achaiah / pywick Goto Github PK

Hi!

Thanks alot for the inclusion of ARiA in your library! The implementation of ARiA in the library is correct, however I feel for ARiA-2 there were substitution errors which led to an incorrect formula. If you were to substitute and reduce (for speedup) you end up getting the following form: ARiA2(x) = xsigmoid(betax)**alpha which can be implemented as

`
class Aria2(nn.Module):
"""
ARiA2 activation function, a special case of ARiA, for ARiA = f(x, 1, 0, 1, 1, b, 1/a)
"""

def __init__(self, a=1.5, b = 1.):
    super(Aria2, self).__init__()
    self.alpha = a
    self.beta = b

def forward(self, x):
    return x * torch.sigmoid(self.beta*x)**(self.alpha)

`

Hi, and thanks for your code
I just try to implement this on keras for a learning purpose, but as i do this i cannot understand block_index, so i think you forgot using for loop or how you pass block_index to it:

class PolyConv2d(nn.Module):
    def forward(self, x, block_index):
        x = self.conv(x)
        bn = self.bn_blocks[block_index]
        x = bn(x)
        x = self.relu(x)
        return x

please let me know if I was wrong about this one

In the file, the dice is defined as a BCE loss, which seems to be wrong.
def init(self, l=0.5, weight_of_focal=1.):
super(BCEDiceFocalLoss, self).init()
# self.bce = BCELoss2d()
# self.dice = SoftDiceLoss()
self.dice = BCELoss2d()
self.focal = FocalLoss(l=l)
self.weight_of_focal = weight_of_focal

https://github.com/achaiah/pywick/blob/master/pywick/losses.py#L434

Pytorch datasets do not require a num_inputs or num_targets field.

https://github.com/achaiah/pywick/blob/master/pywick/modules/module_trainer.py#L387

Hello,

I am trying to run your implementation of DUNet and it fails to initialize the model.

The code I ran was:
`from pywick.models.segmentation import DUNet_Resnet50

img = torch.randn(2, 3, 256, 256)
model = DUNet_Resnet50()
outputs = model(img)
`

which threw this error:

Model file /root/.torch/models/resnet50-25c4b509.pth is not found. Downloading.
Downloading /root/.torch/models/resnet50-25c4b509.zip from https://hangzh.s3.amazonaws.com/encoding/models/resnet50-25c4b509.zip...

RuntimeError Traceback (most recent call last)
in ()
1 img = torch.randn(2, 3, 256, 256)
----> 2 model = DUNet_Resnet50()
3 outputs = model(img)

6 frames
/usr/local/lib/python3.7/dist-packages/pywick/models/segmentation/dunet.py in DUNet_Resnet50(num_classes, **kwargs)
138
139 def DUNet_Resnet50(num_classes=1, **kwargs):
--> 140 return get_dunet(num_classes=num_classes, backbone='resnet50', **kwargs)
141
142

/usr/local/lib/python3.7/dist-packages/pywick/models/segmentation/dunet.py in get_dunet(num_classes, backbone, pretrained, **kwargs)
133 pretrained : bool (default: True) - whether to load pretrained backbone network, that was trained on ImageNet.
134 """
--> 135 model = DUNet(num_classes=num_classes, backbone=backbone, pretrained=pretrained, **kwargs)
136 return model
137

/usr/local/lib/python3.7/dist-packages/pywick/models/segmentation/dunet.py in init(self, num_classes, pretrained, backbone, aux, **kwargs)
26
27 def init(self, num_classes, pretrained=True, backbone='resnet101', aux=False, **kwargs):
---> 28 super(DUNet, self).init(num_classes, pretrained=pretrained, aux=aux, backbone=backbone, **kwargs)
29 self.head = _DUHead(2144, **kwargs)
30 self.dupsample = DUpsampling(256, num_classes, scale_factor=8, **kwargs)

