Mask Prediction
Solution for submission 174985
A detailed solution for submission 174985 submitted for challenge Mask Prediction
jakub_bartczuk
In [91]:
%%capture
!pip install aicrowd-cli
%load_ext aicrowd.magic
In [ ]:
!python main_train.py --model_type fasterrcnn --config_path fasterrcnn.yaml
In [ ]:
!rm -rf data !mkdir data %aicrowd ds dl -c mask-prediction -o data
!unzip data/train.zip -d data/train > /dev/null !unzip data/val.zip -d data/val > /dev/null !unzip data/test.zip -d data/test > /dev/null
yolov3.yaml
datasets: train: metadata_path: train_metadata.json dataset_format: deepsense valid: metadata_path: val_metadata.json dataset_format: deepsense classnames:
- N95
- surgical
- cloth
- KN95 input_height: 256 input_width: 256 experiment_dirpath: experiments/yolov3_training num_train_iters: 2000 valid_freq_iters: 200 verbose_freq_iters: 100 lr_step_size: 200 timestamped: True train_transformations: ["resize(256,256)"] valid_transformations: ["resize(256,256)"]
In [64]:
#Reading the file
import pandas as pd
import numpy as np
import os
from pathlib import Path
# Image Reading & Preprocessing
from PIL import Image, ImageDraw
import cv2
import matplotlib.pyplot as plt
import numpy as np
# Misc.
from tqdm.notebook import tqdm
from sklearn.ensemble import RandomForestClassifier
In [65]:
train_images = 'data/train'
val_images = 'data/val'
test_images = 'data/test'
In [66]:
train_df = pd.read_csv("data/train.csv")
val_df = pd.read_csv("data/val.csv")
In [67]:
train_df.head()
Out[67]:
In [68]:
from ast import literal_eval
In [69]:
def create_bboxes(row):
bbox_coords = literal_eval(row['bbox'])
bbox = {"xmin": bbox_coords[0], "ymin": bbox_coords[1], "xmax": bbox_coords[2], "ymax": bbox_coords[3]}
bbox['classname'] = row['masktype']
bbox['probability'] = None
return bbox
In [70]:
def create_item_metadata(image_path, row):
item_metadata = {
'image_path': str(Path(image_path) / (row['ImageID'] + ".jpg")),
'namedboxes': [create_bboxes(row)],
'namedmasks': [],
'namedkeypoints': [],
'image_classnames': []
}
return item_metadata
In [71]:
def create_dsvision_metadata(image_path, metadata_df):
items = []
for idx, row in metadata_df.iterrows():
item_metadata = create_item_metadata(image_path, row)
items.append(item_metadata)
return items
In [72]:
train_dsvision_metadata = create_dsvision_metadata(train_images, train_df)
val_dsvision_metadata = create_dsvision_metadata(val_images, val_df)
In [73]:
import json
In [74]:
json.dump(train_dsvision_metadata, open("train_metadata.json", "w"))
json.dump(val_dsvision_metadata, open("val_metadata.json", "w"))
In [75]:
for i in train_df['masktype'].unique():
print(" - "+ i)
In [76]:
import deepsense_vision
In [77]:
from deepsense_vision.datasets.deepsense_dataset import DeepsenseDataset
from deepsense_vision.io import load_json
from deepsense_vision.models.fasterrcnn.fasterrcnn_model import FasterRCNNModel
from deepsense_vision.models.yolov3.yolov3_model import YOLOv3Model
from deepsense_vision.models.patchnet.patchnet_model import PatchNetModel
In [78]:
model = FasterRCNNModel()
In [79]:
!ls experiments/
In [80]:
import pathlib
def load_results_from_config_path(experiment_path):
results_path = pathlib.Path(experiment_path) / "checkpoint" / "valid_metrics.json"
if os.path.exists(results_path):
return json.load(open(str(results_path), "r"))['global']
