Real EsrGan X2 Model Conversion

[lazy-nurd]

Env

• GPU : RTX 3090
• OS : Ubuntu20.04
• Cuda version : 11.4
• TensorRT version : 22

About this repo

• master
• real-esrgan

Your problem

I have successfully converted X4 with outscale : 4 model to tensorrt, and I have seen good improvements in overall inference speed and memory footprint. But I want to convert RealESRGAN_x2plus to tensorrt with outscale 2. X2 plus uses pixel unshuffle before first convolution which seems very difficult to me to implement this in C++.

Please let me know how I can convert X2 to tensorrt or please add support for X2 conversion in your repository.

I am looking forward to hear from you. Thanks

[wang-xinyu]

@yester31 Can you please suggest?

[yester31]

added below code(Pixel unshuffle code) before conv_first and use RealESRGAN_x2plus weight

if (OUT_SCALE == 2) {
    // Pixel unshuffle.
    int h = INPUT_H / OUT_SCALE;
    int w = INPUT_W / OUT_SCALE;
    auto attn_shuffle = network->addShuffle(*prep);

    Dims shape_dims;
    std::vector<int> reshape_dims = { INPUT_C, h, OUT_SCALE, w, OUT_SCALE };
    shape_dims.nbDims = (int)reshape_dims.size();
    memcpy(shape_dims.d, reshape_dims.data(), reshape_dims.size() * sizeof(int));
    attn_shuffle->setReshapeDimensions(shape_dims);

    std::vector<int> trans_dims{ 0, 2, 4, 1, 3 };
    Permutation f_trans_dims; memcpy(f_trans_dims.order, trans_dims.data(), trans_dims.size() * sizeof(int));
    attn_shuffle->setSecondTranspose(f_trans_dims);

    auto attn_shuffle2 = network->addShuffle(*attn_shuffle->getOutput(0));
    attn_shuffle2->setReshapeDimensions(Dims3{ INPUT_C * OUT_SCALE * OUT_SCALE, h, w});
    prep = attn_shuffle2->getOutput(0);
}

[lazy-nurd]

This solved the issue. Thank you so much

[lazy-nurd]

Hello, I have been successful in x2 conversion. But I want to infer the engine in Python and the results engine is giving are very odd.

This is the script I am using for inferencing.

import os
import time
import cv2
import numpy as np
import pycuda.autoinit
import pycuda.driver as cuda
import tensorrt as trt
from PIL import Image


class HostDeviceMem(object):
    def __init__(self, host_mem, device_mem):
        self.host = host_mem
        self.device = device_mem

    def __str__(self):
        return "Host:\n" + str(self.host) + "\nDevice:\n" + str(self.device)

    def __repr__(self):
        return self.__str__()


def load_engine(trt_runtime, engine_path):
    with open(engine_path, "rb") as f:
        engine_data = f.read()
    engine = trt_runtime.deserialize_cuda_engine(engine_data)
    return engine


def allocate_buffers(engine, batch_size=1):

    inputs = []
    outputs = []
    bindings = []
    stream = cuda.Stream()

    for binding in engine:
        size = trt.volume(engine.get_binding_shape(binding)) * batch_size
        dtype = trt.nptype(engine.get_binding_dtype(binding))

        # Allocate host and device buffers
        host_mem = cuda.pagelocked_empty(size, dtype)
        device_mem = cuda.mem_alloc(host_mem.nbytes)
        # Append the device buffer to device bindings.
        bindings.append(int(device_mem))
        # Append to the appropriate list.
        if engine.binding_is_input(binding):
            inputs.append(HostDeviceMem(host_mem, device_mem))
        else:
            outputs.append(HostDeviceMem(host_mem, device_mem))
    return inputs, outputs, bindings, stream

def do_inference(context, bindings, inputs, outputs, stream, batch_size=1):
    # Transfer input data to the GPU.
    [cuda.memcpy_htod_async(inp.device, inp.host, stream) for inp in inputs]
    # Run inference.
    context.execute_async(
        batch_size=1, bindings=bindings, stream_handle=stream.handle
    )
    # Transfer predictions back from the GPU.
    [cuda.memcpy_dtoh_async(out.host, out.device, stream) for out in outputs]
    # Synchronize the stream
    stream.synchronize()
    # Return only the host outputs.
    return [out.host for out in outputs]

def preprocess_image(input_image_path):
    image_raw = cv2.imread(input_image_path)
    return image_raw

def process_image(arr, w, h):
    image = Image.fromarray(np.uint8(arr))

    image_resized = image.resize(size=(w, h), resample=Image.BILINEAR)
    img_np = np.array(image_resized)
    # HWC -> CHW
    img_np = img_np.transpose((2, 0, 1))
    # Normalize to [0.0, 1.0] interval (expected by model)
    img_np = (1.0 / 255.0) * img_np
    print(img_np.shape)
    img_np = img_np.ravel()
    return img_np


def predict(image):
    img = preprocess_image(image)
    print(img.shape)

    np.copyto(inputs[0].host, img.ravel())

    inference_start_time = time.time()

    # Fetch output from the model
    output = do_inference(
        context, bindings=bindings, inputs=inputs, outputs=outputs, stream=stream
    )

    # Output inference time
    print(output)

    # And return results
    return output


# -------------- MODEL PARAMETERS FOR DETECTNET_V2 --------------------------------
model_h = 1536
model_w = 1536

import ctypes

PLUGIN_LIBRARY = "build/libmyplugins.so"
ctypes.CDLL(PLUGIN_LIBRARY)

# TensorRT logger singleton
TRT_LOGGER = trt.Logger(trt.Logger.WARNING)
trt_engine_path = 'build/real-esrgan_f32.engine'

trt_runtime = trt.Runtime(TRT_LOGGER)
trt_engine = load_engine(trt_runtime, trt_engine_path)

# This allocates memory for network inputs/outputs on both CPU and GPU
inputs, outputs, bindings, stream = allocate_buffers(trt_engine)

# Execution context is needed for inference
context = trt_engine.create_execution_context()

output = predict('image.jpg')[0]
output = output.reshape(3072,3072,3)
print(output.shape)
print(output.dtype)

cv2.imwrite('output.jpg', output)`

But the results looks very weird.( Attached sample below ).

