Contents
Model generalization
%matplotlib ipympl
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import pickle
from itertools import product
from functools import reduce
import colorspacious
from matplotlib.colors import LinearSegmentedColormap
from ipywidgets import Select, SelectMultiple, IntSlider, ToggleButton, Label, Layout
from IPython.display import display
from ipywidgets import HBox, VBox
plt.rcParams["legend.frameon"] = False
plt.rcParams["legend.fontsize"] = 8
plt.rcParams["legend.title_fontsize"] = 8
plt.rcParams["xtick.labelsize"] = 8
plt.rcParams["ytick.labelsize"] = 8def reduction(series, mode):
match mode:
case "mean":
return series.mean()
case "min":
return series.min()
case "max":
return series.max()
case "median":
return series.median()
case _:
return ValueError
def n_image_select():
return SelectMultiple(
options=[64, 128, 256, 512],
value=[64, 128, 256, 512],
rows=4,
# description="Number of training images",
disabled=False,
layout={"width": "100px", "height":"88px"},
)
def pretrain_lr_select():
return SelectMultiple(
options=["1e-3", "1e-4"],
value=["1e-3", "1e-4"],
rows=2,
# description="Pretrain LRs",
disabled=False,
layout={"width": "100px"},
)
def transfer_lr_select():
return SelectMultiple(
options=["1e-4", "1e-5"],
value=["1e-4", "1e-5"],
rows=2,
# description="Transfer LRs",
disabled=False,
layout={"width": "100px"},
)
def weight_freezing_select():
return SelectMultiple(
options=["none", "decoder", "encoder"],
value=["none", "decoder", "encoder"],
# description="Weight freezing",
rows=3,
disabled=False,
layout={"width": "100px"},
)
def reduction_select():
return Select(
options=["min", "mean", "median", "max"],
value="mean",
rows=4,
# description="Data Reduction",
disabled=False,
layout={"width": "100px"},
)
def transform(c, s, l):
res = colorspacious.cspace_convert(c[:3], "sRGB1", "JCh")
res[0] *= l
res[1] *= s
return colorspacious.cspace_convert(res, "JCh", "sRGB1")
def normalize_endpoints(cmin, cmax, J, C):
cmin = colorspacious.cspace_convert(cmin[:3], "sRGB1", "JCh")
cmax = colorspacious.cspace_convert(cmax[:3], "sRGB1", "JCh")
cmin[0], cmax[0] = J, J
cmin[1], cmax[1] = C, C
cmin = colorspacious.cspace_convert(cmin, "JCh", "sRGB1")
cmax = colorspacious.cspace_convert(cmax, "JCh", "sRGB1")
return cmin, cmax
def get_diverging_colormap():
dmap_endpoint = sns.diverging_palette(250, 30, l=60, s=100, center="dark", sep=1, as_cmap=True)
color_min = dmap_endpoint(0)
color_max = dmap_endpoint(255)
color_min, color_max = normalize_endpoints(color_min, color_max, 50, 70)
L = 1.68
S = 2.5e-3
N = 256
cleft = [
transform(color_min, s, l) for s, l in zip(1 - np.geomspace(S, 1, N), np.linspace(1, L, N))
]
cright = [
transform(color_max, s, l) for s, l in zip(1 - np.geomspace(1, S, N), np.linspace(L, 1, N))
]
cleft[0] = color_min
cright[-1] = color_max
dmap = LinearSegmentedColormap.from_list(
"cmap_name",
# [color_min, color_center, color_max]
np.clip(cleft + cright, 0, 1),
1024,
)
return dmapwith open("mini_df.pkl", "rb") as f:
df = pickle.load(f)def get_heatmap_data(df, reduction_mode, filter_zero):
all_defoci = np.arange(-250, 300, 50)
if filter_zero:
defoci_set = set(all_defoci).difference([0.0])
else:
defoci_set = set(all_defoci)
p_heatmap = np.zeros((5, 11))
t_heatmap = np.zeros((5, 11))
pg_heatmap = np.zeros((5, 11))
