image_size"],
|
|
]
|
|
)
|
|
config = self.get_config()
|
|
|
|
return config, pixel_values
|
|
|
|
def prepare_config_and_inputs_for_common(self):
|
|
config_and_inputs = self.prepare_config_and_inputs()
|
|
config, pixel_values = config_and_inputs
|
|
input_ids = ids_tensor([self.batch_size, self.seq_length], config.text_config.vocab_size - 1) + 1
|
|
attention_mask = input_ids.ne(1).to(torch_device)
|
|
# we are giving 3 images let's make sure we pass in 3 image tokens
|
|
input_ids[:, 1] = config.image_token_index
|
|
inputs_dict = {
|
|
"pixel_values": pixel_values,
|
|
"input_ids": input_ids,
|
|
"attention_mask": attention_mask,
|
|
}
|
|
return config, inputs_dict
|
|
|
|
|
|
@require_torch
|
|
# Copied from transformers.tests.models.llava.test_modeling_llava.LlavaForConditionalGenerationModelTest with Llava->VipLlava
|
|
class VipLlavaForConditionalGenerationModelTest(ModelTesterMixin, unittest.TestCase):
|
|
"""
|
|
Model tester for `VipLlavaForConditionalGeneration`.
|
|
"""
|
|
|
|
all_model_classes = (VipLlavaForConditionalGeneration,) if is_torch_available() else ()
|
|
fx_compatible = False
|
|
test_pruning = False
|
|
test_resize_embeddings = True
|
|
test_head_masking = False
|
|
|
|
def setUp(self):
|
|
self.model_tester = VipLlavaVisionText2TextModelTester(self)
|
|
self.config_tester = ConfigTester(self, config_class=VipLlavaConfig, has_text_modality=False)
|
|
|
|
@unittest.skip(
|
|
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
|
|
)
|
|
def test_training_gradient_checkpointing(self):
|
|
pass
|
|
|
|
@unittest.skip(
|
|
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
|
|
)
|
|
def test_training_gradient_checkpointing_use_reentrant(self):
|
|
pass
|
|
|
|
@unittest.skip(
|
|
reason="This architecure seem to not compute gradients properly when using GC, check: https://github.com/huggingface/transformers/pull/27124"
|
|
)
|
|
def test_training_gradient_checkpointing_use_reentrant_false(self):
|
|
pass
|
|
|
|
# Copied from tests.test_modeling_common.ModelTesterMixin.test_resize_tokens_embeddings with config.vocab_size->config.text_config.vocab_size
|
|
def test_resize_tokens_embeddings(self):
|
|
(
|
|
original_config,
|
|
inputs_dict,
|
|
) = self.model_tester.prepare_config_and_inputs_for_common()
|
|
if not self.test_resize_embeddings:
|
|
return
|
|
|
|
for model_class in self.all_model_classes:
|
|
config = copy.deepcopy(original_config)
|
|
model = model_class(config)
|
|
model.to(torch_device)
|
|
|
|
if self.model_tester.is_training is False:
|
|
model.eval()
|
|
|
|
model_vocab_size = config.text_config.vocab_size
|
|
# Retrieve the embeddings and clone theme
|
|
model_embed = model.resize_token_embeddings(model_vocab_size)
|
|
cloned_embeddings = model_embed.weight.clone()
|
|
|
|
# Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
|
|
model_embed = model.resize_token_embeddings(model_vocab_size + 10)
|
|
self.assertEqual(model.config.text_config.vocab_size, model_vocab_size + 10)
|
|
# Check that it actually resizes the embeddings matrix
|
|
self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] + 10)
|
|
# Check that the model can still do a forward pass successfully (every parameter should be resized)
|
|
model(**self._prepare_for_class(inputs_dict, model_class))
|
|
|
|
# Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
|
|
model_embed = model.resize_token_embeddings(model_vocab_size - 15)
|
|
self.assertEqual(model.config.text_config.vocab_size, model_vocab_size - 15)
|
|
# Check that it actually resizes the embeddings matrix
|
|
self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] - 15)
|
|
|
|
# Check that the model can still do a forward pass successfully (every parameter should be resized)
|
|
# Input ids should be clamped to the maximum size of the vocabulary
|
|
inputs_dict["input_ids"].clamp_(max=model_vocab_size - 15 - 1)
|
|
|
|
# make sure that decoder_input_ids are resized as well
|
|
if "decoder_input_ids" in inputs_dict:
|
|
inputs_dict["decoder_input_ids"].clamp_(max=model_vocab_size - 15 - 1)
|
|
model(**self._prepare_for_class(inputs_dict, model_class))
