|
- import argparse
- import math
- import os
- import struct
- import sys
- import time
- import glob
-
- import numpy as np
- import torch
- import torch.nn.functional as F
- from PIL import Image
-
- # import Util.AE as AE
- import AE
- import Model.model as model
- from Model.context_model import Weighted_Gaussian
- from Util.metrics import evaluate
- from Util import torch_msssim
- from Util.block_metric import check_RD_GEO # blcok based metric
- from Util.config import dict
- from Util.generate_substitute import SubstituteGenerator
-
- os.environ['LRU_CACHE_CAPACITY'] = '1'
- ############################## Load Configuration Parameters ########################
- GPU = dict['GPU']
- SAVE_REC = dict['SAVE_REC']
- USE_GEO = dict['USE_GEO']
- USE_VR_MODEL = dict['USE_VR_MODEL']
- USE_PREPROCESSING = dict['USE_PREPROCESSING']
- if USE_PREPROCESSING:
- num_steps = dict['num_steps']
- block_width = dict['CTU_size']
- block_height = dict['CTU_size']
-
- # index - [0-15]
- if USE_VR_MODEL:
- models = ["mse_VR_low", "mse_VR_high", "msssim_VR_low", "msssim_VR_high"]
- max_lambdas = [64, 256, 1.28, 6.40]
- else:
- models = ["mse200", "mse400", "mse800", "mse1600", "mse3200", "mse6400", "mse12800", "mse25600",
- "msssim4", "msssim8", "msssim16", "msssim32", "msssim64", "msssim128", "msssim320", "msssim640"]
- #model_dir = "/model/ljp105/NIC_v01_baseline/"
- #VR_model_dir = "/model/ljp105/NIC_v02_VR_models"
- # @torch.no_grad()
- def inference_rd(im_dirs, out_dir, model_dir, model_index, lambda_rd_ori):
- if USE_VR_MODEL:
- lambda_rd_max = max_lambdas[model_index]
- if lambda_rd_ori > 1.2 * lambda_rd_max:
- lambda_rd_ori = 1.2 * lambda_rd_max
- lambda_rd_nom = lambda_rd_ori / lambda_rd_max
- lambda_rd_nom_scaled = int(lambda_rd_nom / 1.2 * pow(2, 16))
- lambda_rd_nom_used = lambda_rd_nom_scaled / pow(2, 16) * 1.2
- lambda_rd_numpy = np.zeros((1, 1), np.float32)
- lambda_rd_numpy[0, 0] = lambda_rd_nom_used
- lambda_rd = torch.Tensor(lambda_rd_numpy)
- M, N2 = 192, 128
- if (model_index == 1) or (model_index == 3):
- M, N2 = 256, 192
- image_comp = model.Image_coding(3, M, N2, M, M // 2)
- context = Weighted_Gaussian(M)
-
- log_name = os.path.join(out_dir, models[model_index]+'_lmbda'+str(int(lambda_rd_ori*100)) + '_test_' + str(block_width) + '_RD.txt')
- else:
- M, N2 = 192, 128
- if (model_index == 6) or (model_index == 7) or (model_index == 14) or (model_index == 15):
- M, N2 = 256, 192
- image_comp = model.Image_coding(3, M, N2, M, M//2)
- context = Weighted_Gaussian(M)
- lambda_rd = None
- log_name = os.path.join(out_dir,models[model_index]+'_test_'+str(block_width)+'_RD.txt')
-
- if USE_PREPROCESSING:
- lmbda_list = [200, 400, 800, 1600, 3200, 6400, 12800, 25600, 4, 8, 16, 32, 64, 128, 320, 640]
- stepsize_list = [150, 75, 30, 10, 10, 5, 3, 1, 100, 10, 7, 5, 1, 1, 1, 0.3]
-
- if USE_VR_MODEL:
- lmbda = lambda_rd_ori * 100
- step_size= stepsize_list[lmbda_list.index(lmbda)]
- reconstruction_metric = 'mse' if model_index <= 1 else 'msssim'
- else:
- step_size = stepsize_list[model_index]
- lmbda = lmbda_list[model_index]
