classify

#!/usr/bin/env python3

from fastbook import *
from fastai.vision import *
import argparse
import sys
import os
from contextlib import redirect_stdout
#from PIL import Image

import torch

def main():
	global args, out

	out = sys.stdout
	# out = debogrify_stdout()

	parser = argparse.ArgumentParser(description='Classify images with a trained neural network.')
	parser.add_argument('-v', '--verbose', action='store_true', help='debug')
	parser.add_argument('--path', type=str, default='.', help='path to model')
	parser.add_argument('--model', type=str, default='export.pkl', help='model')
	parser.add_argument('--cpu', action='store_true', help='run on cpu')
	parser.add_argument('--batch', type=int, default=100, help='batch size')
	parser.add_argument('--move', type=float, help='move files with certainty above the threshold')
#	parser.add_argument('--train', action='store_true', help='train the model')
#	parser.add_argument('--watch', type=str, help='folder to watch for images to classify')
#	parser.add_argument('--folder', type=str, default='.', help='folder with images to classify (or train)')
#	parser.add_argument('--certain', type=float, default=0.9, help='minimum certainty to classify image')

	args = parser.parse_args()

	if args.verbose:
		print(torch.cuda.get_device_name(0), file=sys.stderr)

	model_path = Path(args.path)/Path(args.model)
	if not model_path.exists():
		model_path = Path.home()/Path(args.model)

	if args.verbose:
		print(model_path, file=sys.stderr)

	learn = load_learner(model_path)
	if not args.cpu:
		try:
			learn.dls.to('cuda')
		except Exception as e:
			print(e, file=sys.stderr)

	if args.verbose:
		print(f'{learn.dls.device=}', file=sys.stderr)

	if args.batch == 1:
		predict_one_at_time(learn)
	else:
		predict_in_batches(learn)


#def debogrify_stdout():
#	""" deal with some bogus output to stdout from pytorch or something """
#	stdout_fd = sys.stdout.fileno()
#	stderr_fd = sys.stderr.fileno()
#	new_stdout_fd = os.dup(stdout_fd)
#	os.dup2(stderr_fd, stdout_fd)
#	out = os.fdopen(new_stdout_fd, "w")
#	return out


def predict_one_at_time(learn):
	for filename in sys.stdin:
		filename = filename.rstrip()
		pred,i,probs = learn.predict(filename)
		print(f"{probs[i]:.10f}\t{pred}\t{filename}", file=out)
#		for a in zip(learn.dls.vocab,[f'{x:.10f}' for x in probs]):
#			print(a)


def predict_in_batches(learn):
	if args.verbose:
		print(f'batch size {args.batch}', file=sys.stderr)
	batch = []
	for filename in sys.stdin:
		filename = filename.rstrip()
		batch.append(filename)
		if len(batch) >= args.batch:
			predict_batch(learn, batch)
			batch = []
	if len(batch) > 0:
		predict_batch(learn, batch)
		batch = []


def predict_batch(learn, batch):
	vocab = learn.dls.vocab
	with redirect_stdout(sys.stderr):
		# In some versions the following call writes the progress bar to stdout. We can't have that!
		dl = learn.dls.test_dl(batch)
		preds_all = learn.get_preds(dl=dl)
	for i in range(0, len(batch)):
		filename = batch[i]
		preds = preds_all[0][i]
		i = preds.argmax(dim=0)
		label = vocab[i]
		prob = preds[i]
		print(f"{prob:.10f}\t{label}\t{filename}", file=out)
		if args.move is not None and prob >= args.move:
			Path(label).mkdir(exist_ok=True)
			shutil.move(filename, label)

# TODO load ai_helper.py
#def is_cat(x): return x[0].isupper()

main()