tg_heatmap = np.zeros((5, 11))
common_defoci = [-100, -50, 0, 50, 100]
all_defoci = np.arange(-250, 300, 50)
for i, starting_df in enumerate(common_defoci):
p_df = df.query(f"pretrain_defocus == {starting_df}")
pg_heatmap[i, :] = np.std(
[reduction(p_df[f"best_pretrain_val_{p}"], reduction_mode) for p in defoci_set]
)
for j, transfer_df in enumerate(all_defoci):
p_heatmap[i, j] = reduction(
p_df[f"best_pretrain_val_{transfer_df}"], reduction_mode
)
t_df = p_df.query(f"transfer_defocus=={transfer_df}")
tg_heatmap[i, j] = np.std(
[
reduction(t_df[f"best_transfer_performance_{p}"], reduction_mode)
for p in defoci_set
]
)
if len(t_df) > 0:
t_heatmap[i, j] = reduction(
t_df[f"best_transfer_performance_{transfer_df}"], reduction_mode
)
else:
t_heatmap[i, j] = p_heatmap[i, j]
tg_heatmap[i, j] = pg_heatmap[i, j]
return p_heatmap, t_heatmap, pg_heatmap, tg_heatmap
def plot_bars(
df,
tg_heatmap,
p_heatmap,
pg_heatmap,
marker_pairs,
bar_axes,
heatmap_axis,
reduction_mode,
gmax,
smax,
):
all_defoci = np.arange(-250, 300, 50)
pmap = {k: v for k, v in zip(np.arange(-100, 150, 50), range(5))}
tmap = {k: v for k, v in zip(np.arange(-250, 300, 50), range(11))}
res_markers = []
for df_pair, ax, marker in zip(marker_pairs, bar_axes, ("o", "s")):
bar_pdf, bar_tdf = df_pair
p_idx, t_idx = pmap[bar_pdf], tmap[bar_tdf]
r_df = df.query(f"pretrain_defocus == {bar_pdf} and transfer_defocus == {bar_tdf}")
tbars = [
reduction(r_df[f"best_transfer_performance_{p}"], reduction_mode) for p in all_defoci
]
c = cmap(pg_heatmap[p_idx, t_idx] / gmax)
ax.bar((all_defoci - 15) / 10, p_heatmap[p_idx, :], color=c, width=1.5, align="edge")
shift_val = tg_heatmap[p_idx, t_idx] - pg_heatmap[p_idx, t_idx]
if (abs(shift_val) / smax) < 0.2:
c = dmap(0.7) if np.sign(shift_val) == 1 else dmap(0.3)
else:
c = dmap(((shift_val) / smax) / 2 + 0.5)
ax.bar(all_defoci / 10, tbars, width=1.5, color=c, align="edge")
## Add markers to heat maps
heatmap_axis.scatter(
t_idx + 0.25, p_idx - 0.25, s=30, marker=marker, color=c, edgecolors=(0.1, 0.1, 0.1)
)
ax.text(
# -10,
# 0.09,
0.4,
0.9,
f"{df_pair[0] // 10}nm $\\rightarrow$ {df_pair[1] // 10}nm",
fontsize=8,
ha="center",
transform=ax.transAxes,
)
if df_pair[0] == df_pair[1]:
ax.text(0.4, 0.8, "Choose new pair!", fontsize=6, ha="center", color=(0.8, 0.1, 0.0), transform=ax.transAxes,)
else:
c = dmap(0.7) if np.sign(shift_val) == 1 else dmap(0.3)
style = {
'x':32.5,
'y':ax.get_ylim()[1]/2,
'rotation':90,
'ha':'left',
'va':'center',
'color':c,
}
if shift_val < 0:
# ax.text(32.5, ax.get_ylim()[1]/2, 'Improves OOD', rotation=90, ha='center', va='center', color=c)
ax.text(s=' Improves OOD', **style)
else:
ax.text(s=' Worsens OOD', **style)
res_markers.append({'s':80, 'marker':marker, 'color':c, 'edgecolors':(0.1, 0.1, 0.1)})
return res_markers
### Interactive setttings
### Visualization settings
cmap = sns.color_palette("flare_r", as_cmap=True)
dmap = get_diverging_colormap()
filter_zero = True
gmin = 0.0
gmax = 7.5e-2
smax = 7.5e-2
### Figure set up
dpi = 100
# ratio = 2.1
# height = 3.4
# width = ratio * height
width = 7.2
height = 3.4
dpi = 84
fig_style = {
"figsize": (width, height), # inches
"constrained_layout": True,
"dpi":dpi,
}
fig = plt.figure(**fig_style)
gs = fig.add_gridspec(2, 12, wspace=0.1)
fig_ptgen = fig.add_subfigure(gs[:, :6])