- reconstruction_metric = 'mse' if model_index <= 7 else 'msssim'
-
- substitute_generator = SubstituteGenerator(model=image_comp, context_model=context, llambda=lmbda,
- num_steps=num_steps, step_size=step_size,
- reconstruct_metric=reconstruction_metric,
- )
- ######################### Load Model #########################
- image_comp.load_state_dict(torch.load(
- os.path.join(model_dir, models[model_index] + r'.pkl'), map_location='cpu'))
- context.load_state_dict(torch.load(
- os.path.join(model_dir, models[model_index] + r'p.pkl'), map_location='cpu'))
- if GPU:
- image_comp.cuda()
- context.cuda()
- #################### Compress Each Image ###################
- bpp_list = []
- rgb_psnr_list = []
- rgb_msssim_list = []
- for im_dir in im_dirs:
- dec_time = 0
- enc_dec_time_start = time.time()
- bin_dir = os.path.join(out_dir,'enc.bin')
- # bin_dir = os.path.join(out_dir, im_dir.split('/')[-1].replace('.png','enc.bin'))
- rec_dir = os.path.join(out_dir, im_dir.split('/')[-1].replace('.png','dec.png'))
- file_object = open(bin_dir, 'wb')
- ######################### Read Image #########################
- img = Image.open(im_dir)
- source_img = np.array(img)
- img = source_img/255.0
- H, W, _ = img.shape
- print(img.shape)
- num_pixels = H * W
- C = 3
- Head = struct.pack('2HB2?H', H, W, model_index, USE_GEO, USE_VR_MODEL, block_width)
- file_object.write(Head)
- if USE_VR_MODEL:
- Head_lmbda = struct.pack('H', lambda_rd_nom_scaled)
- file_object.write(Head_lmbda)
- out_img = np.zeros([H, W, C]) # recon image
- H_offset = 0
- W_offset = 0
- ######################### Spliting Image #########################
- Block_Num_in_Width = int(np.ceil(W / block_width))
- Block_Num_in_Height = int(np.ceil(H / block_height))
- img_block_list = []
- for i in range(Block_Num_in_Height):
- for j in range(Block_Num_in_Width):
- img_block_list.append(img[i * block_height:np.minimum((i + 1) * block_height, H),j * block_width:np.minimum((j + 1) * block_width,W),...])
-
- ######################### Padding Image #########################
- Block_Idx = 0
- for img in img_block_list: # Traverse CTUs
- block_H = img.shape[0]
- block_W = img.shape[1]
- tile = 64.
- block_H_PAD = int(tile * np.ceil(block_H / tile))
- block_W_PAD = int(tile * np.ceil(block_W / tile))
- im = np.zeros([block_H_PAD, block_W_PAD, 3], dtype='float32')
- im[:block_H, :block_W, :] = img[:, :, :3]
- im = torch.FloatTensor(im)
- im = im.permute(2, 0, 1).contiguous()
- im = im.view(1, C, block_H_PAD, block_W_PAD)
- if GPU:
- im = im.cuda()
- if USE_VR_MODEL:
- lambda_rd = lambda_rd.cuda()
- print('====> Encoding Image:', im_dir, "%dx%d" % (block_H, block_W), 'to', out_dir, " Block Idx: %d" % (Block_Idx))
- Block_Idx += 1
- # begin processing CTU
- im_block_list = []
- im_block_loc_list = []
- im_block_list.append(im) # list size = 1
- im_block_loc_list.append([0,0,block_H_PAD,block_W_PAD])
-
- for im_block_loc, im_block in zip(im_block_loc_list, im_block_list): # Traverse CUs
- ############################ Geometric Flip and rotate ########################
- if USE_GEO:
