diffusers/tests/pipelines/pag/test_pag_pixart_sigma.py

# coding=utf-8
# Copyright 2024 HuggingFace Inc.
#
# 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 inspect
import tempfile
import unittest

import numpy as np
import torch
from transformers import AutoTokenizer, T5EncoderModel

import diffusers
from diffusers import (
    AutoencoderKL,
    DDIMScheduler,
    PixArtSigmaPAGPipeline,
    PixArtSigmaPipeline,
    PixArtTransformer2DModel,
)
from diffusers.utils import logging
from diffusers.utils.testing_utils import (
    CaptureLogger,
    enable_full_determinism,
    print_tensor_test,
    torch_device,
)

from ..pipeline_params import (
    TEXT_TO_IMAGE_BATCH_PARAMS,
    TEXT_TO_IMAGE_IMAGE_PARAMS,
    TEXT_TO_IMAGE_PARAMS,
)
from ..test_pipelines_common import PipelineTesterMixin, assert_mean_pixel_difference, to_np


enable_full_determinism()


class PixArtSigmaPAGPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
    pipeline_class = PixArtSigmaPAGPipeline
    params = TEXT_TO_IMAGE_PARAMS.union({"pag_scale", "pag_adaptive_scale"})
    params = set(params)
    params.remove("cross_attention_kwargs")
    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
    required_optional_params = PipelineTesterMixin.required_optional_params

    def get_dummy_components(self):
        torch.manual_seed(0)
        transformer = PixArtTransformer2DModel(
            sample_size=8,
            num_layers=2,
            patch_size=2,
            attention_head_dim=8,
            num_attention_heads=3,
            caption_channels=32,
            in_channels=4,
            cross_attention_dim=24,
            out_channels=8,
            attention_bias=True,
            activation_fn="gelu-approximate",
            num_embeds_ada_norm=1000,
            norm_type="ada_norm_single",
            norm_elementwise_affine=False,
            norm_eps=1e-6,
        )
        torch.manual_seed(0)
        vae = AutoencoderKL()

        scheduler = DDIMScheduler()
        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")

        tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")

        components = {
            "transformer": transformer.eval(),
            "vae": vae.eval(),
            "scheduler": scheduler,
            "text_encoder": text_encoder,
            "tokenizer": tokenizer,
        }
        return components

    def get_dummy_inputs(self, device, seed=0):
        if str(device).startswith("mps"):
            generator = torch.manual_seed(seed)
        else:
            generator = torch.Generator(device=device).manual_seed(seed)
        inputs = {
            "prompt": "A painting of a squirrel eating a burger",
            "generator": generator,
            "num_inference_steps": 2,
            "guidance_scale": 1.0,
            "pag_scale": 3.0,
            "use_resolution_binning": False,
            "output_type": "np",
        }
        return inputs

    def test_pag_disable_enable(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()

        # base  pipeline (expect same output when pag is disabled)
        pipe = PixArtSigmaPipeline(**components)
        pipe = pipe.to(device)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        del inputs["pag_scale"]
        assert (
            "pag_scale" not in inspect.signature(pipe.__call__).parameters
        ), f"`pag_scale` should not be a call parameter of the base pipeline {pipe.__class__.__name__}."
        out = pipe(**inputs).images[0, -3:, -3:, -1]

        # pag disabled with pag_scale=0.0
        components["pag_applied_layers"] = ["blocks.1"]
        pipe_pag = self.pipeline_class(**components)
        pipe_pag = pipe_pag.to(device)
        pipe_pag.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        inputs["pag_scale"] = 0.0
        out_pag_disabled = pipe_pag(**inputs).images[0, -3:, -3:, -1]

        # pag enabled
        pipe_pag = self.pipeline_class(**components)
        pipe_pag = pipe_pag.to(device)
        pipe_pag.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        out_pag_enabled = pipe_pag(**inputs).images[0, -3:, -3:, -1]

        assert np.abs(out.flatten() - out_pag_disabled.flatten()).max() < 1e-3
        assert np.abs(out.flatten() - out_pag_enabled.flatten()).max() > 1e-3

    def test_pag_applied_layers(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()

        # base pipeline
        pipe = self.pipeline_class(**components)
        pipe = pipe.to(device)
        pipe.set_progress_bar_config(disable=None)