gen_gs = fig_ptgen.add_gridspec(2, 12)
ax_pgen = fig_ptgen.add_subplot(gen_gs[0, 0])
ax_tgen = fig_ptgen.add_subplot(gen_gs[0, 1:])
ax_gen_shift = fig_ptgen.add_subplot(gen_gs[1, 1:])
fig_bars = fig.add_subfigure(gs[:, 8:])
bars_gs = fig_bars.add_gridspec(2, 1)
ax_circ = fig_bars.add_subplot(bars_gs[0])
ax_square = fig_bars.add_subplot(bars_gs[1])
fig_cbars = fig.add_subfigure(gs[:, 6:8])
cbars_gs = fig_cbars.add_gridspec(100, 8)
ax_gen_cbar = fig_cbars.add_subplot(cbars_gs[8:38, 0:3])
ax_shift_cbar = fig_cbars.add_subplot(cbars_gs[55:85, 0:3])
ndata_select = n_image_select()
ptlr_select = pretrain_lr_select()
tllr_select = transfer_lr_select()
freeze_select = weight_freezing_select()
rmode_select = reduction_select()
filter_zero_button = ToggleButton(
value=True, description="Filter zero defocus", style={'font_size':'8pt'}, layout={'width':'125px'},
)
sq_ptdf_slider = IntSlider(
value=-5,
min=-10,
max=10,
step=5,
# description="(Square) Pretrain defocus (nm):",
disabled=False,
continuous_update=False,
orientation="horizontal",
readout=True,
readout_format="d",
layout={'width':'150px'}
)
sq_tldf_slider = IntSlider(
value=10,
min=-25,
max=25,
step=5,
# description="(Square) Transfer defocus (nm):",
disabled=False,
continuous_update=False,
orientation="horizontal",
readout=True,
readout_format="d",
layout={'width':'250px'}
)
cr_ptdf_slider = IntSlider(
value=-5,
min=-10,
max=10,
step=5,
# description=" Pretrain defocus (nm):",
disabled=False,
continuous_update=False,
orientation="horizontal",
readout=True,
readout_format="d",
layout={'width':'150px'}
)
cr_tldf_slider = IntSlider(
value=-10,
min=-25,
max=25,
step=5,
# description="Transfer defocus (nm):",
disabled=False,
continuous_update=False,
orientation="horizontal",
readout=True,
readout_format="d",
layout={'width':'250px'}
)
center_align = Layout(align_items="center") # , width='50%')
small_font = {'font_size':'8pt'}
sq_slider_box = VBox(
[
Label("Adjust square marker"),
Label("Pretrain defocus (nm)", style=small_font),
sq_ptdf_slider,
Label("Transfer defocus (nm)", style=small_font),
sq_tldf_slider,
],
layout=center_align,
)
cr_slider_box = VBox(
[
Label("Adjust circle marker"),
Label("Pretrain defocus (nm)", style=small_font),
cr_ptdf_slider,
Label("Transfer defocus (nm)", style=small_font),
cr_tldf_slider,
],
layout=center_align,
)
ptlr_box = VBox([Label("Pretrain LR"), ptlr_select], layout=center_align)
tllr_box = VBox([Label("Transfer LR"), tllr_select], layout=center_align)
ndata_box = VBox([Label("# of training images"), ndata_select], layout=center_align)
freeze_box = VBox([Label("Weight Freezing"), freeze_select], layout=center_align)
rmode_box = VBox([Label("Data reduction"), rmode_select, filter_zero_button], layout={'align_items':'center', 'width':'150px'})
left_box = VBox([ndata_box, freeze_box], layout={'align_items':'center', 'width':'150px'})
right_box = VBox(
[
ptlr_box,
tllr_box,
],
layout={'align_items':'center', 'width':'150px'},
)
menu_box = HBox([left_box, right_box, rmode_box])
slider_box = HBox([cr_slider_box, sq_slider_box])
full_box = VBox([menu_box, slider_box], )
display(full_box)
def update_plot(*args):
filter_zero = filter_zero_button.value
N_points = ndata_select.value
weight_freezing = freeze_select.value
pretrain_lrs = ptlr_select.value