@wang-xinyu @yester31 Can you please help

[lazy-nurd]

@yester31 @wang-xinyu If you guys can help

[Fu0511]

Hello, I have been successful in x2 conversion. But I want to infer the engine in Python and the results engine is giving are very odd.

This is the script I am using for inferencing.

import os
import time
import cv2
import numpy as np
import pycuda.autoinit
import pycuda.driver as cuda
import tensorrt as trt
from PIL import Image


class HostDeviceMem(object):
    def __init__(self, host_mem, device_mem):
        self.host = host_mem
        self.device = device_mem

    def __str__(self):
        return "Host:\n" + str(self.host) + "\nDevice:\n" + str(self.device)

    def __repr__(self):
        return self.__str__()


def load_engine(trt_runtime, engine_path):
    with open(engine_path, "rb") as f:
        engine_data = f.read()
    engine = trt_runtime.deserialize_cuda_engine(engine_data)
    return engine


def allocate_buffers(engine, batch_size=1):

    inputs = []
    outputs = []
    bindings = []
    stream = cuda.Stream()

    for binding in engine:
        size = trt.volume(engine.get_binding_shape(binding)) * batch_size
        dtype = trt.nptype(engine.get_binding_dtype(binding))

        # Allocate host and device buffers
        host_mem = cuda.pagelocked_empty(size, dtype)
        device_mem = cuda.mem_alloc(host_mem.nbytes)
        # Append the device buffer to device bindings.
        bindings.append(int(device_mem))
        # Append to the appropriate list.
        if engine.binding_is_input(binding):
            inputs.append(HostDeviceMem(host_mem, device_mem))
        else:
            outputs.append(HostDeviceMem(host_mem, device_mem))
    return inputs, outputs, bindings, stream

def do_inference(context, bindings, inputs, outputs, stream, batch_size=1):
    # Transfer input data to the GPU.
    [cuda.memcpy_htod_async(inp.device, inp.host, stream) for inp in inputs]
    # Run inference.
    context.execute_async(
        batch_size=1, bindings=bindings, stream_handle=stream.handle
    )
    # Transfer predictions back from the GPU.
    [cuda.memcpy_dtoh_async(out.host, out.device, stream) for out in outputs]
    # Synchronize the stream
    stream.synchronize()
    # Return only the host outputs.
    return [out.host for out in outputs]

def preprocess_image(input_image_path):
    image_raw = cv2.imread(input_image_path)
    return image_raw

def process_image(arr, w, h):
    image = Image.fromarray(np.uint8(arr))

    image_resized = image.resize(size=(w, h), resample=Image.BILINEAR)
    img_np = np.array(image_resized)
    # HWC -> CHW
    img_np = img_np.transpose((2, 0, 1))
    # Normalize to [0.0, 1.0] interval (expected by model)
    img_np = (1.0 / 255.0) * img_np
    print(img_np.shape)
    img_np = img_np.ravel()
    return img_np


def predict(image):
    img = preprocess_image(image)
    print(img.shape)

    np.copyto(inputs[0].host, img.ravel())

    inference_start_time = time.time()

    # Fetch output from the model
    output = do_inference(
        context, bindings=bindings, inputs=inputs, outputs=outputs, stream=stream
    )

    # Output inference time
    print(output)

    # And return results
    return output


# -------------- MODEL PARAMETERS FOR DETECTNET_V2 --------------------------------
model_h = 1536
model_w = 1536

import ctypes

PLUGIN_LIBRARY = "build/libmyplugins.so"
ctypes.CDLL(PLUGIN_LIBRARY)

# TensorRT logger singleton
TRT_LOGGER = trt.Logger(trt.Logger.WARNING)
trt_engine_path = 'build/real-esrgan_f32.engine'

trt_runtime = trt.Runtime(TRT_LOGGER)
trt_engine = load_engine(trt_runtime, trt_engine_path)

# This allocates memory for network inputs/outputs on both CPU and GPU
inputs, outputs, bindings, stream = allocate_buffers(trt_engine)

# Execution context is needed for inference
context = trt_engine.create_execution_context()

output = predict('image.jpg')[0]
output = output.reshape(3072,3072,3)
print(output.shape)
print(output.dtype)

cv2.imwrite('output.jpg', output)`

But the results looks very weird.( Attached sample below ).

@wang-xinyu @yester31 Can you please help

I met the same problem (the results are odd). Have you solved?

[wang-xinyu]

Do we support X2 in this repo? I thought we only support this one https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth

[yester31]

It was officially made into an x4 model version. And later on request, how to use the x2 model version was explained in a previous comment https://github.com/wang-xinyu/tensorrtx/issues/1085#issuecomment-1229422132.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.