        # "attn1" should apply to all self-attention layers.
        all_self_attn_layers = [k for k in pipe.transformer.attn_processors.keys() if "attn1" in k]
        pag_layers = ["blocks.0", "blocks.1"]
        pipe._set_pag_attn_processor(pag_applied_layers=pag_layers, do_classifier_free_guidance=False)
        assert set(pipe.pag_attn_processors) == set(all_self_attn_layers)

    def test_pag_inference(self):
        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
        components = self.get_dummy_components()

        pipe_pag = self.pipeline_class(**components)
        pipe_pag = pipe_pag.to(device)
        pipe_pag.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(device)
        image = pipe_pag(**inputs).images
        image_slice = image[0, -3:, -3:, -1]
        print_tensor_test(image_slice)

        assert image.shape == (
            1,
            8,
            8,
            3,
        ), f"the shape of the output image should be (1, 8, 8, 3) but got {image.shape}"
        expected_slice = np.array([0.6499, 0.3250, 0.3572, 0.6780, 0.4453, 0.4582, 0.2770, 0.5168, 0.4594])

        max_diff = np.abs(image_slice.flatten() - expected_slice).max()
        self.assertLessEqual(max_diff, 1e-3)

    # Copied from tests.pipelines.pixart_sigma.test_pixart.PixArtSigmaPipelineFastTests.test_save_load_optional_components
    def test_save_load_optional_components(self):
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)
        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(torch_device)

        prompt = inputs["prompt"]
        generator = inputs["generator"]
        num_inference_steps = inputs["num_inference_steps"]
        output_type = inputs["output_type"]

        (
            prompt_embeds,
            prompt_attention_mask,
            negative_prompt_embeds,
            negative_prompt_attention_mask,
        ) = pipe.encode_prompt(prompt)

        # inputs with prompt converted to embeddings
        inputs = {
            "prompt_embeds": prompt_embeds,
            "prompt_attention_mask": prompt_attention_mask,
            "negative_prompt": None,
            "negative_prompt_embeds": negative_prompt_embeds,
            "negative_prompt_attention_mask": negative_prompt_attention_mask,
            "generator": generator,
            "num_inference_steps": num_inference_steps,
            "output_type": output_type,
            "use_resolution_binning": False,
        }

        # set all optional components to None
        for optional_component in pipe._optional_components:
            setattr(pipe, optional_component, None)

        output = pipe(**inputs)[0]

        with tempfile.TemporaryDirectory() as tmpdir:
            pipe.save_pretrained(tmpdir)
            pipe_loaded = self.pipeline_class.from_pretrained(tmpdir, pag_applied_layers=["blocks.1"])
            pipe_loaded.to(torch_device)
            pipe_loaded.set_progress_bar_config(disable=None)

        for optional_component in pipe._optional_components:
            self.assertTrue(
                getattr(pipe_loaded, optional_component) is None,
                f"`{optional_component}` did not stay set to None after loading.",
            )

        inputs = self.get_dummy_inputs(torch_device)

        generator = inputs["generator"]
        num_inference_steps = inputs["num_inference_steps"]
        output_type = inputs["output_type"]

        # inputs with prompt converted to embeddings
        inputs = {
            "prompt_embeds": prompt_embeds,
            "prompt_attention_mask": prompt_attention_mask,
            "negative_prompt": None,
            "negative_prompt_embeds": negative_prompt_embeds,
            "negative_prompt_attention_mask": negative_prompt_attention_mask,
            "generator": generator,
            "num_inference_steps": num_inference_steps,
            "output_type": output_type,
            "use_resolution_binning": False,
        }

        output_loaded = pipe_loaded(**inputs)[0]

        max_diff = np.abs(to_np(output) - to_np(output_loaded)).max()
        self.assertLess(max_diff, 1e-4)

    # Because the PAG PixArt Sigma has `pag_applied_layers`.
    # Also, we shouldn't be doing `set_default_attn_processor()` after loading
    # the pipeline with `pag_applied_layers`.
    def test_save_load_local(self, expected_max_difference=1e-4):
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)

        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(torch_device)
        output = pipe(**inputs)[0]

        logger = logging.get_logger("diffusers.pipelines.pipeline_utils")
        logger.setLevel(diffusers.logging.INFO)

        with tempfile.TemporaryDirectory() as tmpdir:
            pipe.save_pretrained(tmpdir, safe_serialization=False)

            with CaptureLogger(logger) as cap_logger:
                pipe_loaded = self.pipeline_class.from_pretrained(tmpdir, pag_applied_layers=["blocks.1"])

            for name in pipe_loaded.components.keys():
                if name not in pipe_loaded._optional_components:
                    assert name in str(cap_logger)

            pipe_loaded.to(torch_device)
            pipe_loaded.set_progress_bar_config(disable=None)

        inputs = self.get_dummy_inputs(torch_device)
        output_loaded = pipe_loaded(**inputs)[0]

        max_diff = np.abs(to_np(output) - to_np(output_loaded)).max()
        self.assertLess(max_diff, expected_max_difference)

