Data Augmentation with cGAN

Requirements

• Pytorch
• Numpy
• Pandas
• Seaborn
• Sklearn

CGAN

Training

To train the CGAN is necessary to create a configuration file and then execute:

python3 -m train_cgan --run_name <model name> --config <path/to/configuration/file.yaml>

NOTE: examples of configuration file are contained in config/cgan/ folder.

Testing

To test the CGAN execute:

python3 -m test_cgan --config <path/to/configuration/file.yaml>

NOTE: for each model specify the same configuration file used during its training.

Classificator

Training

To train and test the classificator it's required to create a configuration file and then execute:

python3 -m train_classifier --run_name <model name> --config <path/to/configuration/file.yaml>

NOTE: examples of configuration file are contained in config/classifier/ folder.

Testing

To test a classificator model execute:

python3 -m test_classifier --model_name <model name> --config <path/to/configuration/file.yaml>

NOTES:

• For each model specify the same configuration file used during its training;
• As model name specify only the file name and do not write the complete path.

Results

Graphs visualization

To visualize the obtained results, both for CGAN and classifier, execute:

tensorboard --logdir=<path/to/tensorboard/folder/>

FID scores registered

Each model has been used to generate a new CIFAR10 dataset, that has been compared with the original one class by class through the FID score. Here are the obtained values.

