def my_model(features, labels, mode, params):
predictions, attn_weights = predict(features, params, transformer)
estimator_spec = tf.estimator.EstimatorSpec(mode, predictions={"predictions":predictions})
print(transformer.summary())
return estimator_spec
def main():
# Basic usage: python predict.py /path/to/image checkpoint
parser = argparse.ArgumentParser(description='Optional add-ons.')
parser.add_argument('image_filepath')
parser.add_argument('checkpoint')
parser.add_argument('--topk', action='store', type=int)
parser.add_argument('--gpu', action='store_true')
parser.add_argument('--category_names', action='store')
args = parser.parse_args()
topk = args.topk if args.topk else 1
device = torch.device('cuda') if (args.gpu | torch.cuda.is_available()) else torch.device('cpu')
# assign model from checkpoint
model = ph.load_checkpoint(args.checkpoint, device)
# assign predicted top "k" probabilities and classes
probs, classes = ph.predict(args.image_filepath, model, topk)
# get category class mapping to names for flowers
if args.category_names:
with open(args.category_names, 'r') as f:
cat_to_name = json.load(f)
# hard coded
else:
with open('cat_to_name.json', 'r') as g:
cat_to_name = json.load(g)
# convert class ID to a category name
cat_class = cat_to_name[classes[0]]
print('Flower Name: ', cat_class)
print('Top Probability(ies): ', probs[:args.topk])
parser.add_argument("--gpu", help="Will you like to predict on gpu", action="store_true", default=False)
# Parse all the arguments
args = parser.parse_args()
# Extract parameters
image_dir = args.image_dir
checkpoint = args.checkpoint
top_k = int(args.top_k) if args.top_k else 1
# Process Image
image = predict_helper.process_image(image_dir)
# Load model from checkpoint
model, class_to_idx = predict_helper.load_checkpoint(checkpoint)
# Give image to model to predict output
top_prob, top_class = predict_helper.predict(image, model, class_to_idx, top_k, args.gpu)
# Check if the category names were provided so category names are printed instead
if args.category_names:
# Read in the categories
cat_to_name = predict_helper.read_category(args.category_names)
new_class = []
for i in top_class:
new_class.append(cat_to_name[i])
top_class = new_class
# Print the results
print("The model is ", top_prob*100, "% certain that the image has a predicted class of ", top_class)
from PIL import Image
in_args = phelp.get_input_args()
print(in_args)
with open(in_args.category_names, 'r') as f:
cat_to_name = json.load(f)
imagepath = in_args.input
model_loaded = phelp.load_checkpoint(in_args.ckpdir, in_args.arch)
if in_args.gpu:
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
if (device == "cpu"):
print("not found gpu,.... running on cpu")
else:
device = "cpu"
probability, index = phelp.predict(imagepath,
model_loaded,
topk=in_args.topk,
device=device)
pil_image = Image.open(imagepath)
folder = [
model_loaded.class_to_idx[int(index[0][i])] for i in range(len(index[0]))
]
name = [cat_to_name[str(folder[i])] for i in range(len(folder))]
pb = list(probability[0].cpu().detach().numpy())
print("1 st class name={}...... Probability={}".format(name[0], pb[0]))
print("2 nd class name={}...... Probability={}".format(name[1], pb[1]))
print("3 rd class name={}...... Probability={}".format(name[2], pb[2]))
print("4 fth class name={}..... Probability={}".format(name[3], pb[3]))
print("5 fith class name={}.... Probability={}".format(name[4], pb[4]))
baseFilePath = "../../_analysis/" + directory + "/"
# print(baseFilePath + "\n")
# TODO: ordered dict in function?
segmentsDict = file_helper.getSegmentDict(baseFilePath)
#print("\n".join(segments))
# Loop segments
limit = 10
i = 0
for segment, segmentBasePath in segmentsDict.items():
# Predict one segment
segmentPngPath = segmentBasePath + ".png"
score = predict_helper.predict(segmentPngPath, segment)
# print(predictionDict)
# Create report and cleanup
if (score >= threshold):
segmentPrint = segment + " " + str(score) + "\n"
print(segmentPrint) # debug
html += segmentPrint
html += report_helper.audioEmbed(segment)
html += report_helper.spectrogram(segment)
else:
print(segment + " below threshold")
# ELSE delete PNG & MP3
# Limit how many segments handled (debug)
segmentMeta["file_id"] = fileId
segmentMeta["fileDirectory"] = directory
segmentMeta["segmentStartUTC"] = fileData[
"recordDateStartUTC"] + datetime.timedelta(
0, segmentMeta["segmentStartSeconds"])
if onlyAnalyse:
print(segmentMeta) # debug
spectroFilePath = "../.." + rootDirectory + "/" + segmentMeta[
"fileDirectory"] + "/" + segmentMeta["spectroFilename"]
print(", PATH: " + spectroFilePath) # debug
score = predict_helper.predict(spectroFilePath,
segmentMeta["spectroFilename"])
if score >= threshold:
print("above " + str(score) + "\n")
# html += "\n\n" + segmentMeta["baseAudioFilename"] + "\n"
report.addPositiveSegment(segmentMeta, score)
else:
print("below " + str(score) + "\n")
report.addNegativeSegment(score)
mp3FilePath = "../.." + rootDirectory + "/" + segmentMeta[
"fileDirectory"] + "/" + segmentMeta["finalAudioFilename"]
os.remove(mp3FilePath)
os.remove(spectroFilePath)
# If analyzing, skip saving to database.
my_parser.add_argument('checkpoint_path',
metavar='checkpoint_path',
type=str,
help='the checkpoint_path where model is saved')
my_parser.add_argument('--top_k', action='store', type=int, default=5)
my_parser.add_argument('--category_names',
action='store',
type=str,
default='./cat_to_name.json')
my_parser.add_argument('--gpu', action='store_true')
# Execute the parse_args() method
args = my_parser.parse_args()
path_to_image = args.path_to_image
checkpoint_path = args.checkpoint_path
top_k = args.top_k
category_names = args.category_names
gpu = args.gpu
print(args)
if not os.path.exists(path_to_image):
print('The path_to_image specified does not exist')
sys.exit()
if not os.path.exists(checkpoint_path):
print('The checkpoint_path specified does not exist')
sys.exit()
predict_helper.predict(path_to_image, checkpoint_path, top_k, category_names,
gpu)