    # We shouldn't be setting `set_default_attn_processor` here.
    def test_attention_slicing_forward_pass(
        self, test_max_difference=True, test_mean_pixel_difference=True, expected_max_diff=1e-3
    ):
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)

        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        generator_device = "cpu"
        inputs = self.get_dummy_inputs(generator_device)
        output_without_slicing = pipe(**inputs)[0]

        pipe.enable_attention_slicing(slice_size=1)
        inputs = self.get_dummy_inputs(generator_device)
        output_with_slicing1 = pipe(**inputs)[0]

        pipe.enable_attention_slicing(slice_size=2)
        inputs = self.get_dummy_inputs(generator_device)
        output_with_slicing2 = pipe(**inputs)[0]

        if test_max_difference:
            max_diff1 = np.abs(to_np(output_with_slicing1) - to_np(output_without_slicing)).max()
            max_diff2 = np.abs(to_np(output_with_slicing2) - to_np(output_without_slicing)).max()
            self.assertLess(
                max(max_diff1, max_diff2),
                expected_max_diff,
                "Attention slicing should not affect the inference results",
            )

        if test_mean_pixel_difference:
            assert_mean_pixel_difference(to_np(output_with_slicing1[0]), to_np(output_without_slicing[0]))
            assert_mean_pixel_difference(to_np(output_with_slicing2[0]), to_np(output_without_slicing[0]))

    # Because we have `pag_applied_layers` we cannot direcly apply
    # `set_default_attn_processor`
    def test_dict_tuple_outputs_equivalent(self, expected_slice=None, expected_max_difference=1e-4):
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)

        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)

        generator_device = "cpu"
        if expected_slice is None:
            output = pipe(**self.get_dummy_inputs(generator_device))[0]
        else:
            output = expected_slice

        output_tuple = pipe(**self.get_dummy_inputs(generator_device), return_dict=False)[0]

        if expected_slice is None:
            max_diff = np.abs(to_np(output) - to_np(output_tuple)).max()
        else:
            if output_tuple.ndim != 5:
                max_diff = np.abs(to_np(output) - to_np(output_tuple)[0, -3:, -3:, -1].flatten()).max()
            else:
                max_diff = np.abs(to_np(output) - to_np(output_tuple)[0, -3:, -3:, -1, -1].flatten()).max()

        self.assertLess(max_diff, expected_max_difference)

    # Same reason as above
    def test_inference_batch_single_identical(
        self,
        batch_size=2,
        expected_max_diff=1e-4,
        additional_params_copy_to_batched_inputs=["num_inference_steps"],
    ):
        components = self.get_dummy_components()
        pipe = self.pipeline_class(**components)

        pipe.to(torch_device)
        pipe.set_progress_bar_config(disable=None)
        inputs = self.get_dummy_inputs(torch_device)
        # Reset generator in case it is has been used in self.get_dummy_inputs
        inputs["generator"] = self.get_generator(0)

        logger = logging.get_logger(pipe.__module__)
        logger.setLevel(level=diffusers.logging.FATAL)

        # batchify inputs
        batched_inputs = {}
        batched_inputs.update(inputs)

        for name in self.batch_params:
            if name not in inputs:
                continue

            value = inputs[name]
            if name == "prompt":
                len_prompt = len(value)
                batched_inputs[name] = [value[: len_prompt // i] for i in range(1, batch_size + 1)]
                batched_inputs[name][-1] = 100 * "very long"

            else:
                batched_inputs[name] = batch_size * [value]

        if "generator" in inputs:
            batched_inputs["generator"] = [self.get_generator(i) for i in range(batch_size)]

        if "batch_size" in inputs:
            batched_inputs["batch_size"] = batch_size

        for arg in additional_params_copy_to_batched_inputs:
            batched_inputs[arg] = inputs[arg]

        output = pipe(**inputs)
        output_batch = pipe(**batched_inputs)

        assert output_batch[0].shape[0] == batch_size

        max_diff = np.abs(to_np(output_batch[0][0]) - to_np(output[0][0])).max()
        assert max_diff < expected_max_diff

    # Because we're passing `pag_applied_layers` (type of List) in the components as well.
    def test_components_function(self):
        init_components = self.get_dummy_components()
        init_components = {k: v for k, v in init_components.items() if not isinstance(v, (str, int, float, list))}

        pipe = self.pipeline_class(**init_components)

        self.assertTrue(hasattr(pipe, "components"))
        self.assertTrue(set(pipe.components.keys()) == set(init_components.keys()))