removed duplicate functions from util

representativity/util.py

@@ -0,0 +1,180 @@
+import numpy as np
+import torch
+import slicegan
+from scipy import stats, ndimage
+from matplotlib import pyplot as plt
+
+import core
+
+
+def load_generator(Project_path):
+    img_size, img_channels, scale_factor = 64, 1, 1
+    z_channels = 16
+    lays = 6
+    dk, gk = [4] * lays, [4] * lays
+    ds, gs = [2] * lays, [2] * lays
+    df, gf = [img_channels, 64, 128, 256, 512, 1], [
+        z_channels,
+        512,
+        256,
+        128,
+        64,
+        img_channels,
+    ]
+    dp, gp = [1, 1, 1, 1, 0], [2, 2, 2, 2, 3]
+
+    ## Create Networks
+    netD, netG = slicegan.networks.slicegan_nets(
+        Project_path, False, "grayscale", dk, ds, df, dp, gk, gs, gf, gp
+    )
+    netG = netG()
+    netG = netG.cuda()
+    return netG
+
+
+def generate_image(netG, slice_dim=0, lf=50, threed=False, reps=50):
+
+    netG.eval()
+    imgs = []
+
+    for _ in range(reps):
+        if (_ % 50) == 0 and _ != 0:
+            print(f"generating image {_}")
+        noise = torch.randn(1, 16, lf if threed else 4, lf, lf)
+        noise.transpose_(2, slice_dim + 2)
+        noise = noise.cuda()
+        img = netG(noise, threed, slice_dim)
+        img = slicegan.util.post_proc(img)
+        img.transpose_(0, slice_dim)
+        if not threed:
+            imgs.append(img[0])
+        else:
+            imgs.append(img.cpu())
+    img = torch.stack(imgs, 0)
+    return img.float()
+
+
+def angular_img(img):
+    base_len, l = img.shape[0:2]
+    img = img.cpu().numpy()
+    plt.imshow(img[0, :100, :100])
+    plt.show()
+    img_rot = ndimage.rotate(img, base_len / l * 90, axes=(1, 0), reshape=False)
+    for i in range(img_rot.shape[0]):
+        print(f"slice {i}")
+        plt.imshow(img_rot[i, :100, :100])
+        plt.show()
+        plt.imshow(img_rot[i, -100:, -100:])
+        plt.show()
+    return img_rot
+
+
+def stat_analysis_error(img, pf, edge_lengths):  # TODO see if to delete this or not
+    img_dims = [np.array((l,) * (len(img.shape) - 1)) for l in edge_lengths]
+    err_exp = real_image_stats(img, edge_lengths, pf)
+    real_cls = core.fit_statisical_cls_from_errors(err_exp, img_dims, pf)
+    # TODO different size image 1000 vs 1500
+    return real_cls
+
+
+def real_image_stats(img, ls, pf, repeats=4000, conf=0.95):
+    """Calculates the error of the stat. analysis for different edge lengths.
+    The error is calculated by the std of the mean of the subimages divided by the pf.
+    params:
+    img: the image to calculate the error for (Should be a stack of images).
+    ls: the edge lengths to calculate the error for.
+    pf: the phase fraction of the image.
+    repeats: the number of repeats for each edge length.
+    conf: the confidence level for the error."""
+    dims = len(img[0].shape)
+    errs = []
+    for l in ls:
+        pfs = []
+        n_pos_ims = int(np.prod(img.shape) / l**dims)
+        repeats = n_pos_ims * 2
+        # print(f'one im repeats = {repeats} for l = {l}')
+        if dims == 1:
+            for _ in range(repeats):
+                bm, xm = img.shape
+                x = torch.randint(0, xm - l, (1,))
+                b = torch.randint(0, bm, (1,))
+                crop = img[b, x : x + l]
+                pfs.append(torch.mean(crop).cpu())
+        elif dims == 2:
+            for _ in range(repeats):
+                bm, xm, ym = img.shape
+                x = torch.randint(0, xm - l, (1,))
+                y = torch.randint(0, ym - l, (1,))
+                b = torch.randint(0, bm, (1,))
+                crop = img[b, x : x + l, y : y + l]
+                pfs.append(torch.mean(crop).cpu())
+        else:  # 3D
+            for _ in range(repeats):
+                bm, xm, ym, zm = img.shape
+                x = torch.randint(0, xm - l, (1,))
+                y = torch.randint(0, ym - l, (1,))
+                z = torch.randint(0, zm - l, (1,))
+                b = torch.randint(0, bm, (1,))
+                crop = img[b, x : x + l, y : y + l, z : z + l]
+                pfs.append(torch.mean(crop).cpu())
+        pfs = np.array(pfs)
+        ddof = 1  # for unbiased std
+        std = np.std(pfs, ddof=ddof)
+        errs.append(100 * (stats.norm.interval(conf, scale=std)[1] / pf))
+    return errs
+
+
+def bernouli_from_cls(cls, pf, img_size, conf=0.95):
+    ns = core.n_samples_from_dims([np.array(img_size)], cls)
+    return core.bernouli(pf, ns, conf)
+
+
+# fit_cls now fit_statistical_cls_from_errors
+
+# ns_from_dims now n_samples_from_dims
+
+# test_cls_set now test_all_cls_in_range
+
+
+def tpc_fit(x, a, b, c):
+    return a * np.e ** (-b * x) + c
+
+
+def percentage_error(y_true, y_pred):
+    return (y_true - y_pred) / y_true
+
+
+def mape(y_true, y_pred):  # mean absolute percentage error
+    return np.mean(np.abs(percentage_error(y_true, y_pred)))
+
+
+def mape_linear_objective(params, y_pred, y_true):
+    y_pred_new = linear_fit(y_pred, *params)
+    return mape(y_true, y_pred_new)
+
+
+def linear_fit(x, m, b):
+    return m * x + b
+
+
+def optimize_error_conf_pred(bern_conf, total_conf, std_bern, std_model, pf):
+    model_conf = total_conf / bern_conf
+    err_bern = stats.norm.interval(bern_conf, scale=std_bern)[1]
+    one_side_error_model = model_conf * 2 - 1
+    err_model = stats.norm.interval(one_side_error_model, scale=std_model)[1]
+    return err_bern * (1 + err_model)
+
+
+def optimize_error_n_pred(bern_conf, total_conf, std_model, pf, err_targ):
+    model_conf = total_conf / bern_conf
+    z1 = stats.norm.interval(bern_conf)[1]
+    one_side_error_model = model_conf * 2 - 1
+    err_model = stats.norm.interval(one_side_error_model, scale=std_model)[1]
+    num = (err_model + 1) ** 2 * (1 - pf) * z1**2 * pf
+    den = (err_targ) ** 2  # TODO go over the calcs and see if this is right
+    return num / den
+
+
+# renamed calc_autocorrelation_orthant -> autocorrelation_orthant
+
+# renamed one_img_stat_analysis_error -> stat_analysis_error_classic