RunName	AllDataset	Airplane	Automobile	Bird	Cat	Deer	Dog	Frog	Horse	Ship	Truck
basic_cGAN - epoch30	25.0814	43.8033	51.2340	50.1187	35.8533	38.5345	50.2043	51.5834	50.0480	48.3488	51.9783
basic_cGAN - epoch100	19.7900	39.9388	43.6565	45.3907	32.5473	33.6034	51.6767	35.6730	39.1974	36.5174	40.2165
basic_cGAN_BCEWithLogits - epoch30	26.1962	46.6924	59.9631	53.5438	37.7538	47.6463	56.2596	41.2485	49.8263	43.2729	59.8866
basic_cGAN_BCEWithLogits - epoch100	18.3587	39.3256	47.1962	41.2120	28.7295	28.8009	42.4902	35.0227	35.2195	39.2759	43.8909
basic_cGAN_NOINIT - epoch30	28.2961	49.6503	55.5690	52.5145	41.0350	48.4470	57.8165	49.0617	66.2571	42.2449	53.7113
basic_cGAN_NOINIT - epoch100	20.6019	42.3565	48.9903	43.9808	35.1413	32.7063	50.8310	37.9150	46.3769	38.1915	45.6485
basic_cGAN_NONORM - epoch30	54.6617	85.8362	83.3916	93.3328	65.6713	74.7721	80.4029	82.7151	92.3878	75.4047	77.9362
basic_cGAN_NONORM - epoch100	33.9906	60.6132	62.9891	64.2217	47.2884	46.7887	61.3682	49.3307	62.6472	49.1161	57.0115
basic_cGAN_SGD - epoch30	230.954	354.929	272.712	335.822	365.622	351.331	403.674	324.052	325.617	336.270	342.140
basic_cGAN_SGD - epoch100	214.799	365.356	308.137	356.645	316.402	296.593	317.441	316.219	306.465	302.288	367.789
basic_cGAN_SmoothL1loss - epoch30	25.2347	47.6879	47.8939	51.5360	38.2146	45.1380	56.9448	40.9715	56.9873	44.1011	51.9278
basic_cGAN_SmoothL1loss - epoch100	17.5960	36.6824	44.6527	39.5682	27.6457	33.0219	43.7501	31.2807	35.8994	35.7174	44.7159
cGAN_SmoothL1loss_bs128 - epoch30	28.8892	50.3035	65.5648	53.7833	37.6564	45.9779	58.8280	45.3261	62.7498	51.2253	93.3664
cGAN_SmoothL1loss_bs128 - epoch100	18.4161	37.6183	52.4211	39.3867	30.0114	30.3955	42.6866	30.4660	39.6675	36.7077	84.9438
cGAN_SmoothL1loss_Adamax - epoch30	56.9044	75.2484	105.268	85.8292	85.4688	86.5913	103.241	86.7148	128.169	64.7571	86.9369
cGAN_SmoothL1loss_Adamax - epoch100	29.5539	47.2033	83.6270	53.9002	45.3708	44.6316	57.7528	60.4991	59.0338	62.2606	66.2551
cGAN_SmoothL1loss_NAdam - epoch30	26.9565	43.4741	53.5850	56.4188	41.4548	46.8531	64.0583	68.0111	59.6425	34.2241	53.8282
cGAN_SmoothL1loss_NAdam - epoch100	20.2487	39.7137	44.5671	44.1415	32.1569	34.8685	47.4983	72.6950	40.1842	39.2248	43.1007
basic_wGAN - epoch30	48.2562	68.4371	75.0996	71.9104	54.4870	83.9637	81.9094	84.2123	83.1763	54.3745	72.1841
basic_wGAN - epoch100	28.1914	54.6635	58.3065	49.8910	38.9040	37.1997	52.3032	38.4660	42.6299	41.8378	53.9019
cGAN_64_bs128 - epoch30	204.184	222.695	228.378	349.715	283.556	274.376	255.650	265.553	235.954	249.086	310.055
cGAN_64_bs128 - epoch100	152.912	217.718	224.458	231.542	219.200	198.162	230.474	222.847	218.450	222.499	238.959
custom_cGAN - epoch30	35.8927	58.0080	68.7365	69.9182	50.5328	72.1439	70.0219	68.9697	88.6795	62.6273	80.5999
custom_cGAN - epoch100	26.6001	46.4214	52.4751	59.7820	50.5737	49.3311	83.9058	71.9584	51.5222	56.8851	52.5101
custom_wGAN - epoch30	69.4120	93.0382	102.035	104.180	98.7037	92.2057	129.500	137.826	119.934	66.3337	94.8316
custom_wGAN - epoch100	42.3130	62.3584	80.7449	79.4685	59.9742	69.1960	85.8635	61.7054	68.5446	54.6184	72.8860
custom_cGAN_SmoothL1loss - epoch30	45.7718	61.5156	98.6816	68.1409	69.3301	61.7120	81.4490	69.4798	92.9643	56.3719	94.6059
custom_cGAN_SmoothL1loss - epoch100	18.4845	37.1320	40.4285	38.8271	31.3358	32.3157	47.3368	43.6842	35.2573	34.4679	40.3287
custom_cGAN_2 - epoch30	104.208	159.158	150.377	174.049	154.885	177.240	186.384	210.636	175.558	136.752	133.861
custom_cGAN_2 - epoch100	73.0790	111.615	184.046	175.774	139.551	151.019	128.772	141.181	174.255	126.151	210.390

Accuracy registered

Each generated dataset has been used to perform a classification task. So, it has been performed a fine tuning last 5 epochs over a pretrained AlexNet. The training has been performed using the generated datasets both as training set and as augmentation of the original CIFAR10 training set, for which has been used the 70% of images. The reference value for the accuracy over the original model was 9.1 %.

Here are the obtained values.

