disco / inference.py
menghanxia's picture
fixed the randomness issue
2a10b61
raw
history blame
4.86 kB
import os, glob, sys, logging
import argparse, datetime, time
import numpy as np
import cv2
from PIL import Image
import torch
import torch.nn as nn
import torch.nn.functional as F
from models import model, basic
from utils import util
def setup_model(checkpt_path, device="cuda"):
#print('--------------', torch.cuda.is_available())
"""Load the model into memory to make running multiple predictions efficient"""
colorLabeler = basic.ColorLabel(device=device)
colorizer = model.AnchorColorProb(inChannel=1, outChannel=313, enhanced=True, colorLabeler=colorLabeler)
colorizer = colorizer.to(device)
#checkpt_path = "./checkpoints/disco-beta.pth.rar"
assert os.path.exists(checkpt_path), "No checkpoint found!"
data_dict = torch.load(checkpt_path, map_location=torch.device('cpu'))
colorizer.load_state_dict(data_dict['state_dict'])
colorizer.eval()
return colorizer, colorLabeler
def resize_ab2l(gray_img, lab_imgs, vis=False):
H, W = gray_img.shape[:2]
reszied_ab = cv2.resize(lab_imgs[:,:,1:], (W,H), interpolation=cv2.INTER_LINEAR)
if vis:
gray_img = cv2.resize(lab_imgs[:,:,:1], (W,H), interpolation=cv2.INTER_LINEAR)
return np.concatenate((gray_img[:,:,np.newaxis], reszied_ab), axis=2)
else:
return np.concatenate((gray_img, reszied_ab), axis=2)
def prepare_data(rgb_img, target_res):
rgb_img = np.array(rgb_img / 255., np.float32)
lab_img = cv2.cvtColor(rgb_img, cv2.COLOR_RGB2LAB)
org_grays = (lab_img[:,:,[0]]-50.) / 50.
lab_img = cv2.resize(lab_img, target_res, interpolation=cv2.INTER_LINEAR)
lab_img = torch.from_numpy(lab_img.transpose((2, 0, 1)))
gray_img = (lab_img[0:1,:,:]-50.) / 50.
ab_chans = lab_img[1:3,:,:] / 110.
input_grays = gray_img.unsqueeze(0)
input_colors = ab_chans.unsqueeze(0)
return input_grays, input_colors, org_grays
def colorize_grayscale(colorizer, color_class, rgb_img, hint_img, n_anchors, is_high_res, is_editable, device="cuda"):
n_anchors = int(n_anchors)
n_anchors = max(n_anchors, 3)
n_anchors = min(n_anchors, 14)
target_res = (512,512) if is_high_res else (256,256)
input_grays, input_colors, org_grays = prepare_data(rgb_img, target_res)
input_grays = input_grays.to(device)
input_colors = input_colors.to(device)
if is_editable:
print('>>>:editable mode')
sampled_T = -1
_, input_colors, _ = prepare_data(hint_img, target_res)
input_colors = input_colors.to(device)
pal_logit, ref_logit, enhanced_ab, affinity_map, spix_colors, hint_mask = colorizer(input_grays, \
input_colors, n_anchors, sampled_T)
else:
print('>>>:automatic mode')
sampled_T = 0
pal_logit, ref_logit, enhanced_ab, affinity_map, spix_colors, hint_mask = colorizer(input_grays, \
input_colors, n_anchors, sampled_T)
pred_labs = torch.cat((input_grays,enhanced_ab), dim=1)
lab_imgs = basic.tensor2array(pred_labs).squeeze(axis=0)
lab_imgs = resize_ab2l(org_grays, lab_imgs)
lab_imgs[:,:,0] = lab_imgs[:,:,0] * 50.0 + 50.0
lab_imgs[:,:,1:3] = lab_imgs[:,:,1:3] * 110.0
rgb_output = cv2.cvtColor(lab_imgs[:,:,:], cv2.COLOR_LAB2RGB)
return (rgb_output*255.0).astype(np.uint8)
def predict_anchors(colorizer, color_class, rgb_img, n_anchors, is_high_res, is_editable, device="cuda"):
n_anchors = int(n_anchors)
n_anchors = max(n_anchors, 3)
n_anchors = min(n_anchors, 14)
target_res = (512,512) if is_high_res else (256,256)
input_grays, input_colors, org_grays = prepare_data(rgb_img, target_res)
input_grays = input_grays.to(device)
input_colors = input_colors.to(device)
sampled_T, sp_size = 0, 16
pal_logit, ref_logit, enhanced_ab, affinity_map, spix_colors, hint_mask = colorizer(input_grays, \
input_colors, n_anchors, sampled_T)
pred_probs = pal_logit
guided_colors = color_class.decode_ind2ab(ref_logit, T=0)
guided_colors = basic.upfeat(guided_colors, affinity_map, sp_size, sp_size)
anchor_masks = basic.upfeat(hint_mask, affinity_map, sp_size, sp_size)
marked_labs = basic.mark_color_hints(input_grays, guided_colors, anchor_masks, base_ABs=None)
lab_imgs = basic.tensor2array(marked_labs).squeeze(axis=0)
lab_imgs = resize_ab2l(org_grays, lab_imgs, vis=True)
lab_imgs[:,:,0] = lab_imgs[:,:,0] * 50.0 + 50.0
lab_imgs[:,:,1:3] = lab_imgs[:,:,1:3] * 110.0
rgb_output = cv2.cvtColor(lab_imgs[:,:,:], cv2.COLOR_LAB2RGB)
return (rgb_output*255.0).astype(np.uint8)