- _, _, geo_index, _ = check_RD_GEO(im_block, lambda_rd, image_comp, context, model_index)
- i_rot =int(geo_index % 4)
- if geo_index < 4:
- im_block = torch.rot90(im_block, k=i_rot, dims=[2, 3])
- else:
- im_block = torch.rot90(torch.flip(im_block, dims=[2]), k=i_rot, dims=[2, 3])
-
- ############################ Preprocessing to find a substitute ########################
- if USE_PREPROCESSING:
- if USE_VR_MODEL:
- im_block = substitute_generator.perturb(orig_image=im_block, lambda_rd=lambda_rd)
- else:
- im_block = substitute_generator.perturb(orig_image=im_block)
-
- with torch.no_grad():
- y_main, y_hyper = image_comp.encoder(im_block, lambda_rd)
- y_main_q= torch.round(y_main)
- y_main_q = torch.Tensor(y_main_q.cpu().numpy().astype(np.int))
- if GPU:
- y_main_q = y_main_q.cuda()
- dec_time_start = time.time()
- rec_img_block = image_comp.decoder(y_main_q, lambda_rd)
- ############################ Reverse Geometric Flip and Rotate ########################
- if USE_GEO:
- if geo_index < 4:
- rec_img_block = torch.rot90(rec_img_block,k=4-i_rot, dims=[2, 3])
- else:
- rec_img_block = torch.flip(torch.rot90(rec_img_block,k=4-i_rot, dims=[2, 3]), dims=[2])
-
- ################################### Reconstruct Image #######################################
- output_ = torch.clamp(rec_img_block, min=0., max=1.0)
- out = output_.data[0].cpu().numpy()
- out = out.transpose(1, 2, 0)
-
- out_img[H_offset : H_offset + block_H, W_offset : W_offset + block_W, :] = out[:block_H, :block_W, :]
- dec_time += (time.time()-dec_time_start)
-
- y_hyper_q, xp2 = image_comp.factorized_entropy_func(y_hyper, 2)
- y_hyper_q = torch.Tensor(y_hyper_q.cpu().numpy().astype(np.int))
- if GPU:
- y_hyper_q = y_hyper_q.cuda()
-
- hyper_dec = image_comp.p(image_comp.hyper_dec(y_hyper_q))
- xp3, params_prob = context(y_main_q, hyper_dec)
-
- # Main Arith Encode
- Datas = torch.reshape(y_main_q, [-1]).cpu().numpy().astype(np.int).tolist()
- Max_Main = max(Datas)
- Min_Main = min(Datas)
- sample = np.arange(Min_Main, Max_Main+1+1) # [Min_V - 0.5 , Max_V + 0.5]
- _, c, h, w = y_main_q.shape
- print("Main Channel:", c)
- sample = torch.FloatTensor(np.tile(sample, [1, c, h, w, 1]))
- if GPU:
- sample = sample.cuda()
-
- # 3 gaussian
- prob0, mean0, scale0, prob1, mean1, scale1, prob2, mean2, scale2 = [
- torch.chunk(params_prob, 9, dim=1)[i].squeeze(1) for i in range(9)]
- del params_prob
- # keep the weight summation of prob == 1
- probs = torch.stack([prob0, prob1, prob2], dim=-1)
- del prob0, prob1, prob2
-
- probs = F.softmax(probs, dim=-1)
- # process the scale value to positive non-zero
- scale0 = torch.abs(scale0)
- scale1 = torch.abs(scale1)
- scale2 = torch.abs(scale2)
- scale0[scale0 < 1e-6] = 1e-6
- scale1[scale1 < 1e-6] = 1e-6
- scale2[scale2 < 1e-6] = 1e-6
-
- m0 = torch.distributions.normal.Normal(mean0, scale0)
- m1 = torch.distributions.normal.Normal(mean1, scale1)
- m2 = torch.distributions.normal.Normal(mean2, scale2)
- lower = torch.zeros(1, c, h, w, Max_Main-Min_Main+2)
- for i in range(sample.shape[4]):
- # print("CDF:", i)
- lower0 = m0.cdf(sample[:, :, :, :, i]-0.5)