|
|
|
|
# Check that adding and removing tokens has not modified the first part of the embedding matrix.
|
|
models_equal = True
|
|
for p1, p2 in zip(cloned_embeddings, model_embed.weight):
|
|
if p1.data.ne(p2.data).sum() > 0:
|
|
models_equal = False
|
|
|
|
self.assertTrue(models_equal)
|
|
|
|
config = copy.deepcopy(original_config)
|
|
model = model_class(config)
|
|
model.to(torch_device)
|
|
|
|
model_vocab_size = config.text_config.vocab_size
|
|
model.resize_token_embeddings(model_vocab_size + 10, pad_to_multiple_of=1)
|
|
self.assertTrue(model.config.text_config.vocab_size + 10, model_vocab_size)
|
|
|
|
model_embed = model.resize_token_embeddings(model_vocab_size, pad_to_multiple_of=64)
|
|
self.assertTrue(model_embed.weight.shape[0] // 64, 0)
|
|
|
|
self.assertTrue(model_embed.weight.shape[0], model.config.text_config.vocab_size)
|
|
self.assertTrue(model.config.text_config.vocab_size, model.vocab_size)
|
|
|
|
model_embed = model.resize_token_embeddings(model_vocab_size + 13, pad_to_multiple_of=64)
|
|
self.assertTrue(model_embed.weight.shape[0] // 64, 0)
|
|
|
|
# Check that resizing a model to a multiple of pad_to_multiple leads to a model of exactly that size
|
|
target_dimension = 128
|
|
model_embed = model.resize_token_embeddings(target_dimension, pad_to_multiple_of=64)
|
|
self.assertTrue(model_embed.weight.shape[0], target_dimension)
|
|
|
|
with self.assertRaisesRegex(
|
|
ValueError,
|
|
"Asking to pad the embedding matrix to a multiple of `1.3`, which is not and integer. Please make sure to pass an integer",
|
|
):
|
|
model.resize_token_embeddings(model_vocab_size, pad_to_multiple_of=1.3)
|
|
|
|
# Copied from tests.test_modeling_common.ModelTesterMixin.test_resize_embeddings_untied with config.vocab_size->config.text_config.vocab_size
|
|
def test_resize_embeddings_untied(self):
|
|
(
|
|
original_config,
|
|
inputs_dict,
|
|
) = self.model_tester.prepare_config_and_inputs_for_common()
|
|
if not self.test_resize_embeddings:
|
|
return
|
|
|
|
original_config.tie_word_embeddings = False
|
|
|
|
# if model cannot untied embeddings -> leave test
|
|
if original_config.tie_word_embeddings:
|
|
return
|
|
|
|
for model_class in self.all_model_classes:
|
|
config = copy.deepcopy(original_config)
|
|
model = model_class(config).to(torch_device)
|
|
|
|
# if no output embeddings -> leave test
|
|
if model.get_output_embeddings() is None:
|
|
continue
|
|
|
|
# Check that resizing the token embeddings with a larger vocab size increases the model's vocab size
|
|
model_vocab_size = config.text_config.vocab_size
|
|
model.resize_token_embeddings(model_vocab_size + 10)
|
|
self.assertEqual(model.config.text_config.vocab_size, model_vocab_size + 10)
|
|
output_embeds = model.get_output_embeddings()
|
|
self.assertEqual(output_embeds.weight.shape[0], model_vocab_size + 10)
|
|
# Check bias if present
|
|
if output_embeds.bias is not None:
|
|
self.assertEqual(output_embeds.bias.shape[0], model_vocab_size + 10)
|
|
# Check that the model can still do a forward pass successfully (every parameter should be resized)
|
|
model(**self._prepare_for_class(inputs_dict, model_class))
|
|
|
|
# Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
|
|
model.resize_token_embeddings(model_vocab_size - 15)
|
|
self.assertEqual(model.config.text_config.vocab_size, model_vocab_size - 15)
|
|
# Check that it actually resizes the embeddings matrix
|
|
output_embeds = model.get_output_embeddings()
|
|
self.assertEqual(output_embeds.weight.shape[0], model_vocab_size - 15)
|
|
# Check bias if present
|
|
if output_embeds.bias is not None:
|
|
self.assertEqual(output_embeds.bias.shape[0], model_vocab_size - 15)
|
|
# Check that the model can still do a forward pass successfully (every parameter should be resized)
|
|
# Input ids should be clamped to the maximum size of the vocabulary
|
|