else:
return None
In [81]:
!ls 20220210_124238_344077_faster_rcnn_training_1000/
In [82]:
load_results_from_config_path("experiments/20220210_124238_344077_faster_rcnn_training_1000")
Out[82]:
In [83]:
import glob
In [84]:
experiment_results_jsons = {
config_path: load_results_from_config_path(config_path)
for config_path in glob.glob("experiments/*")
}
In [85]:
experiment_results_jsons
Out[85]:
In [86]:
experiment_results = [
{"experiment": path, **{k: v for (k,v) in results.items() if "mAP" in k}}
for (path, results) in experiment_results_jsons.items()
if not results is None
]
In [87]:
experiments_df = pd.DataFrame.from_records(experiment_results).sort_values("mAP", ascending=False)
In [88]:
experiments_df.index = experiments_df['experiment']
In [89]:
experiments_df.sort_values("mAP", ascending=False)
Out[89]:
In [28]:
#pathlib.Path("experiments/20220210_100929_544292_faster_rcnn_training_1000")
best_model_dir = pathlib.Path(experiments_df['mAP'].idxmax())
best_model_path = best_model_dir / "checkpoint" / "fasterrcnn.pt"
In [29]:
model.load_checkpoint(str(best_model_path))
In [30]:
model.to_gpu()
In [31]:
from pathlib import Path
In [ ]:
test_image_paths = list(map(str, Path("data/test/").glob("*")))
In [ ]:
import skimage.io
import skimage.transform
In [ ]:
test_image_paths[0]
In [ ]:
test_img = skimage.io.imread(test_image_paths[0])
In [ ]:
test_img.shape
In [ ]:
plt.imshow(test_img)
In [ ]:
model.config.inference_preprocessing_longer_axis_size = 512
model.config.inference_preprocessing_padding_min_image_height = 512
model.config.inference_preprocessing_padding_min_image_width = 512
model.config.inference_preprocessing_longer_axis_size = 512
model.config.inference_sliding_window_height = 512
model.config.inference_sliding_window_width = 512
In [ ]:
test_img.shape
In [ ]:
test_img.transpose(-1, 0, 1).shape
In [ ]:
pred = model.predict_from_array(test_img)
In [ ]:
pred = model.predict_from_filepath(test_image_paths[1])
In [ ]:
pred.visualize()
In [ ]:
import tqdm
import numpy as np
In [ ]:
def convert_box_coords_to_int(box):
for key in ['xmin', 'xmax', 'ymin', 'ymax']:
box[key] = int(np.round(box[key]))
In [ ]:
raw_predictions = [
[box.to_dict() for box in model.predict_from_filepath(str(p)).namedboxes]
for p in tqdm.tqdm_notebook(test_image_paths)
]
In [ ]:
for pred in raw_predictions:
for b in pred:
convert_box_coords_to_int(b)
In [ ]:
raw_predictions[0]
In [ ]:
predictions = [b[0] if len(b) > 0 else raw_predictions[0][0] for b in raw_predictions]
In [ ]:
def box_to_str(box):
return "[{}, {}, {}, {}]".format(box['xmin'], box['ymin'], box['xmax'], box['ymax'])
In [ ]:
str_predictions = [box_to_str(b) for b in predictions]
cls_predictions = [b['classname'] for b in predictions]
In [ ]:
idxs = [os.path.splitext(os.path.basename(p))[0] for p in test_image_paths]
In [ ]:
preds_df = pd.DataFrame.from_records(zip(idxs, str_predictions, cls_predictions))
In [ ]:
preds_df.columns = ['ImageID', 'bbox', 'masktype']
In [ ]:
preds_df.to_csv('submission.csv', index=False)
In [ ]:
!cp Masks_DS_Vision.ipynb notebook.ipynb
!zip -r submission.zip notebook.ipynb submission.csv
In [ ]:
!zip -r submission.zip submission/
In [ ]:
#%aicrowd login
In [ ]:
!aicrowd submission create -c mask-prediction -f submission.zip
In [ ]:
Content
Comments
You must login before you can post a comment.