- lower1 = m1.cdf(sample[:, :, :, :, i]-0.5)
- lower2 = m2.cdf(sample[:, :, :, :, i]-0.5)
- lower[:, :, :, :, i] = probs[:, :, :, :, 0]*lower0 + \
- probs[:, :, :, :, 1]*lower1+probs[:, :, :, :, 2]*lower2
- del probs, lower0, lower1, lower2
-
- precise = 16
- cdf_m = lower.data.cpu().numpy()*((1 << precise) - (Max_Main -
- Min_Main + 1)) # [1, c, h, w ,Max-Min+1]
- cdf_m = cdf_m.astype(np.int32) + sample.cpu().numpy().astype(np.int32) - Min_Main
- cdf_main = np.reshape(cdf_m, [len(Datas), -1])
-
- # Cdf[Datas - Min_V]
- Cdf_lower = list(map(lambda x, y: int(y[x - Min_Main]), Datas, cdf_main))
- # Cdf[Datas + 1 - Min_V]
- Cdf_upper = list(map(lambda x, y: int(
- y[x - Min_Main]), Datas, cdf_main[:, 1:]))
- AE.encode_cdf(Cdf_lower, Cdf_upper, "main.bin")
- FileSizeMain = os.path.getsize("main.bin")
- #TotalFileSizeMain += FileSizeMain
- #print("main.bin: %d bytes" % (FileSizeMain))
-
- # Hyper Arith Encode
- Min_V_HYPER = torch.min(y_hyper_q).cpu().numpy().astype(np.int).tolist()
- Max_V_HYPER = torch.max(y_hyper_q).cpu().numpy().astype(np.int).tolist()
- _, c, h, w = y_hyper_q.shape
- # print("Hyper Channel:", c)
- Datas_hyper = torch.reshape(
- y_hyper_q, [c, -1]).cpu().numpy().astype(np.int).tolist()
- # [Min_V - 0.5 , Max_V + 0.5]
- sample = np.arange(Min_V_HYPER, Max_V_HYPER+1+1)
- sample = np.tile(sample, [c, 1, 1])
- sample_tensor = torch.FloatTensor(sample)
- if GPU:
- sample_tensor = sample_tensor.cuda()
- lower = torch.sigmoid(image_comp.factorized_entropy_func._logits_cumulative(
- sample_tensor - 0.5, stop_gradient=False))
- cdf_h = lower.data.cpu().numpy()*((1 << precise) - (Max_V_HYPER -
- Min_V_HYPER + 1)) # [N1, 1, Max-Min+1]
- cdf_h = cdf_h.astype(np.int) + sample.astype(np.int) - Min_V_HYPER
- cdf_hyper = np.reshape(np.tile(cdf_h, [len(Datas_hyper[0]), 1, 1, 1]), [
- len(Datas_hyper[0]), c, -1])
-
- # Datas_hyper [256, N], cdf_hyper [256,1,X]
- Cdf_0, Cdf_1 = [], []
- for i in range(c):
- Cdf_0.extend(list(map(lambda x, y: int(
- y[x - Min_V_HYPER]), Datas_hyper[i], cdf_hyper[:, i, :]))) # Cdf[Datas - Min_V]
- Cdf_1.extend(list(map(lambda x, y: int(
- y[x - Min_V_HYPER]), Datas_hyper[i], cdf_hyper[:, i, 1:]))) # Cdf[Datas + 1 - Min_V]
- AE.encode_cdf(Cdf_0, Cdf_1, "hyper.bin")
- FileSizeHyper = os.path.getsize("hyper.bin")
- #TotalFileSizeHyper += FileSizeHyper
- #print("hyper.bin: %d bytes" % (FileSizeHyper))
- if USE_GEO:
- Head_block = struct.pack('4h2IB', Min_Main, Max_Main, Min_V_HYPER, Max_V_HYPER,
- FileSizeMain, FileSizeHyper, geo_index)
- else:
- Head_block = struct.pack('4h2I', Min_Main, Max_Main, Min_V_HYPER, Max_V_HYPER,
- FileSizeMain, FileSizeHyper)
- file_object.write(Head_block) # CU information
- # cat Head_Infor and 2 files together
- # Head = [FileSizeMain,FileSizeHyper,H,W,Min_Main,Max_Main,Min_V_HYPER,Max_V_HYPER,model_index]
- # print("Head Info:",Head)
- with open("main.bin", 'rb') as f:
- bits = f.read()
- file_object.write(bits)
- with open("hyper.bin", 'rb') as f:
- bits = f.read()
- file_object.write(bits)
- del im, im_block_list, im_block_loc_list
- W_offset += block_W
- if W_offset >= W:
