Deepforest Model Training

DeepForest uses deep learning object detection networks to predict bounding boxes corresponding to individual trees in RGB imagery. DeepForest is built on the retinanet model from the torchvision package and designed to make training models for tree detection simpler.

Prebuilt models are always improved by adding data from the target area. In our work, we found that even an hour of carefully selected hand annotations can greatly improve accuracy and precision. We believe that at least some fine-tuning of pre-built models is worthwhile for most scientific applications. Using off-the-shelf models for training, we found that 5-10 epochs were sufficient.

The dataset we used to test to train our data was urban-tree-detection-data 

Libraries we used to achieve it:

from pytorch_lightning.loops import dataloader

import pandas as pd

import torch

from torch.utils.data import Dataset, DataLoader

import deepforest

from deepforest import main

import pytorch_lightning as pl

import os

import rasterio

import glob

We had many CSV files to train our model on the dataset which had only two columns x and y present representing the location of tree crowns but the deepforest model takes input as the bounding box, not coordinates. so we thought of the robust idea that by creating a bounding box from coordinates by the error of 1 and it also should have image_path in our CSV file, following is the code snippet for it

    # making data feasible to train on our model

df = pd.read_csv(annotationFileTemp)

df['image_path'] = image_path

df['xmin']=df['x'] - 1

df['xmax']=df['x'] + 1

df['ymin']=df['y'] - 1

df['ymax']=df['y'] + 1

df['label']="Tree"

df.to_csv("/content/data/annotation.csv")

where annotationFileTemp is the original CSV file that only contains 2D coordinates, So now our dataset is ready.

We tell the config that we want to train on this csv file, and that the images are in the same directory. If images are in a separate folder, change the root_dir.

    # model configuration

annotationFile="/content/data/annotation.csv"

self.model.config["gpus"] = "-1"

self.model.config["train"]["epochs"] = 5

self.model.config["epochs"] = 5

self.model.config["train"]["csv_file"] = annotationFile

self.model.config["score_thresh"] = 0.4

self.model.config["train"]["root_dir"] = os.path.dirname(annotationFile)

self.model.config["train"]["fast_dev_run"] = False

One can get a whole set of parameters config and one can also pass any additional PyTorch lightning argument to the trainer. 

To begin training, we create a PyTorch-lightning trainer and call a trainer.fit on the model object directly on itself. While this might look a touch awkward, it is useful for exposing the PyTorch lightning functionality.

We have to create a trainer and then fit the model in that trainer

model.create_trainer()

model.use_release()

model.trainer.fit(model)

But apparently, on training, we were getting an error stating as

"ReferenceError: weakly-referenced object no longer exists", it was because of the model.trainer so what we did was created PyTorch lightening trainer outside of deepforest and ignored create_trainer and then fit the model in the trainer

# creating trainer to train the model

trainer= pl.Trainer( max_epochs=5,

gpus="-1",

enable_checkpointing=False,

accelerator='gpu',

fast_dev_run= False)

# training model on the respective data

trainer.fit(model=self.model)  

But as I stated first we had multiple CSV files so we applied the loop to the training of the model so every time it got trained more and more on multiple CSV files

training=Training()

counter=0

# Giving our model respective data files to train

for file in glob.iglob(f"/content/urban-tree-detection-data/csv/*"):

if counter<100:

file_name=""

try:

file_name=file

file_name=f"/content/urban-tree-detection-data/images/{file[39:-4]}.tif"

training.training_of_model(testing_csv=file, image_path=file_name)

except:

print(f"File has some invalid data, Filename : {file}")

Link to the colab notebook: Colab Notebook