• I am using p2.x.large machine on aws .It has 12gb gpu memory

• For batch size > 1 I see high gpu utility and gpu memory usage .Thus ending up with out of memory issues

• With batch size = 1 ,an epoch takes 4 hrs ,thus 3000 epochs = 500 days :)

• Is there a pretrained model which I can use for training

• Is such High usage normal.I see that each image is 10kb and mask is 2.5kb.I am using chainer for the first time .Am i doing something wrong .

The URL for the pre-trained model fcn-8s-pascalcontext_W_and_b.pkl 'https://googledrive.com/host/0BxSyYt1jT6LhUlhITjdicDFyNHM' not accessible .I also checked in https://github.com/shiba24/pretrained-model-collections

When using all the suggested parameters, but CPU, instead of GPU and the initial model in pkl format provided by you, I get this error:

loading VGG model... ('training datasets: ', 9788, 'test datasets: ', 516) Epoch 1 : training... 0it [00:00, ?it/s]Create Iterator settings

Traceback (most recent call last): File "train.py", line 102, in model, optimizer, train_mean_loss, train_ac = train(model, optimizer, MiniBatchLoader, train_mean_loss, train_ac) File "train.py", line 31, in train optimizer.update(model, x, t) File "/usr/local/lib/python2.7/dist-packages/chainer/optimizer.py", line 392, in update loss = lossfun(_args, *_kwds) File "/home/tamme/Desktop/humanparts/model.py", line 86, in call self.accuracy = self.calculate_accuracy(h, t) File "/home/tamme/Desktop/humanparts/model.py", line 168, in calculate_accuracy mask = truths != -1 File "/usr/local/lib/python2.7/dist-packages/chainer/variable.py", line 463, in ne raise NotImplementedError() NotImplementedError

Is it a bug or have I done something incorrectly? I don't seem to find a related problem in google either.

I am now working to reproduce some experiments reported in the fcn paper. https://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf

Compared to the layer-by-layer refinement training reported in the paper, the current model is trained end-to-end. It is interesting to see how much differences it make. One thing I observed is that the bilinear interpolation initialization for deconvolution layer is essential, otherwise the model can not be properly trained.

I will also include some pretrained models if needed.

I add some important changes to the initialization. Now the current model is successfully trained on a simplified setting (> 90 pixel accuracy at 20 epochs), which is the binary segmentation of human silhouette. I will try the four parts and fourteen parts segmentation later.

• I am using a p2 instance on aws .I ran python train.py -g 0 -b 5 -e 3000 -l on -s on

And I am getting

loading VGG model... ('training datasets: ', 9788, 'test datasets: ', 516) Epoch 1 : training... 0it [00:00, ?it/s]Create Iterator settings 1it [00:04, 4.19s/it]Traceback (most recent call last): File "train.py", line 102, in model, optimizer, train_mean_loss, train_ac = train(model, optimizer, MiniBatchLoader, train_mean_loss, train_ac) File "train.py", line 27, in train for X, y in tqdm(MiniBatchLoader): File "/usr/local/lib/python2.7/dist-packages/tqdm/_tqdm.py", line 816, in iter for obj in iterable: File "/home/ubuntu/deep-learning-for-human-part-discovery-in-images/data.py", line 83, in next minibatch_X, minibatch_y = self.process_batch(minibatch_X, minibatch_y) File "/home/ubuntu/deep-learning-for-human-part-discovery-in-images/data.py", line 145, in process_batch reshaped_X = np.transpose(self.standardize(processed_X), (0, 3, 1, 2)) # n_batch, n_channel, h, w File "/usr/local/lib/python2.7/dist-packages/numpy/core/fromnumeric.py", line 556, in transpose return transpose(axes) ValueError: axes don't match array

• Update - I am not able to reproduce the issue .Everything seems to be working fine now

loading VGG model... Traceback (most recent call last): File "train.py", line 81, in model = HumanPartsNet(n_class=15) File "/home/leisther/Downloads/deep-learning-for-human-part-discovery-in-images-master/model.py", line 53, in init upsample_pool1=L.Convolution2D(64, self.n_class, ksize=1, stride=1, pad=0, wscale=0.01), File "/home/leisther/anaconda2/envs/venv/lib/python2.7/site-packages/chainer/links/connection/convolution_2d.py", line 120, in init deterministic="deterministic argument is not supported anymore. " File "/home/leisther/anaconda2/envs/venv/lib/python2.7/site-packages/chainer/utils/argument.py", line 19, in parse_kwargs raise TypeError(message) TypeError: init() got unexpected keyword argument(s) 'wscale'

Hello, I'm having an issue regarding the dataset preparation process here is the error:

loading VGG model... training datasets: 9273 test datasets: 1031 scanning all images for human part labels ... 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 10304/10304 [07:30<00:00, 22.87it/s] training datasets: 0 test datasets: 0 found 3785 images Epoch 1 : training... 0it [00:00, ?it/s]Traceback (most recent call last): File "train.py", line 113, in model, optimizer, train_mean_loss, train_ac, train_IoU = train(model, optimizer, MiniBatchLoader, train_mean_loss, train_ac, train_IoU) File "train.py", line 27, in train for X, y in tqdm(MiniBatchLoader): File "C:\Users\Lenovo\AppData\Local\Programs\Python\Python36\lib\site-packages\tqdm_tqdm.py", line 962, in iter for obj in iterable: TypeError: iter() returned non-iterator of type 'MiniBatchLoader'

could anyone help me please it's URGENT

Hi, I am a newbie to ML and just followed the readme to download the FCN model, without any idea using it. My goal is to use the pre-trained model and see how it works, because my GPU is around 6GB which seems not enough for training. I googled the pkl file and it seems to be something working with python pickle model, so I:

python ActivePython 2.7.10.12 (ActiveState Software Inc.) based on Python 2.7.10 (default, Aug 21 2015, 12:07:58) [MSC v.1500 64 bit (AMD64)] on win32 Type "help", "copyright", "credits" or "license" for more information.

import pickle loaded_model = pickle.load(open('fcn-8s-pascalcontext_W_and_b.pkl', 'rb'))

and get the dictionary "loaded model", with a bunch of numbers, while the keys are:

loaded_model.keys() ['conv4_3_W', 'conv5_1_b', 'conv1_2_b', 'upsample_W', 'conv5_2_b', 'conv1_1_W', 'score-pool3_W', 'conv5_2_W', 'conv5_3_W', 'conv1_1_b', 'conv4_3_b', 'conv5_3_b', 'conv5_1_W', 'conv1_2_W', 'conv3_2_W', 'conv4_2_b', 'conv4_1_b', 'upscore2_W', 'conv3_3_W', 'conv2_1_b', 'conv3_1_b', 'conv2_2_W', 'fc6_b', 'score-pool4_W', 'fc7_b', 'score59_W', 'conv2_2_b', 'fc6_W', 'upsample-fused-16_W', 'score59_b', 'fc7_W', 'conv4_1_W', 'conv3_2_b', 'conv4_2_W', 'score-pool4_b', 'conv3_3_b', 'conv3_1_W', 'conv2_1_W', 'score-pool3_b']

Then stuck here, not knowing what to do next. Could anyone please help telling how to use the pre-trained model? Any hints would be appreciative.