diff --git a/confic.py b/confic.py
index 634bd5e..3016472 100644
--- a/confic.py
+++ b/confic.py
@@ -1,7 +1,7 @@
import torch
import pathlib
-BATCH_SIZE = 4
+BATCH_SIZE = 32
RESIZE_TO = 416
NUM_EPOCHS = 20
NUM_WORKERS = 4
diff --git a/data/generate_dataset.py b/data/generate_dataset.py
index 8b59eff..e3e0e91 100644
--- a/data/generate_dataset.py
+++ b/data/generate_dataset.py
@@ -114,10 +114,30 @@ def bboxes_from_file(times_v, fish_freq, rise_idx, rise_size, fish_baseline_freq
dt_bbox = right_time_bound - left_time_bound
df_bbox = upper_freq_bound - lower_freq_bound
- left_time_bound -= dt_bbox * 0.1
- right_time_bound += dt_bbox * 0.1
- lower_freq_bound -= df_bbox * 0.1
- upper_freq_bound += df_bbox * 0.1
+
+ # embed()
+ # quit()
+ # left_time_bound -= dt_bbox + 0.01 * (t1 - t0)
+ # right_time_bound += dt_bbox + 0.01 * (t1 - t0)
+ # lower_freq_bound -= df_bbox + 0.01 * (f1 - f0)
+ # upper_freq_bound += df_bbox + 0.01 * (f1 - f0)
+
+ left_time_bound -= 0.01 * (t1 - t0)
+ right_time_bound += 0.05 * (t1 - t0)
+ lower_freq_bound -= 0.01 * (f1 - f0)
+ upper_freq_bound += 0.05 * (f1 - f0)
+
+ # embed()
+ # quit()
+ mask2 = ((left_time_bound >= t0) &
+ (right_time_bound <= t1) &
+ (lower_freq_bound >= f0) &
+ (upper_freq_bound <= f1)
+ )
+ left_time_bound = left_time_bound[mask2]
+ right_time_bound = right_time_bound[mask2]
+ lower_freq_bound = lower_freq_bound[mask2]
+ upper_freq_bound = upper_freq_bound[mask2]
x0 = np.array((left_time_bound - t0) / (t1 - t0) * width, dtype=int)
x1 = np.array((right_time_bound - t0) / (t1 - t0) * width, dtype=int)
@@ -129,7 +149,7 @@ def bboxes_from_file(times_v, fish_freq, rise_idx, rise_size, fish_baseline_freq
right_time_bound,
lower_freq_bound,
upper_freq_bound,
- x0, x1, y0, y1])
+ x0, y0, x1, y1])
# test_s = ['a', 'a', 'a', 'a']
tmp_df = pd.DataFrame(
# index= [pic_save_str for i in range(len(left_time_bound))],
@@ -142,15 +162,53 @@ def bboxes_from_file(times_v, fish_freq, rise_idx, rise_size, fish_baseline_freq
return bbox_df
def main(args):
+ def development_fn():
+ fig_title = (f'{Path(args.folder).name}__{t0:.0f}s-{t1:.0f}s__{f0:4.0f}-{f1:4.0f}Hz').replace(' ', '0')
+ fig = plt.figure(figsize=(7, 7), num=fig_title)
+ gs = gridspec.GridSpec(1, 2, width_ratios=(8, 1), wspace=0, left=0.1, bottom=0.1, right=0.9,
+ top=0.95) # , bottom=0, left=0, right=1, top=1
+ ax = fig.add_subplot(gs[0, 0])
+ cax = fig.add_subplot(gs[0, 1])
+ im = ax.imshow(s_trans.squeeze(), cmap='gray', aspect='auto', origin='lower',
+ extent=(times[t_idx0], times[t_idx1] + t_res, freq[f_idx0], freq[f_idx1] + f_res))
+ fig.colorbar(im, cax=cax, orientation='vertical')
+
+
+ cols = ['image', 't0', 't1', 'f0', 'f1', 'x0', 'y0', 'x1', 'y1']
+ dev_df = pd.DataFrame(columns=cols)
+
+ dev_df = bboxes_from_file(times_v, fish_freq, rise_idx, rise_size, fish_baseline_freq_time, fish_baseline_freq,
+ fig_title, dev_df, cols, (7*256), (7*256), t0, t1, f0, f1)
+
+ # embed()
+ # quit()
+ time_freq_bbox = torch.as_tensor(dev_df.loc[:, ['t0', 'f0', 't1', 'f1']].values.astype(np.float32))
+
+ for bbox in time_freq_bbox:
+ Ct0, Cf0, Ct1, Cf1 = bbox
+ ax.add_patch(
+ Rectangle((Ct0, Cf0), Ct1-Ct0, Cf1-Cf0, fill=False, color="white", linewidth=2, zorder=10)
+ )
+ # for enu in range(len(left_time_bound)):
+ # if np.isnan(right_time_bound[enu]):
+ # continue
+ # ax.add_patch(