No Augmentation

RunName	Accuracy
basic_cGAN - epoch30	57.75 %
basic_cGAN - epoch100	63.66 %
basic_cGAN_BCEWithLogits - epoch30	57.36 %
basic_cGAN_BCEWithLogits - epoch100	62.76 %
basic_cGAN_NOINIT - epoch30	55.90 %
basic_cGAN_NOINIT - epoch100	61.26 %
basic_cGAN_NONORM - epoch30	52.05 %
basic_cGAN_NONORM - epoch100	63.38 %
basic_cGAN_SGD - epoch30	9.86 %
basic_cGAN_SGD - epoch100	15.75 %
basic_cGAN_SmoothL1loss - epoch30	55.27 %
basic_cGAN_SmoothL1loss - epoch100	63.26 %
adamax_bs128_basic_cGAN_SmoothL1loss - epoch30	54.93 %
adamax_bs128_basic_cGAN_SmoothL1loss - epoch100	54.91 %
adamax_basic_cGAN_SmoothL1loss - epoch30	54.92 %
adamax_basic_cGAN_SmoothL1loss - epoch100	64.27 %
nadam_basic_cGAN_SmoothL1loss - epoch30	53.20 %
nadam_basic_cGAN_SmoothL1loss - epoch100	60.84 %
cGAN_SmoothL1loss_bs128 - epoch30	29.02 %
cGAN_SmoothL1loss_bs128 - epoch100	64.05 %
cGAN_SmoothL1loss_Adamax - epoch30	24.39 %
cGAN_SmoothL1loss_Adamax - epoch100	38.45 %
adamax_basic_cGAN_SmoothL1loss_NAdam - epoch30	54.75 %
adamax_basic_cGAN_SmoothL1loss_NAdam - epoch100	56.96 %
basic_wGAN - epoch30	52.54 %
basic_wGAN - epoch100	66.28 %
cGAN_64_bs128 - epoch30	16.97 %
cGAN_64_bs128 - epoch100	26.91 %
custom_cGAN - epoch30	42.74 %
custom_cGAN - epoch100	57.42 %
custom_wGAN - epoch30	35.85 %
custom_wGAN - epoch100	54.04 %
custom_cGAN_SmoothL1loss - epoch30	44.47 %
custom_cGAN_SmoothL1loss - epoch100	62.13 %
custom_cGAN_2 - epoch30	22.38 %
custom_cGAN_2 - epoch100	24.77 %

Augmentation

RunName	Accuracy
basic_cGAN - epoch30	80.13 %
basic_cGAN - epoch100	78.99 %
basic_cGAN_BCEWithLogits - epoch30	79.29 %
basic_cGAN_BCEWithLogits - epoch100	78.66 %
basic_cGAN_NOINIT - epoch30	78.81 %
basic_cGAN_NOINIT - epoch100	79.85 %
basic_cGAN_NONORM - epoch30	79.77 %
basic_cGAN_NONORM - epoch100	80.48 %
basic_cGAN_SGD - epoch30	81.80 %
basic_cGAN_SGD - epoch100	81.39 %
basic_cGAN_SmoothL1loss - epoch30	78.98 %
basic_cGAN_SmoothL1loss - epoch100	80.08 %
adamax_basic_cGAN_SmoothL1loss - epoch30	81.38 %
adamax_basic_cGAN_SmoothL1loss - epoch100	81.50 %
nadam_basic_cGAN_SmoothL1loss - epoch30	79.81 %
nadam_basic_cGAN_SmoothL1loss - epoch100	78.83 %
adamax_bs128_basic_cGAN_SmoothL1loss - epoch30	81.47 %
adamax_bs128_basic_cGAN_SmoothL1loss - epoch100	81.59 %
cGAN_SmoothL1loss_bs128 - epoch30	80.35 %
cGAN_SmoothL1loss_bs128 - epoch100	79.46 %
cGAN_SmoothL1loss_Adamax - epoch30	80.12 %
cGAN_SmoothL1loss_Adamax - epoch100	79.78 %
adamax_basic_cGAN_SmoothL1loss_NAdam - epoch30	81.68 %
adamax_basic_cGAN_SmoothL1loss_NAdam - epoch100	81.91 %
basic_wGAN - epoch30	79.23 %
basic_wGAN - epoch100	80.30 %
cGAN_64_bs128 - epoch30	81.11 %
cGAN_64_bs128 - epoch100	80.87 %
custom_cGAN - epoch30	79.39 %
custom_cGAN - epoch100	79.91 %
custom_wGAN - epoch30	79.15 %
custom_wGAN - epoch100	79.86 %
custom_cGAN_SmoothL1loss - epoch30	78.92 %
custom_cGAN_SmoothL1loss - epoch100	79.61 %
custom_cGAN_2 - epoch30	80.04 %
custom_cGAN_2 - epoch100	80.91 %