- W_offset = 0
- H_offset += block_H
- file_object.close()
-
- out_img = np.round(out_img * 255.0)
- out_img = out_img.astype('uint8')
- out_img = out_img[:H, :W, :]
- if SAVE_REC:
- img = Image.fromarray(out_img)
- img.save(rec_dir)
- # calculate bpp, psnr, msssim
- with open(bin_dir, "rb") as f:
- bpp = len(f.read())*8./num_pixels
- f.close()
- [rgb_psnr, rgb_msssim, yuv_psnr,y_msssim]=evaluate(source_img,out_img)
- bpp_list.append(bpp)
- rgb_psnr_list.append(rgb_psnr)
- rgb_msssim_list.append(rgb_msssim)
-
- class_name = im_dir.split('/')[-2]
- image_name = im_dir.split('/')[-1].replace('.png','')
- enc_dec_time = time.time() - enc_dec_time_start
- enc_time = enc_dec_time - dec_time
-
- print(class_name+'/'+image_name+'\t'+str(bpp)+'\t'+str(rgb_psnr)+'\t'+str(rgb_msssim)+'\t'+str(-10*np.log10(1-rgb_msssim))+
- '\t'+str(yuv_psnr)+'\t'+str(y_msssim)+'\t'+str(-10*np.log10(1-y_msssim))+
- '\t'+str(enc_time)+'\t'+str(dec_time)+'\n')
-
- with open(log_name, 'a') as f:
- f.write(class_name+'/'+image_name+'\t'+str(bpp)+'\t'+str(rgb_psnr)+'\t'+str(rgb_msssim)+'\t'+str(-10*np.log10(1-rgb_msssim))+
- '\t'+str(yuv_psnr)+'\t'+str(y_msssim)+'\t'+str(-10*np.log10(1-y_msssim))+
- '\t'+str(enc_time)+'\t'+str(dec_time)+'\n')
- f.close()
- del out_img
- mean_bpp = np.mean(bpp_list)
- mean_psnr = np.mean(rgb_psnr_list)
- mean_msssim = np.mean(rgb_msssim_list)
- with open(log_name, 'a') as f:
- f.write("mean of bpp:" + str(mean_bpp) + '\n' + "mean of psnr:" + str(mean_psnr) +'\n' + "mean of msssim:" + str(mean_msssim))
- f.close()
- print("mean of bpp:", mean_bpp)
- print("mean of psnr:", mean_psnr)
- print("mean of msssim:", mean_msssim)
-
- if __name__ == '__main__':
- parser = argparse.ArgumentParser()
- parser.add_argument("-i", "--input", type=str, required=True, help="Input Image")
- parser.add_argument("-o", "--output", type=str, required=True, help="Output Bin(encode)/Image(decode)")
- parser.add_argument("-m_dir", "--model_dir", type=str, required=True, help="Directory containing trained models")
- parser.add_argument("-m", "--model", type=int, default=0, help="Model Index [0-9]")
- parser.add_argument("--lambda_rd", type=float, default=16, help="Input lambda for variable-rate models")
- # parser.add_argument('--encode', dest='coder_flag', action='store_true')
- # parser.add_argument('--decode', dest='coder_flag', action='store_false')
- # parser.add_argument("--block_width", type=int, default=2048, help="coding block width")
- # parser.add_argument("--block_height", type=int, default=1024, help="coding block height")
- parser.add_argument("--number", type=int, default=24, help="Number of Coding Images")
- args = parser.parse_args()
-
- num = 0
- test_image_paths = []
- dirs = os.listdir(args.input)
- for dir in dirs:
- path = os.path.join(args.input, dir)
- if os.path.isdir(path):
- test_image_paths += glob.glob(path + '/*.png')[0:args.number]
- if os.path.isfile(path):
- if num < args.number:
- test_image_paths.append(path)
- num += 1
-
- inference_rd(test_image_paths, args.output, args.model_dir, args.model, args.lambda_rd)
|