inputs_dict["input_ids"].clamp_(max=model_vocab_size - 15 - 1)
|
|
if "decoder_input_ids" in inputs_dict:
|
|
inputs_dict["decoder_input_ids"].clamp_(max=model_vocab_size - 15 - 1)
|
|
# Check that the model can still do a forward pass successfully (every parameter should be resized)
|
|
model(**self._prepare_for_class(inputs_dict, model_class))
|
|
|
|
# Copied from tests.test_modeling_common.ModelTesterMixin.test_tie_model_weights with config.vocab_size->config.text_config.vocab_size
|
|
def test_tie_model_weights(self):
|
|
if not self.test_torchscript:
|
|
return
|
|
|
|
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
|
|
|
def check_same_values(layer_1, layer_2):
|
|
equal = True
|
|
for p1, p2 in zip(layer_1.weight, layer_2.weight):
|
|
if p1.data.ne(p2.data).sum() > 0:
|
|
equal = False
|
|
return equal
|
|
|
|
for model_class in self.all_model_classes:
|
|
config.torchscript = True
|
|
model_not_tied = model_class(config)
|
|
if model_not_tied.get_output_embeddings() is None:
|
|
continue
|
|
|
|
config_tied = copy.deepcopy(config)
|
|
config_tied.torchscript = False
|
|
model_tied = model_class(config_tied)
|
|
params_tied = list(model_tied.parameters())
|
|
# Check that the embedding layer and decoding layer are the same in size and in value
|
|
# self.assertTrue(check_same_values(embeddings, decoding))
|
|
|
|
# Check that after resize they remain tied.
|
|
model_tied.resize_token_embeddings(config.text_config.vocab_size + 10)
|
|
params_tied_2 = list(model_tied.parameters())
|
|
self.assertEqual(len(params_tied_2), len(params_tied))
|
|
|
|
|
|
@require_torch
|
|
class VipLlavaForConditionalGenerationIntegrationTest(unittest.TestCase):
|
|
def setUp(self):
|
|
self.processor = AutoProcessor.from_pretrained("llava-hf/vip-llava-7b-hf")
|
|
|
|
def tearDown(self):
|
|
gc.collect()
|
|
torch.cuda.empty_cache()
|
|
|
|
@slow
|
|
@require_bitsandbytes
|
|
def test_small_model_integration_test(self):
|
|
model_id = "llava-hf/vip-llava-7b-hf"
|
|
|
|
model = VipLlavaForConditionalGeneration.from_pretrained(model_id, load_in_4bit=True)
|
|
processor = AutoProcessor.from_pretrained(model_id)
|
|
|
|
url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/compel-neg.png"
|
|
|
|
image = Image.open(requests.get(url, stream=True).raw)
|
|
prompt = "USER: <image>\nCan you please describe this image?\nASSISTANT:"
|
|
|
|
inputs = processor(prompt, image, return_tensors="pt").to(torch_device, torch.float16)
|
|
|
|
outputs = model.generate(**inputs, max_new_tokens=10)
|
|
|
|
EXPECTED_OUTPUT = "USER: <image> \nCan you please describe this image?\nASSISTANT: The image features a brown and white cat sitting on"
|
|
self.assertEqual(processor.decode(outputs[0], skip_special_tokens=True), EXPECTED_OUTPUT)
|
|
|
|
@slow
|
|
@require_torch_gpu
|
|
def test_vipllava_merge_inputs_error_bug(self):
|
|
# This is a reproducer of https://github.com/huggingface/transformers/pull/28333 and makes sure it does not happen anymore
|
|
model_id = "llava-hf/vip-llava-7b-hf"
|
|
model = VipLlavaForConditionalGeneration.from_pretrained(
|
|
model_id, torch_dtype=torch.float16, low_cpu_mem_usage=True
|
|
).to(torch_device)
|
|
|
|
# Simulate some user inputs
|
|
pixel_values = torch.randn(
|
|
(2, 3, 336, 336),
|
|
dtype=torch.float,
|
|
device=torch_device,
|
|
)
|
|
input_ids = torch.tensor(
|
|
[
|
|
[32001, 32001, 1, 15043, 7084, 32000, 29871, 13, 7900],
|
|
[1, 15043, 7084, 29901, 29871, 32000, 29871, 13, 7900],
|
|
],
|
|
dtype=torch.long,
|
|
device=torch_device,
|
|
)
|
|
attention_mask = torch.tensor(
|
|
[[0, 0, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1]],
|
|
dtype=torch.long,
|
|
device=torch_device,
|
|
)
|
|
|
|
# Make sure that the loss is properly computed
|
|
loss = model(
|
|
pixel_values=pixel_values,
|
|
input_ids=input_ids,
|
|
attention_mask=attention_mask,
|
|
labels=input_ids,
|
|
).loss
|
|
loss.backward()
|