+ # Rectangle((left_time_bound[enu], lower_freq_bound[enu]),
+ # (right_time_bound[enu] - left_time_bound[enu]),
+ # (upper_freq_bound[enu] - lower_freq_bound[enu]),
+ # fill=False, color="white", linewidth=2, zorder=10)
+ # )
+ plt.show()
+
min_freq = 200
max_freq = 1500
d_freq = 200
- freq_overlap = 50
- d_time = 60*15
- time_overlap = 60*5
+ freq_overlap = 25
+ d_time = 60*10
+ time_overlap = 60*1
if not os.path.exists(os.path.join('train', 'bbox_dataset.csv')):
- cols = ['image', 't0', 't1', 'f0', 'f1', 'x0', 'x1', 'y0', 'y1']
+ cols = ['image', 't0', 't1', 'f0', 'f1', 'x0', 'y0', 'x1', 'y1']
bbox_df = pd.DataFrame(columns=cols)
else:
@@ -161,10 +219,15 @@ def main(args):
for f in pd.unique(bbox_df['image']):
eval_files.append(f.split('__')[0])
- folders = [args.folder]
+ folders = list(f.parent for f in Path(args.folder).rglob('fill_times.npy'))
+
+ # embed()
+ # quit()
for enu, folder in enumerate(folders):
print(f'DataSet generation from {folder} | {enu+1}/{len(folders)}')
+ if not (folder/'analysis'/'rise_idx.npy').exists():
+ continue
freq, times, spec, EODf_v, ident_v, idx_v, times_v, fish_freq, rise_idx, rise_size, fish_baseline_freq, fish_baseline_freq_time = (
load_data(folder))
@@ -174,14 +237,15 @@ def main(args):
np.arange(0, times[-1], d_time),
np.arange(min_freq, max_freq, d_freq)
),
- total=int(((max_freq-min_freq)//d_freq) * (times[-1] // d_time))
+ total=int((((max_freq-min_freq)//d_freq)+1) * ((times[-1] // d_time)+1))
)
for t0, f0 in pic_base:
+
t1 = t0 + d_time + time_overlap
f1 = f0 + d_freq + freq_overlap
- present_freqs = EODf_v[(~np.isnan(ident_v)) &
+ present_freqs = EODf_v[(~np.isnan(ident_v)) &
(t0 <= times_v[idx_v]) &
(times_v[idx_v] <= t1) &
(EODf_v >= f0) &
@@ -204,73 +268,10 @@ def main(args):
fish_baseline_freq_time, fish_baseline_freq,
pic_save_str, bbox_df, cols, width, height, t0, t1, f0, f1)
else:
- fig_title = (f'{Path(args.folder).name}__{t0:.0f}s-{t1:.0f}s__{f0:4.0f}-{f1:4.0f}Hz').replace(' ', '0')
- fig = plt.figure(figsize=(10, 7), num=fig_title)
- gs = gridspec.GridSpec(1, 2, width_ratios=(8, 1), wspace=0, left=0.1, bottom=0.1, right=0.9, top=0.95) # , bottom=0, left=0, right=1, top=1
- ax = fig.add_subplot(gs[0, 0])
- cax = fig.add_subplot(gs[0, 1])
- im = ax.imshow(s_trans.squeeze(), cmap='gray', aspect='auto', origin='lower',
- extent=(times[t_idx0], times[t_idx1] + t_res, freq[f_idx0], freq[f_idx1] + f_res))
- fig.colorbar(im, cax=cax, orientation='vertical')
-
-
- times_v_idx0, times_v_idx1 = np.argmin(np.abs(times_v - t0)), np.argmin(np.abs(times_v - t1))
- for id_idx in range(len(fish_freq)):
- ax.plot(times_v[times_v_idx0:times_v_idx1], fish_freq[id_idx][times_v_idx0:times_v_idx1], marker='.', color='k', markersize=4)
- rise_idx_oi = np.array(rise_idx[id_idx][
- (rise_idx[id_idx] >= times_v_idx0) &
- (rise_idx[id_idx] <= times_v_idx1) &
- (rise_size[id_idx] >= 10)], dtype=int)
- rise_size_oi = rise_size[id_idx][(rise_idx[id_idx] >= times_v_idx0) &
- (rise_idx[id_idx] <= times_v_idx1) &
- (rise_size[id_idx] >= 10)]
-
- ax.plot(times_v[rise_idx_oi], fish_freq[id_idx][rise_idx_oi], 'o', color='tab:red')
-
- if len(rise_idx_oi) > 0:
- closest_baseline_idx = list(map(lambda x: np.argmin(np.abs(fish_baseline_freq_time - x)), times_v[rise_idx_oi]))
- closest_baseline_freq = fish_baseline_freq[id_idx][closest_baseline_idx]
