How to get the dimensionality when num_classes=0

[sachinruk]

I was wondering if there was a way to get the shape of the output without using a sample test image. Specifically, is there a way to get this information for when we do:

model = timm.create_model(model_name, num_classes=0)
# one way to get dim would be following.
dim = model(torch.randn(1, 3, 400, 400)).shape[1]
# however I am hoping that this information is obtainable in a better way (without test image)

TIA

[rwightman]

pytorch-image-models/tests/test_models.py

Lines 284 to 304 in 648aaa4

	outputs = model.forward_features(input_tensor)
	outputs_pre = model.forward_head(outputs, pre_logits=True)
	if isinstance(outputs, (tuple, list)):
	# cannot currently verify multi-tensor output.
	pass
	else:
	if feat_dim is None:
	feat_dim = -1 if outputs.ndim == 3 else 1
	assert outputs.shape[feat_dim] == model.num_features
	assert outputs_pre.shape[1] == model.head_hidden_size
	# test forward after deleting the classifier, output should be poooled, size(-1) == model.num_features
	model.reset_classifier(0)
	model.to(torch_device)
	outputs = model.forward(input_tensor)
	if isinstance(outputs, (tuple, list)):
	outputs = outputs[0]
	if feat_dim is None:
	feat_dim = -1 if outputs.ndim == 3 else 1
	assert outputs.shape[feat_dim] == model.head_hidden_size, 'pooled num_features != config'
	assert outputs.shape == outputs_pre.shape

pytorch-image-models/tests/test_models.py

Lines 213 to 243 in 648aaa4

	# test forward_features (always unpooled) & forward_head w/ pre_logits
	outputs = model.forward_features(input_tensor)
	outputs_pre = model.forward_head(outputs, pre_logits=True)
	assert outputs.shape[spatial_axis[0]] == pool_size[0], f'unpooled feature shape {outputs.shape} != config'
	assert outputs.shape[spatial_axis[1]] == pool_size[1], f'unpooled feature shape {outputs.shape} != config'
	assert outputs.shape[feat_axis] == model.num_features, f'unpooled feature dim {outputs.shape[feat_axis]} != model.num_features {model.num_features}'
	assert outputs_pre.shape[1] == model.head_hidden_size, f'pre_logits feature dim {outputs_pre.shape[1]} != model.head_hidden_size {model.head_hidden_size}'
	# test forward after deleting the classifier, output should be poooled, size(-1) == model.num_features
	model.reset_classifier(0)
	model.to(torch_device)
	outputs = model.forward(input_tensor)
	assert len(outputs.shape) == 2
	assert outputs.shape[1] == model.head_hidden_size, f'feature dim w/ removed classifier {outputs.shape[1]} != model.head_hidden_size {model.head_hidden_size}'
	assert outputs.shape == outputs_pre.shape, f'output shape of pre_logits {outputs_pre.shape} does not match reset_head(0) {outputs.shape}'
	# test model forward after removing pooling and classifier
	if not isinstance(model, EARLY_POOL_MODELS):
	model.reset_classifier(0, '') # reset classifier and disable global pooling
	model.to(torch_device)
	outputs = model.forward(input_tensor)
	assert len(outputs.shape) == 4
	assert outputs.shape[spatial_axis[0]] == pool_size[0] and outputs.shape[spatial_axis[1]] == pool_size[1]
	# test classifier + global pool deletion via __init__
	if 'pruned' not in model_name and not isinstance(model, EARLY_POOL_MODELS):
	model = create_model(model_name, pretrained=False, num_classes=0, global_pool='').eval()
	model.to(torch_device)
	outputs = model.forward(input_tensor)
	assert len(outputs.shape) == 4
	assert outputs.shape[spatial_axis[0]] == pool_size[0] and outputs.shape[spatial_axis[1]] == pool_size[1]