What's the expected result of Classification/PyT notebook?

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

Hi, I ran the notebook in the below image, (image classification, pytorch, pretrained tiny fan model on imagenet, cats and dogs dataset).

image
image2308×108 13.5 KB

But it seems that it converges to the accuracy ~50% which is the chance level.

I’m currently changing training config such as optimizer, lr, and batch size. While I’m on it, I’m curious how accurate the model would be expected from the notebook example.

Do you have any results on this?

2

I just much lower the lr and got a well trained model.

3

Thanks for the info. Could you share the latest spec file? Thanks.

4

Hi, I already overwrote the spec file, I can share spec json file.

The only thing I changed was lower lr to 1e-5, it shows slow converge at early epochs. So bit higher lr might be working as well.


{"model": {"type": "ImageClassifier", "backbone": {"type": "fan_tiny_8_p4_hybrid", "drop_path": 0.1}, "neck": null, "head": {"type": "FANLinearClsHead", "num_classes": 2, "in_channels": 192, "loss": {"type": "CrossEntropyLoss", "loss_weight": 1.0, "use_soft": false}, "topk": [1], "head_init_scale": 1}, "init_cfg": {"type": "Pretrained", "checkpoint": "/workspace/tao-experiments/pretrained_fan_hybrid_tiny/pretrained_fan_classification_imagenet_vfan_hybrid_tiny/fan_hybrid_tiny.pth", "prefix": null}, "train_cfg": {"augments": null}}, "dataset": {"img_norm_cfg": {"mean": [123.675, 116.28, 103.53], "std": [58.395, 57.12, 57.375], "to_rgb": true}, "data": {"samples_per_gpu": 8, "workers_per_gpu": 2, "train": {"type": "ImageNet", "data_prefix": "/data/cats_dogs_dataset/training_set/training_set/", "pipeline": [{"type": "LoadImageFromFile"}, {"type": "RandomResizedCrop", "size": 224}, {"type": "RandomFlip", "flip_prob": 0.5, "direction": "horizontal"}, {"type": "Normalize", "mean": [123.675, 116.28, 103.53], "std": [58.395, 57.12, 57.375], "to_rgb": true}, {"type": "ImageToTensor", "keys": ["img"]}, {"type": "ToTensor", "keys": ["gt_label"]}, {"type": "Collect", "keys": ["img", "gt_label"]}], "classes": "/data/cats_dogs_dataset/classes.txt"}, "val": {"type": "ImageNet", "data_prefix": "/data/cats_dogs_dataset/val_set/val_set", "ann_file": null, "pipeline": [{"type": "LoadImageFromFile"}, {"type": "Resize", "size": [256, -1]}, {"type": "CenterCrop", "crop_size": 224}, {"type": "Normalize", "mean": [123.675, 116.28, 103.53], "std": [58.395, 57.12, 57.375], "to_rgb": true}, {"type": "ImageToTensor", "keys": ["img"]}, {"type": "Collect", "keys": ["img"]}], "classes": "/data/cats_dogs_dataset/classes.txt"}, "test": {"type": "ImageNet", "data_prefix": "/data/cats_dogs_dataset/val_set/val_set", "ann_file": null, "pipeline": [{"type": "LoadImageFromFile"}, {"type": "Resize", "size": [256, -1]}, {"type": "CenterCrop", "crop_size": 224}, {"type": "Normalize", "mean": [123.675, 116.28, 103.53], "std": [58.395, 57.12, 57.375], "to_rgb": true}, {"type": "ImageToTensor", "keys": ["img"]}, {"type": "Collect", "keys": ["img"]}], "classes": "/data/cats_dogs_dataset/classes.txt"}}, "sampler": null}, "train": {"exp_config": {"manual_seed": 49, "MASTER_ADDR": "127.0.0.1", "MASTER_PORT": 631}, "validate": false, "train_config": {"checkpoint_config": {"interval": 1, "by_epoch": true}, "optimizer": {"type": "SGD", "lr": 1e-05, "weight_decay": 0.05}, "paramwise_cfg": null, "optimizer_config": {"grad_clip": null}, "lr_config": {"policy": "CosineAnnealing", "min_lr": 0.001, "warmup": "linear", "warmup_iters": 5, "warmup_ratio": 0.01, "warmup_by_epoch": true}, "runner": {"type": "TAOEpochBasedRunner", "max_epochs": 100}, "logging": {"interval": 500, "log_dir": "logs"}, "evaluation": {"interval": 1, "metric": "accuracy"}, "find_unused_parameters": false, "resume_training_checkpoint_path": null, "validate": true, "load_from": null, "custom_hooks": [{"type": "EMAHook", "momentum": 4e-05, "priority": "ABOVE_NORMAL"}]}, "num_gpus": 1, "results_dir": null}, "evaluate": {"num_gpus": 1, "batch_size": 1, "checkpoint": null, "trt_engine": null, "exp_config": {"manual_seed": 47, "MASTER_ADDR": "127.0.0.1", "MASTER_PORT": 631}, "topk": 1, "results_dir": null}, "inference": {"num_gpus": 1, "batch_size": 1, "checkpoint": null, "trt_engine": null, "exp_config": {"manual_seed": 47, "MASTER_ADDR": "127.0.0.1", "MASTER_PORT": 631}, "results_dir": null}, "gen_trt_engine": {"results_dir": null, "gpu_id": 0, "onnx_file": null, "trt_engine": null, "input_channel": 3, "input_width": 224, "input_height": 224, "opset_version": 12, "batch_size": -1, "verbose": false, "tensorrt": {"data_type": "FP32", "workspace_size": 1024, "min_batch_size": 1, "opt_batch_size": 1, "max_batch_size": 1}}, "export": {"verify": false, "opset_version": 12, "checkpoint": null, "input_channel": 3, "input_width": 224, "input_height": 224, "onnx_file": null, "results_dir": null}, "results_dir": "/results/classification_experiment"}
5

Thanks for the info. Glad to know it works.