/usr/local/lib/python3.7/dist-packages/pywick/models/segmentation/da_basenets/segbase.py in init(self, num_classes, pretrained, aux, backbone, **kwargs)
21 self.nclass = num_classes
22 if backbone == 'resnet50':
---> 23 self.pretrained = resnet50_v1s(pretrained=pretrained, **kwargs)
24 elif backbone == 'resnet101':
25 self.pretrained = resnet101_v1s(pretrained=pretrained, **kwargs)

/usr/local/lib/python3.7/dist-packages/pywick/models/segmentation/da_basenets/resnetv1b.py in resnet50_v1s(pretrained, model_root, **kwargs)
236 if pretrained:
237 from .model_store import get_resnet_file
--> 238 model.load_state_dict(torch.load(get_resnet_file('resnet50', root=model_root)), strict=False)
239 return model
240

/usr/local/lib/python3.7/dist-packages/pywick/models/segmentation/da_basenets/model_store.py in get_resnet_file(name, root)
49 download(_url_format.format(repo_url=repo_url, file_name=file_name),
50 path=zip_file_path,
---> 51 overwrite=True)
52 with zipfile.ZipFile(zip_file_path) as zf:
53 zf.extractall(root)

/usr/local/lib/python3.7/dist-packages/pywick/models/segmentation/da_basenets/download.py in download(url, path, overwrite, sha1_hash)
66 r = requests.get(url, stream=True)
67 if r.status_code != 200:
---> 68 raise RuntimeError("Failed downloading url %s"%url)
69 total_length = r.headers.get('content-length')
70 with open(fname, 'wb') as f:

RuntimeError: Failed downloading url https://hangzh.s3.amazonaws.com/encoding/models/resnet50-25c4b509.zip

The same happens with other DUNet backbones. I am not sure if other models have similar issues, but it seems to be a simple access problem.

Hi ,

Thank you for creating this wonderful package!

I am having trouble in creating data model pipe line. Could you please create a simple example using say cat/dog data and describe how data pipe line would work using pywick. I think this would be helpful to people who want to quick start with this package.

Thanks

The assignment of self.alpha is missing in the constructor.

Hello, I don't know why the Focalloss need to add "epsilon = 1e-10".

Hello, first of all, I would like to appreciate the work you are doing. I was trying to use the FocalLoss for Multi-Class problem and tried using FocalLoss2. It shows nan values for train and validation loss, and when I keep the gamma to 2, the training fails after one epoch while when gamma = 0(equal to cross_entropy_loss) it is working but still the loss values are nan. Please help me

I wrote the following custom metric,

class F1Score(pywick.metrics.Metric):
    def __init__(self, micmac='micro'):
        super().__init__()
        self._name = f'f1_{micmac}'
        self.micmac = micmac
        self.correct = []
        self.preds = []

    def reset(self):
        self.correct = []
        self.preds = []

    def __call__(self, inputs, y_pred, y_true, is_val=False):
        self.correct.extend(y_true.detach().to('cpu').numpy().tolist())
        self.preds.extend(y_pred.detach().to('cpu').numpy().argmax(axis=1).tolist())
        f1 = sklearn.metrics.f1_score(self.correct, self.preds, average=self.micmac)
        return f1

and use it like so

metrics = [F1Score('micro'), CategoricalAccuracySingleInput(top_k=1)]
trainer.compile(optimizer='adam',
                criterion='cross_entropy',
                initializers=initializers,
                metrics=metrics)

But while running trainer.fit_loader(...), I get the following output

Epoch 1/1:   7%|▋         | 7/98 [00:12<02:46,  1.83s/ batches, loss=0.6508, top_1:acc=76.04, learn_rates=[0.001]]

No mention of the custom metric anywhere!

But running trainer.eval_loader() gives,

{'val_f1_micro': 0.7514395393474088,
 'val_loss': 0.6613252784975586,
 'val_top_1:acc_metric': 75.14395393474088}

Which works fine, so what's going on with fit_loader()?

Hi there :)

I would like to ping you for a possible collaboration on an open source project, if you are curios, just ping me at [email protected]

Thank you

Have a wonderful day

Francesco

This code base is the most exciting repo I recently find in github!!
Keep on this awesome work!