transfer_lrs = tllr_select.value
reduction_mode = rmode_select.value
active_lr_pairs = tuple(product(pretrain_lrs, transfer_lrs))
marker_pairs = [ # Point and click
(cr_ptdf_slider.value*10, cr_tldf_slider.value*10),
(sq_ptdf_slider.value*10, sq_tldf_slider.value*10),
]
lr_query = reduce(
lambda x, y: f"{x} or {y}",
(f"(pretrain_lr == {x} and transfer_lr == {y})" for x, y in active_lr_pairs),
)
plot_df = df.query(
f"N_tl_training_points in {N_points} and freeze_option in {weight_freezing}"
).query(lr_query)
p_heatmap, _, pg_heatmap, tg_heatmap = get_heatmap_data(plot_df, reduction_mode, filter_zero)
### Plotting
ax_tgen.clear()
gen_cim = ax_pgen.matshow(pg_heatmap[:, 0][:, None], cmap=cmap, vmin=gmin, vmax=gmax)
ax_tgen.matshow(tg_heatmap, cmap=cmap, vmin=gmin, vmax=gmax)
shift_cim = ax_gen_shift.matshow(
tg_heatmap - pg_heatmap,
cmap=dmap,
vmin=-smax,
vmax=smax,
)
tc = fig.colorbar(gen_cim, cax=ax_gen_cbar, aspect=1)
fig.colorbar(shift_cim, cax=ax_shift_cbar)
ax_circ.clear()
ax_square.clear()
res_markers = plot_bars(
df,
tg_heatmap,
p_heatmap,
pg_heatmap,
marker_pairs,
(ax_circ, ax_square),
ax_tgen,
reduction_mode,
gmax,
smax,
)
### Formatting
for ax in [ax_tgen, ax_gen_shift]:
ax.set_xticks([1, 3, 5, 7, 9])
ax.set_xticklabels([-20, -10, 0, 10, 20], rotation=0, fontsize=8)
ax.xaxis.set_ticks_position("bottom")
for ax in [ax_pgen, ax_gen_shift]:
ax.set_yticks([0, 2, 4])
ax.set_yticklabels([-10, 0, 10], fontsize=8)
for ax, pair, marker in zip([ax_circ, ax_square], marker_pairs, res_markers):
pt_df = pair[0] / 10
if pt_df < 5:
ax.set_ylim([0, 0.1])
sy = 0.09
elif pt_df == 5:
ax.set_ylim([0, 0.2])
sy = 0.18
elif pt_df == 10:
ax.set_ylim([0, 0.3])
sy = 0.27
ax.set_xlim([-30, 30])
ax.scatter(25, sy, **marker)
ax_pgen.set_xticks([])
ax_tgen.set_yticks([])
ax_tgen.set_title("Generalization: \nAfter transfer learning", fontsize=10)
ax_gen_shift.set_title("Generalization shift", fontsize=10)
ax_pgen.set_title("After PT", fontsize=10)
fig_ptgen.supylabel("Pretrain defocus (nm)", fontsize=10)
fig_bars.supylabel("Loss", fontsize=10)
fig_bars.suptitle(" ")
ax_square.set_xlabel("Defocus (nm)", fontsize=10)
ax_gen_shift.set_xlabel("Transfer defocus (nm)", fontsize=10)
ax_gen_cbar.set_ylabel("Std. Dev. of Losses", fontsize=8, x=0.65)
ax_gen_cbar.set_title(" ")
ax_shift_cbar.set_ylabel(r"$\underset{\text{PT} \rightarrow \text{TL}}{\Delta}$Std. Dev. of Losses" + "\n" + r"$\leftarrow$ Improves", fontsize=8, x=0.65)
# ax_shift_cbar.set_ylabel("Std. Dev. of Losses", fontsize=8, x=0.65)
# ax_shift_cbar.set_title(" ")
# ax_shift_cbar.text(0, -0.1, 'Improves', fontsize=8, va='top')
# ax_shift_cbar.text(0, 0.1, 'Worsens', fontsize=8, va='bottom',)
_ = [
plt.setp(ax.spines.values(), linewidth=1.25)
for ax in (
ax_pgen,
ax_tgen,
ax_gen_shift,
ax_circ,
ax_square,
ax_gen_cbar,
ax_shift_cbar,
)
]
update_plot()
ndata_select.observe(update_plot, "value")
ptlr_select.observe(update_plot, "value")
tllr_select.observe(update_plot, "value")
rmode_select.observe(update_plot, "value")
freeze_select.observe(update_plot, "value")
filter_zero_button.observe(update_plot, "value")
sq_ptdf_slider.observe(update_plot, "value")
sq_tldf_slider.observe(update_plot, "value")
cr_ptdf_slider.observe(update_plot, "value")
cr_tldf_slider.observe(update_plot, "value")Loading...
Loading...