-
- upper_freq_bound = closest_baseline_freq + rise_size_oi
- lower_freq_bound = closest_baseline_freq
-
- left_time_bound = times_v[rise_idx_oi]
- right_time_bound = np.zeros_like(left_time_bound)
-
- for enu, Ct_oi in enumerate(times_v[rise_idx_oi]):
- Crise_size = rise_size_oi[enu]
- Cblf = closest_baseline_freq[enu]
-
- rise_end_t = times_v[(times_v > Ct_oi) & (fish_freq[id_idx] < Cblf + Crise_size * 0.37)]
- if len(rise_end_t) == 0:
- right_time_bound[enu] = np.nan
- else:
- right_time_bound[enu] = rise_end_t[0]
-
- dt_bbox = right_time_bound - left_time_bound
- df_bbox = upper_freq_bound - lower_freq_bound
- left_time_bound -= dt_bbox*0.1
- right_time_bound += dt_bbox*0.1
- lower_freq_bound -= df_bbox*0.1
- upper_freq_bound += df_bbox*0.1
-
- print(f'f0: {lower_freq_bound}')
- print(f'f1: {upper_freq_bound}')
- print(f't0: {left_time_bound}')
- print(f't1: {right_time_bound}')
-
- for enu in range(len(left_time_bound)):
- if np.isnan(right_time_bound[enu]):
- continue
- ax.add_patch(
- Rectangle((left_time_bound[enu], lower_freq_bound[enu]),
- (right_time_bound[enu] - left_time_bound[enu]),
- (upper_freq_bound[enu] - lower_freq_bound[enu]),
- fill=False, color="white", linewidth=2, zorder=10)
- )
- plt.show()
+ development_fn()
if not args.dev:
+ print('save')
bbox_df.to_csv(os.path.join('train', 'bbox_dataset.csv'), columns=cols, sep=',')
if __name__ == '__main__':
diff --git a/datasets.py b/datasets.py
index a9b8617..54e992c 100644
--- a/datasets.py
+++ b/datasets.py
@@ -102,8 +102,8 @@ if __name__ == '__main__':
for s, t in zip(samples, targets):
fig, ax = plt.subplots()
ax.imshow(s.permute(1, 2, 0), aspect='auto')
- for (x0, x1, y0, y1), l in zip(t['boxes'], t['labels']):
- print(x0, x1, y0, y1, l)
+ for (x0, y0, x1, y1), l in zip(t['boxes'], t['labels']):
+ print(x0, y0, x1, y1, l)
ax.add_patch(
Rectangle((x0, y0),
(x1 - x0),
diff --git a/inference.py b/inference.py
new file mode 100644
index 0000000..83a34c3
--- /dev/null
+++ b/inference.py
@@ -0,0 +1,37 @@
+import numpy as np
+import torch
+import torchvision.transforms.functional as F
+import glob
+import os
+from PIL import Image
+
+from model import create_model
+from confic import NUM_CLASSES, DEVICE, CLASSES, OUTDIR
+
+from IPython import embed
+from tqdm.auto import tqdm
+
+if __name__ == '__main__':
+ model = create_model(num_classes=NUM_CLASSES)
+ checkpoint = torch.load(f'{OUTDIR}/best_model.pth', map_location=DEVICE)
+ model.load_state_dict(checkpoint["model_state_dict"])
+ model.to(DEVICE).eval()
+
+ DIR_TEST = 'data/train'
+ test_images = glob.glob(f"{DIR_TEST}/*.png")
+
+ detection_threshold = 0.8
+
+ frame_count = 0
+ total_fps = 0
+
+ for i in tqdm(np.arange(len(test_images))):
+ image_name = test_images[i].split(os.path.sep)[-1].split('.')[0]
+
+ img = Image.open(test_images[i])
+ img_tensor = F.to_tensor(img.convert('RGB')).unsqueeze(dim=0)
+
+ with torch.inference_mode():
+ outputs = model(img_tensor.to(DEVICE))
+
+ print(len(outputs[0]['boxes']))
\ No newline at end of file
diff --git a/train.py b/train.py
index 33ea54e..b46fc13 100644
--- a/train.py
+++ b/train.py
@@ -45,6 +45,8 @@ def validate(test_loader, model, val_loss):
targets = [{k: v.to(DEVICE) for k, v in t.items()} for t in targets]
+ embed()
+ quit()
with torch.inference_mode():
loss_dict = model(images, targets)